content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
Mass-loss due to various processes during the lifetime of a star will increase the semi-major axis of a wide binary \citep{DOM:63, VG:88}, what we refer to here as orbital separation amplification. It is expected that mass-loss during the stellar lifetime, in particular Post-Main Sequence (PMS) mass-loss, will statistically distort a frequency distribution of weakly interacting wide binary (Fragile Binary, FB) separations. Visual confirmation of such processes would provide tests of current theories of stellar evolution and fundamental information on the dynamics of Galactic structures \citep{W:88}.
As discussed by \cite{OS:07} and \cite{CHAN:07}, FB's provide insight into several astrophysically important problems that cannot be easily addressed by other means. In particular, such fragile pairs provide valuable constraints on the Galactic tidal potential, stellar and ISM perturbations, and PMS mass-loss, all of which will affect orbital separations. A number of early papers have addressed this topic: \cite{EGG:65} studied the colors and magnitudes of 228 visual binaries (VBs), deriving an $M_v$ vs. $R-I$ calibration from trigonometric parallaxes. \cite{VD:80} showed that, using proper motions from micrometer measurements, a pair could be deemed gravitationally bound and thus, through statistical studies and interferometry, orbital parameters could be determined. \cite{TRIM:74} performed a statistical study of the observed populations of binary systems (798 VBs) and was one of the first to note the peak in the mass ratio (q) distribution at $q\sim1$.
In recent years, computing power has improved to the point that modeling the evolution of whole stellar binary populations is now possible. Examples include: \cite{W:88, VG:88}, and more recently \cite{HTAP:01}. Prior work has shown evidence for orbital separation amplification by as much as half an order of magnitude in binaries, where at least one of the partners has undergone PMS mass loss \citep{HAD:66, G:86, WO:92}. Considering that the majority of all stars are born with one or more companions as a result of the initial conditions required for a stable system \citep{M:94}, the importance of understanding how mass-loss affects the evolution of binary systems cannot be over stated. In addition, the initial-to-final mass relation (IFMR) for WD stars is at present very poorly constrained \citep{WEID:93, ZH:11}. Proper modeling of FB orbital separation amplification can provide a valuable independent estimate of this important relation.
The distribution of $WD+MS$ pair separations is most strongly affected by the PMS (i.e. red giant phase) mass-loss. FB systems typically have component separations of $\sim10^3 AU$ \citep{ZI:84}; this results in negligible tidal interactions and mass transfer between companions. Such pairs evolve as if they are two separate but coeval stars. Once mass-loss has ended, orbital evolution is limited to the effects of external forces such as stellar encounters, giant molecular cloud encounters, gravitational tidal forces, and rare encounters with other stars \citep{CHAN:04,JIAN:10} .
In this study, which updates and improves upon our previous work \citep{JOHN:07}, we compare a theoretically generated population of FB angular separations to the observed angular separation distributions of FBs found in a large-scale survey. We constructed a simulation code, which we refer to here as the Double Star (DS) program, to generate the initial parameters and distributions of the tested population. The DS program evolves the binary components using the Single Star Evolution (SSE) code \citep{HPT:00} over the lifetime of the simulation. The DS program is also responsible for the evolution of individual orbital elements and translation of the modeled physical properties into observable features. The results of the translation is a population of theoretical FBs, with observable properties.
\section{Observational Data}
It should be possible to produce a statistically significant frequency distribution of angular separations for FBs with various constituents, by sampling a large survey such as the Luyten Double Star (LDS) Catalog \citep{LDS:69}. While the LDS catalog, with 6124 pairs and multiples, does not have the volume of more recent surveys such as the Sloan Digital Sky Survey (SDSS), it provides a well-observed large sample to compare to our theoretical models. The identification of FB candidates requires selection of an upper bound separation. Luyten found pairs separated by as much as $400"$. Due to the POSS survey plate scale, Luyten's minimum angular separation limit was $\sim0.5"$. We will use this sample as a first test of our programme, while we note that further samples of fragile binaries are currently being constructed (e.g., \citealt{SES:08, QUIN:09, LONG:10, ZHA:11}).
\section{Initial Parameters}
To build a synthetic model of the local Galactic FB population, we first proposed a set of initial parameters that defined the individual components of our binaries, as well as the binary system separations. We used a combination of published functions from previous theoretical models and observational surveys. When there were no such references, or when the functions published were not appropriate for FBs, we implemented physically reasonable distributions.
\subsection{Mass Generating Function}
Following \cite{KTG:93} we assumed a Salpeter distribution of ZAMS stars using a variation of Eggleton's mass generating function (Equation \ref{eq:a1}).
\begin{equation}\label{eq:a1}
M(x) = 0.33( \frac{1.0} {(1.0 - x)^{0.75} + 0.04(1.0 - x)^{0.25}} - \frac{1.0}{1.04} (1.0 - x)^2 )
\end{equation}
Where x is a uniform randomly distributed number and M(x) is the randomly generated, restricted, mass of a ZAMS star. For this project, we initially assumed a minimum mass of $0.07M_{\odot}$(spectral type-T), given by \cite{CPH:03}. From this we produced an initial distribution of primary masses (Figure \ref{fig:a}).
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{InitialFinalMassPrimaryFig1.pdf}
\caption{Histogram of the distribution of primary masses from initial generation (in red with 45 deg. hatching) and after evolution and orbital expansion (in blue with -45 deg. hatching). The difference in population size is resulting from the dissolution of some binaries from the original population. Each bin is of size $0.05$ in $\log{(M_\odot)}$}\label{fig:a}
\end{figure}
\subsection{Mass Ratio}
The mass ratio, $q$, is one of the most poorly determined variables in a wide binary system. Current observed distributions fall into two categories: (1) a distribution that peaks at $q\sim1$ \citep{TRIM:87} or (2) a bimodal distribution with peaks at $q\sim 0.2$ and $q\sim 0.7$ \citep{HALB:03}. The former are explained in terms of turbulence and fragmentation during the binary formation process \citep{LUC:79}. This tends to result in near equal clumps of would-be stars forming out of a single GMC. Bimodal distributions seem to be created by selection effects (spectroscopic methods vs. common proper motion) and evolutionary effects (binary capture, disruption, mass-loss, mass-transfer,etc.), for references of either see \cite{KRU:67, P:71, SHU:81}. For the purpose of finding an initial mass ratio function (IMRF), we will tend to use the discussion of distributions which tend towards a maximum of one in the set 0$\leq q \leq$1(unity-like), as these will be the most likely candidates to explain a mass-ratio distribution created by a GMC to stellar binary creation process. After multiple trials with various models, we chose a random uniform distribution, where 0$\leq q \leq$1 and $\lambda$ is a constant (Equation \ref{eq:b1}).
\begin{equation}\label{eq:b1}
f(q) = \lambda
\end{equation}
A constant mass ratio distribution function has the added advantage of producing an undistorted secondary mass distribution, implying that the same physical processes produce the primary and secondary components. Using an acceptance-rejection method of random variate generation \citep{LAW:05}, we produced the distribution of secondary masses seen in Figure \ref{fig:b}. In addition we restricted all lower mass components to $M_{star}\geq0.07M_{\odot}$. This has the added effect of producing a tendency for $q\sim1$, as seen in Figure \ref{fig:c}, in accord with prior observational studies.
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{InitialFinalMassSecondaryFig2.pdf}
\caption{Histogram of the distribution of secondary masses from initial generation (in red with 45 deg. hatching) and after evolution and orbital expansion (in blue with -45 deg. hatching). The difference in population size resulted from the dissolution of some binaries from the original population.Each bin is of size $0.05$ in $\log{(M_\odot)}$}\label{fig:b}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{InitialFinalMassRatioFig3.pdf}
\caption{Histogram of the distribution of mass ratios from initial generation (in red with 45 deg. hatching) and after evolution and orbital expansion (in blue with -45 deg. hatching). The difference in population size resulted from the dissolution of some binaries from the original population. Bin size is of size $0.05$ in $q$-ratio}\label{fig:c}
\end{figure}
\subsection{Orbital Parameter Distribution Functions}
In this paper, we focused on the development of two specific orbital element distributions, specifically the distribution of semi-major axis (separation) and the distribution of eccentricity. From these parameters, we develop elliptical binary mechanics for each binary in our initial population. The statistical characteristics of the separation distribution for the un-evolved binaries, as well as the distribution characteristics for the primary masses, secondary masses, and periods is given in Table \ref{tab:a} at the end of this section.
\subsubsection{Initial Separation Function}
The separations of binary systems is a difficult observable parameter, requiring distance, inclination, eccentricity and mass. One of the first and most often used references to separation distribution was given by \cite{O:24}, and implemented by \cite{ZI:84} and \cite{P:88}. It is a simple uniform log distribution:
\begin{equation}\label{eq:c1}
f(a) = \log(a)
\end{equation}
Where $a \geq 100 A.U.$, approximately the closest separation where post-MS mass exchanges are unlikely to have occured. We transform equation \ref{eq:c1} into a random variate separation generating function (Equation \ref{eq:d1}).
\begin{equation}\label{eq:d1}
a(X) = 10^{2.0X+2.0}
\end{equation}
Where $0 \leq X \leq 1$ and $X$ is a uniformly distributed random number. Similarly, we used the separation and masses now assigned to each pair, along with Kepler's $3^{rd}$ law to calculate the period for each binary.
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{InitialFinalSeparationFig4.pdf}
\caption{Histogram of the distribution of separation from initial generation (in red with 45 deg. hatching) and after evolution and orbital expansion (in blue with -45 deg. hatching). The difference in population size resulted from the dissolution of some binaries from the original population. Each bin is of size 0.05 in $\log{(A.U.)}$}\label{fig:d}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{InitialFinalPeriodFig5.pdf}
\caption{Histogram of the distribution of periods from initial generation (in red with 45 deg. hatching) and after evolution and orbital expansion (in blue with -45 deg. hatching). The difference in population size resulted from the dissolution of some binaries from the original population. Each bin is of size 0.1 in $\log{(Myr)}$}\label{fig:e}
\end{figure}
In addition to the physical limit of $a\geq100 A.U.$ for non-interacting pairs, we also imposed observed limitations on our binary pairs. We rely on the \cite{ZI:84} definition of CPMB/FB, VB, and SBs and limit the initial smallest separation that a common proper motion binary can have as $100 A.U.$. This is wide enough that the two components of the pair are individually resolvable and the effects of Roche Lobe interaction and component-to-component tidal interactions are minimal compared to Galactic effects and stellar GMC encounter perturbations.
We additionally imposed a maximum of $10^4 A.U.$ , which appears to be the maximum theoretical separation in wide pairs \citep{WW:87} imposed by gravitational disruptions of the Galaxy and stellar perturbations. An algorithm was developed to impose this upper limit continuously over the evolution time frame and will be discussed below in section 4.3.
\subsubsection{Eccentricity}
The simulation derives the initial distribution of eccentricities following \cite{HEG:75}. The initial distribution of eccentricity for wide binary pairs is given in Equation \ref{eq:e1}.
\begin{equation}\label{eq:e1}
f(e) = 2e
\end{equation}
We transform this distribution into the random variate generating function (Equation \ref{eq:f1}).
\begin{equation}\label{eq:f1}
e(X) = \sqrt{X}
\end{equation}
Where $0 \leq e < 1$ and $X$ is $0 \leq X \leq 1$, a uniformly distributed random number. This distribution was derived based on assumptions of a Maxwellian particle distribution. It becomes apparent from the generating function that binaries, with this eccentricity distribution, are not favored to have initial circular orbits. In fact, using this distribution one can easily see that the expected initial density of circular orbits is zero. For wide binary pairs we can convince ourselves that this is true, as these pairs tend to have coeval components but have had little or no interaction beyond the gravitational binding which makes them a pair (as discussed in the mass ratio section). This creation via loose interaction would tend to result in more initially eccentric orbits.
\begin{table}[h!]
\caption{Statistical Characteristics of the Un-evolved Distributions}\label{tab:a}
\begin{center}
\begin{tabular}{c|cccccc}
\multicolumn{7}{c}{} \\
\hline
Metric & Average & StdDev & Skew & Kurtosis & Min & Max \\
\hline
Mass 1($\log{M_{\odot}}$)& -0.28& 0.41& 0.58& 0.79& -1.10& 1.93\\
Mass 2($\log{M_{\odot}}$) & -0.55& 0.40& 0.92& 1.03& -1.10& 1.84\\
Separation ($\log{A.U.}$) & 3.00& 0.58& -0.002& -1.20& 2.00& 4.00\\
Period ($\log{Myr}$) & -1.46& 0.826& -0.008& -1.08& -4.04& 0.39\\
\hline
\end{tabular}
\end{center}
\end{table}
\subsection{Birth Rate}
Our experiments incorporated an initial stellar burst of $10^2$ binaries with \cite{WK:04} constant repeating burst model of the stellar birth rate, $10^2$ binaries being created every $10 Myrs$, resulting in a total initial population of 100,100 binaries. All bursts have identical initial factors and distributions, resulting in a continuous and smooth distribution \citep{HPT:00} in the initial population. The only references that challenge this approach, suggest that metallicity may play a role in distorting the initial distribution function. However, the region of space that we are attempting to model is limited to the solar neighborhood (i.e. distances under $100 pc$), FBs in this region tend to have solar metallicitys. We therefore assumed solar metallicity for all ZAMS stars. Future work may address a more realistic history of star formation rate, dervied for the solar neighbourhood \citep{HER:00}.
\subsection{Color Estimation}
The goal of this project is to compute a theoretical population of FBs that accurately compared to observed angular separation distributions. This requires a transformation of our physical characteristics for each binary system into observables such as color, magnitude and angular separation. We employed a mix of \cite{KV:06, K:79, HAW:02} temperature vs. color relations for MS and WD stars, to produce the $ugriz(J)$ absolute magnitudes required for comparison.
We divided stars into four categories: MLT, WD, MS, and other. MLT refers to the spectral type of the star: M, L or T, corresponding to effective temperatures from $\sim3800K$ to $1000K$ \citep{R:00, R:02}. The spectral type of a star below $3400K$ is assumed to be dependent only on temperature. We justify this estimation with three points: (1) Reliable statistics are not available for surface gravities of low mass stars (early-T and late-L); (2) The current classification scheme for MLT dwarfs is based mostly on line features. A number of flux ratios ($H_2O, CO_2, CH_4,$etc) show linear relations with spectral type \citep{BURG:02}; (3) \cite{STE:01} used a hybrid approach to show that temperatures and spectral types scale linearly with each other for L-type dwarfs.
Next, we used a combination of spectral type and temperature relations to transform the observed absolute magnitudes by \cite{HAW:02} in $r^\star i^\star z^\star J$ to produce a temperature vs. absolute magnitude relation. We then used this relation to produce colors and errors for all MLT stars. Neither the LDS nor the SDSS survey provides apparent $J$ magnitudes. Future surveys however will produce near-infrared colors (y). At such a time our code can be updated and expanded.
For WD and MS stars, we use models provided by \cite{KV:06} and \cite{K:79}, respectively. These atmospheric models are grids of color as functions of temperature and gravity. We employed a 2nd/3rd order unevenly spaced 2-dim interpolation subroutine to determine both expected absolute magnitude and error in absolute magnitude given $T$ and $\log(g)$. As before, we estimated $u^\star g^\star r^\star i^\star z^\star$ colors. This allowed simple estimates of distance modulus for each pair.
Models used to compute discrete grids are available\footnote{http://www.stsci.edu/hst/observatory/cdbs/k93models.html}, derived from the Kurucz stellar atmosphere models. For MS stars temperatures range from $3,170K\leq T \leq 42,000K$ and $0\leq \log(g) \leq5$. For WD stars, two sets of grids were used: a fine resolution with range $7,000K\leq T \leq84,000K$ and $7\leq \log(g) \leq9.5$ in $0.5$ intervals and a broad resolution set with $4,500K\leq T \leq84,000K$ and $7\leq \log(g) \leq9$ with $1.0$ gravity intervals. The fine resolution set was used for stellar temperatures with $7,000K \leq T \leq84,000K$. For temperatures above $84,000K$ and below $7,000K$, the broad resolution set was used.
\subsection{Distance}
In order to compute the distance modulus, a variety of models were tried, including a synthetic Galactic population model that includes distance as a function of mass \citep{RRD:03}. Initially, we produced a distance model that best fits the observed LDS distribution with the given parameters. The adopted distance distribution had the form:
\begin{equation}\label{eq:g1}
f(d) = \log(d)
\end{equation}
This equation was then transformed into a random variate separation generating function:
\begin{equation}\label{eq:h1}
a(X) = 10^{3.0X+1.0}
\end{equation}
Where $0 < X < 1$ and $X$ is a uniformly distributed random number. We expect this will, through random evolutionary processes, produce the synthetic lognormal distribution of distances that \cite{RRD:03} obtained. We also assume that binary and single star distance distributions are identical.
\subsection{Stellar Evolution Code}
One of the two major tasks in this project is the ability to accurately evolve a theoretical star while keeping track of its mass, luminosity, and temperature; as previous analysis used analytical, rather than physically-motivated, mass-loss rates (e.g. \citealt{DOM:63, DOB:98, RAHO:09}). For a population of binary stars this requires substantial computational resources. As mentioned above, the Single Star Evolution (SSE) program \citep{HPT:00} was adapted for this effort. The basis of the SSE is a set of comprehensive analytical formulae, that accurately track the evolution of a star with given initial mass and metallicity. The goal of the SSE program was to allow for the rapid calculation of a number of stellar properties for a given stellar lifetime, thus facilitating the production of a population synthesis model that could produce a statistically significant sample of binaries in a reasonable computational timeframe.
The key to the SSE code's accuracy and speed is its process style, which employs a combination of interpolation formulae and tabular construction. The HPT group uses tabulated stellar models \citep{SSMM:92, CMMSS:93, MSM:98}. \cite{PSH:98} later expanded the range of mass and metallicity of the tables incorporated in the SSE program. Analytic approximations to the movement of a star in the Hertzsprung-Russell diagram (HRD) are a simplistic approach to a complex problem \citep{EFT:89}. However, supporting these calculations with tabulated stellar models allows a more precise estimation of physical properties and greatly decreases computational time.
The SSE does not take into account stars with solar masses lower then $0.2M_\odot$. For stars with masses lower then this limit we adopted the \cite{BCAH:98} low-mass stellar evolutionary models. We do not expect such low mass stars to undergo stellar mass evolution, but we do expect small changes in other stellar parameters over time. Using these evolutionary models we were able to determine the initial and final values of temperature and gravity based on initial mass and metallicity.
We selected the following parameters for the SSE code: no wind-loss, a Reimers Coefficient for mass-loss of 0.5, and a constant metallicity distribution of 0.1. This set of parameters provides an initialization point from which to develop further analysis, while still generating realistic results. Future research may include analysis (ANOVA) to determine of optimal parameter distributions and values for comparison of theoretical and observed population distributions. The authors of the SSE code have subsequently developed a follow-up algorithm, the Binary Star Evolutionary code. For development purposes, this new program was not used, as the primary concern of the DS-code is the generation of evolved binary populations and information, relating to fragile coeval binaries. The authors feel that wrapping the elliptical orbital computations, stellar population generating functions, and apparent magnitude translations around the original SSE code, allows for the maximum amount of computational efficiency and provides the necessary forensic test points for further experimentation, specifically in the generation of an ANOVA experiment for optimal parameter and distribution selection.
\subsection{Orbital Evolution}
In the present model, we assumed isotropic non-conservative mass-loss for both stars, both during the MS and PMS lifetimes. Since we modeled what could be considered two phases of mass-loss, slow (MS) and sudden (PMS), we found a two stage evolutionary code was needed. Following \cite{VG:88, HAD:66, KOP:78}, we considered the effect double, non-interacting, mass-loss, has on the orbital parameters of the binary. For this experiment we modified versions of \cite{VG:88} equations into the simulation code presented here. We allowed no accretion onto either star from its companion and no accretion from the ISM or any outside bodies. The addition of perturbation forces from exterior sources beyond those addressed within this work, such as loose tertiary components, dense stellar populations, or other “chance” encounters were not modeled here, however we would recommend \cite{HEG:96} as starting research in the matter.
We can show via adjustment of the mass-loss orbital parameter equations that the perturbing force due to an isotropic variation of mass is tangential to the orbit, thus only the planar orbital elements will be perturbed. This assumption holds when the mass-loss timescale of the binary is much greater than the orbital timescale (period). These perturbation equations are given as a function of time, the logarithm of the total system mass, the eccentric anomaly, and the eccentricity are given as the equations \ref{eq:i1} and \ref{eq:j1}.
\begin{equation}\label{eq:i1}
\frac{d\log{a}}{dt} = -\frac{1+e\cos{E}}{1-e\cos{E}}\frac{d\log{M}}{dt}
\end{equation}
\begin{equation}\label{eq:j1}
\frac{de}{dt} = -\frac{(1-e^2)*\cos{E}}{1-e\cos{E}}*\frac{d\log{M}}{dt}
\end{equation}
Where $\frac{d\log{M}}{dt}$ is the rate of mass-loss over the time period of interest (iteration), and $E$ is the eccentric anomaly. Note, that when expressing these equations in terms of the eccentric anomaly, Kepler's equations is no longer valid and should be replaced with the equation,
\begin{equation}\label{eq:k1}
\frac{dE}{dt} = \left( \frac{2\pi}{P} + \frac{d\log{M}}{dt}\frac{\sin{E}}{e}\right)*\left(1-e*\cos{E}\right)^{-1}
\end{equation}
The effect that this MS mass-loss has on the binary system is documented by \cite{VER:69}, \cite{DOM:63}, \cite{DOB:98} and \cite{RAHO:09}. Here, Verhulst focused the majority of the research on the effect a Jeans-Eddington mass-loss relation would analytically have on the eccentricity of the binary. It was shown that one can expect periodic variation about a mean eccentricity for a binary with isotropic continuous mass-loss depending on the parameters of the Jeans-Eddington mass-loss relationship. \cite{HAD:63} demonstrated similar effects, showing that with decreasing system mass the eccentricity of the system will sinusoidally increase. More recently \cite{VEA:11} demonstrated orbital parameter evolution for low q-ratio binaries ($q < 0.1$, exoplanets and smaller bodies). They showed that, at least for these conditions, the effect of mass-loss, varies greatly depending on the initial eccentricity distribution.
Given that our mass-loss is a function of the stellar evolution code designed into the SSE, with the additional consideration of the low mass evolution models provided by \citep{BCAH:98}, we expect a similar result (increasing non-linear, sinusoidal eccentricity as a function of mass-loss). However we expect that the numerical (and probabilistic) nature of our experiment would demonstrate a different (and potentially more complete) relationship. For situations where mass-loss occurs on a timescale less than that of the period of the binary, we follow the “sudden” mass-loss equations \ref{eq:l1} and \ref{eq:m1} outlined by \cite{HIL:83}.
\begin{equation}\label{eq:l1}
a = \frac{a_0}{2} \left( \frac{1 - \frac{\Delta M}{M}}{\frac{1}{2} - (\frac{a_0}{r} ) *\frac{\Delta M}{M}} \right)
\end{equation}
\begin{equation}\label{eq:m1}
e = \left(1 - (1-e_0^2)*\left[\frac{1 - \frac{2a_0}{r}\frac{\Delta M}{M}}{(1 - \frac{\Delta M}{M})^2}\right]\right)^{0.5}
\end{equation}
We have further transformed the original equations into a function of the mass-loss ratio over a given defined time period (delta mass over original mass). The increase in separation and eccentricity, is a function of how close to periapsis or apoapsis the binary is at the time of mass-loss. It can be shown that for binary separation distances less than the semi-major axis (which is dependent on the eccentricity); sudden mass-loss can drive the binary together (decrease orbital separation).
Similarly, as stated by \citeauthor{HIL:83}, if mass-loss occurs in a binary with a very eccentric orbit, the average post-explosion eccentricity will be less than the pre-explosion eccentricity. Conversely, given the position of the binary, the separation can be greatly amplified, resulting in populations of $WD+MS$ and $WD+WD$ pairs with separations on average larger than their $MS+MS$ counterparts. We can demonstrate for given eccentricity and location in the orbit, post-main sequence mass-loss can result in a perturbation of the orbital eccentricity beyond a bound elliptical orbit, i.e., $e \geq 1$. This is demonstrated in Equation \ref{eq:n1}.
\begin{equation}\label{eq:n1}
\frac{r}{2a} < \frac{\Delta M}{M}
\end{equation}
It is apparent, that increasing eccentricity can pose a problem for the binary system. The potential for binary separation can result from increasing the eccentricity of the binary beyond $(e \geq 1)$. Thus, we expect that the increase in eccentricity, and therefore the associated orbital velocities, in combination with the wide orbital amplitudes, will result in a number of binary systems being perturbed out of their orbits (dissolution via external perturbation). It should be noted that decreases in eccentricity resulting from sudden PMS mass-loss, coupled with the lack of mass-loss associated with white dwarf components, can result in the effective circularization of the orbit. This is apparent from the disproportionate number of binaries in the $WD+MS$ and $WD+WD$ population evolving to a final eccentricity under 0.1 in our simulations.
To model the effect high eccentricity has on the binary system, specifically on the dissolution rate of the binary populations resulting from isotropic mass-loss and galactic tidal forces, we have implemented a function that flags evolved binary systems which have eccentricity greater than 0.95. This, combined with the \cite{AL:72} tidal limit provides a first order approximation to dissolution effects on the binary systems. Other improvements are outlined in the Binary Dissolution section
\subsection{Binary Dissolution}
Even a simple model for the orbital evolution of wide binaries must include the influences of stellar encounters, Giant Molecular Clouds (GMCs) and the Galactic tidal field \citep{WEIN:88}. For the wide separations of FBs, even minor forces such as these operating over the course of millions of years, can affect the binary separations, eventually dissolving them. \cite{AL:72} described a tidal limit for wide binaries, where the gravitational forces from Galactic tidal perturbations outweigh the gravitational attraction between the binaries themselves(Equation \ref{eq:o1}).
\begin{equation}\label{eq:o1}
a_T \sim 2.07 \times 10^5 A.U. \left( \frac{M_1+M_2}{M_\odot}\right)^{1/3}
\end{equation}
This relation does not include either GMC or stellar encounters, but it does provide an initial cut off for galactic tidal disruption as a function of mass. We take the \cite{AL:72} equation as a first approximation to a dissolution algorithm. A more realistic approach would include all disruption effects. A rigorous treatment is beyond the scope of the current project. For a first-order approximation to the dissolution time problem, we used the \cite{WW:87} Galactic tidal perturbation formula for a $2M_\odot$ binary system with $q = 1$. The expected half-life ($t_{1/2}$) of an FB system with no consideration for the GMC encounter is given by Equation \ref{eq:p1}.
\begin{equation}\label{eq:p1}
t_{1/2} = 4.3 \times 10^3 Myr \left(\frac{a}{2.07\times10^4}\right)^{-1.4}
\end{equation}
For wide binaries with consideration for the GMC local density of $3.6\times10^8 pc^{-3}$, this relation becomes Equation \ref{eq:q1}.
\begin{equation}\label{eq:q1}
t_{1/2} = 1.32 \times 10^3 Myr \left(\frac{a}{2.07\times10^4}\right)^{-1.4}
\end{equation}
We define $t_{1/2}$ as the $50\%$ likelihood survival time for a binary with separation $a$ in $A.U.$ While this cannot be used to estimate the expansion caused by a GMC and perturbations, using a binomial distribution selection program we produced a dissolution algorithm to remove binaries from the sample that likely would have dissolved in the evolutionary time frame given. An evenly distributed binomial random variable with two possible outcomes ($50/50$) was used to simulate $t_{1/2}$.
Lastly we chose a simple ratio of survival times to linearly scale the probability mechanism. Binaries with a separation greater than the $50\%$ likelihood survival time will have a ratio greater than one and a better chance of dissociating (i.e. $50/50$ turns into $75/25$), while binaries with a separation less than the half point limit, will have a ratio less than one and a worse chance of dissociating (i.e. $50/50$ turns into $25/75$). For example: if a binary has a separation of $10^4 A.U.$ its half-life is $3655 Myr$. If a binary evolves for $3655 Myr$, there will be a $50\%$ chance of it being selected to dissolve. If the binary evolves for $10,000 Myrs$ then: $\frac{3655}{10000} = 0.37$ and $50\ast0.37\sim18$. Therefore instead of a $50\%$ chance of our binary surviving, there is an $18\%$ chance of it surviving. Binaries that have been selected using our binomal random number generator, are removed in the posterior population samples.
We note again that the dissolution algorithm is not a continuous function; its purpose is to generate in the final sample the expected number of dissolved binaries in a given population. The linear estimation of what should be an exponential (Poisson) function, will cause the dissolution algorithm to under-sample the number of dissolved binaries in our distribution. The \textit{dissolve} or \textit{survive} approach also does not allow for external forces and perturbations to expand the binaries. However, as stated above, the residual term from our model vs. observation comparison will provide some insight to the degree of which the sum of these external forces will effect our model. Binaries that have been removed are not considered in the following analysis.
\section{Results}
Rather then modeling individual binaries, we attempted to produce theoretical separation frequency distributions to compare with observational statistics. Histogram bin sizes were chosen to be similar for like quantities across the various cases. We then transformed our model parameters into model observations and compared our results with the observed \cite{LDS:69} separation distribution. The statistical characteristics of the distribution of primary masses, secondary masses, separations and periods for the evolved binaries is given in Table \ref{tab:b}.
\subsection{Mass Graphs}
For comparison purposes, we plotted orbital frequency distributions for both the initial and posterior populations. Figure \ref{fig:a} and Figure \ref{fig:b} include not only the initial burst binaries but also those born in bursts later on. Thus, not all binaries formed at $t_0 = 10^4$ in the initial sample, while the posterior populations have identical $t_{final}$.
Figure \ref{fig:a} and Figure \ref{fig:b} are a representation of the same sample at an age of $\sim10 Gyr$. In Figure \ref{fig:b}, the random q-ratio has produced, in the secondaries, a distribution similar to the primaries. There are two artifacts in our secondary mass population that may not be physically realistic. First is the cutoff at $0.07M_\odot$, resulting from adopting the limit suggested by \cite{CPH:03}. However, two points mitigate this arbitrary cutoff: (1) as shown by \citeauthor{CPH:03}, the mechanism of star formation does have a minimum mass, and (2) imposing our q values to be a number between 0 and 1 causes the average secondary mass value to be less than the primary, as expected. According to \citeauthor{CPH:03}, the initial mass function at lower masses is poorly determined. Moreover, the LDS and SDSS surveys have detected relatively few M, L, and T dwarfs, particularly in binary systems, and not yet enough to warrant an attempt to model this region.
A look at the posterior distributions in Figures \ref{fig:a} and \ref{fig:b} shows the impact of mass loss. Peaks in the distribution near the $0.5M_{\odot}$ to $1.0M_{\odot}$ range and a decrease of high mass stars is a sign that some of the MS stars have evolved to less massive WD stars, or Neutron Stars and Black Holes (NS and BH). At $10 Gyr$, in some pairs both components have evolved to $\sim0.6M_{\odot}$ ($WD+WD$).
\begin{table}[h!]
\caption{Statistical Characteristics of Evolved Distributions.}\label{tab:b}
\begin{center}
\begin{tabular}{c|cccccc}
\multicolumn{7}{c}{}\\
\hline
Metric & Average & StdDev & Skew & Kurtosis & Min & Max \\
\hline
Mass 1 $(\log{M_{\odot}})$& -0.36& 0.30& -0.40& -0.36& -1.10& 1.01\\
Mass 2 $(\log{M_{\odot}})$ & -0.58& 0.33& 0.27& 10.990& -1.10& 0.81\\
Separation $\log{A.U.}$& 2.97& 0.55& 0.069& -1.04& 1.74& 4.88\\
Period $\log{Myr}$ & -1.48& 0.84& 0.075& -1.02& -3.60& 1.11\\
\hline
\end{tabular}
\end{center}
\end{table}
\subsection{Mass Ratio}
Figure \ref{fig:c} is a histogram of the mass ratio distribution for the initial population of binaries and the posterior population of binaries. There is a trend towards $q\sim1$ in the initial population, as expected from the mass ratio generating function. A unique by-product of evolution is evident in the posterior population of the $q$-distribution: high mass MS stars have evolved into lower mass WDs. Now the secondary has a higher mass than the WD, resulting in a distribution of binaries with $q > 1$. This time, however, the $q > 1$ range represents the region where the primary has undergone PMS mass-loss and now has less mass than the secondary, i.e. one or both of the components of the binary have evolved into degenerate stars (WD, NS, or BH). The additional evolution to WD of the secondary would likely only cause a density change in the population above $q=1$ or in the compact region near $q = 0.9$.
\subsection{Separation and Period}
Figures \ref{fig:d} and \ref{fig:e} present comparisons of the initial and posterior populations of binaries, the separations (in $\log{(A.U.)}$) and the period distributions (in $\log{(Myrs)}$). The sharp observed initial distribution of separations is a result of the initial parameters: a cutoff at $10^4 A.U.$ (GMC birth cloud), and a lower cutoff ($100 A.U.$) based on the definition of what qualifies as a FB. We impose an \"{O}pik distribution on top of this to produce the desired frequency distribution. The scatter is a result of binning and low number density. At higher densities ($10^5$ binaries and greater), this scatter becomes negligible. Our initial distributions are leptokurtic (i.e. flat-top), which is expected in $\log{(a)}$ due to our initial assumptions of a uniform log distribution. The skewness of the initial distribution $\sim0$ results from the flat sides and little or no discernable tail of the distribution. We see similar characteristics in the initial $\log{(P)}$ distributions(Figure \ref{fig:e}), however the edges of the distribution have been rounded out resulting from the effect the mass distribution has on the generated distribution. Our initial limits in the separation distribution results in the min ($a > 10^2 A.U.$) and max ($a < 10^4 A.U.$), respectively.
Comparisons to posterior distributions in Figure \ref{fig:d} and Figure \ref{fig:e} demonstrate that the maximum separation is now $\log{(A.U.)} = 4.88$ ($75,850 A.U.$). Two other parameters have also changed: the shift to higher separations has depleted the number of close binaries and increased the higher separation binaries. This has effectively rounded out the histogram of separtions (kurtosis $= -1.04$) and narrowed the histogram, (standard deviation $= 0.55$) compared to the standard deviation of $0.58$ in the initial distribution of Figure \ref{fig:d}. Figure \ref{fig:e} shows a similar trend in variance, average and kurtosis. The period has the addition of a randomly generated mass factored into its calculation; this in turn has affected the kurtosis and skewness which now reflect a more normal distribution, per the central limiting theorem (CLT), as well as making the curve appear smoother.
\subsection{Analysis by Binary Category}
Figures \ref{fig:f} to \ref{fig:i} are overlay plots of the three binary classes $MS+MS$, $WD+MS$ and $WD+WD$. All distrbutions are derivitives of the posterior population distribution described in the previous section; the "other" catagory refers to binaries with either component not in the MS or WD phase (RG, BH, NS...etc), the breakdown of population distribution types is given in Table \ref{tab:c}. The statistical properties of each of the three samples are given in Tables \ref{tab:d} - \ref{tab:f}.
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{FinalAllPrimaryFig6.pdf}
\caption{Histogram of the distribution of primary masses from the MS+MS populaton (in red with 45 deg. hatching), from the WD+MS population (in green with 135 deg. hatching) and the WD+WD population (in blue with 0 deg. hatching). Note the mixed colors represent overlaping distributions, bin sizes are in 0.05 $\log{(M_{\odot})}$ intervals.}\label{fig:f}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{FinalAllQFig7.pdf}
\caption{Histogram of the distribution of mass ratios from the MS+MS populaton (in red with 45 deg. hatching), from the WD+MS population (in green with 135 deg. hatching) and the WD+WD population (in blue with 0 deg. hatching). Note the mixed colors represent overlaping distributions, bin sizes are in 0.05 intervals.}\label{fig:g}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{FinalAllSeparationFig8.pdf}
\caption{Histogram of the distribution of separation from the MS+MS populaton (in red with 45 deg. hatching), from the WD+MS population (in green with 135 deg. hatching) and the WD+WD population (in blue with 0 deg. hatching). Note the mixed colors represent overlaping distributions, bin sizes are in 0.05 $\log{(A.U.)}$ intervals.}\label{fig:h}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{FinalAllPeriodFig9.pdf}
\caption{Histogram of the distribution of period from theMS+MS populaton (in red with 45 deg. hatching), from the WD+MS population (in green with 135 deg. hatching) and the WD+WD population (in blue with 0 deg. hatching). Note the mixed colors represent overlaping distributions, bin sizes are in 0.1 $\log{(Myr)}$ intervals.}\label{fig:i}
\end{figure}
\begin{table}[h!]
\caption{Distribution of Evolved Binaries in Posterior Population. (Population size is 90,523).}\label{tab:c}
\begin{center}
\begin{tabular}{c|cccccc}
\multicolumn{5}{c}{} \\
\hline
Metric & $MS+MS$ & $WD+MS$ & $WD+WD$ & Other \\
\hline
Count & 76,886& 7,327& 4,791& 1,519\\
Percent & 84.94\%& 8.09\%& 5.29\%& 1.68\%\\
\hline
\end{tabular}
\end{center}
\end{table}
\subsubsection{MS+MS}
The $MS+MS$ parameters are relatively unchanged and appear almost identical to the initial distribution plots in $q$-ratio, $\log{(Myr)}$ and $\log{(A.U.)}$. However, the $MS+MS$ mass plots do show missing upper mass stars. These, of course, have been transferred to the other groups as WD, BH, NS, or other stars. Similarly, a good number of higher separation binaries have dissolved, resulting in the observed change in skewness (observed asymmetry in separation distribution). Of the given model (90,523 binaries), 76,886 ($84.94\%$) binaries remain in the $MS+MS$ category.
\begin{table}[h!]
\caption{Statistical Characteristics of Evolved MS+MS Distributions}\label{tab:d}
\begin{center}
\begin{tabular}{c|cccccc}
\multicolumn{7}{c}{}\label{tab:d} \\
\hline
Metric & Average & StdDev & Skew & Kurtosis & Min & Max \\
\hline
$\log{a} (A.U.)$ & 2.94& 0.553& 0.072& -1.12& 2.00& 4.00\\
$\log{P} (Myr)$ & -1.49& 0.841& 0.066& -1.05& -3.28& 0.39\\
\hline
\end{tabular}
\end{center}
\end{table}
\subsubsection{WD+MS}
We observe the peak mass distribution of the primary in the $WD+MS$ (i.e., the $WD$) at slightly less then unity ($WD$ with masses of less then one solar mass). In addition, the secondary is by definition an $MS$ star (has not evolved yet). The secondary mass distribution has a peak at $\sim0.7_\odot$, the approximate turn-off-mass for Disk stars, thus restricting the $q$-ratio distribution. The histogram for $WD+MS$ $q$-ratio values (Figure \ref{fig:g} ) shows a number of $q > 0$ value binaries, reflecting the fact that some original secondaries are now more massive than their evolved primaries.
\begin{table}[h!]
\caption{Statistical Characteristics of Evolved WD+MS Distributions}\label{tab:e}
\begin{center}
\begin{tabular}{c|cccccc}
\multicolumn{7}{c}{ }\\
\hline
Metric & Average & StdDev & Skew & Kurtosis & Min & Max \\
\hline
$\log{a}(A.U.)$ & 3.13& 0.489& 0.262& -0.90& 2.24& 4.48\\
$\log{P}(Myr)$ & -1.40& 0.735& 0.246& -0.89& -2.78& 0.72\\
\hline
\end{tabular}
\end{center}
\end{table}
The $WD+MS$ histogram of separations shows an increase in excess kurtosis, i.e., the movement of low period and small separations towards higher values. This causes the more normal distribution curve that is seen ($kurtosis \Longrightarrow 0$). Additionally, the average separation has increased to $\log{(A.U.)} = 3.1\pm0.5$, i.e., $WD+MS$ pairs display an average separation $a = 477 A.U.$ greater than the original $MS+MS$ distribution $\log{(A.U.)}=2.9\pm0.5$. Thus, orbital amplification due to $PMS$ mass-loss is shown to have increased the average separation of the population. We note, however, that this is not the order of magnitude difference that \cite{G:86} expected. Perhaps our models have underestimated the effects of environmental perturbations (GMC, chance encounters, Galactic tidal forces). There are 7327 $WD+MS$ binaries in the final sample (i.e., $\sim8.1\%$).
\subsubsection{WD+WD}
To date, there are only about two dozen \citep{SION:91} double degenerate systems known, and certainly not as many as shown here ($N = 4791, \sim5.3\%$). The most plausible reason for this lack of discovery of $WD+WD$ binaries results from either one or both of the components being too dim to be detected in the surveys currently available. The histogram of the $q$-ratio (Figure \ref{fig:g}) show a tendency to nearly identical primary and secondary masses. This is to be expected with the narrow range of white dwarf final masses, resulting from the SSE's IFMR calculations.
\begin{table}[h!]
\caption{Statistical Characteristics of Evolved WD+WD Distributions}\label{tab:f}
\begin{center}
\begin{tabular}{c|cccccc}
\multicolumn{7}{c}{}\\
\hline
Metric & Average & StdDev & Skew & Kurtosis & Min & Max \\
\hline
$\log{a}(A.U.)$ & 3.15& 0.500& 0.316& -0.81& 2.27& 4.66\\
$\log{P}(Myr)$ & -1.36& 0.749& 0.308& -0.82& -2.74& 0.9\\
\hline
\end{tabular}
\end{center}
\end{table}
As predicted, the secondary companion (lower initial mass) produces a lower mass $WD$ than the primary, causing the peak at $q = 1$ (Figure \ref{fig:g}. Similar to the $WD+MS$ histogram of separations, the $WD+WD$ histogram shows an increase in kurtosis, i.e. loss of short period and small separations, evolving towards higher values. This causes the more normal distribution curve that is seen ($kurtosis\Longrightarrow0$) compared to the MS+MS population. The increase is the result of continued evolution of the binaries. The post-MS sudden mass-loss of the binary secondary influences the separation and period distributions a second time, pushing the uniform distribution parameters closer to a more random, Gaussian distribution. Additionally, average separation has increased from $\log{(A.U.)}=3.13$ for the $WD+MS$, to $\log{(A.U.)}=3.15$ for the $WD+WD$, i.e. an average separation increase of $63.6 A.U.$. We also note that the maximum separation has increased to $\log{(A.U.)}=4.66$ range, much greater than the initial $\log{(A.U.)}=4.0$ maximum imposed on the initial distribution.
\subsection{Comparison to Observed Data}
We begin our comparison by testing a variety of distance models against the LDS $WD+MS$ sample. For our models, we used a $10^4Myr$ evolutionary time with $10^2$ initial binaries (at $t = 0$) and a $\frac{100binaries}{10Myr}$ birth rate (i.e. constant average birth rate, \citealt{HPT:00}). For each component, we have roughly determined absolute $ugriz$ magnitudes using the \cite{K:79} models. We found temperatures and gravities from the SSE output or from \cite{CPH:03} evolutionary models for solar metallicity low-mass stars. Using the luminosities and separations, we applied a random distance model derived by assuming a uniform synthetic Solar Neighborhood population to approximate angular separation and apparent magnitude. This yielded a theoretical population histogram, restricted by the physical observation limits of the LDS. The distribution of the subset of modelled observed evolved binaries is given in Table \ref{tab:g}. Note that, in comparison to Table \ref{tab:c}, there are fewer $MS+MS$ observed than in the simulated population, an effect likely caused by the dimness of some low mass companions. Similarly, the percentage of $WD+MS$ pairs in the observed distribution is higher than in total posterior population. The latter is likely caused by the expected relative brightness of the binary pair; as mass ratio tends towards one, a pair with a $WD$ as a primary is likely to have a secondary with mass just slightly less than the progenitor of the $WD$, resulting in a brighter pair on average, than observed in the $MS+MS$ distribution.
\begin{table}[h!]
\caption{Distribution of Theoretically Observed Evolved Binarys in Postieror Population. Total observed population size is 29,6335}\label{tab:g}
\begin{center}
\begin{tabular}{c|cccccc}
\multicolumn{5}{c}{} \\
\hline
Metric & $MS+MS$ & $WD+MS$ & $WD+WD$ & Other \\
\hline
Count & 23,507& 3,801& 1,442& 885\\
Percent& 79.3\%& 12.8\%& 4.9\%& 3.0\%\\
Fraction of Posterior & 30.6\% & 51.9\% & 30.1\% & 58.3\% \\
\hline
\end{tabular}
\end{center}
\end{table}
\subsection{Distance Model Comparison}
The comparison of our model and observed angular separation distributions is shown in Figures \ref{fig:k} and \ref{fig:l}. We expect the observational data, which is a much smaller sample ($\sim6124$ binaries) to have more random scatter than the modeled data (WD+MS population $ = 1670$ binaries). Although the agreement is good dispite the scatter in the LDS sample, there are discrepancies (that we intend to address in a later paper).
\begin{itemize}
\item The kurtosis of the model is is smaller than in observed sample. Perhaps wide binaries at smaller separations were under-observed by Luyten. Alternatively, our model's lower limit on separation could be unrealistic.
\item The scatter in the observed data is $1\sigma$ greater than the modeled data.
\item The mean and skewness of both data sets is similar; this suggests that our basic model is sound, at least as a first order approximation.
\end{itemize}
\section{Discussion}
The distribution of binary angular separations in our model is similar to the observed distribution(i.e., the LDS sample). The LDS sample has a lognormal separation distribution that most likely stems from a log-uniform distance distribution. The estimated range of the LDS sample is between $1<\log{d}<3$ (where distance is in parsecs), similar to our model distance range. Initially we can compare the precentage of $WD+MS$ pairs observed in the LDS sample ($511/6124 = 0.08$) with the precentage of modeled $WD+MS$ pairs ($.12$) as well as the population of $WD+WD$ pairs observed in the LDS sample ($24/6124 = 0.003$) with the population of $WD+WD$ modeled ($0.04$). These distribution numbers suggest either we have underestimated the number of binaries lost to dissolution, or over-estimated the brightiness of some of the pairs. Likewise, these comparisons may also indicate where the LDS survey is under-sampled (i.e., the WD+WD pairs).
Figures \ref{fig:j} and \ref{fig:l}, present the results from our comparison of $MS+MS$ LDS and model pairs as well as $WD+MS$ LDS and model pairs. The figures show the cumulative distribution functions of the populations. Note that the separations are given in $\log(s")$ and have been put into 0.2 interval bins. Our initial hypothesis that PMS mass-loss significantly distorts a distribution of separations is apparent in Figure \ref{fig:k}, the offset in the CDF curves reflects an increase in average angular separation.
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{CDFModelComparisonFig10.pdf}
\caption{Comparison of observed angular separation CDF for MS+MS models and WD+MS models. Red Dashed Line: MS+MS, Blue Solid Line: WD+MS}\label{fig:j}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{CDFMSMSObservedFig11.pdf}
\caption{Plot of CDF of Model to Observed Cumulative Angular Separation Distributions for MS+MS pairs. Red Dashed Line: Observed, Blue Solid Line: Model}\label{fig:k}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{CDFWDMSObservedFig12.pdf}
\caption{Plot of CDF of Model to Observed Cumulative Angular Separation Distributions for WD+MS pairs. Red Dashed Line: Observed, Blue Solid Line: Model}\label{fig:l}
\end{figure}
We use the Cram\'{e}r-Von Mises goodness-of-fit test for comparison of the two-population cumulative distributions \citep{AN:62}, given in equations \ref{eq:r} and equation \ref{eq:s}. Here, goodness-of-fit is defined as the hypothesis $H_0: \bar{r} = 0.0$ where $\bar{r} = F_N(x) - G_M(x)$, where $F_N(x)$ and $G_M(x)$ are the cumulative distribution functions of samples of size $N$ and $M$, respectively. The Cram\'{e}r-Von Mises goodness-of-fit statistic here is defined by $T$.
\begin{equation}\label{eq:r}
T = \frac{NM}{N+M}\int_\infty^\infty [F_N(x) - G_M(x)]^2dH_{N+M}(x)
\end{equation}
\begin{equation}\label{eq:s}
T = \frac{NM}{(N+M)^2}*{\sum^N_{i=1}[F_N(x_i) - G_M(x_i)]^2 + \sum^M_{j=1}[F_N(y_j) - G_M(y_j)]^2}
\end{equation}
For the $MS+MS$ model and observation population set the values are $N = 26$ and $M = 26$. For the $WD+MS$ model and observation population set the values are also $N = 26$ and $M = 26$ as we have used the same spacing for both empirical distribution functions. These empirical distribution functions are based on the cumulative distributions generated for the populations with a bin interval of 0.2 between $0.4<\log{(s")}<2.5$. We compute T values for the comparison of $MS+MS$ populations ($T = 0.1343$) and the comparison of $WD+MS$ populations ($T = 0.0418$). Based on these statistics we estimate the limiting distribution, $a_1(T)$, of T for our models as in \cite{AN:52}:
\begin{equation}\label{eq:t}
a_1(T) = \lim_{n\to \infty} Pr\{n\omega^2 \leq T\}
\end{equation}
For the two comparisons, from the \cite{AN:52} tables $a_{MS+MS} > 0.5$ and $a_{WD+MS} < 0.5$, i.e., the probability of the $MS+MS$ model being drawn from the same distribution as the $MS+MS$ observed population is greater than 10\%, while the probability of the $WD+MS$ model being drawn from the same distribution as the $WD+MS$ observed population is at least 5\%. Given our original hypothesis $H_0: \bar{r} = 0.0$, the $MS+MS$ distribution comparison hypothesis fails, and the $WD+MS$ comparison hypothesis is accepted. We conclude then, that the distributions of observed and modeled $WD+MS$ angular separations are similar, but not for the $MS+MS$ distributions.
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{CDFResidualEstimatesFig13.pdf}
\caption{Plot of the Residual Square Model to Observed Cumulative Angular Separation Distributions for MS+MS pairs (in blue open circles) and WD+MS pairs (in red closed circles)}\label{fig:m}
\end{figure}
We further demonstrate the relationship between the models by graphing the residual of the angular separation distributions in Figure \ref{fig:m}. The residuals plotted have no particular linear trend, nor offset. This suggests either a difference in spread or mean. There are however, offsets in the $MS+MS$ comparison. These offsets cause the failure of the null hypothesis for the $MS+MS$ comparison. This suggests: (1) there are additional factors for which we have not taken into consideration (such as inclination) and/or (2) small systematic errors associated with our model.
\section{Conclusions}
We have shown that PMS mass loss significantly shifts and reshapes the distribution of $MS+MS$ separations. The semi-major axis of a wide binary is a difficult parameter to determine, requiring distance, inclination, period and other measurements. It is our hypothesis that a distribution of angular separations, preserves information about the effect of $PMS$ main-losses on the binary system. We take as proof, the following observation: Both our $MS+MS$ and $WD+MS$ models compare reasonably well to the observed LDS distributions. Our model, however, has only been affected by $PMS$ mass loss and a crude attempt has been made to account for other disrupting factors. This assumption and other first order approximation in the creation of the model, likely affect the goodness-of-fit test for the $MS+MS$ observed populations.
Figures \ref{fig:k} and \ref{fig:l} show the cumulative distribution of $\log{(s")}$ in the $MS+MS$ and $WD+MS$ pairs. The difference in observational and theoretical distributions is apparent.This difference is a function of both the incompleteness of the survey and secondary effects for which we have not taken into consideration (such as inclination or external perturbations). For example, observations and models have shown that metallicity may play a much larger role in the amount of post Main-Sequence mass-loss (see \citealt{ZH:11}).
Future work should include a more realistic model for stellar formation history rather than the constant birth function implemented here. Furthermore, this stellar birth model should be used to drive the initial conditions of both the metallicity and the mass distribution function. In view of the trends in the metallicity, especially as a function of stellar age \citep{ZH:11}, additional effect on the mass-loss of the population as a function of time would be expected. We have demonstrated in this paper, and it has been demonstrated in other simulations, that post-MS mass-loss on average results in an increase in semi-major axis for visual and fragile binaries. Changes to the metallicity history would be expected to influence orbital amplification as a function of time, in turn affecting the final angular separation distribution. Our photometric colors were based on a consistent metallicity for the population of binaries. Adjustments to the metallicity history would affect the associated color indices, as well as the apparent magnitude of the binary components. Again, this would effect the visual angular separation distribution.
The observed $MS+MS$ and $WD+MS$ distributions are affected by observational bias and/or physical orbital amplification. If orbital amplification results in an increase in orbitla separation, we conclude that binaries with small separations have expanded to middle separation binaries, middle binary separations have increased slightly, and binaries that were already wide may have been lost to dissociation. Our current simplistic model describes the orbital frequency distribution of FBs reasonably well. In general, our biggest issue is population size of the observed distribution. The statistically small observed population causes a large scatter in the frequency distribution, or rather a larger scatter than in the model distribution.
The DS code is currently deficient at both extremes of stellar mass ( i.e., $M < 0.07M_{\odot}$ and $M > 90M_\odot$ ). We do not have an evolutionary code applicable to these regimes, nor does our code include metallicity variation. However, such variance is not of major important in the modeling of disk stars. Future work will require such an extension to the SSE code and extrapolation of the Kurucz color/atmosphere models. Low-mass models are available and have been taken into account for masses as low as $M\sim0.075M_{\odot}$ \citep{BCAH:98}. While their contributions to the current observed sample of FBs are small, in order to produce a realistic model of the evolved population, they should be taken into account. This lower mass limit does not directly affect the current project.
\section{Acknowledgments}
All graphs in this paper have been made with the R-project program. The authors thank Jarrod Hurley for providing them a copy of the fast stellar evolutionary code SSE. We would also like to thank Dr. Matt Wood and Dr. St\'{e}phane Vennes for their support and advice. This project was supported in part by NASA grant Y701296 (T.D.O.) and NSF grants AST 02-06115 (T.D.O.) and AST0807919.
\section{Introduction}
Mass-loss due to various processes during the lifetime of a star will increase the semi-major axis of a wide binary \citep{DOM:63, VG:88}, what we refer to here as orbital separation amplification. It is expected that mass-loss during the stellar lifetime, in particular Post-Main Sequence (PMS) mass-loss, will statistically distort a frequency distribution of weakly interacting wide binary (Fragile Binary, FB) separations. Visual confirmation of such processes would provide tests of current theories of stellar evolution and fundamental information on the dynamics of Galactic structures \citep{W:88}.
As discussed by \cite{OS:07} and \cite{CHAN:07}, FB's provide insight into several astrophysically important problems that cannot be easily addressed by other means. In particular, such fragile pairs provide valuable constraints on the Galactic tidal potential, stellar and ISM perturbations, and PMS mass-loss, all of which will affect orbital separations. A number of early papers have addressed this topic: \cite{EGG:65} studied the colors and magnitudes of 228 visual binaries (VBs), deriving an $M_v$ vs. $R-I$ calibration from trigonometric parallaxes. \cite{VD:80} showed that, using proper motions from micrometer measurements, a pair could be deemed gravitationally bound and thus, through statistical studies and interferometry, orbital parameters could be determined. \cite{TRIM:74} performed a statistical study of the observed populations of binary systems (798 VBs) and was one of the first to note the peak in the mass ratio (q) distribution at $q\sim1$.
In recent years, computing power has improved to the point that modeling the evolution of whole stellar binary populations is now possible. Examples include: \cite{W:88, VG:88}, and more recently \cite{HTAP:01}. Prior work has shown evidence for orbital separation amplification by as much as half an order of magnitude in binaries, where at least one of the partners has undergone PMS mass loss \citep{HAD:66, G:86, WO:92}. Considering that the majority of all stars are born with one or more companions as a result of the initial conditions required for a stable system \citep{M:94}, the importance of understanding how mass-loss affects the evolution of binary systems cannot be over stated. In addition, the initial-to-final mass relation (IFMR) for WD stars is at present very poorly constrained \citep{WEID:93, ZH:11}. Proper modeling of FB orbital separation amplification can provide a valuable independent estimate of this important relation.
The distribution of $WD+MS$ pair separations is most strongly affected by the PMS (i.e. red giant phase) mass-loss. FB systems typically have component separations of $\sim10^3 AU$ \citep{ZI:84}; this results in negligible tidal interactions and mass transfer between companions. Such pairs evolve as if they are two separate but coeval stars. Once mass-loss has ended, orbital evolution is limited to the effects of external forces such as stellar encounters, giant molecular cloud encounters, gravitational tidal forces, and rare encounters with other stars \citep{CHAN:04,JIAN:10} .
In this study, which updates and improves upon our previous work \citep{JOHN:07}, we compare a theoretically generated population of FB angular separations to the observed angular separation distributions of FBs found in a large-scale survey. We constructed a simulation code, which we refer to here as the Double Star (DS) program, to generate the initial parameters and distributions of the tested population. The DS program evolves the binary components using the Single Star Evolution (SSE) code \citep{HPT:00} over the lifetime of the simulation. The DS program is also responsible for the evolution of individual orbital elements and translation of the modeled physical properties into observable features. The results of the translation is a population of theoretical FBs, with observable properties.
\section{Observational Data}
It should be possible to produce a statistically significant frequency distribution of angular separations for FBs with various constituents, by sampling a large survey such as the Luyten Double Star (LDS) Catalog \citep{LDS:69}. While the LDS catalog, with 6124 pairs and multiples, does not have the volume of more recent surveys such as the Sloan Digital Sky Survey (SDSS), it provides a well-observed large sample to compare to our theoretical models. The identification of FB candidates requires selection of an upper bound separation. Luyten found pairs separated by as much as $400"$. Due to the POSS survey plate scale, Luyten's minimum angular separation limit was $\sim0.5"$. We will use this sample as a first test of our programme, while we note that further samples of fragile binaries are currently being constructed (e.g., \citealt{SES:08, QUIN:09, LONG:10, ZHA:11}).
\section{Initial Parameters}
To build a synthetic model of the local Galactic FB population, we first proposed a set of initial parameters that defined the individual components of our binaries, as well as the binary system separations. We used a combination of published functions from previous theoretical models and observational surveys. When there were no such references, or when the functions published were not appropriate for FBs, we implemented physically reasonable distributions.
\subsection{Mass Generating Function}
Following \cite{KTG:93} we assumed a Salpeter distribution of ZAMS stars using a variation of Eggleton's mass generating function (Equation \ref{eq:a1}).
\begin{equation}\label{eq:a1}
M(x) = 0.33( \frac{1.0} {(1.0 - x)^{0.75} + 0.04(1.0 - x)^{0.25}} - \frac{1.0}{1.04} (1.0 - x)^2 )
\end{equation}
Where x is a uniform randomly distributed number and M(x) is the randomly generated, restricted, mass of a ZAMS star. For this project, we initially assumed a minimum mass of $0.07M_{\odot}$(spectral type-T), given by \cite{CPH:03}. From this we produced an initial distribution of primary masses (Figure \ref{fig:a}).
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{InitialFinalMassPrimaryFig1.pdf}
\caption{Histogram of the distribution of primary masses from initial generation (in red with 45 deg. hatching) and after evolution and orbital expansion (in blue with -45 deg. hatching). The difference in population size is resulting from the dissolution of some binaries from the original population. Each bin is of size $0.05$ in $\log{(M_\odot)}$}\label{fig:a}
\end{figure}
\subsection{Mass Ratio}
The mass ratio, $q$, is one of the most poorly determined variables in a wide binary system. Current observed distributions fall into two categories: (1) a distribution that peaks at $q\sim1$ \citep{TRIM:87} or (2) a bimodal distribution with peaks at $q\sim 0.2$ and $q\sim 0.7$ \citep{HALB:03}. The former are explained in terms of turbulence and fragmentation during the binary formation process \citep{LUC:79}. This tends to result in near equal clumps of would-be stars forming out of a single GMC. Bimodal distributions seem to be created by selection effects (spectroscopic methods vs. common proper motion) and evolutionary effects (binary capture, disruption, mass-loss, mass-transfer,etc.), for references of either see \cite{KRU:67, P:71, SHU:81}. For the purpose of finding an initial mass ratio function (IMRF), we will tend to use the discussion of distributions which tend towards a maximum of one in the set 0$\leq q \leq$1(unity-like), as these will be the most likely candidates to explain a mass-ratio distribution created by a GMC to stellar binary creation process. After multiple trials with various models, we chose a random uniform distribution, where 0$\leq q \leq$1 and $\lambda$ is a constant (Equation \ref{eq:b1}).
\begin{equation}\label{eq:b1}
f(q) = \lambda
\end{equation}
A constant mass ratio distribution function has the added advantage of producing an undistorted secondary mass distribution, implying that the same physical processes produce the primary and secondary components. Using an acceptance-rejection method of random variate generation \citep{LAW:05}, we produced the distribution of secondary masses seen in Figure \ref{fig:b}. In addition we restricted all lower mass components to $M_{star}\geq0.07M_{\odot}$. This has the added effect of producing a tendency for $q\sim1$, as seen in Figure \ref{fig:c}, in accord with prior observational studies.
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{InitialFinalMassSecondaryFig2.pdf}
\caption{Histogram of the distribution of secondary masses from initial generation (in red with 45 deg. hatching) and after evolution and orbital expansion (in blue with -45 deg. hatching). The difference in population size resulted from the dissolution of some binaries from the original population.Each bin is of size $0.05$ in $\log{(M_\odot)}$}\label{fig:b}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{InitialFinalMassRatioFig3.pdf}
\caption{Histogram of the distribution of mass ratios from initial generation (in red with 45 deg. hatching) and after evolution and orbital expansion (in blue with -45 deg. hatching). The difference in population size resulted from the dissolution of some binaries from the original population. Bin size is of size $0.05$ in $q$-ratio}\label{fig:c}
\end{figure}
\subsection{Orbital Parameter Distribution Functions}
In this paper, we focused on the development of two specific orbital element distributions, specifically the distribution of semi-major axis (separation) and the distribution of eccentricity. From these parameters, we develop elliptical binary mechanics for each binary in our initial population. The statistical characteristics of the separation distribution for the un-evolved binaries, as well as the distribution characteristics for the primary masses, secondary masses, and periods is given in Table \ref{tab:a} at the end of this section.
\subsubsection{Initial Separation Function}
The separations of binary systems is a difficult observable parameter, requiring distance, inclination, eccentricity and mass. One of the first and most often used references to separation distribution was given by \cite{O:24}, and implemented by \cite{ZI:84} and \cite{P:88}. It is a simple uniform log distribution:
\begin{equation}\label{eq:c1}
f(a) = \log(a)
\end{equation}
Where $a \geq 100 A.U.$, approximately the closest separation where post-MS mass exchanges are unlikely to have occured. We transform equation \ref{eq:c1} into a random variate separation generating function (Equation \ref{eq:d1}).
\begin{equation}\label{eq:d1}
a(X) = 10^{2.0X+2.0}
\end{equation}
Where $0 \leq X \leq 1$ and $X$ is a uniformly distributed random number. Similarly, we used the separation and masses now assigned to each pair, along with Kepler's $3^{rd}$ law to calculate the period for each binary.
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{InitialFinalSeparationFig4.pdf}
\caption{Histogram of the distribution of separation from initial generation (in red with 45 deg. hatching) and after evolution and orbital expansion (in blue with -45 deg. hatching). The difference in population size resulted from the dissolution of some binaries from the original population. Each bin is of size 0.05 in $\log{(A.U.)}$}\label{fig:d}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{InitialFinalPeriodFig5.pdf}
\caption{Histogram of the distribution of periods from initial generation (in red with 45 deg. hatching) and after evolution and orbital expansion (in blue with -45 deg. hatching). The difference in population size resulted from the dissolution of some binaries from the original population. Each bin is of size 0.1 in $\log{(Myr)}$}\label{fig:e}
\end{figure}
In addition to the physical limit of $a\geq100 A.U.$ for non-interacting pairs, we also imposed observed limitations on our binary pairs. We rely on the \cite{ZI:84} definition of CPMB/FB, VB, and SBs and limit the initial smallest separation that a common proper motion binary can have as $100 A.U.$. This is wide enough that the two components of the pair are individually resolvable and the effects of Roche Lobe interaction and component-to-component tidal interactions are minimal compared to Galactic effects and stellar GMC encounter perturbations.
We additionally imposed a maximum of $10^4 A.U.$ , which appears to be the maximum theoretical separation in wide pairs \citep{WW:87} imposed by gravitational disruptions of the Galaxy and stellar perturbations. An algorithm was developed to impose this upper limit continuously over the evolution time frame and will be discussed below in section 4.3.
\subsubsection{Eccentricity}
The simulation derives the initial distribution of eccentricities following \cite{HEG:75}. The initial distribution of eccentricity for wide binary pairs is given in Equation \ref{eq:e1}.
\begin{equation}\label{eq:e1}
f(e) = 2e
\end{equation}
We transform this distribution into the random variate generating function (Equation \ref{eq:f1}).
\begin{equation}\label{eq:f1}
e(X) = \sqrt{X}
\end{equation}
Where $0 \leq e < 1$ and $X$ is $0 \leq X \leq 1$, a uniformly distributed random number. This distribution was derived based on assumptions of a Maxwellian particle distribution. It becomes apparent from the generating function that binaries, with this eccentricity distribution, are not favored to have initial circular orbits. In fact, using this distribution one can easily see that the expected initial density of circular orbits is zero. For wide binary pairs we can convince ourselves that this is true, as these pairs tend to have coeval components but have had little or no interaction beyond the gravitational binding which makes them a pair (as discussed in the mass ratio section). This creation via loose interaction would tend to result in more initially eccentric orbits.
\begin{table}[h!]
\caption{Statistical Characteristics of the Un-evolved Distributions}\label{tab:a}
\begin{center}
\begin{tabular}{c|cccccc}
\multicolumn{7}{c}{} \\
\hline
Metric & Average & StdDev & Skew & Kurtosis & Min & Max \\
\hline
Mass 1($\log{M_{\odot}}$)& -0.28& 0.41& 0.58& 0.79& -1.10& 1.93\\
Mass 2($\log{M_{\odot}}$) & -0.55& 0.40& 0.92& 1.03& -1.10& 1.84\\
Separation ($\log{A.U.}$) & 3.00& 0.58& -0.002& -1.20& 2.00& 4.00\\
Period ($\log{Myr}$) & -1.46& 0.826& -0.008& -1.08& -4.04& 0.39\\
\hline
\end{tabular}
\end{center}
\end{table}
\subsection{Birth Rate}
Our experiments incorporated an initial stellar burst of $10^2$ binaries with \cite{WK:04} constant repeating burst model of the stellar birth rate, $10^2$ binaries being created every $10 Myrs$, resulting in a total initial population of 100,100 binaries. All bursts have identical initial factors and distributions, resulting in a continuous and smooth distribution \citep{HPT:00} in the initial population. The only references that challenge this approach, suggest that metallicity may play a role in distorting the initial distribution function. However, the region of space that we are attempting to model is limited to the solar neighborhood (i.e. distances under $100 pc$), FBs in this region tend to have solar metallicitys. We therefore assumed solar metallicity for all ZAMS stars. Future work may address a more realistic history of star formation rate, dervied for the solar neighbourhood \citep{HER:00}.
\subsection{Color Estimation}
The goal of this project is to compute a theoretical population of FBs that accurately compared to observed angular separation distributions. This requires a transformation of our physical characteristics for each binary system into observables such as color, magnitude and angular separation. We employed a mix of \cite{KV:06, K:79, HAW:02} temperature vs. color relations for MS and WD stars, to produce the $ugriz(J)$ absolute magnitudes required for comparison.
We divided stars into four categories: MLT, WD, MS, and other. MLT refers to the spectral type of the star: M, L or T, corresponding to effective temperatures from $\sim3800K$ to $1000K$ \citep{R:00, R:02}. The spectral type of a star below $3400K$ is assumed to be dependent only on temperature. We justify this estimation with three points: (1) Reliable statistics are not available for surface gravities of low mass stars (early-T and late-L); (2) The current classification scheme for MLT dwarfs is based mostly on line features. A number of flux ratios ($H_2O, CO_2, CH_4,$etc) show linear relations with spectral type \citep{BURG:02}; (3) \cite{STE:01} used a hybrid approach to show that temperatures and spectral types scale linearly with each other for L-type dwarfs.
Next, we used a combination of spectral type and temperature relations to transform the observed absolute magnitudes by \cite{HAW:02} in $r^\star i^\star z^\star J$ to produce a temperature vs. absolute magnitude relation. We then used this relation to produce colors and errors for all MLT stars. Neither the LDS nor the SDSS survey provides apparent $J$ magnitudes. Future surveys however will produce near-infrared colors (y). At such a time our code can be updated and expanded.
For WD and MS stars, we use models provided by \cite{KV:06} and \cite{K:79}, respectively. These atmospheric models are grids of color as functions of temperature and gravity. We employed a 2nd/3rd order unevenly spaced 2-dim interpolation subroutine to determine both expected absolute magnitude and error in absolute magnitude given $T$ and $\log(g)$. As before, we estimated $u^\star g^\star r^\star i^\star z^\star$ colors. This allowed simple estimates of distance modulus for each pair.
Models used to compute discrete grids are available\footnote{http://www.stsci.edu/hst/observatory/cdbs/k93models.html}, derived from the Kurucz stellar atmosphere models. For MS stars temperatures range from $3,170K\leq T \leq 42,000K$ and $0\leq \log(g) \leq5$. For WD stars, two sets of grids were used: a fine resolution with range $7,000K\leq T \leq84,000K$ and $7\leq \log(g) \leq9.5$ in $0.5$ intervals and a broad resolution set with $4,500K\leq T \leq84,000K$ and $7\leq \log(g) \leq9$ with $1.0$ gravity intervals. The fine resolution set was used for stellar temperatures with $7,000K \leq T \leq84,000K$. For temperatures above $84,000K$ and below $7,000K$, the broad resolution set was used.
\subsection{Distance}
In order to compute the distance modulus, a variety of models were tried, including a synthetic Galactic population model that includes distance as a function of mass \citep{RRD:03}. Initially, we produced a distance model that best fits the observed LDS distribution with the given parameters. The adopted distance distribution had the form:
\begin{equation}\label{eq:g1}
f(d) = \log(d)
\end{equation}
This equation was then transformed into a random variate separation generating function:
\begin{equation}\label{eq:h1}
a(X) = 10^{3.0X+1.0}
\end{equation}
Where $0 < X < 1$ and $X$ is a uniformly distributed random number. We expect this will, through random evolutionary processes, produce the synthetic lognormal distribution of distances that \cite{RRD:03} obtained. We also assume that binary and single star distance distributions are identical.
\subsection{Stellar Evolution Code}
One of the two major tasks in this project is the ability to accurately evolve a theoretical star while keeping track of its mass, luminosity, and temperature; as previous analysis used analytical, rather than physically-motivated, mass-loss rates (e.g. \citealt{DOM:63, DOB:98, RAHO:09}). For a population of binary stars this requires substantial computational resources. As mentioned above, the Single Star Evolution (SSE) program \citep{HPT:00} was adapted for this effort. The basis of the SSE is a set of comprehensive analytical formulae, that accurately track the evolution of a star with given initial mass and metallicity. The goal of the SSE program was to allow for the rapid calculation of a number of stellar properties for a given stellar lifetime, thus facilitating the production of a population synthesis model that could produce a statistically significant sample of binaries in a reasonable computational timeframe.
The key to the SSE code's accuracy and speed is its process style, which employs a combination of interpolation formulae and tabular construction. The HPT group uses tabulated stellar models \citep{SSMM:92, CMMSS:93, MSM:98}. \cite{PSH:98} later expanded the range of mass and metallicity of the tables incorporated in the SSE program. Analytic approximations to the movement of a star in the Hertzsprung-Russell diagram (HRD) are a simplistic approach to a complex problem \citep{EFT:89}. However, supporting these calculations with tabulated stellar models allows a more precise estimation of physical properties and greatly decreases computational time.
The SSE does not take into account stars with solar masses lower then $0.2M_\odot$. For stars with masses lower then this limit we adopted the \cite{BCAH:98} low-mass stellar evolutionary models. We do not expect such low mass stars to undergo stellar mass evolution, but we do expect small changes in other stellar parameters over time. Using these evolutionary models we were able to determine the initial and final values of temperature and gravity based on initial mass and metallicity.
We selected the following parameters for the SSE code: no wind-loss, a Reimers Coefficient for mass-loss of 0.5, and a constant metallicity distribution of 0.1. This set of parameters provides an initialization point from which to develop further analysis, while still generating realistic results. Future research may include analysis (ANOVA) to determine of optimal parameter distributions and values for comparison of theoretical and observed population distributions. The authors of the SSE code have subsequently developed a follow-up algorithm, the Binary Star Evolutionary code. For development purposes, this new program was not used, as the primary concern of the DS-code is the generation of evolved binary populations and information, relating to fragile coeval binaries. The authors feel that wrapping the elliptical orbital computations, stellar population generating functions, and apparent magnitude translations around the original SSE code, allows for the maximum amount of computational efficiency and provides the necessary forensic test points for further experimentation, specifically in the generation of an ANOVA experiment for optimal parameter and distribution selection.
\subsection{Orbital Evolution}
In the present model, we assumed isotropic non-conservative mass-loss for both stars, both during the MS and PMS lifetimes. Since we modeled what could be considered two phases of mass-loss, slow (MS) and sudden (PMS), we found a two stage evolutionary code was needed. Following \cite{VG:88, HAD:66, KOP:78}, we considered the effect double, non-interacting, mass-loss, has on the orbital parameters of the binary. For this experiment we modified versions of \cite{VG:88} equations into the simulation code presented here. We allowed no accretion onto either star from its companion and no accretion from the ISM or any outside bodies. The addition of perturbation forces from exterior sources beyond those addressed within this work, such as loose tertiary components, dense stellar populations, or other “chance” encounters were not modeled here, however we would recommend \cite{HEG:96} as starting research in the matter.
We can show via adjustment of the mass-loss orbital parameter equations that the perturbing force due to an isotropic variation of mass is tangential to the orbit, thus only the planar orbital elements will be perturbed. This assumption holds when the mass-loss timescale of the binary is much greater than the orbital timescale (period). These perturbation equations are given as a function of time, the logarithm of the total system mass, the eccentric anomaly, and the eccentricity are given as the equations \ref{eq:i1} and \ref{eq:j1}.
\begin{equation}\label{eq:i1}
\frac{d\log{a}}{dt} = -\frac{1+e\cos{E}}{1-e\cos{E}}\frac{d\log{M}}{dt}
\end{equation}
\begin{equation}\label{eq:j1}
\frac{de}{dt} = -\frac{(1-e^2)*\cos{E}}{1-e\cos{E}}*\frac{d\log{M}}{dt}
\end{equation}
Where $\frac{d\log{M}}{dt}$ is the rate of mass-loss over the time period of interest (iteration), and $E$ is the eccentric anomaly. Note, that when expressing these equations in terms of the eccentric anomaly, Kepler's equations is no longer valid and should be replaced with the equation,
\begin{equation}\label{eq:k1}
\frac{dE}{dt} = \left( \frac{2\pi}{P} + \frac{d\log{M}}{dt}\frac{\sin{E}}{e}\right)*\left(1-e*\cos{E}\right)^{-1}
\end{equation}
The effect that this MS mass-loss has on the binary system is documented by \cite{VER:69}, \cite{DOM:63}, \cite{DOB:98} and \cite{RAHO:09}. Here, Verhulst focused the majority of the research on the effect a Jeans-Eddington mass-loss relation would analytically have on the eccentricity of the binary. It was shown that one can expect periodic variation about a mean eccentricity for a binary with isotropic continuous mass-loss depending on the parameters of the Jeans-Eddington mass-loss relationship. \cite{HAD:63} demonstrated similar effects, showing that with decreasing system mass the eccentricity of the system will sinusoidally increase. More recently \cite{VEA:11} demonstrated orbital parameter evolution for low q-ratio binaries ($q < 0.1$, exoplanets and smaller bodies). They showed that, at least for these conditions, the effect of mass-loss, varies greatly depending on the initial eccentricity distribution.
Given that our mass-loss is a function of the stellar evolution code designed into the SSE, with the additional consideration of the low mass evolution models provided by \citep{BCAH:98}, we expect a similar result (increasing non-linear, sinusoidal eccentricity as a function of mass-loss). However we expect that the numerical (and probabilistic) nature of our experiment would demonstrate a different (and potentially more complete) relationship. For situations where mass-loss occurs on a timescale less than that of the period of the binary, we follow the “sudden” mass-loss equations \ref{eq:l1} and \ref{eq:m1} outlined by \cite{HIL:83}.
\begin{equation}\label{eq:l1}
a = \frac{a_0}{2} \left( \frac{1 - \frac{\Delta M}{M}}{\frac{1}{2} - (\frac{a_0}{r} ) *\frac{\Delta M}{M}} \right)
\end{equation}
\begin{equation}\label{eq:m1}
e = \left(1 - (1-e_0^2)*\left[\frac{1 - \frac{2a_0}{r}\frac{\Delta M}{M}}{(1 - \frac{\Delta M}{M})^2}\right]\right)^{0.5}
\end{equation}
We have further transformed the original equations into a function of the mass-loss ratio over a given defined time period (delta mass over original mass). The increase in separation and eccentricity, is a function of how close to periapsis or apoapsis the binary is at the time of mass-loss. It can be shown that for binary separation distances less than the semi-major axis (which is dependent on the eccentricity); sudden mass-loss can drive the binary together (decrease orbital separation).
Similarly, as stated by \citeauthor{HIL:83}, if mass-loss occurs in a binary with a very eccentric orbit, the average post-explosion eccentricity will be less than the pre-explosion eccentricity. Conversely, given the position of the binary, the separation can be greatly amplified, resulting in populations of $WD+MS$ and $WD+WD$ pairs with separations on average larger than their $MS+MS$ counterparts. We can demonstrate for given eccentricity and location in the orbit, post-main sequence mass-loss can result in a perturbation of the orbital eccentricity beyond a bound elliptical orbit, i.e., $e \geq 1$. This is demonstrated in Equation \ref{eq:n1}.
\begin{equation}\label{eq:n1}
\frac{r}{2a} < \frac{\Delta M}{M}
\end{equation}
It is apparent, that increasing eccentricity can pose a problem for the binary system. The potential for binary separation can result from increasing the eccentricity of the binary beyond $(e \geq 1)$. Thus, we expect that the increase in eccentricity, and therefore the associated orbital velocities, in combination with the wide orbital amplitudes, will result in a number of binary systems being perturbed out of their orbits (dissolution via external perturbation). It should be noted that decreases in eccentricity resulting from sudden PMS mass-loss, coupled with the lack of mass-loss associated with white dwarf components, can result in the effective circularization of the orbit. This is apparent from the disproportionate number of binaries in the $WD+MS$ and $WD+WD$ population evolving to a final eccentricity under 0.1 in our simulations.
To model the effect high eccentricity has on the binary system, specifically on the dissolution rate of the binary populations resulting from isotropic mass-loss and galactic tidal forces, we have implemented a function that flags evolved binary systems which have eccentricity greater than 0.95. This, combined with the \cite{AL:72} tidal limit provides a first order approximation to dissolution effects on the binary systems. Other improvements are outlined in the Binary Dissolution section
\subsection{Binary Dissolution}
Even a simple model for the orbital evolution of wide binaries must include the influences of stellar encounters, Giant Molecular Clouds (GMCs) and the Galactic tidal field \citep{WEIN:88}. For the wide separations of FBs, even minor forces such as these operating over the course of millions of years, can affect the binary separations, eventually dissolving them. \cite{AL:72} described a tidal limit for wide binaries, where the gravitational forces from Galactic tidal perturbations outweigh the gravitational attraction between the binaries themselves(Equation \ref{eq:o1}).
\begin{equation}\label{eq:o1}
a_T \sim 2.07 \times 10^5 A.U. \left( \frac{M_1+M_2}{M_\odot}\right)^{1/3}
\end{equation}
This relation does not include either GMC or stellar encounters, but it does provide an initial cut off for galactic tidal disruption as a function of mass. We take the \cite{AL:72} equation as a first approximation to a dissolution algorithm. A more realistic approach would include all disruption effects. A rigorous treatment is beyond the scope of the current project. For a first-order approximation to the dissolution time problem, we used the \cite{WW:87} Galactic tidal perturbation formula for a $2M_\odot$ binary system with $q = 1$. The expected half-life ($t_{1/2}$) of an FB system with no consideration for the GMC encounter is given by Equation \ref{eq:p1}.
\begin{equation}\label{eq:p1}
t_{1/2} = 4.3 \times 10^3 Myr \left(\frac{a}{2.07\times10^4}\right)^{-1.4}
\end{equation}
For wide binaries with consideration for the GMC local density of $3.6\times10^8 pc^{-3}$, this relation becomes Equation \ref{eq:q1}.
\begin{equation}\label{eq:q1}
t_{1/2} = 1.32 \times 10^3 Myr \left(\frac{a}{2.07\times10^4}\right)^{-1.4}
\end{equation}
We define $t_{1/2}$ as the $50\%$ likelihood survival time for a binary with separation $a$ in $A.U.$ While this cannot be used to estimate the expansion caused by a GMC and perturbations, using a binomial distribution selection program we produced a dissolution algorithm to remove binaries from the sample that likely would have dissolved in the evolutionary time frame given. An evenly distributed binomial random variable with two possible outcomes ($50/50$) was used to simulate $t_{1/2}$.
Lastly we chose a simple ratio of survival times to linearly scale the probability mechanism. Binaries with a separation greater than the $50\%$ likelihood survival time will have a ratio greater than one and a better chance of dissociating (i.e. $50/50$ turns into $75/25$), while binaries with a separation less than the half point limit, will have a ratio less than one and a worse chance of dissociating (i.e. $50/50$ turns into $25/75$). For example: if a binary has a separation of $10^4 A.U.$ its half-life is $3655 Myr$. If a binary evolves for $3655 Myr$, there will be a $50\%$ chance of it being selected to dissolve. If the binary evolves for $10,000 Myrs$ then: $\frac{3655}{10000} = 0.37$ and $50\ast0.37\sim18$. Therefore instead of a $50\%$ chance of our binary surviving, there is an $18\%$ chance of it surviving. Binaries that have been selected using our binomal random number generator, are removed in the posterior population samples.
We note again that the dissolution algorithm is not a continuous function; its purpose is to generate in the final sample the expected number of dissolved binaries in a given population. The linear estimation of what should be an exponential (Poisson) function, will cause the dissolution algorithm to under-sample the number of dissolved binaries in our distribution. The \textit{dissolve} or \textit{survive} approach also does not allow for external forces and perturbations to expand the binaries. However, as stated above, the residual term from our model vs. observation comparison will provide some insight to the degree of which the sum of these external forces will effect our model. Binaries that have been removed are not considered in the following analysis.
\section{Results}
Rather then modeling individual binaries, we attempted to produce theoretical separation frequency distributions to compare with observational statistics. Histogram bin sizes were chosen to be similar for like quantities across the various cases. We then transformed our model parameters into model observations and compared our results with the observed \cite{LDS:69} separation distribution. The statistical characteristics of the distribution of primary masses, secondary masses, separations and periods for the evolved binaries is given in Table \ref{tab:b}.
\subsection{Mass Graphs}
For comparison purposes, we plotted orbital frequency distributions for both the initial and posterior populations. Figure \ref{fig:a} and Figure \ref{fig:b} include not only the initial burst binaries but also those born in bursts later on. Thus, not all binaries formed at $t_0 = 10^4$ in the initial sample, while the posterior populations have identical $t_{final}$.
Figure \ref{fig:a} and Figure \ref{fig:b} are a representation of the same sample at an age of $\sim10 Gyr$. In Figure \ref{fig:b}, the random q-ratio has produced, in the secondaries, a distribution similar to the primaries. There are two artifacts in our secondary mass population that may not be physically realistic. First is the cutoff at $0.07M_\odot$, resulting from adopting the limit suggested by \cite{CPH:03}. However, two points mitigate this arbitrary cutoff: (1) as shown by \citeauthor{CPH:03}, the mechanism of star formation does have a minimum mass, and (2) imposing our q values to be a number between 0 and 1 causes the average secondary mass value to be less than the primary, as expected. According to \citeauthor{CPH:03}, the initial mass function at lower masses is poorly determined. Moreover, the LDS and SDSS surveys have detected relatively few M, L, and T dwarfs, particularly in binary systems, and not yet enough to warrant an attempt to model this region.
A look at the posterior distributions in Figures \ref{fig:a} and \ref{fig:b} shows the impact of mass loss. Peaks in the distribution near the $0.5M_{\odot}$ to $1.0M_{\odot}$ range and a decrease of high mass stars is a sign that some of the MS stars have evolved to less massive WD stars, or Neutron Stars and Black Holes (NS and BH). At $10 Gyr$, in some pairs both components have evolved to $\sim0.6M_{\odot}$ ($WD+WD$).
\begin{table}[h!]
\caption{Statistical Characteristics of Evolved Distributions.}\label{tab:b}
\begin{center}
\begin{tabular}{c|cccccc}
\multicolumn{7}{c}{}\\
\hline
Metric & Average & StdDev & Skew & Kurtosis & Min & Max \\
\hline
Mass 1 $(\log{M_{\odot}})$& -0.36& 0.30& -0.40& -0.36& -1.10& 1.01\\
Mass 2 $(\log{M_{\odot}})$ & -0.58& 0.33& 0.27& 10.990& -1.10& 0.81\\
Separation $\log{A.U.}$& 2.97& 0.55& 0.069& -1.04& 1.74& 4.88\\
Period $\log{Myr}$ & -1.48& 0.84& 0.075& -1.02& -3.60& 1.11\\
\hline
\end{tabular}
\end{center}
\end{table}
\subsection{Mass Ratio}
Figure \ref{fig:c} is a histogram of the mass ratio distribution for the initial population of binaries and the posterior population of binaries. There is a trend towards $q\sim1$ in the initial population, as expected from the mass ratio generating function. A unique by-product of evolution is evident in the posterior population of the $q$-distribution: high mass MS stars have evolved into lower mass WDs. Now the secondary has a higher mass than the WD, resulting in a distribution of binaries with $q > 1$. This time, however, the $q > 1$ range represents the region where the primary has undergone PMS mass-loss and now has less mass than the secondary, i.e. one or both of the components of the binary have evolved into degenerate stars (WD, NS, or BH). The additional evolution to WD of the secondary would likely only cause a density change in the population above $q=1$ or in the compact region near $q = 0.9$.
\subsection{Separation and Period}
Figures \ref{fig:d} and \ref{fig:e} present comparisons of the initial and posterior populations of binaries, the separations (in $\log{(A.U.)}$) and the period distributions (in $\log{(Myrs)}$). The sharp observed initial distribution of separations is a result of the initial parameters: a cutoff at $10^4 A.U.$ (GMC birth cloud), and a lower cutoff ($100 A.U.$) based on the definition of what qualifies as a FB. We impose an \"{O}pik distribution on top of this to produce the desired frequency distribution. The scatter is a result of binning and low number density. At higher densities ($10^5$ binaries and greater), this scatter becomes negligible. Our initial distributions are leptokurtic (i.e. flat-top), which is expected in $\log{(a)}$ due to our initial assumptions of a uniform log distribution. The skewness of the initial distribution $\sim0$ results from the flat sides and little or no discernable tail of the distribution. We see similar characteristics in the initial $\log{(P)}$ distributions(Figure \ref{fig:e}), however the edges of the distribution have been rounded out resulting from the effect the mass distribution has on the generated distribution. Our initial limits in the separation distribution results in the min ($a > 10^2 A.U.$) and max ($a < 10^4 A.U.$), respectively.
Comparisons to posterior distributions in Figure \ref{fig:d} and Figure \ref{fig:e} demonstrate that the maximum separation is now $\log{(A.U.)} = 4.88$ ($75,850 A.U.$). Two other parameters have also changed: the shift to higher separations has depleted the number of close binaries and increased the higher separation binaries. This has effectively rounded out the histogram of separtions (kurtosis $= -1.04$) and narrowed the histogram, (standard deviation $= 0.55$) compared to the standard deviation of $0.58$ in the initial distribution of Figure \ref{fig:d}. Figure \ref{fig:e} shows a similar trend in variance, average and kurtosis. The period has the addition of a randomly generated mass factored into its calculation; this in turn has affected the kurtosis and skewness which now reflect a more normal distribution, per the central limiting theorem (CLT), as well as making the curve appear smoother.
\subsection{Analysis by Binary Category}
Figures \ref{fig:f} to \ref{fig:i} are overlay plots of the three binary classes $MS+MS$, $WD+MS$ and $WD+WD$. All distrbutions are derivitives of the posterior population distribution described in the previous section; the "other" catagory refers to binaries with either component not in the MS or WD phase (RG, BH, NS...etc), the breakdown of population distribution types is given in Table \ref{tab:c}. The statistical properties of each of the three samples are given in Tables \ref{tab:d} - \ref{tab:f}.
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{FinalAllPrimaryFig6.pdf}
\caption{Histogram of the distribution of primary masses from the MS+MS populaton (in red with 45 deg. hatching), from the WD+MS population (in green with 135 deg. hatching) and the WD+WD population (in blue with 0 deg. hatching). Note the mixed colors represent overlaping distributions, bin sizes are in 0.05 $\log{(M_{\odot})}$ intervals.}\label{fig:f}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{FinalAllQFig7.pdf}
\caption{Histogram of the distribution of mass ratios from the MS+MS populaton (in red with 45 deg. hatching), from the WD+MS population (in green with 135 deg. hatching) and the WD+WD population (in blue with 0 deg. hatching). Note the mixed colors represent overlaping distributions, bin sizes are in 0.05 intervals.}\label{fig:g}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{FinalAllSeparationFig8.pdf}
\caption{Histogram of the distribution of separation from the MS+MS populaton (in red with 45 deg. hatching), from the WD+MS population (in green with 135 deg. hatching) and the WD+WD population (in blue with 0 deg. hatching). Note the mixed colors represent overlaping distributions, bin sizes are in 0.05 $\log{(A.U.)}$ intervals.}\label{fig:h}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{FinalAllPeriodFig9.pdf}
\caption{Histogram of the distribution of period from theMS+MS populaton (in red with 45 deg. hatching), from the WD+MS population (in green with 135 deg. hatching) and the WD+WD population (in blue with 0 deg. hatching). Note the mixed colors represent overlaping distributions, bin sizes are in 0.1 $\log{(Myr)}$ intervals.}\label{fig:i}
\end{figure}
\begin{table}[h!]
\caption{Distribution of Evolved Binaries in Posterior Population. (Population size is 90,523).}\label{tab:c}
\begin{center}
\begin{tabular}{c|cccccc}
\multicolumn{5}{c}{} \\
\hline
Metric & $MS+MS$ & $WD+MS$ & $WD+WD$ & Other \\
\hline
Count & 76,886& 7,327& 4,791& 1,519\\
Percent & 84.94\%& 8.09\%& 5.29\%& 1.68\%\\
\hline
\end{tabular}
\end{center}
\end{table}
\subsubsection{MS+MS}
The $MS+MS$ parameters are relatively unchanged and appear almost identical to the initial distribution plots in $q$-ratio, $\log{(Myr)}$ and $\log{(A.U.)}$. However, the $MS+MS$ mass plots do show missing upper mass stars. These, of course, have been transferred to the other groups as WD, BH, NS, or other stars. Similarly, a good number of higher separation binaries have dissolved, resulting in the observed change in skewness (observed asymmetry in separation distribution). Of the given model (90,523 binaries), 76,886 ($84.94\%$) binaries remain in the $MS+MS$ category.
\begin{table}[h!]
\caption{Statistical Characteristics of Evolved MS+MS Distributions}\label{tab:d}
\begin{center}
\begin{tabular}{c|cccccc}
\multicolumn{7}{c}{}\label{tab:d} \\
\hline
Metric & Average & StdDev & Skew & Kurtosis & Min & Max \\
\hline
$\log{a} (A.U.)$ & 2.94& 0.553& 0.072& -1.12& 2.00& 4.00\\
$\log{P} (Myr)$ & -1.49& 0.841& 0.066& -1.05& -3.28& 0.39\\
\hline
\end{tabular}
\end{center}
\end{table}
\subsubsection{WD+MS}
We observe the peak mass distribution of the primary in the $WD+MS$ (i.e., the $WD$) at slightly less then unity ($WD$ with masses of less then one solar mass). In addition, the secondary is by definition an $MS$ star (has not evolved yet). The secondary mass distribution has a peak at $\sim0.7_\odot$, the approximate turn-off-mass for Disk stars, thus restricting the $q$-ratio distribution. The histogram for $WD+MS$ $q$-ratio values (Figure \ref{fig:g} ) shows a number of $q > 0$ value binaries, reflecting the fact that some original secondaries are now more massive than their evolved primaries.
\begin{table}[h!]
\caption{Statistical Characteristics of Evolved WD+MS Distributions}\label{tab:e}
\begin{center}
\begin{tabular}{c|cccccc}
\multicolumn{7}{c}{ }\\
\hline
Metric & Average & StdDev & Skew & Kurtosis & Min & Max \\
\hline
$\log{a}(A.U.)$ & 3.13& 0.489& 0.262& -0.90& 2.24& 4.48\\
$\log{P}(Myr)$ & -1.40& 0.735& 0.246& -0.89& -2.78& 0.72\\
\hline
\end{tabular}
\end{center}
\end{table}
The $WD+MS$ histogram of separations shows an increase in excess kurtosis, i.e., the movement of low period and small separations towards higher values. This causes the more normal distribution curve that is seen ($kurtosis \Longrightarrow 0$). Additionally, the average separation has increased to $\log{(A.U.)} = 3.1\pm0.5$, i.e., $WD+MS$ pairs display an average separation $a = 477 A.U.$ greater than the original $MS+MS$ distribution $\log{(A.U.)}=2.9\pm0.5$. Thus, orbital amplification due to $PMS$ mass-loss is shown to have increased the average separation of the population. We note, however, that this is not the order of magnitude difference that \cite{G:86} expected. Perhaps our models have underestimated the effects of environmental perturbations (GMC, chance encounters, Galactic tidal forces). There are 7327 $WD+MS$ binaries in the final sample (i.e., $\sim8.1\%$).
\subsubsection{WD+WD}
To date, there are only about two dozen \citep{SION:91} double degenerate systems known, and certainly not as many as shown here ($N = 4791, \sim5.3\%$). The most plausible reason for this lack of discovery of $WD+WD$ binaries results from either one or both of the components being too dim to be detected in the surveys currently available. The histogram of the $q$-ratio (Figure \ref{fig:g}) show a tendency to nearly identical primary and secondary masses. This is to be expected with the narrow range of white dwarf final masses, resulting from the SSE's IFMR calculations.
\begin{table}[h!]
\caption{Statistical Characteristics of Evolved WD+WD Distributions}\label{tab:f}
\begin{center}
\begin{tabular}{c|cccccc}
\multicolumn{7}{c}{}\\
\hline
Metric & Average & StdDev & Skew & Kurtosis & Min & Max \\
\hline
$\log{a}(A.U.)$ & 3.15& 0.500& 0.316& -0.81& 2.27& 4.66\\
$\log{P}(Myr)$ & -1.36& 0.749& 0.308& -0.82& -2.74& 0.9\\
\hline
\end{tabular}
\end{center}
\end{table}
As predicted, the secondary companion (lower initial mass) produces a lower mass $WD$ than the primary, causing the peak at $q = 1$ (Figure \ref{fig:g}. Similar to the $WD+MS$ histogram of separations, the $WD+WD$ histogram shows an increase in kurtosis, i.e. loss of short period and small separations, evolving towards higher values. This causes the more normal distribution curve that is seen ($kurtosis\Longrightarrow0$) compared to the MS+MS population. The increase is the result of continued evolution of the binaries. The post-MS sudden mass-loss of the binary secondary influences the separation and period distributions a second time, pushing the uniform distribution parameters closer to a more random, Gaussian distribution. Additionally, average separation has increased from $\log{(A.U.)}=3.13$ for the $WD+MS$, to $\log{(A.U.)}=3.15$ for the $WD+WD$, i.e. an average separation increase of $63.6 A.U.$. We also note that the maximum separation has increased to $\log{(A.U.)}=4.66$ range, much greater than the initial $\log{(A.U.)}=4.0$ maximum imposed on the initial distribution.
\subsection{Comparison to Observed Data}
We begin our comparison by testing a variety of distance models against the LDS $WD+MS$ sample. For our models, we used a $10^4Myr$ evolutionary time with $10^2$ initial binaries (at $t = 0$) and a $\frac{100binaries}{10Myr}$ birth rate (i.e. constant average birth rate, \citealt{HPT:00}). For each component, we have roughly determined absolute $ugriz$ magnitudes using the \cite{K:79} models. We found temperatures and gravities from the SSE output or from \cite{CPH:03} evolutionary models for solar metallicity low-mass stars. Using the luminosities and separations, we applied a random distance model derived by assuming a uniform synthetic Solar Neighborhood population to approximate angular separation and apparent magnitude. This yielded a theoretical population histogram, restricted by the physical observation limits of the LDS. The distribution of the subset of modelled observed evolved binaries is given in Table \ref{tab:g}. Note that, in comparison to Table \ref{tab:c}, there are fewer $MS+MS$ observed than in the simulated population, an effect likely caused by the dimness of some low mass companions. Similarly, the percentage of $WD+MS$ pairs in the observed distribution is higher than in total posterior population. The latter is likely caused by the expected relative brightness of the binary pair; as mass ratio tends towards one, a pair with a $WD$ as a primary is likely to have a secondary with mass just slightly less than the progenitor of the $WD$, resulting in a brighter pair on average, than observed in the $MS+MS$ distribution.
\begin{table}[h!]
\caption{Distribution of Theoretically Observed Evolved Binarys in Postieror Population. Total observed population size is 29,6335}\label{tab:g}
\begin{center}
\begin{tabular}{c|cccccc}
\multicolumn{5}{c}{} \\
\hline
Metric & $MS+MS$ & $WD+MS$ & $WD+WD$ & Other \\
\hline
Count & 23,507& 3,801& 1,442& 885\\
Percent& 79.3\%& 12.8\%& 4.9\%& 3.0\%\\
Fraction of Posterior & 30.6\% & 51.9\% & 30.1\% & 58.3\% \\
\hline
\end{tabular}
\end{center}
\end{table}
\subsection{Distance Model Comparison}
The comparison of our model and observed angular separation distributions is shown in Figures \ref{fig:k} and \ref{fig:l}. We expect the observational data, which is a much smaller sample ($\sim6124$ binaries) to have more random scatter than the modeled data (WD+MS population $ = 1670$ binaries). Although the agreement is good dispite the scatter in the LDS sample, there are discrepancies (that we intend to address in a later paper).
\begin{itemize}
\item The kurtosis of the model is is smaller than in observed sample. Perhaps wide binaries at smaller separations were under-observed by Luyten. Alternatively, our model's lower limit on separation could be unrealistic.
\item The scatter in the observed data is $1\sigma$ greater than the modeled data.
\item The mean and skewness of both data sets is similar; this suggests that our basic model is sound, at least as a first order approximation.
\end{itemize}
\section{Discussion}
The distribution of binary angular separations in our model is similar to the observed distribution(i.e., the LDS sample). The LDS sample has a lognormal separation distribution that most likely stems from a log-uniform distance distribution. The estimated range of the LDS sample is between $1<\log{d}<3$ (where distance is in parsecs), similar to our model distance range. Initially we can compare the precentage of $WD+MS$ pairs observed in the LDS sample ($511/6124 = 0.08$) with the precentage of modeled $WD+MS$ pairs ($.12$) as well as the population of $WD+WD$ pairs observed in the LDS sample ($24/6124 = 0.003$) with the population of $WD+WD$ modeled ($0.04$). These distribution numbers suggest either we have underestimated the number of binaries lost to dissolution, or over-estimated the brightiness of some of the pairs. Likewise, these comparisons may also indicate where the LDS survey is under-sampled (i.e., the WD+WD pairs).
Figures \ref{fig:j} and \ref{fig:l}, present the results from our comparison of $MS+MS$ LDS and model pairs as well as $WD+MS$ LDS and model pairs. The figures show the cumulative distribution functions of the populations. Note that the separations are given in $\log(s")$ and have been put into 0.2 interval bins. Our initial hypothesis that PMS mass-loss significantly distorts a distribution of separations is apparent in Figure \ref{fig:k}, the offset in the CDF curves reflects an increase in average angular separation.
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{CDFModelComparisonFig10.pdf}
\caption{Comparison of observed angular separation CDF for MS+MS models and WD+MS models. Red Dashed Line: MS+MS, Blue Solid Line: WD+MS}\label{fig:j}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{CDFMSMSObservedFig11.pdf}
\caption{Plot of CDF of Model to Observed Cumulative Angular Separation Distributions for MS+MS pairs. Red Dashed Line: Observed, Blue Solid Line: Model}\label{fig:k}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{CDFWDMSObservedFig12.pdf}
\caption{Plot of CDF of Model to Observed Cumulative Angular Separation Distributions for WD+MS pairs. Red Dashed Line: Observed, Blue Solid Line: Model}\label{fig:l}
\end{figure}
We use the Cram\'{e}r-Von Mises goodness-of-fit test for comparison of the two-population cumulative distributions \citep{AN:62}, given in equations \ref{eq:r} and equation \ref{eq:s}. Here, goodness-of-fit is defined as the hypothesis $H_0: \bar{r} = 0.0$ where $\bar{r} = F_N(x) - G_M(x)$, where $F_N(x)$ and $G_M(x)$ are the cumulative distribution functions of samples of size $N$ and $M$, respectively. The Cram\'{e}r-Von Mises goodness-of-fit statistic here is defined by $T$.
\begin{equation}\label{eq:r}
T = \frac{NM}{N+M}\int_\infty^\infty [F_N(x) - G_M(x)]^2dH_{N+M}(x)
\end{equation}
\begin{equation}\label{eq:s}
T = \frac{NM}{(N+M)^2}*{\sum^N_{i=1}[F_N(x_i) - G_M(x_i)]^2 + \sum^M_{j=1}[F_N(y_j) - G_M(y_j)]^2}
\end{equation}
For the $MS+MS$ model and observation population set the values are $N = 26$ and $M = 26$. For the $WD+MS$ model and observation population set the values are also $N = 26$ and $M = 26$ as we have used the same spacing for both empirical distribution functions. These empirical distribution functions are based on the cumulative distributions generated for the populations with a bin interval of 0.2 between $0.4<\log{(s")}<2.5$. We compute T values for the comparison of $MS+MS$ populations ($T = 0.1343$) and the comparison of $WD+MS$ populations ($T = 0.0418$). Based on these statistics we estimate the limiting distribution, $a_1(T)$, of T for our models as in \cite{AN:52}:
\begin{equation}\label{eq:t}
a_1(T) = \lim_{n\to \infty} Pr\{n\omega^2 \leq T\}
\end{equation}
For the two comparisons, from the \cite{AN:52} tables $a_{MS+MS} > 0.5$ and $a_{WD+MS} < 0.5$, i.e., the probability of the $MS+MS$ model being drawn from the same distribution as the $MS+MS$ observed population is greater than 10\%, while the probability of the $WD+MS$ model being drawn from the same distribution as the $WD+MS$ observed population is at least 5\%. Given our original hypothesis $H_0: \bar{r} = 0.0$, the $MS+MS$ distribution comparison hypothesis fails, and the $WD+MS$ comparison hypothesis is accepted. We conclude then, that the distributions of observed and modeled $WD+MS$ angular separations are similar, but not for the $MS+MS$ distributions.
\begin{figure}[h!]
\centering
\includegraphics[scale=1.0]{CDFResidualEstimatesFig13.pdf}
\caption{Plot of the Residual Square Model to Observed Cumulative Angular Separation Distributions for MS+MS pairs (in blue open circles) and WD+MS pairs (in red closed circles)}\label{fig:m}
\end{figure}
We further demonstrate the relationship between the models by graphing the residual of the angular separation distributions in Figure \ref{fig:m}. The residuals plotted have no particular linear trend, nor offset. This suggests either a difference in spread or mean. There are however, offsets in the $MS+MS$ comparison. These offsets cause the failure of the null hypothesis for the $MS+MS$ comparison. This suggests: (1) there are additional factors for which we have not taken into consideration (such as inclination) and/or (2) small systematic errors associated with our model.
\section{Conclusions}
We have shown that PMS mass loss significantly shifts and reshapes the distribution of $MS+MS$ separations. The semi-major axis of a wide binary is a difficult parameter to determine, requiring distance, inclination, period and other measurements. It is our hypothesis that a distribution of angular separations, preserves information about the effect of $PMS$ main-losses on the binary system. We take as proof, the following observation: Both our $MS+MS$ and $WD+MS$ models compare reasonably well to the observed LDS distributions. Our model, however, has only been affected by $PMS$ mass loss and a crude attempt has been made to account for other disrupting factors. This assumption and other first order approximation in the creation of the model, likely affect the goodness-of-fit test for the $MS+MS$ observed populations.
Figures \ref{fig:k} and \ref{fig:l} show the cumulative distribution of $\log{(s")}$ in the $MS+MS$ and $WD+MS$ pairs. The difference in observational and theoretical distributions is apparent.This difference is a function of both the incompleteness of the survey and secondary effects for which we have not taken into consideration (such as inclination or external perturbations). For example, observations and models have shown that metallicity may play a much larger role in the amount of post Main-Sequence mass-loss (see \citealt{ZH:11}).
Future work should include a more realistic model for stellar formation history rather than the constant birth function implemented here. Furthermore, this stellar birth model should be used to drive the initial conditions of both the metallicity and the mass distribution function. In view of the trends in the metallicity, especially as a function of stellar age \citep{ZH:11}, additional effect on the mass-loss of the population as a function of time would be expected. We have demonstrated in this paper, and it has been demonstrated in other simulations, that post-MS mass-loss on average results in an increase in semi-major axis for visual and fragile binaries. Changes to the metallicity history would be expected to influence orbital amplification as a function of time, in turn affecting the final angular separation distribution. Our photometric colors were based on a consistent metallicity for the population of binaries. Adjustments to the metallicity history would affect the associated color indices, as well as the apparent magnitude of the binary components. Again, this would effect the visual angular separation distribution.
The observed $MS+MS$ and $WD+MS$ distributions are affected by observational bias and/or physical orbital amplification. If orbital amplification results in an increase in orbitla separation, we conclude that binaries with small separations have expanded to middle separation binaries, middle binary separations have increased slightly, and binaries that were already wide may have been lost to dissociation. Our current simplistic model describes the orbital frequency distribution of FBs reasonably well. In general, our biggest issue is population size of the observed distribution. The statistically small observed population causes a large scatter in the frequency distribution, or rather a larger scatter than in the model distribution.
The DS code is currently deficient at both extremes of stellar mass ( i.e., $M < 0.07M_{\odot}$ and $M > 90M_\odot$ ). We do not have an evolutionary code applicable to these regimes, nor does our code include metallicity variation. However, such variance is not of major important in the modeling of disk stars. Future work will require such an extension to the SSE code and extrapolation of the Kurucz color/atmosphere models. Low-mass models are available and have been taken into account for masses as low as $M\sim0.075M_{\odot}$ \citep{BCAH:98}. While their contributions to the current observed sample of FBs are small, in order to produce a realistic model of the evolved population, they should be taken into account. This lower mass limit does not directly affect the current project.
\section{Acknowledgments}
All graphs in this paper have been made with the R-project program. The authors thank Jarrod Hurley for providing them a copy of the fast stellar evolutionary code SSE. We would also like to thank Dr. Matt Wood and Dr. St\'{e}phane Vennes for their support and advice. This project was supported in part by NASA grant Y701296 (T.D.O.) and NSF grants AST 02-06115 (T.D.O.) and AST0807919.
|
\section{Introduction}
Magnetic nanostructures are intensively studied nowadays for fundamental and practical purposes\cite{Hillebrands2001.vol.1}. Understanding the magnetism in low-dimensions presents a fundamental interest, while achieving smaller data storage media and magnetic memory is of utmost technological importance. Large periodic arrays of magnetic nanowires have been easily produced with inexpensive techniques like electrodeposition\cite{Pirota200418}. For such structures, the manipulation of magnetization in very short times is very important for high-speed applications and requires a complete understanding of the nature of magnetic excitations, magnons or spin-waves, and their dependence on geometry.
The spin-wave spectrum for ellipsoidal samples is known for some time\cite{PhysRev.105.390,PhysRev.110.1295}. In uniform magnetized ellipsoidal samples the demagnetizing field is uniform. As many of the magnetic nanostructures studied nowadays have in general a non ellipsoidal form, the demagnetizing field inside them is nonuniform even if they are uniformly magnetized\cite{Joseph}. Thus the shape can drastically affect the dynamic properties and the spectrum of the spin-waves will be modified (see Ref.\cite{Demokritov2008} and references therein). As the dimension of the wires decreases, the exchange interaction becomes important. We consider here that both dipolar and exchange interactions contribute to the spin-wave spectrum. Obtaining a general theory is challenging, but approximate analytical solutions and numerical results can be derived for certain geometries.
Reducing the dimensions of magnetic nanostructures is the natural trend to obtain smaller devices. Until now, only very long nanowires were studied with diameter-to-length ratio (d/L) $\ll$ 1. In this paper, we investigate nanowires of moderate aspect ratio magnetized along the axial direction. We develop the analytical theory and present numerical results of the spatial distribution of spin waves modes in this type of nanostructures. We will outline the important features of taking into account the in-homogeneity of the demagnetizing field for the confined geometry under study. The previous studies of cylindrical geometry neglected the nonuniformity of the demagnetizing fields considering the cylinders semi-infinite\cite{Joseph2,PhysRevB.63.134439}.
\section{Analytical model}
We start with a ferromagnetic cylindrical sample, magnetized to saturation along the axial direction (\textit{z} direction) which ensures that the longitudinal component of \textbf{M} is much larger than the transverses ones, \textbf{M} = ($m_{x}, m_{y}$, M$_{s}$). A static magnetic field \textbf{H} is applied along z and an rf field \textbf{h}(\textbf{r},t)=\textit{h}(\textbf{r})e$^{-i\omega t}$ is applied in perpendicular direction, $\textbf{H}_{ext} = (h_{x}, h_{y}, H)$. We assume that the rf field and the saturation magnetization M$_{s}$ are uniform in the sample, even if the state of uniform magnetization is not actually realized in this geometry. Two main approaches are used for solving the spectrum of spin-waves in confined structures. In both, one solves simultaneously the linearized Landau-Lifschitz equation of motion for the magnetization together with the Maxwell equations satisfying the electromagnetic boundary conditions and the exchange boundary conditions\cite{Kalinikosbook}. The two methods are equivalent. We choose here the method of magnetic potential where the relation \textbf{m}(\textbf{h}) is found first from the equation of motion and then we search for the solutions of Maxwell equations which satisfy the boundary conditions.
The Landau-Lifschitz (LL) equation of motion for the magnetization neglecting damping is:
\begin{equation}
\label{eq1}
\frac{d\textbf{M}}{dt}(\textbf{r},t) = -\gamma\textbf{M}\times\textbf{H}_{eff}
\end{equation}
\noindent where $\gamma$ is the gyromagnetic ratio. The effective field inside the sample represents the sum of the external applied field $\textbf{H}_{ext}$, the exchange field $\textbf{H}_{exch}=D\nabla^{2}\textbf{M}$ and the demagnetizing field $\textbf{H}_{demag}= - \widehat{\textbf{N}}\textbf{M}$ (excluding crystal anisotropy):
\begin{equation}
\label{eq2}
\textbf{H}_{eff}=\textbf{H}_{ext}+\textbf{H}_{exch}+\textbf{H}_{demag}
\end{equation}
\noindent where D is the exchange stiffness and $\widehat{\textbf{N}}$ the demagnetizing tensor field. For non-ellipsoidal bodies the demagnetizing tensor field (in the first order) is function of position and is defined in the Fourier space as in Ref.\cite{Tandon20049}:
\begin{equation}
\label{eq3}
N_{ij}(\textbf{k}) = \frac{D(\textbf{k})}{k^2}k_ik_j
\end{equation}
\noindent with D(\textbf{k}) the shape function. The LL equation (Eq.(\ref{eq1})) describe an uniform precession of the magnetization around the effective field $\textbf{H}_{eff}$. To find the normal modes of the magnetization we search for non zero solutions of the LL equation. Considering the dynamical magnetization uniform in the sample we find:
\begin{align}
\label{eq4}
i\omega m_x &= \gamma m_y[H_i-D\nabla^2]-\gamma M_sh_y \nonumber \\
-i\omega m_y &= \gamma m_x[H_i-D\nabla^2]-\gamma M_sh_x
\end{align}
\noindent with $N_{zz}$ and $N_{rr}$ the longitudinal (dc) and transversal (ac) demagnetizing factors and H$_{i}$=H--4$\pi$(N$_{zz}$--N$_{rr}$)M$_{s}$. Here, we'll consider as first order approximation that the demagnetizing tensor field can be diagonalized as $N_{zz}$+2$N_{rr}$=4$\pi$.
The demagnetizing field is determined by Maxwell equations in the magnetostatic limit:
\begin{align}
\nabla\times\textbf{H}_{int} & = 0 \label{eq5}\\
\nabla \textbf{B}=\nabla(\textbf{H}_{int} + 4\pi\textbf{M}) & = 0 \label{eq6}
\end{align}
\noindent where the internal field is $\textbf{H}_{int}=\textbf{h}-N_{rr}\textbf{m}+H-N_{zz}M_s$. From Eq.(\ref{eq5}), the magnetic field may be written as $\textbf{h} = -\nabla\Psi_m$, where $\Psi_m$ is the scalar magnetic potential. Replacing the scalar potential in Eq.(\ref{eq6}) we obtain:
\begin{equation}
\label{eq7}
-\nabla^2\Psi_m+(4\pi-N_{rr})(\frac{\partial m_x}{\partial x}+\frac{\partial m_y}{\partial y})-M_s\frac{\partial N_{zz}}{\partial z}=0
\end{equation}
Eq.(\ref{eq7}) has the form of an anisotropic Laplace equation or Walker type equation for cylinders with the last term coming from the variation of the demagnetization (internal) field near edges. In the center of the cylinder, where the demagnetization field is almost constant (here the longitudinal demagnetization field is zero) the Walker equation reduces to the classical equation for cylinders\cite{PhysRevB.63.134439}. We considered here only the variation of the longitudinal demagnetizing factor $N_{zz}$ along z (see Appendix for more details). The nonuniformity of $N_{zz}$ strongly affects the mode frequencies as each mode feels a different demagnetizing factor. The full analysis should also take into account the variation of N$_{rr}$ in the radial direction r, which we neglect here as Nrr is almost constant with r due to the fact that our cylinders have a reduced aspect ratio (diameter-to-length inferior of 0.1) and such an analysis will introduce a high degree of analytical complexity.
\begin{figure}[t]
\center
\includegraphics[width=7cm]{Fig1.eps}
\caption{\label{Fig.1} Variation of the demagnetizing tensor components (here normalized to $N_{zz}$+2$N_{rr}$=1) along the axial direction for a cylinder with length-diameter (L/d) ratio of 10: N$_{zz}$ solid line, N$_{rr}$ dashed line and $\partial$N$_{zz}$/$\partial$z triangles. Only half of the cylinder is shown with largest variation near the edge.}
\end{figure}
We state that we made here the simplifying assumptions: the static magnetization is considered as constant and independent of the external field and that the dc and ac demagnetizing fields are the same and can be calculated with the expressions given in Eq.(\ref{eq3}). Even if the non-diagonal elements of the demagnetizing tensor field are not zero, they can be overlooked as they are negligible (the N$_{rz}$ component is inferior of 10$^{-4}$). The variation of the demagnetizing factors\cite{Tandon20049} along the axial direction is shown in the Fig.\ref{Fig.1}.
Choosing the scalar potential of the form $\Psi_m(r,\varphi,z)$=J$_{n}$(qr)exp(in$\phi$+ik$_{z}$z) where J$_{n}(qr)$ is the Bessel function of order \textit{n}, we obtain the transcendental equation (see Appendix):
\begin{multline}
\label{eq8}
\omega^{2}(k_{z}^{2}+q^{2}-M_{s}\frac{\partial N_{zz}}{\partial z})=\gamma^{2}D^{2}(k_{z}^{2}+q^{2})^{3}+\gamma^{2}D(k_{z}^{2}+q^{2})^{2}\\
\times\left[ 2H_{i}+(4\pi-N_{rr})M_{s}\right]+\gamma^{2}(k_{z}^{2}+q^{2})H_{i}\left[ H_{i}+(4\pi-N_{rr})M_{s}\right] \\
-k_{z}^{2}\gamma^{2}M_{s}(4\pi-N_{rr})H_{i}-\gamma^{2}M_{s}D(4\pi-N_{rr})k_{z}^{2}(k_{z}^{2}+q^{2})\\
-\frac{\partial N_{zz}}{\partial z}\gamma^{2}M_{s}H_{i}^{2} +\frac{\partial^{3}N_{zz}}{\partial z^{3}}2\gamma^{2}M_{s}DH_{i}-\gamma^{2}D^{2}M_{s}\frac{\partial^{5}N_{zz}}{\partial z^{5}}
\end{multline}
In the limit of very long wavelength, when the wave vector has all components zero we derive the following expression:
\begin{multline}
\label{eq9}
-\frac{\partial N_{zz}}{\partial z}\omega_0^{2}=-\frac{\partial N_{zz}}{\partial z}\gamma^{2}H_{i}^{2} +\frac{\partial^{3}N_{zz}}{\partial z^{3}}2\gamma^{2}DH_{i}-\\
\gamma^{2}D^{2}\frac{\partial^{5}N_{zz}}{\partial z^{5}}
\end{multline}
\noindent which takes the form of the Kittel uniform mode\cite{PhysRev.71.270.2} if we neglect the higher exchange terms (derivative of order three and five) and keep only the first derivative term: $\omega_0=\gamma H_{i}$. However, no real uniform mode exists in cylinders of finite aspect ratio as the internal field varies near edges. In the case of a nonuniform magnetic field we usually assume that a spin wave can propagate with continuously changing wave vector $k_z$(z)\cite{Schlomann}. At certain points (surface), the wave vector becomes zero and the condition of quasi-static variation of the internal field is no longer satisfied thus the solution of the equation of motion has to be found with variable parameters. This surface is called turning surface by analogy with quantum mechanics. The Eq.(\ref{eq9}) at the turning surface has no physical meaning. A localized spin wave is excited with changing wave vector propagating from the turning surface in the direction of the decreasing internal field. This effect called spin-wave well was demonstrated in stripes recently\cite{PhysRevLett.88.047204}.
For samples of finite size, the dispersion relation is shape dependent. The complete spin-wave spectrum and the modes shape depend strongly on the boundary conditions. We expect some degree of pinning at the edges, and this degree of pinning is determined by a competition between dipolar magnetostatic energy and exchange energy. Apart from the Maxwell boundary conditions, continuity of the magnetization and the magnetic potential, usually a Rado-Weertmann type of boundary condition\cite{JPhysChemSolids.11.315,PhysRevB.72.014463} is considered at the cylinder surface:
\begin{equation}
\label{eq10}
\frac{\partial m}{\partial z}\pm p\times m=0
\end{equation}
\noindent at z=$\pm$L/2 with L the length of the cylinder and p the so-called pinning parameter which is determined by the effective surface anisotropy K$_{s}$ and the exchange stiffness constant D: p$\simeq$K$_{s}$/D. This condition implies stationary waves in the z direction with the particular solution of the type $m_{z}$=A$\sin k_{z}z$+B$\cos k_{z}z$.
The frequency of an eigenmode is constant throughout the magnet and the spin wave vector varies to accommodate the changes in the internal field profile. The dynamic magnetization has a plane-wave character. In almost all the volume of the cylinder, between the turning surfaces at both ends (excluding the edge domains), the variation of the demagnetization field is small and usually the averaged value of the components of demagnetizing field is used. The derivatives of the N$_{zz}$ in the Eq.(\ref{eq8}) are negligible. For moderate aspect ratio cylinders we can use a simple and intuitive model where the dynamical magnetization takes the form:
\begin{equation}
\label{eq11}
m_{ln}(r,z) \sim J_{n}(qr)\mbox{cos}(k_{z}^{l}z),
\end{equation}
\noindent where k$_{z}^{l}$ is the longitudinal wave number for the \textit{l}th longitudinal mode (backward geometry) and \textit{n} indexes the radial modes. This relation is valid between the turning surfaces, where the internal field is almost constant. In this region, stationary waves on the axial and radial directions are formed and we can have mixed modes. The distance between the turning surfaces, or effective length $\Delta$z, is considered to be approximately 0.8L (with L the length of the cylinder). The dispersion curves are characterized by two quantization numbers \textit{n} and \textit{l}. The quantization parameter \textit{l} should be taken with care, it provides here a qualitative description of longitudinal modes which is to be compared with the micromagnetic simulations. We can use the mean value of the wave vector of each mode which can be evaluated as: k$_{z}^{l}=l\pi$/$\Delta$z, where $\Delta$z is the effective length where the mode is localized in the sample. The spectrum of the SW consists of a series of dispersion curves characterized by two quantization numbers \textit{n} and \textit{l} and depend on the ratio of the cylinder dimensions. Changing the aspect ratio of the cylinders induces a change in the values of wavenumbers (and also a change in the demagnetizing field). For example, if the aspect ratio is diminished, q becomes much larger than k$_{z}$. The principal effect is a reduced influence of k$_{z}$ on frequency or applied field and thus the longitudinal modes form a continuous band. When d/L=0.1, the difference between the first and the fifth longitudinal mode is of 0.8kOe at fixed frequency (20GHz). The same difference is only of 0.03kOe when d/L=0.02, while the difference in the radial modes is much larger.
\subsection{Spatial distribution of edge modes}
The terms which contain the derivatives of the demagnetizing field that appear in the Eq.(\ref{eq8}) have an increased influence near the edges of the cylinder where their variation is most important. These terms will provide a correction to the frequencies of the edge modes. To quantify this correction we numerically calculated the frequencies of the edge modes for a Nickel cylinder of aspect ratio d/L=0.1 (L=300nm) with and without the additional terms at fixed magnetic field. For example, without correction we obtain a frequency of 13.79GHz for the edge mode at 1.5kOe, a 48\% difference from the 9.3GHz value obtained with the correction terms. The frequencies obtained for the edge modes by the analytical formula with the correction terms are represented in the Fig.\ref{Fig.2}(b) along with the values obtained from micromagnetic simulations discussed in the next section. These values are similar which imply that the correction terms are necessary to calculate exactly the frequencies of edge modes in finite size nanowires.
In Fig.\ref{Fig.2}(a), we show the simulated 2D spatial distribution of the edge mode near surface for a DC magnetic field of 2.9kOe. The spatial distribution is obtained from the permitted values of the wavenumber near surface (positive). Our method is the following: we first numerically determine the frequencies of the edge modes near boundaries assuming that it does not propagate into the center of the cylinder. Next, we fix the frequency at the value obtained in the first step (at a given DC magnetic field) and we calculate the spatial distribution of the allowed values for the wavenumber in the edge region. From the plot we observe that the edge mode is not restricted to the actual surface but goes underneath the surface in parabolic fashion.
\begin{figure}[t]
\center
\includegraphics[width=7cm]{Fig2.eps}\\
\caption{\label{Fig.2} (a) Spatial distribution of an edge mode calculated from the analytical expression for a cylinder with diameter-to-length ratio of 0.1 (L=300nm) in a false color code. Red (positive) represents a permitted value for the wavenumber. (b) Edge mode frequencies determined from the analytical expression (circles) and from micromagnetic simulation (rectangles) at different DC magnetic fields. }
\end{figure}
\section{Micromagnetic simulation}
In order to understand the mode structure and to compare the analytical results of the previous section, a micromagnetic simulation was conducted with the nmag package\cite{Fischbacher} for a cylinder with an aspect ration d/L=0.1 (L=300nm). The cylinder was discretized with a cell size of 5nm. First, the magnetization was relaxed to equilibrium using a large damping parameter $\alpha$=0.5. The equilibrium configuration was then excited adiabatically with a small 20GHz rf magnetic field in the perpendicular direction, while a constant magnetic field was applied along the cylinder axis. The values used for the parameters are the same as used in the analytical simulation and are those of Ni: D=2$\times$10$^{-9}$Gcm$^{2}$, $\gamma$=188.5GHz/T (g factor of 2.15) and M$_{s}$=480 emu/cm$^{3}$\cite{PhysRevB.63.134439,PhysRevB.64.144421} and a damping parameter of 0.015. We subsequently computed the spatial distribution of the oscillation of the dynamic magnetization doing Fourier transforms on a number of cells along the z and x axes of the cylinder. For comparison with previous results, we kept fixed the frequency of the rf magnetic field and we varied the amplitude of the DC magnetic field.
To determine the spatial distribution of the longitudinal modes in an cylindrical nanowire we calculated the amplitude of oscillation of the magnetization at different magnetic fields. The values of the static magnetic field were chosen to correspond to longitudinal modes as calculated with Eq.(\ref{eq8}). In Fig.\ref{Fig.3}(a), we show the average of the magnetization profiles along the z-axis for one DC magnetic field which should correspond to the 5th longitudinal spin wave mode determined from the analytical model. These profiles are calculated doing Fourier transforms on the magnetization for each nm in the z direction and in the x direction and extracting the values at 20GHz. Basically, we compute the amplitude of the dynamical magnetization variation in a xz-plane passing through the center of the cylinder (Fig.\ref{Fig.3}(b)) and then we average between the curves in the x direction. This corresponds with the average of a contour plot in a xz-plane. As observed, the calculated profile corresponds well with the one expected at this applied external field from the analytical model. The spins which form the longitudinal mode precess are mainly in the effective length $\Delta$z. Another mode appear at the edges of the cylinder where the spins precess at a lower frequency.
\begin{figure}[t]
\center
\includegraphics[width=7cm]{Fig3.eps}\\
\caption{\label{Fig.3} (a) Average magnetization profile along z axis of a cylinder with diameter-to-length ratio of 0.1 (L=300nm). The average was calculated from 30 magnetization profiles taken each nm in the x direction in the plane y=0; (b) 2D color projection in the xz plane of the 30 magnetization profiles. The maximal amplitude of the dynamical magnetization (red) precess at 20GHz with a DC magnetic field applied (2.9kOe) corresponding to the 5th longitudinal mode.}
\end{figure}
In Fig.\ref{Fig.4}, the 2D spatial distribution of the magnetization is shown corresponding to the edge modes obtained at two different DC magnetic fields. The xz plane shown passes through the center of the cylinder, rotating it 360$^{\circ}$ around its axis will provide the 3D profile of the mode. The panel (b) corresponds to an edge mode with spins precessing at a lower frequency of 13.25GHz (DC field 2.9kOe) than the longitudinal mode shown in the Fig.\ref{Fig.3}(b) where the spins precess at 20GHz. The spatial distribution of edge modes comes into agreement with that calculated in the previous section (Fig.\ref{Fig.2}(a)). The frequencies of the edge modes obtained from the microwave simulation are shown in the Fig.\ref{Fig.2}(b). The values are similar with those calculated with the Eq.(\ref{eq8}).
\begin{figure}[t]
\centering
\includegraphics[width=7cm]{Fig4.eps}\\
\caption{\label{Fig.4} 2D color projection in the xz plane of two edge modes of a cylinder with diameter-to-length ratio of 0.1 (L=300nm) at two different DC magnetic fields: (a) 1.5kOe and (b) 2.9kOe.}
\end{figure}
In Fig.\ref{Fig.5}, a snapshot of the 3D profile of the y component of magnetization is shown after 30ps. We clearly observe the variation of the dynamical magnetization along the z direction (long axis). The mode is 3D longitudinal mode as expected even though some hybridization effects may exists, the standing waves picture being simplistic\cite{PhysRevB.70.054409}. The edge mode is clearly seen, it corresponds to the spins rotating at the surface (in plane) and precessing at lower frequency than the bulk modes. The edge mode is localized and it is always present due to the magnetic field inhomogeneity at the boundaries. The longitudinal modes which precess at 20GHz decay in time, becoming zero after 1.5ns.
\begin{figure}
\center
\includegraphics[width=7cm]{Fig5.eps}
\caption{\label{Fig.5} 3D profile of the evolution of the y component of magnetization is shown at a given instant in time (30ps). The applied DC magnetic field is 2.9kOe and the frequency of the rf field is 20GHz.}
\end{figure}
\section{Discussion and conclusion}
The spatial distribution of longitudinal and edge modes in cylinders of moderate aspect ratio was calculated in this paper. It provides new insight on the spatial distribution of spin waves in cylindrical nanowires. We used a 2D analytical model and obtained the spin wave modes frequencies and profiles for the edge modes. The validity of the 2D model was tested through comparison with 3D micromagnetic results. The values of the resonances agree quantitatively although the profiles of the edge modes from 3D micromagnetics are not exactly parabolic and seem to penetrate on a longer distance beneath the surface (30-50nm). The difference between the two results are probably due to the assumptions and approximations that where made: in the analytical model the magnetization is considered uniform in the whole cylinder and a correction in the z direction is obtained but the variation of the radial demagnetizing field is not considered (which can be an useful extension of the theory) while in the 3D micromagnetics we break the cylinder into small tetrahedral cells (discretization length of 4.8nm) and consider that the magnetization is continuous and vary linearly and then the demagnetizing field is computed.
The nanowires used in the simulation have finite diameter-to-length ratio (d/L=0.1) and thus the longitudinal modes are well separated and their spatial distribution was resolved. We tried to resolve the spatial distribution of longitudinal modes for a cylinder with d/L=0.02 without success, as the modes are hybridized (closer in frequency) and form a quasi-continuum band. Furthermore, as the demagnetizing field is inhomogeneous it confines the spin waves at boundaries of the cylinder: the edge modes are always present in finite aspect ratio cylinders. These modes can be observed in experiments. Until now, only modes of an ensemble of nanowire have been investigated\cite{PhysRevB.64.144421}. To measure the spectrum of an individual nanowires a particularly well suited technique is the (Ferro)Magnetic Resonance Force Microscopy\cite{Zhang, Klein}. This local probe technique can validate our results.
\section*{Acknowledgments}
I would like to thank O. Klein, V. Dolocan, M. Franchin, H. Fangohr for their help with the simulations and L. Raymond and J.M. Debierre for use of the Lafite server.
|
\section{Introduction}
\label{sec:intro}
\vspace*{\closersection}
Chiral perturbation theory (chPT) \cite{Gasser:1983yg,Gasser:1984gg} is a widely used tool in many phenomenological applications and also helpful to guide an extrapolation to lighter quark masses in lattice-QCD simulations. Here we will report on a determination of the NLO low-energy constants (LECs) $\bar{l}_3$ and $\bar{l}_4$ which appear in the light quark mass dependence of the pseudo-scalar meson masses and decay constants in SU(2) chPT.
We analyze configurations generated by the Wuppertal-Budapest Collaboration \cite{Aoki:2006we,Aoki:2006br,Aoki:2005vt,Aoki:2009sc,Borsanyi:2010bp,Borsanyi:2010cj} using the Symanzik glue and 2-fold stout-smeared staggered fermion action for a 2+1 flavor QCD-simulation. The mass of the single flavor has been kept at the value of the physical strange quark mass, whereas the two degenerate lighter quark masses have been varied such that light meson masses in the range of 135 to 440 MeV were simulated. The simulations were performed at five different gauge couplings $\beta$, resulting in lattice scales between 0.7 and 2.0 GeV (see next section for details on how the scale has been determined). Table~\ref{tab:ensembles} summarizes some of the parameters of the simulations.
The 2-fold stout-smeared version of the staggered quark action has been proven to be advantageous \cite{Borsanyi:2010bp} in reducing the inevitable taste-breaking of staggered fermion formulations. Therefore, in this work we only consider the pseudo-scalar mesons with taste matrix $\gamma_5$ when measuring meson masses or decay constants. Details of the computation of these quantities will be reported in a forthcoming publication.
\begin{table}[b]
\begin{center}
\begin{tabular}{cccc}
$\beta$ & $1/a$ [GeV] & $m_l/m_l^{\rm phys}$ (approx.) & $(L/a)^3\times(T/a)$\\\hline
3.45 & 0.69 & 1.0, 3.0, 5.0, 7.0, 9.0 & $24^3\times32$ -- $12^3\times28$\\
3.55 & 0.91 & 1.0, 3.5, 5.0, 7.0, 9.0 & $24^3\times32$ -- $12^3\times28$\\
3.67 & 1.31 & 1.0, 4.0, 6.0, 7.5, 9.5 & $32^3\times48$ -- $14^3\times32$\\
3.75 & 1.62 & 1.0, 4.0, 6.0, 8.0, 10.0 & $40^3\times64$ -- $16^3\times32$\\
3.85 & 2.04 & 1.0, 1.4, 2.0, 4.0, 6.0, 8.0, 10.0 & $48^3\times64$ -- $24^3\times
48$
\end{tabular}
\end{center}
\vspace*{\closercaption}
\caption{Simulated lattice ensembles: gauge coupling $\beta$, lattice spacing $1/a$, simulated quark masses $m_l$, and range of lattice sizes.}
\label{tab:ensembles}
\vspace*{\afterTable}
\end{table}
\section{Scale setting and physical quark masses}
\label{sec:scale_mud}
\vspace*{\closersection}
To set the scale at each simulated gauge coupling $\beta$ and identify the physical point, i.e.\ the average up/down quark mass $m_l^{\rm phys}=(m_{\rm u}+m_{\rm d})/2$ corresponding to a pion in the isospin limit with an estimated mass of $M_\pi=134.8\,{\rm MeV}$ \cite{Colangelo:2010et}, we use a two-step procedure. First, we extrapolate the ratio $(aM_{ll})^2/(af_{ll})^2$ of the squared meson masses and decay constants to $M_\pi^2/f_\pi^2 = (134.8\,{\rm MeV} / 130.41\,{\rm MeV})^2 = 1.06846$, where we also used the PDG-value $f_\pi=130.41\,{\rm MeV}$ \cite{Nakamura:2010zzi}. In that way $am_l^{\rm phys}$ is obtained. In the second step, we extrapolate $af_{ll}$ to this quark mass value and obtain the lattice scale with the help of the PDG-value for $f_\pi$. For the extrapolation we used two different ans\"atze: a quadratic and a rational (linear in numerator and denominator) fit form. An example of these extrapolations is shown for the ensembles at $\beta=3.85$ in Fig.~\ref{fig:set_mud_scale}. There, like for all other $\beta$-values as well, the heaviest quark mass point has been excluded, resulting in a fit range of approx.\ $am_{l}/am_l^{\rm phys} \leq 8.0$ (corresponding to $M_{ll}\leq390\,{\rm MeV}$). We stress that here, like in the chiral fits to be discussed below, the data has been corrected for finite volume effects beforehand, by means of using the two- and three-loop resummed formulae of \cite{Colangelo:2005gd} for the pion decay constants and masses, respectively. Our spatial lattice volumes $L^3$ are in the range $(4.7\,{\rm fm})^3$ -- $(6.8\,{\rm fm})^3$ with a minimal $M_{ll}L\approx 3.2$, ensuring that the finite volume corrections within our fit ranges are at most at the order of 1 per cent for the decay constants and even less for the meson masses.
By fixing $1/a$ and $am_l^{\rm phys}$ in the way described above, the meson masses and decay constants show no discretization effects at all directly at the physical point and we can assume those effects to be small (since of higher order in the quark masses and/or lattice spacing) in the vicinity of the physical point, i.e.\ in the mass range covered by our fits. Such discretization effects, of course, are present in other observables, which are not considered in this work.
\begin{figure}[t!]
\begin{center}
\includegraphics*[angle=-90,width=.49\textwidth]{Plots/plot_ratio_b3.85.ps}%
\includegraphics*[angle=-90,width=.49\textwidth]{Plots/plot_fPi_b3.85.ps}%
\end{center}
\vspace*{\closercaption}
\caption{{\it Left panel:} ratio $(aM_{ll})^2/(af_{ll})^2$ extrapolated to $M_\pi^2/f_\pi^2=1.06846$ to obtain $am_l^{\rm phys}$, {\it right panel:} $af_{ll}$ extrapolated to $am_l^{\rm phys}$ to obtain $1/a$; both at $\beta=3.85$.}
\label{fig:set_mud_scale}
\end{figure}
\vspace*{\afterFigure}
\section{Fits to NLO SU(2) chPT}
\label{subsec:fits.nlo}
\vspace*{\closersection}
The quark mass dependence of the finite-volume corrected data for the meson masses and decay constants is fitted simultaneously at different $\beta$-values using the NLO-SU(2) chPT formulae
\begin{eqnarray}
M_{ll}^2 &=& \left(\frac{1}{a}\right)^2 (aM_{ll})^2 \;=\; \chi_l\,\left[1\,+\,\frac{\chi_l}{16\pi^2 f^2}\log\frac{\chi_l}{\Lambda_3^2}\right]\,, \\
f_{ll} &=& \left(\frac1{a}\right) (af_{ll}) \;=\; f\left[1\,-\,\frac{\chi_l}{8\pi^2f^2}\log\frac{\chi_l}{\Lambda_4^2}\right]\,,\\
\chi_l &=& 2B\,m_l \;=\; (2Bm_l^{\rm phys})\,\frac{am_l}{am_l^{\rm phys}}\,,
\end{eqnarray}
where we made use of the already determined $1/a$ and $am_l^{\rm phys}$ to scale the quark masses and the meson masses and decay constants measured in lattice units. This fit has four free parameters: two NLO low-energy scales $\Lambda_3$, $\Lambda_4$, the decay constant in the SU(2) chiral limit $f$ and the renormalization scheme-independent combination $(2Bm_l^{\rm phys})$ of the LO low-energy constant $B$ and the physical quark mass $m_l^{\rm phys}$.
We would like to point out that the chiral fit formulae do not include any taste breaking effects, i.e., we did not use staggered chPT. This seems justified to us, since we are only considering $\gamma_5$-taste mesons as mentioned above and use these to define our scaling trajectory at the physical point. In other words, since the meson mass and decay constant at the physical point were used to set the quark masses and lattice scales, no discretization or taste breaking effects are present in the chPT formulae for $M_{ll}^2$ and $f_{ll}$ as discussed above. Furthermore, taste breaking effects are reduced anyway by the choice of the fermion action as mentioned above.
The top panels of Fig.~\ref{fig:fits_NLO} show the combined fits including the data at all lattice spacings and for meson masses in the range of 135 to 390 MeV. (Here and for all following plots we mark data points included in the fit by circles, while those not included in the fit are marked by diamonds.) As one can already see by eye, the description of the data by the fit is not satisfactory, resulting in a $\chi^2/{\rm d.o.f.}\approx 4.3$. As expected, the fit quality measured, e.g., by $\chi^2/{\rm d.o.f.}$ improves continuously when reducing the upper bound of the meson mass range. The middle panel of Fig.~\ref{fig:fits_NLO} shows the fit to all meson masses in the range $135\,{\rm MeV}\leq M_{ll}\leq275\,{\rm MeV}$ giving an acceptable $\chi^2/{\rm d.o.f.}\approx 1.0$. A similar improvement can be achieved by excluding the two coarsest lattice ensembles from the fit, i.e., limiting $1/a\geq 1.3\,{\rm GeV}$. The bottom panel of Fig.~\ref{fig:fits_NLO} shows an example of such a fit with $135\,{\rm MeV}\leq M_{ll}\leq 340\,{\rm MeV}$ resulting in $\chi^2/{\rm d.o.f.}\approx 1.7$ (this number has to be compared to $\chi^2/{\rm d.o.f.}\approx 2.6$ for the same mass range and using all $\beta$). Applying both kinds of cuts, i.e.\ $135\,{\rm MeV}\leq M_{ll}\leq275\,{\rm MeV}$ and $1/a\geq1.3\,{\rm GeV}$, eventually gives a $\chi^2/{\rm d.o.f.} \approx 0.8$.
\begin{figure}[ht!]
\begin{center}
\includegraphics*[angle=-90,width=.49\textwidth]{Plots/fpi_bAll_m1-8.ps}%
\includegraphics*[angle=-90,width=.49\textwidth]{Plots/mpisqrdiv_bAll_m1-8.ps}\\
\includegraphics*[angle=-90,width=.49\textwidth]{Plots/fpi_bAll_m1-4.ps}%
\includegraphics*[angle=-90,width=.49\textwidth]{Plots/mpisqrdiv_bAll_m1-4.ps}\\
\includegraphics*[angle=-90,width=.49\textwidth]{Plots/fpi_bFine_m1-6.ps}%
\includegraphics*[angle=-90,width=.49\textwidth]{Plots/mpisqrdiv_bFine_m1-6.ps}\\
\end{center}
\vspace*{\closercaption}
\caption{Combined NLO SU(2) chPT fits with various fit ranges. The {\it left panels} show the decay constants $f_{ll}$, the {\it right panels} the squared meson masses $M_{ll}^2$ divided by the quark mass ratio $am_l/am_l^{\rm phys}$. The fit ranges are: {\it top:} all $\beta$, $135\,{\rm MeV} \leq M_{ll} \leq 390\,{\rm MeV}$, {\it middle:} all $\beta$, $135\,{\rm MeV} \leq M_{ll} \leq 275\,{\rm MeV}$, {\it bottom:} only $1/a\geq 1.35\,{\rm GeV}$, $135\,{\rm MeV} \leq M_{ll} \leq 340\,{\rm MeV}$.}
\label{fig:fits_NLO}
\end{figure}
\vspace*{\afterFigure}
Here we are mainly interested in the SU(2) low-energy constants $\bar{l}_3$ and $\bar{l}_4$ which are related to the low-energy scales $\Lambda_3$, $\Lambda_4$, respectively. Therefore, in Fig.~\ref{fig:summary_lbar3_lbar4_f} we show the fitted values for these parameters obtained with different fit ranges. We also display the ratio $f_\pi/f$ as obtained from the various fits. Whereas for $\bar{l}_3$ (left panel), if at all, one could identify a shift in the result depending on whether or not the two coarsest lattices ensemble are excluded, for $\bar{l}_4$ and $f_\pi/f$ one observes a clear dependency on the fitted mass range, while the influence of excluding coarser lattice ensembles seems to have only a marginal effect. Eventually, we quote as our result for the low-energy constants the central value and statistical error obtained from the fit range $135\,{\rm MeV}\leq M_{ll}\leq275\,{\rm MeV}$, $1/a\geq 1.3\,{\rm GeV}$ and take the variation with respect to that value from other fits including the nearly physical points (data marked by asterisks in Fig.~\ref{fig:summary_lbar3_lbar4_f}) as our estimate for the systematic error, so that we obtain:
\begin{equation}
\label{eq:LECs_fit}
\bar{l}_3\;=\;2.90(11)_{\rm stat}(17)_{\rm syst}\,,\;\;\;\bar{l}_4\;=\;4.04(04)_{\rm stat}(13)_{\rm syst}\,,\;\;\;f_\pi/f\;=\;1.0627(07)_{\rm stat}(24)_{\rm syst}\,.
\end{equation}
\begin{figure}[t!]
\begin{center}
\includegraphics*[width=.33\textwidth]{Plots/lbar3.ps}%
\includegraphics*[width=.33\textwidth]{Plots/lbar4.ps}%
\includegraphics*[width=.33\textwidth]{Plots/fratio.ps}
\end{center}
\vspace*{\closercaption}
\caption{LECs obtained from NLO SU(2) chPT fits with different fit ranges: {\it left panel:} $\bar{l}_3$, {\it middle panel:} $\bar{l}_4$, {\it right panel:} $f_\pi/f$. {\it Blue points} denote fits where $1/a\geq1.35\,{\rm GeV}$, {\it red points} fits where all $\beta$ are included. Fits including the nearly physical points are marked by an {\it asterisk}. The {\it solid}, {\it dashed} and {\it dashed-dotted lines} display the central value, statistical and combined (stat.\ and syst.) error, resp., of our quoted results.}
\label{fig:summary_lbar3_lbar4_f}
\end{figure}
\vspace*{\afterFigure}
\begin{figure}[t!]
\begin{center}
\includegraphics*[angle=-90,width=.49\textwidth]{Plots/fpi_bFine_m3-6.ps}%
\includegraphics*[angle=-90,width=.49\textwidth]{Plots/fpi_bFine_m3-8.ps}
\end{center}
\vspace*{\closercaption}
\caption{Two examples for NLO SU(2) chPT fits excluding the nearly physical points (only $f_{ll}$ is shown here). {\it Left panel:} $230\,{\rm MeV} \leq M_{ll} \leq 340\,{\rm MeV}$, {\it right panel:} $230\,{\rm MeV} \leq M_{ll} \leq 390\,{\rm MeV}$; both $1/a\geq1.35\,{\rm GeV}$.}
\label{fig:fits_NLO_exclPhys}
\end{figure}
\vspace*{\afterFigure}
Since often lattice data from meson masses larger than the physical $M_\pi$ are extrapolated to the physical point using SU(2) chPT, we also investigated fit ranges excluding the physical point. In Fig.~\ref{fig:fits_NLO_exclPhys} the fits for the meson decay constant are shown for $230\,{\rm MeV}\leq M_{ll} \leq 340\,{\rm MeV}$ (left panel) and $230\,{\rm MeV}\leq M_{ll} \leq 390\,{\rm MeV}$ (right panel). As one can see in the close-up view of the region near the physical point, the value for $f_\pi$ extrapolated from such a fit is below ($f_\pi^{\rm extr.}\approx 128(1)\,{\rm MeV}$) the values simulated near the physical point. As one can see from Fig.~\ref{fig:summary_lbar3_lbar4_f}, also $\bar{l}_4$ and $f_\pi/f$ are significantly changing, once the nearly physical points are excluded from the fit range.
\section{Fits to NNLO SU(2) chPT}
\label{subsec:fit.nnlo}
\vspace*{\closersection}
Extending the SU(2) chPT fit formulae for the meson masses and decay constants to NNLO (e.g.\ cf.~\cite{Colangelo:2010et}), in our set-up three new fit parameters have to be added: a combination of the NLO low-energy constants $\bar{l}_1$, $\bar{l}_2$: $\bar{l}_{12}=(7\bar{l}_1+8\bar{l}_2)/15$ and two parameters for NNLO-LECs $k_m$, $k_f$. Again fitting our data for the meson masses and decay constants at various $\beta$ simultaneously now using the NNLO fit formulae without any constraints on the fit parameters (7 in total) leads to an unnatural order of the NLO- compared to the NNLO-contribution as can be seen from the left panel of Fig.~\ref{fig:fits_NNLO}. There the black line denotes the full fit up to NNLO and the red line only the contribution up to NLO, the large difference between the two being the NNLO-contribution. The situation can be improved by using priors for some of the fit parameters, e.g., using a phenomenological estimate for $\bar{l}_{12}=2.1\pm0.3$ as can be obtained from values quoted for $\bar{l}_1$, $\bar{l}_2$ in \cite{Colangelo:2001df}. A fit using such a prior is shown in the right panel of Fig.~\ref{fig:fits_NNLO}, which describes the data well and has a reasonable ordering of the NLO- compared to NNLO-contribution. Still we refrain from using NNLO-chPT as long as we do not have enough data in the light quark mass region to constrain such fits without having to rely on additional input used for priors on the fit parameters. But it is reassuring to us, that by using such priors, a NNLO-fit results in NLO-LECs comparable to those found in our NLO-fits.
\begin{figure}[t!]
\begin{center}
\includegraphics*[angle=-90,width=.49\textwidth]{Plots/NNLO_fpi_bAll_m1-6.ps}%
\includegraphics*[angle=-90,width=.49\textwidth]{Plots/NNLO_priorl12_fpi_bAll_m1-6.ps}
\end{center}
\vspace*{\closercaption}
\caption{Two examples for NNLO SU(2) chPT fits (only $f_{ll}$ is shown here). {\it Left panel:} without priors, {\it right panel:} using prior $\bar{l}_{12}=2.1\pm0.3$. Both use all $\beta$ and $135\,{\rm MeV} \leq M_{ll} \leq 340\,{\rm MeV}$.}
\label{fig:fits_NNLO}
\end{figure}
\vspace*{\afterFigure}
\section{Conclusions}
\vspace*{\closersection}
From our NLO SU(2) chPT fits to meson masses and decay constants measured on staggered 2+1 flavor lattice simulations of QCD, we quote the following set of LECs (see Eq.~(\ref{eq:LECs_fit})) as our preliminary result:
\[ \bar{l}_3\;=\;2.90\pm0.20\,,\;\;\;\bar{l}_4\;=\;4.04\pm0.14\,,\;\;\;f_\pi/f\;=\;1.0627\pm0.0025\,.\]
These values are in good agreement with other recent lattice determinations of LECs, for example the FLAG-report \cite{Colangelo:2010et} quotes $\bar{l}_3=3.2\pm0.8$ and $f_\pi/f=1.073(15)$ as lattice averages, while due to some tension in the results no value for $\bar{l}_4$ is quoted at the moment. Our findings also agree well with the phenomenological estimates $\bar{l}_3=2.9\pm2.4$ and $\bar{l}_4=4.4\pm0.2$ \cite{Colangelo:2001df} and $f_\pi/f=1.0719\pm0.0052$ \cite{Colangelo:2003hf,Colangelo:2010et}.
For a forthcoming publication we hope to have additional data points available at light quark masses corresponding to meson masses between 135 and 275 MeV. More details about our chiral fits will be reported there as well.
The speaker acknowledges support from the DFG SFB/TR 55 and the EU grant PITN-GA-2009-238353 (ITN STRONGnet).
\vspace*{-.4cm}
|
\section{Introduction}
Powerful concentration and deviation inequalities for suprema of empirical
processes have been derived during the last 20 years. These
inequalities turned out to be crucial for example in the study of
consistency and rates of convergence for many estimators. Unfortunately,
the known inequalities are only valid for bounded empirical processes or under strict
tail assumptions. So, this paper was prompted by the question whether
useful inequalities can be obtained under considerably weaker
assumptions.\\
Let us first set the framework, starting with a brief summary of the known
results for bounded empirical processes, or more
precisely, for empirical processes index by bounded functions. To this end, we
consider independent and identically
distributed random variables $X_1,...,X_n$ and a countable function class $\mathcal F$
such that $\sup_{f\in\mathcal{F}}\Vert
f\Vert_\infty\leq 1$ and $\sup_{f\in\mathcal{F}}|\mathbb{E} f(X_1)|=0$. The
quantity of interest is then $Y:=\sup_{f\in\mathcal{F}}|\frac{1}{n}\sum_{i=1}^n
f(X_i)|$. Bousquet derives in
\citep{Bousquet02} with $\sigma_Y^2:=\sup_{f\in\mathcal{F}}\operatorname{Var}f(X_1)$ and
$\nu:=\sigma_Y^2+2\mathbb{E}[Y]$ the exponential deviation inequality
\begin{equation*}
\mathbb{P}\left(Y-EY\geq \sqrt{2x\nu}+\frac{x}{3}\right)\leq
e^{\text{-}nx}\text{~~~for all~}x>0.
\end{equation*}
Bousquet's proof is a refinement of Rio's proof
in \citep{Rio02} and relies on the entropy
method (see for example \citep[Chapter~5.3]{Massart07}).
Similar exponential bounds
for bounded empirical processes have been found by Klein and Rio
\citep{Klein05} and by Massart \citep{Massart00}. These bounds are slightly
less sharp, but additionally hold for
not necessarily identically distributed random variables and also for
$\text{-Y}$. We finally mention Talagrand's work
\citep{Talagrand96b} that probably provided the spark for the development
in this field.\\
Results are also known for possibly unbounded empirical processes that have
very weak tails. We
consider independent and identically
distributed random variables $X_1,...,X_n$ and a function class $\mathcal
F$ that fulfills the Bernstein conditions, that is, $\sup_{f\in\mathcal{F}}\frac{1}{n}\sum_{i=1}^n\mathbb{E}
|f(X_i)|^m\leq\frac{m!}{2}K^{m-2},~m=2,3,...$ for a constant
$K$. Additionally, we assume
$\sup_{i,f\in\mathcal{F}}|\mathbb{E} f(X_i)|=0$ and $\operatorname{card}\mathcal{F}=p$. B\"uhlmann
and van de Geer then derive in \citep{Buhlmann11} the exponential deviation inequality
\begin{equation*}
\mathbb{P}\left(Y\geq Kx +\sqrt{2x}
+\sqrt{\frac{2\log(2p)}{n}}
+\frac{K\log(2p)}{n}\right)\leq
e^{\text{-}nx}\text{~~~for all~}x>0
\end{equation*}
for $Y$ as above. Other exponential bounds for unbounded empirical
processes are given by Adamczak in
\citep{Adamczak08} and by van de Geer and Lederer in
\citep{vdGeer11b}. These authors assume very weak tails with respect to suitable
Orlicz norms.\\
But what if the empirical process is unbounded and does not fulfill the
strict tail assumptions mentioned above? There is no hope to derive
exponential bounds as above under considerably weaker
assumptions. However, we show in the following that weak moment assumptions
are sufficient to obtain useful moment type
concentration inequalities. For this purpose, we
consider independent, not necessarily identically
distributed random variables $X_1,...,X_n$ and a countable
function class $\mathcal F$ with an envelope that has $p$th moment at
most $M^p$ for a $p\in[1,\infty)$. Our main result,
Theorem~\ref{lemma.ConIn2.FirstCor}, implies then for the quantity of
interest $Y:=\sup_{f\in\mathcal{F}}|\frac{1}{n}\sum_{i=1}^n
(f(X_i)-\mathbb{E} f(X_i))|$, and for $1\leq l\leq p$, $\sigma_Y$ as above, and $(\cdot)_+^l:=\left(\max\{0,\cdot\}\right)^l$
\begin{equation*}
\mathbb{E} \left[Y-(1+\epsilon)\mathbb{E} {Y}\right]_+^l\leq
\left(\left(\frac{64}{\epsilon}+7+\epsilon\right)\frac{lM}{n^{1-\frac{l}{p}}}+\frac{4\sqrt
l{\sigma_Y}}{\sqrt n} \right)^l\text{~~~for all~}\epsilon>0
\end{equation*}
and
\begin{align*}
\mathbb{E} \left[(1-\epsilon)\mathbb{E} {Y}-Y\right]_+^l\leq
\left(\left(\frac{86.4}{\epsilon}+7-\epsilon\right)\frac{lM}{n^{1-\frac{l}{p}}}+\frac{4.7\sqrt
l{\sigma_Y}}{\sqrt n} \right)^l\text{~~~for~}\epsilon\in (0,1].
\end{align*}
\noindent We argue in
Section~\ref{sec.M1} that this result is especially useful in
the common case where the envelope (measured by $M$) is much larger than
the single functions (measured by ${\sigma_Y}$).\\
We close this section with a short outline of the paper. In
Section~\ref{sec.Guideline}, we give the basic definitions and
assumptions. In
Section~\ref{sec.M1}, we then state and discuss the main result. This is
followed by complementary bounds in Section~\ref{sec.M2}. Detailed proofs are finally given in Section~\ref{sec.proofs}.
\section{Random Vectors, Concentration Inequalities and Envelopes}
\label{sec.Guideline}
We are mainly interested in the behavior of \itshape suprema of empirical
processes \normalfont
\begin{equation}
\label{eq.intro.ep}
Y:=\sup_{f\in\mathcal F}\left |\frac{1}{n}\sum_{i=1}^n f(X_i)\right|\text{~~or~~}Y:=\sup_{f\in\mathcal F}\left|\frac{1}{n}\sum_{i=1}^n
\left(f(X_i)-\mathbb{E} f(X_i)\right)\right|
\end{equation}
for large $n$. Here, $X_1,...,X_n$ are independent, not necessarily
identically distributed random variables and $\{f:f\in\mathcal{F}\}$ is a
countable family of
real, measurable functions. In the sequel, we may restrict
ourselves to finitely many functions by virtue of the monotonous convergence
theorem.\\
\itshape Random vectors \normalfont generalize the notion of empirical
processes. Let $\mathcal{Z}_1,...,\mathcal{Z}_n$ be arbitrary probability
spaces and $\{Z_i(j):\mathcal{Z}_i\to \mathbb{R}, 1\leq j\leq N, 1\leq i\leq n\}$
a set of random variables. We then define the random vectors as $Z(j):=(Z_1(j),...,Z_n(j))^T:\mathcal{Z}_1\times
...\times\mathcal{Z}_n\to\mathbb{R}^n$. For convenience, we introduce their mean
as $ P Z(j):=\frac{1}{n}\sum_{i=1}^n\mathbb{E} Z_i(j)$ and their empirical mean as
$ \mathbb{P}_nZ(j):=\frac{1}{n}\sum_{i=1}^nZ_i(j).$ Throughout this paper, we
then consider the generalized formulation of \eqref{eq.intro.ep}
\begin{align}\label{eq.intro.rv}
Z&:=\max_{1\leq j\leq N}|\mathbb{P}_n Z(j)|.
\end{align}
The corresponding results for the empirical
processes \eqref{eq.intro.ep} can be found via $Z_i(j):= f_j(X_i)$ or $Z_i(j):=f_j(X_i)-\mathbb{E}
f_j(X_i)$ for $\mathcal F=\{f_1,...,f_N\}$.\\
\itshape Concentration
inequalities \normalfont are a standard tool to characterize the behavior of
the process \eqref{eq.intro.rv} (and thus of \eqref{eq.intro.ep}). For
$n\to\infty$, the process \eqref{eq.intro.rv} is typically governed by the central limit theorem. In contrast, concentration inequalities bound the deviation
in both directions
from the mean or related quantities for finite $n$. Similarly, \itshape
deviation inequalities \normalfont bound the deviation in one direction
only. Concentration or deviation inequalities - contrarily to
maximal inequalities for example - provide bounds that depend only on $n$ and
moment properties of an envelope and the single functions $f$ (and particularly not on
$N$).\\
Let us finally express the moment properties of the envelope. First, we call $\mathcal{E}:=(\mathcal{E}_1,...,\mathcal{E}_n)^T :\mathcal{Z}_1\times
...\times\mathcal{Z}_n\to\mathbb{R}^n$ an \itshape envelope \normalfont if $|Z_i(j)|\leq \mathcal{E}_i$ for all $1\leq j\leq
N$ and $1\leq i \leq n$ . The basic assumption of this paper is then that
there is a $p\in [1,\infty)$ and an $M>0$ such that
\begin{equation}
\mathbb{E}\mathcal{E}_i^p\leq M^p
\end{equation}
for all
$1\leq i \leq n$. Typically,
the envelope is much larger than the single random vectors. In these cases,
we have $M\gg \sigma:=\sqrt{\max_{1\leq j \leq N}\frac{1}{n}\sum_{i=1}^n\mathbb{E} Z_i(j)^2}$.
\section{Proofs}
\label{sec.proofs}
In this last section we give detailed proofs.
\subsection{Proof of Theorem~\ref{lemma.ConIn2.FirstCor}}
The key idea of our proofs is to introduce an appropriate truncation that
depends on the envelope of the
empirical process. This allows us to split the problem into two parts that can be
treated separately: On the one hand, a part corresponding to a bounded empirical process that can be treated by convexity
arguments and Massart's
results on bounded random vectors \citep{Massart00}. And
on the other hand, a part corresponding to an unbounded empirical process
that can be treated by rather elementary means.\\
For ease of exposition we present some convenient
notation for the truncation first. After deriving two simple auxiliary
results, we then turn
to the main task of this section: We prove Lemma~\ref{thm.ConIn2.wmass2}, a generalization
of Theorem~\ref{lemma.ConIn2.FirstCor}. The main result of this paper, Theorem~\ref{lemma.ConIn2.FirstCor},
is then an easy consequence.\\
A basic tool used in this section is \itshape truncation\normalfont. Before
turning to the proofs, we want to give some additional notation for this
tool. First, we define the \itshape unbounded \normalfont and the \itshape bounded part of the random vectors \normalfont as
\begin{alignat*}{2}
\overline{Z}(j)&:=(\overline{Z}_1(j),...,\overline{Z}_n(j))^T:=(Z_1(j)1_{\{\mathcal{E}_1>K\}},...,Z_n(j)1_{\{\mathcal{E}_n>K\}})^T\\
\underline{Z}(j)&:=(\underline{Z}_1(j),...,\underline{Z}_n(j))^T:=(Z_1(j)1_{\{\mathcal{E}_1\leq
K\}},...,Z_n(j)1_{\{\mathcal{E}_n\leq
K\}})^T.
\end{alignat*}
Similarly, we define
\begin{alignat*}{2}
\overline{\mathcal E}(j)&:=(\overline{\mathcal{E}}_1(j),...,\overline{\mathcal{E}}_n(j))^T:=(\mathcal{E}_1(j)1_{\{\mathcal{E}_1>K\}},...,\mathcal{E}_n(j)1_{\{\mathcal{E}_n>K\}})^T\\
\underline{\mathcal{E}}(j)&:=(\underline{\mathcal{E}}_1(j),...,\underline{\mathcal{E}}_n(j))^T:=(\mathcal{E}_1(j)1_{\{\mathcal{E}_1\leq
K\}},...,\mathcal{E}_n(j)1_{\{\mathcal{E}_n\leq
K\}})^T.
\end{alignat*}
To prevent an overflow of indices, the \itshape truncation level
\normalfont $K>0$ is not included explicitly in the notation. The
truncation level is, however, given at the adequate places so that there
should not be any confusion. Finally, we define the maxima of the truncated random variables as
\begin{align*}
\overline{Z}:=\max_{1\leq j\leq
N}|\mathbb{P}_n \overline{Z}(j)|\text{~~and~~}
\underline{Z}:=\max_{1\leq j\leq
N}|\mathbb{P}_n \underline{Z}(j)|.
\end{align*}
Now we derive two simple auxiliary lemmas.
\begin{lemma}
\label{prop.ConIn2.moment}
Let $l\geq 1$, $W_i:\mathcal{Z}_i\to \mathbb{R}^+_0$ for $1\leq i \leq n$ and
$\mathbb{E} W_i^l\leq 1$. Then,
\begin{equation*}
\mathbb{E}\left[\mathbb{P}_n W\right]^l\leq 1
\end{equation*}
for the corresponding random vector $W$ on the product space.
\end{lemma}
\begin{proof} By the triangle inequality we have
\begin{equation*}
\left(\mathbb{E}\left[\mathbb{P}_n W\right]^l\right)^\frac{1}{l}= \frac{1}{n}\left(\mathbb{E}\left[\sum_{i=1}^n W_i\right]^l\right)^\frac{1}{l}\leq \frac{1}{n}\sum_{i=1}^n\left(\mathbb{E}\left[ W_i^l\right]\right)^\frac{1}{l}\leq 1.
\end{equation*}
\end{proof}
\begin{lemma}
\label{lemma.ConIn.Verg}
Under the assumptions of Theorem~\ref{lemma.ConIn2.FirstCor} it holds
for $K=n^{\frac{l}{p}}M$ that
\begin{equation*}
|\mathbb{E}[\underline{Z}-Z]|\leq \frac{M}{n^{l(1-\frac{1}{p})}}
\end{equation*}
and $\underline\sigma\leq \sigma$ for
\begin{align*}
{\underline{\sigma}}&:=\sqrt{\max_{1\leq j \leq N}\frac{1}{n}\sum_{i=1}^n\operatorname{Var}\underline{Z}_i(j)}.
\end{align*}
\end{lemma}
\begin{proof}
Since $||a|-|b||\leq |a-b|$ for all $a,b\in\mathbb{R}$, it holds that
\begin{align*}
\left|\mathbb{E}[\underline{Z}-Z]\right|&=\left|\mathbb{E}\left[\max_{1\leq j\leq
N}|\mathbb{P}_n \underline{Z}(j)|-\max_{1\leq j\leq
N}|\mathbb{P}_n Z(j)|\right]\right|\\
&\leq\mathbb{E}\left[\max_{1\leq j\leq
N}||\mathbb{P}_n \underline{Z}(j)|-|\mathbb{P}_n Z(j)||\right]\\
&\leq\mathbb{E}\left[\max_{1\leq j\leq
N}|\mathbb{P}_n( \underline{Z}(j)-Z(j))|\right]\\
&=\mathbb{E}\left[\max_{1\leq j\leq
N}|\mathbb{P}_n\overline{Z}|\right]\\
&\leq\mathbb{E}\left[\frac{1}{n}\sum_{i=1}^n\overline{\mathcal{E}}_i\right]
\end{align*}
With H\"older's and Chebyshev's Inequality we obtain for $1\leq i\leq n$
\begin{align*}
\mathbb{E} \overline{\mathcal{E}}_i^l
=&\mathbb{E}\mathcal{E}_i^l 1_{\{\mathcal{E}_i>K\}}\\
\leq&(\mathbb{E}\mathcal{E}_i^p)^\frac{l}{p} (\mathbb{E}
1_{\{\mathcal{E}_i>K\}})^{1-\frac{l}{p}}\\
\leq&(\mathbb{E}\mathcal{E}_i^p)^\frac{l}{p} \left(\frac{\mathbb{E}\mathcal{E}_i^p}{K^p}\right)^{1-\frac{l}{p}}\\
\leq&\frac{M^p}{K^{p-l}}.
\end{align*}
These two results yield then the first assertion. The second assertion is straightforward.
\end{proof}
We can now turn to the harder part of this section. The following lemma is
a generalization of Theorem~\ref{lemma.ConIn2.FirstCor}. The derivation of
Theorem~\ref{lemma.ConIn2.FirstCor} from this and
Lemma~\ref{lemma.ConIn.Verg} is then a simple task.
\begin{lemma}
\label{thm.ConIn2.wmass2}
Let $1\leq l\leq p$ and $\epsilon,K\in\mathbb{R}^+$. Then,
\begin{align*}
\mathbb{E} \left[Z-(1+\epsilon)\mathbb{E}\underline{Z}\right]_+^l\leq \left(
\left(\frac{64}{\epsilon}+5\right)\frac{lK}{n}+\frac{4\sqrt l\underline{\sigma}}{\sqrt
n}+\frac{M^{\frac{p}{l}}}{K^{\frac{p}{l}-1}}\right)^l
\end{align*}
and
\begin{align*}
\mathbb{E} \left[(1-\epsilon)\mathbb{E}\underline{Z}-Z\right]_+^l\leq\left(
\left(\frac{86.4}{\epsilon}+5\right)\frac{lK}{n}+\frac{4.7\sqrt l\underline{\sigma}}{\sqrt
n}+\frac{M^{\frac{p}{l}}}{K^{\frac{p}{l}-1}}\right)^l.
\end{align*}
\end{lemma}
\begin{proof}
The key idea of the proof is to separate the bounded from the unbounded quantities. On the one hand, we develop bounds for $\mathbb{E}\overline{Z}_+^l$ via elementary means. On the other hand, we develop bounds for $\mathbb{E}\left[\underline{Z}-(1+\epsilon)\mathbb{E}\underline{Z}\right]_+^l$ and
$\mathbb{E}\left[(1-\epsilon)\mathbb{E}\underline{Z}-\underline{Z}\right]_+^l$ via convexity arguments and \citep[Theorem~4]{Massart00}. They can then be combined to deduce the result.\\
We start with the proof of the first inequality. First, we split $Z$ in a bounded and an unbounded part
\begin{align*}\nonumber
Z&=\max_{1\leq j \leq N}|\mathbb{P}_n Z(j)|\\\nonumber
&=\max_{1\leq j \leq N}|\mathbb{P}_n( \underline{Z}(j)+\overline{Z}(j))|\\\nonumber
&\leq\max_{1\leq j \leq N}(|\mathbb{P}_n\underline{Z}(j)|+|\mathbb{P}_n\overline{Z}(j)|)\\
&\leq\underline{Z}+\overline{Z}
\end{align*}
and deduce with the triangle inequality that
\begin{align}\nonumber
&\mathbb{E} \left[Z-(1+\epsilon)\mathbb{E}\underline{Z}\right]_+^l\\\nonumber
\leq&\mathbb{E} \left[\underline{Z}+\overline{Z}-(1+\epsilon)\mathbb{E}\underline{Z}\right]_+^l\\\nonumber
\leq&\mathbb{E} \left[(\underline{Z}-(1+\epsilon)\mathbb{E}\underline{Z})_++\overline{Z}_+\right]^l\\
\leq &\left((\mathbb{E}\left[\underline{Z}-(1+\epsilon)\mathbb{E}
\underline{Z}\right]_+^l)^{\frac{1}{l}}+(\mathbb{E}\overline{Z}_+^{l})^\frac{1}{l}\right)^l.
\label{eq.ConIn2.splitting2}
\end{align}
Now, we turn to the development of bounds for $\mathbb{E}\overline{Z}_+^l$. With the help of H\"older's and Chebyshev's Inequalities we
obtain as above for $1\leq i\leq n$
\begin{align*}
\mathbb{E} \overline{\mathcal{E}}_i^l
\leq&\frac{M^p}{K^{p-l}}.
\end{align*}
That is,
\begin{equation*}
\mathbb{E}\left[\frac{\overline{\mathcal{E}}_i}{(\frac{M^p}{K^{p-l}})^\frac{1}{l}}\right]^l\leq 1.
\end{equation*}
We may consequently apply Lemma~\ref{prop.ConIn2.moment} and obtain
\begin{equation}
\label{eq.ConIn2.Chebby}
\mathbb{E} \overline{Z}_+^l=\mathbb{E} \overline{Z}^l\leq \mathbb{E} [\mathbb{P}_n\overline{\mathcal{E}}]^l\leq\frac{M^p}{K^{p-l}}.
\end{equation}
As a next step, we derive bounds for
$\mathbb{E}\left[\underline{Z}-(1+\epsilon)\mathbb{E}\underline{Z}\right]_+^l$. To begin,
we set $ J:=(\frac{32}{\epsilon}+2.5)K$, $\sigma:=\underline\sigma$ and define the
function $g_l:\mathbb{R}^+\to (1,\infty)$ as
\begin{equation*}
g_l (x):=\exp\left(\frac{\sqrt{n}(\sqrt{2\sigma^2+Jx^\frac{1}{l}}-\sqrt{2}\sigma)}{\sqrt{2}J}\right)^2.
\end{equation*}
Note that $g_l$ is strictly increasing and smooth. Moreover, the first and second derivatives are
\begin{align*}
g_l'(x)=&\frac{2n(\sqrt{2\sigma^2+Jx^\frac{1}{l}}-\sqrt{2}\sigma)}{4J^2\sqrt{2\sigma^2+Jx^\frac{1}{l}}}\frac{J}{l}x^{\frac{1}{l}-1}g_l (x)\\
=& \frac{n}{2l J}\left(1-\frac{\sqrt{2}\sigma}{\sqrt{2\sigma^2+Jx^\frac{1}{l}}}\right)x^{\frac{1}{l}-1}g_l (x)
\end{align*}
and
\begin{align*}
g''_l(x)&= \frac{n^2}{4l^2
J^2}\left(1-\frac{\sqrt{2}\sigma}{\sqrt{2\sigma^2+Jx^\frac{1}{l}}}\right)^2x^{\frac{2}{l}-2}g_l (x)\\
&~~~~~~~+ \frac{n}{2l J}\frac{\sqrt{2}\sigma}{2(2\sigma^2+Jx^\frac{1}{l})^\frac{3}{2}}\frac{J}{l}x^{\frac{2}{l}-2}g_l (x) \\
&~~~~~~~+ \frac{n}{2l
J}\left(1-\frac{\sqrt{2}\sigma}{\sqrt{2\sigma^2+Jx^\frac{1}{l}}}\right)\left(\frac{1}{l}-1\right)x^{\frac{1}{l}-2}g_l (x)\\
&\geq \frac{nx^{\frac{2}{l}-2}g_l (x)}{4l^2 J}\mbox{\fontsize{10}{10}\selectfont $\left(1-\frac{\sqrt{2}\sigma}{\sqrt{2\sigma^2+Jx^\frac{1}{l}}}\right)\left(\left(1-\frac{\sqrt{2}\sigma}{\sqrt{2\sigma^2+Jx^\frac{1}{l}}}\right){\displaystyle \frac{n}{J}}+2(1-l)x^{-\frac{1}{l}}\right)$}.
\end{align*}
We now use the lower bound for the second derivative to find an interval, on which the function $g_l$ is convex. To this end, we observe that for $\sigma>0$
\begin{equation}
\label{eq.ConIn2.ineqx}
\left(1-\frac{\sqrt{2}\sigma}{\sqrt{2\sigma^2+Jx^\frac{1}{l}}}\right)\frac{n}{J}+2(1-l)x^{-\frac{1}{l}}\geq 0
\end{equation}
is equivalent to
\begin{equation*}
1-\frac{1}{\sqrt{1+\frac{J}{2\sigma^2}x^\frac{1}{l}}}\geq \frac{2(l-1)J}{n}x^{-\frac{1}{l}}.
\end{equation*}
This can be rewritten with the definition
$u:=\frac{J}{2\sigma^2}x^\frac{1}{l}>0$ as
\begin{equation*}
1-\frac{1}{\sqrt{1+u}}\geq \frac{(l-1)J^2}{n\sigma^2u}
\end{equation*}
and with the definition $C:=\frac{(l
-1)J^2}{n\sigma^2}\geq 0$ as
\begin{equation}
\label{eq.ConIn2.coninu}
1-\frac{1}{\sqrt{1+u}}\geq \frac{C}{u}.
\end{equation}
We assume now, that $u\geq C$. Then,
\begin{alignat*}{2}
&~~~~~~
&1-\frac{1}{\sqrt{1+u}}&\geq \frac{C}{u}\\
&\Leftrightarrow &\sqrt{1+u}\left(1-\frac{C}{u}\right)&\geq 1\\
&\Leftrightarrow &(1+u)(u^2-2Cu+C^2)&\geq u^2\\
&\Leftrightarrow &u^3-2Cu^2+(C^2-2C)u+C^2&\geq 0.
\end{alignat*}
Considering the equality
\begin{equation*}
u^2-2Cu+C^2-2C= 0
\end{equation*}
with roots $\{C\pm \sqrt{2C}\}$, we deduce that
\begin{equation*}
u^3-2Cu^2+(C^2-2C)u+C^2\geq 0
\end{equation*}
for all $u\geq C+\sqrt{2C}$. Consequently, for $u\geq C+\sqrt{2C}$, Inequality \eqref{eq.ConIn2.coninu} holds true. Hence, if we postulate
\begin{equation}
\label{ConIn2Reqx}
\frac{J}{2\sigma^2}x^\frac{1}{l}\geq
\frac{(l
-1)J^2}{n\sigma^2}+\sqrt{\frac{2(l
-1)J^2}{n\sigma^2}}
\end{equation}
Equation \eqref{eq.ConIn2.ineqx} holds true. The postulate \eqref{ConIn2Reqx} is equivalent to
\begin{equation*}
x^\frac{1}{l}\geq \frac{2(l
-1)J}{n}+\frac{\sqrt{8(l
-1)}\sigma}{\sqrt n}
\end{equation*}
and to
\begin{equation*}
x\geq \left(\frac{2(l
-1)J}{n}+\frac{\sqrt{8(l
-1)}\sigma}{\sqrt n}\right)^l=:I.
\end{equation*}
Additionally, note that with this condition on $x$, Equation \eqref{eq.ConIn2.ineqx}
is also true for $\sigma=0$. So we finally derived, since $\frac{nx^{\frac{2}{l}-2}g_l (x)}{4l^2 J}(1-\frac{\sqrt{2}\sigma}{\sqrt{2\sigma^2+Jx^\frac{1}{l}}})$ is positive, that the function $g_l$ is convex on the domain
$(I,\infty)$.\\
This convexity property makes it possible to apply a result of \citep{Massart00}. To show this, we introduce
\begin{equation*}
X:=(\underline{Z}-(1+\epsilon)\mathbb{E}\underline{Z})_+^l
\end{equation*}
and find with Jensen's Inequality and the fact that $g_l$ is increasing
\begin{equation*}
g_l(\mathbb{E} X)\leq g_l(\mathbb{E} [X\vee I])\leq \mathbb{E} g_l(X\vee I).
\end{equation*}
We used here the notation $a\vee
b:=\max\{a,b\}$ for $a,b\in\mathbb{R}$. Massart's Inequality \citep[Theorem~4,
(13)]{Massart00} for bounded random vectors translates then to our setting as
\begin{equation*}
\mathbb{P}(n\underline{Z}\geq (1+\epsilon)n\mathbb{E}\underline{Z}+\sigma\sqrt{8nx}+\left(\frac{32}{\epsilon}+2.5\right)Kx)\leq e^{-x},
\end{equation*}
where $\epsilon, x >0$. This is equivalent to
\begin{equation}
\label{eq.ConIn2.massart}
\mathbb{P}(\underline{Z}\geq (1+\epsilon)\mathbb{E}\underline{Z}+\sigma\sqrt{\frac{8x}{n}}+\frac{J}{n}x)\leq e^{-x}.
\end{equation}
We then deduce (cf. \cite{vdGeer11b})
\begin{align*}
&\mathbb{E}\exp\left(\frac{\sqrt{n}\left(\sqrt{2\sigma^2+J(X\vee I)^{\frac{1}{l}}}-\sqrt{2}\sigma\right)}{\sqrt{2}J}\right)^2\\
=&\int_0^\infty\mathbb{P}\left(\exp\left(\frac{\sqrt{n}\left(\sqrt{2\sigma^2+J(X\vee I)^{\frac{1}{l}}}-\sqrt{2}\sigma\right)}{\sqrt{2}J}\right)^2>
t\right)dt\\
\leq &1+\int_1^\infty\mathbb{P}\left(\exp\left(\frac{\sqrt{n}\left(\sqrt{2\sigma^2+J(X\vee I)^{\frac{1}{l}}}-\sqrt{2}\sigma\right)}{\sqrt{2}J}\right)^2>
t\right)dt\\
=&1+\int_1^\infty\mathbb{P}\left(\sqrt{2\sigma^2+J(X\vee I)^{\frac{1}{l}}}> \sqrt{2}\sigma+\sqrt{\frac{2J^2}{n}\log
t}\right)dt\\
=&1+\int_1^\infty\mathbb{P}\left(J(X\vee I)^{\frac{1}{l}}> 4\sigma\sqrt{\frac{J^2}{n}\log t}+\frac{2J^2}{n}\log
t\right)dt
\end{align*}
and note that
\begin{align*}
JI^{\frac{1}{l}}&< 4\sigma\sqrt{\frac{J^2}{n}\log
t}+\frac{2J^2}{n}\log t\\
\Leftrightarrow~~~ \frac{2(l
-1)J}{n}+\frac{\sqrt{8(l
-1)}\sigma}{\sqrt n}&<4\sigma\sqrt{\frac{\log t}{n}}+ \frac{2J}{n}\log t.
\end{align*}
This is fulfilled if $t\geq e^{l-1}$. Hence, with Massart's Inequality \eqref{eq.ConIn2.massart},
\begin{align*}
&\mathbb{E}\exp\left(\frac{\sqrt{n}\left(\sqrt{2\sigma^2+J(X\vee I)^{\frac{1}{l}}}-\sqrt{2}\sigma\right)}{\sqrt{2}J}\right)^2\\
\leq &1+e^{l-1}-1+\int_{e^{l-1}}^\infty\mathbb{P}\left(X^{\frac{1}{l}}> 4\sigma\sqrt{\frac{\log t}{n}}+\frac{2J}{n}\log
t\right)dt\\
=&e^{l-1}+\int_{e^{l-1}}^\infty\mathbb{P}\left(\underline{Z}>(1+\epsilon)\mathbb{E}\underline{Z}+4\sigma\sqrt{\frac{\log t}{n}}+ \frac{2J}{n}\log
t\right)dt\\
\leq& e^{l-1}+\int_{e^{l-1}}^\infty\exp(-\log t^2)dt < e^{l}.
\end{align*}
In summary, we have
\begin{equation*}
g_l(\mathbb{E} X)< e^l.
\end{equation*}
This is now inverted, observing that for $y \in
(1,\infty)$ such that
\begin{alignat*}{2}
&~~~~~~ &y&=\exp\left( \frac{\sqrt{n}(\sqrt{2\sigma^2+Jx^\frac{1}{l}}-\sqrt{2}\sigma)}{\sqrt{2}J}\right)^2\\
&\Rightarrow&\sqrt{\log{y}}&=\frac{\sqrt{n}(\sqrt{2\sigma^2+Jx^\frac{1}{l}}-\sqrt{2}\sigma)}{\sqrt{2}J}\\
&\Rightarrow&\frac{\sqrt 2 J}{\sqrt
n}\sqrt{\log{y}}+\sqrt{2}\sigma&=\sqrt{2\sigma^2+Jx^\frac{1}{l}}\\
&\Rightarrow&\frac{2J^2}{n}\log{y}+\frac{4\sigma J}{\sqrt n}\sqrt{\log
y}&=Jx^\frac{1}{l}\\
&\Rightarrow&\left(\frac{2J}{n}\log{y}+\frac{4\sigma}{\sqrt n}\sqrt{\log y}\right)^l&=x.
\end{alignat*}
This is now applied with $x=\mathbb{E} X$ to obtain
\begin{equation}
\mathbb{E} X\leq \left(\frac{2lJ}{n}+\frac{4\sqrt{l}\sigma}{\sqrt n}\right)^l.
\label{eq.ConIn2.bounded}
\end{equation}
We are now ready to collect the terms. Inequalities \eqref{eq.ConIn2.splitting2}, \eqref{eq.ConIn2.Chebby} and \eqref{eq.ConIn2.bounded} give
\begin{align*}
\mathbb{E} \left[Z-(1+\epsilon)\mathbb{E}\underline{Z}\right]_+^l
&=\left(\frac{2lJ}{n}+\frac{4\sqrt{l}\sigma}{\sqrt
n} + \frac{M^{\frac{p}{l}}}{K^{\frac{p}{l}-1}}\right)^l.
\end{align*}
Hence, since $ J=(\frac{32}{\epsilon}+2.5)K$,
\begin{align*}
\mathbb{E} \left[Z-(1+\epsilon)\mathbb{E}\underline{Z}\right]_+^l
&\leq \left( \left(\frac{64}{\epsilon}+5\right)\frac{lK}{n}+\frac{4\sqrt{l}\sigma}{\sqrt
n} + \frac{M^{\frac{p}{l}}}{K^{\frac{p}{l}-1}}\right)^l.
\end{align*}
This finishes the proof of the first part of the lemma. For the second part, we note that
\begin{align*}
\underline{Z}&=\max_{1\leq j \leq N}|\mathbb{P}_n \underline{Z}(j)|\\
&=\max_{1\leq j \leq N}|\mathbb{P}_n((Z(j)-\overline{Z}(j))|\\
&\leq\max_{1\leq j \leq N}(|\mathbb{P}_n Z(j)|+|\mathbb{P}_n\overline{Z}(j)|)\\
&\leq Z+\overline{Z}
\end{align*}
and therefore $Z\geq \underline{Z} - \overline{Z}$. Consequently,
\begin{align*}
&\mathbb{E} \left[(1-\epsilon)\mathbb{E}\underline{Z}-Z\right]_+^l\\
\leq&\mathbb{E} \left[(1-\epsilon)\mathbb{E}\underline{Z}-\underline{Z}+\overline{Z}\right]_+^l\\
\leq&\mathbb{E} \left[((1-\epsilon)\mathbb{E}\underline{Z}-\underline{Z})_++\overline{Z}_+\right]^l\\
=&\left(\mathbb{E}[(1-\epsilon)\mathbb{E}
\underline{Z}-\underline{Z}]_+^l)^\frac{1}{l}+(\mathbb{E}\overline{Z}_+^{l})^\frac{1}{l}\right)^l.
\end{align*}
One can then proceed as in the first part and use \citep[Theorem~4,
(14)]{Massart00}.
\end{proof}
\begin{proof}[Proof of Theorem \ref{lemma.ConIn2.FirstCor}]
Set $K=n^\frac{l}{p}M$ in Lemma~\ref{thm.ConIn2.wmass2} and use
Lemma~\ref{lemma.ConIn.Verg} to replace the truncated quantities by the
original ones.
\end{proof}
\begin{comment}
\begin{proof}[Proof of Lemma \ref{lemma.ConIn.Bsp}] We first observe that
\begin{align*}
\mathbb{E} \left[Z-2\mathbb{E}\underline{Z}\right]_+=&\mathbb{E} \left[Z-2\mathbb{E} Z
-2\mathbb{E}[\underline{Z}-Z]\right]_+\\
\geq& \mathbb{E} \left[Z-2\mathbb{E} Z\right]_+
-2(\mathbb{E}\left[\underline{Z}-Z\right])_+\\
\geq& \mathbb{E} \left[Z-2\mathbb{E} Z\right]_+
-\frac{2M}{n^{1-\frac{1}{p}}}
\end{align*}
and similarly
\begin{align*}
\mathbb{E} \left[\frac{1}{2}\mathbb{E}\underline{Z}-Z\right]_+=&\mathbb{E}
\left[\frac{1}{2}\mathbb{E}[\underline{Z}-Z]+\frac{1}{2}\mathbb{E}{Z}-Z\right]_+\\
\geq &\mathbb{E} \left[\frac{1}{2}\mathbb{E}{Z}-Z\right]_+-\frac{M}{2n^{1-\frac{1}{p}}}
\end{align*}
with the help of Lemma~\ref{lemma.ConIn.Verg}. This is then plugged in
Theorem~\ref{lemma.ConIn2.FirstCor} to obtain the result.
\end{proof}
\end{comment}
\subsection{Proof of Theorem~\ref{theorem.Massart2}}
Here, we prove Theorem~\ref{theorem.Massart2} with the help of
symmetrization and Massart \citep{Massart00}. The proof here is considerably shorter than the proof of Theorem~\ref{lemma.ConIn2.FirstCor}. This is because we do not need tedious convexity arguments.
\begin{proof}[Proof of Theorem \ref{theorem.Massart2}]
The trick is to use symmetrization and desymmetrization arguments so that we are
able to use \citep[Theorem 9]{Massart00} in a favorable way.\\
Beforehand, we define $Z_\epsilon:=\max_{1\leq j\leq N}|\frac{1}{n}\sum_{i=1}^n \epsilon_i Z_i(j)|$ with independent Rademacher random variables $\epsilon_i$. Then, we symmetrize according to \citep[Lemma 2.3.6]{vdVaart00} with the function $\Phi(x)=(x-4\mathbb{E} Z)_+^l$ to obtain
\begin{align*}
& \mathbb{E}\left[Z-4\mathbb{E} Z\right]_+^l\leq \mathbb{E}\left[2Z_\epsilon-4\mathbb{E} Z\right]_+^l
\end{align*}
and we desymmetrize with the function $\Phi(x)=x$ to obtain
\begin{align*}
& \mathbb{E}\left[2Z_\epsilon-4\mathbb{E} Z\right]_+^l\leq \mathbb{E}\left[2Z_\epsilon-\mathbb{E} 2Z_\epsilon\right]_+^l.
\end{align*}
Hence,
\begin{align}
\label{eq.Massartzwo1}
\mathbb{E}\left[Z-4\mathbb{E} Z\right]_+^l\leq 2^l\mathbb{E}\me_\epsilon\left[Z_\epsilon-\mathbb{E} Z_\epsilon\right]_+^l,
\end{align}
where we write here and in the following $\mathbb{E}_\epsilon$ for the expectation and $\mathbb{P}_\epsilon$ for the probability w.r.t. the Rademacher random variables. Next, we observe that
\begin{align*}
&\mathbb{E}_\epsilon\left[Z_\epsilon-\mathbb{E} Z_\epsilon\right]_+^l\\
=&\int_0^\infty\mathbb{P}_\epsilon\left(\left(Z_\epsilon-\mathbb{E} Z_\epsilon\right)_+^l>t\right)dt\\
=&\int_0^\infty\mathbb{P}_\epsilon\left(Z_\epsilon>\mathbb{E} Z_\epsilon + t^\frac{1}{l}\right)dt\\
\leq&\int_0^\infty\mathbb{P}_\epsilon\left(\max_{1\leq j\leq N}\frac{1}{n}\sum_{i=1}^n \epsilon_iZ_i(j)>\mathbb{E} \max_{1\leq j\leq N}\frac{1}{n}\sum_{i=1}^n \epsilon_iZ_i(j) + t^\frac{1}{l}\right)dt\\~~~~~~~~~~&+\int_0^\infty\mathbb{P}_\epsilon\left(\max_{1\leq j\leq N}\text{-}\frac{1}{n}\sum_{i=1}^n \epsilon_iZ_i(j)>\mathbb{E} \max_{1\leq j\leq N}\frac{1}{n}\sum_{i=1}^n \epsilon_iZ_i(j) + t^\frac{1}{l}\right)dt\\
\leq&2\int_0^\infty\mathbb{P}_\epsilon\left(\max_{1\leq j\leq N}\frac{1}{n}\sum_{i=1}^n \epsilon_iZ_i(j)>\mathbb{E} \max_{1\leq j\leq N}\frac{1}{n}\sum_{i=1}^n \epsilon_iZ_i(j) + t^\frac{1}{l}\right)dt.
\end{align*}
In a final step, we apply Massart's Inequality \citep[Theorem 9]{Massart00} with
\begin{equation*}
L^2=\max_{1\leq j\leq N}\sum_{i=1}^n\left(2|Z_i(j)|\right)^2\leq
4n\mathbb{P}_n\mathcal{E}^2,
\end{equation*}
where $\mathbb{P}_n\mathcal{E}^2:=\frac{1}{n}\sum_{i=1}^n\mathcal{E}_i^2$. This yields
\begin{align*}
&2\int_0^\infty\mathbb{P}_\epsilon\left(\max_{1\leq j\leq N}\frac{1}{n}\sum_{i=1}^n \epsilon_iZ_i(j)>\mathbb{E} \max_{1\leq j\leq N}\frac{1}{n}\sum_{i=1}^n \epsilon_iZ_i(j) + t^\frac{1}{l}\right)dt\\
\leq&
2\int_0^\infty\exp\left(-\frac{nt^\frac{2}{l}}{8\mathbb{P}_n\mathcal{E}^2}\right)dt\\
=&2\left(\frac{8}{n}\right)^\frac{l}{2}(\mathbb{P}_n\mathcal{E}^2)^\frac{l}{2}\int_0^\infty\exp\left(-t^\frac{2}{l}\right)dt\\
=&2\left(\frac{8}{n}\right)^\frac{l}{2}(\mathbb{P}_n\mathcal{E}^2)^\frac{l}{2}\frac{l\Gamma\left(\frac{l}{2}\right)}{2}.
\end{align*}
With Inequality~\eqref{eq.Massartzwo1} this gives
\begin{align*}
\mathbb{E}\left[Z-4\mathbb{E} Z\right]_+^l\leq 2^ll\left(\frac{8}{n}\right)^\frac{l}{2}\mathbb{E}\left[\mathbb{P}_n\mathcal{E}^2\right]^\frac{l}{2}\Gamma\left(\frac{l}{2}\right).
\end{align*}
Finally, because of Lemma~\ref{prop.ConIn2.moment}, it holds that
\begin{equation*}
\mathbb{E}\left[\mathbb{P}_n\mathcal{E}^2\right]^\frac{l}{2}\leq \mathbb{E}\left[\mathbb{P}_n\mathcal{E}\right]^{l}\leq M^l
\end{equation*}
and hence
\begin{align*}
\mathbb{E}\left[Z-4\mathbb{E} Z\right]_+^l\leq l\Gamma\left(\frac{l}{2}\right)\left(\frac{32}{n}\right)^\frac{l}{2}M^l.
\end{align*}
\end{proof}
\subsection{Proof of Theorem~\ref{sec4.Corrollary1}}
We eventually derive Theorem~\ref{sec4.Corrollary1} using
truncation. After some auxiliary results, we derive Lemma~\ref{lemma.helpwoe}. This Lemma settles the bounded part of the
problem. It is then used to proof Lemma~\ref{theorem.mainwoe} which is a slight
generalization of the main theorem. Finally, we derive
Theorem~\ref{sec4.Corrollary1} as a simple corollary.\\
We begin with two auxiliary lemmas:
\begin{lemma}
\label{lemma.helpfirst}
Let $W$ be a centered random variable with values in $[-A,A]$, $A\geq 0$, such that $\mathbb{E} W^2\leq 1$. Then,
\begin{equation*}
\mathbb{E} e^\frac{W}{A}\leq 1 + \frac{1}{A^2}.
\end{equation*}
\end{lemma}
\begin{pro}
We follow well known ideas (see e.g. \cite[Chapter~14]{Buhlmann11}):
\begin{align*}
\mathbb{E} e^\frac{W}{A}&= 1 + \mathbb{E}\left [ e^\frac{W}{A}-1-\frac{W}{A}\right]\\
&\leq 1 + \mathbb{E}\left[ e^\frac{|W|}{A}-1-\frac{|W|}{A}\right]\\
&=1+\sum_{m=2}^\infty\frac{\mathbb{E} |W|^m}{m!A^m}\\
&\leq 1+\sum_{m=2}^\infty\frac{A^{m-2}}{m!A^m}\\
&\leq 1 +\frac{1}{A^2}.
\end{align*}
\end{pro}
\begin{lemma}
\label{lemma.ConIn4.Combi}
Let $C_m^n:=|\{ (i_1,...,i_m)^T\in\{ 1,...,n \}^m :\forall j\in \{ 1,...,m \} \exists j'\in\{ 1,...,m \},j'\neq j,i_j=i_{j'}\}|$ for $m,n\in\mathbb{N}$. Then,
\begin{equation*}
C_m^n\leq m!\left(\frac{n}{2}\right)^{\lfloor \frac{m}{2}\rfloor}.
\end{equation*}
\end{lemma}
\begin{proof}
The proof of this lemma is a simple counting exercise. We start with the case $m\leq 2$. One finds easily that $C_1^n=0$ and $C_2^n=n$, which completes the case $m\leq 2$. Next, we consider the case $m>2$. To this end, we note that $C_m^1=1$, $C_3^2=2$ and $C_m^2\leq 2^m\leq m!$ for $m>3$. This completes the cases $n\leq 2$. Now, we do an induction in $n$. So we let $n\geq 2$ and find
\begin{align*}
C_m^{n+1}=&C_m^n +\frac{m(m-1)}{2!}C_{m-2}^n\\&+\frac{m(m-1)(m-2)}{3!}C_{m-3}^n
+...+\frac{m(m-1)...3}{(m-2)!}C_{2}^n+1.
\end{align*}
By induction, this yields
\begin{align*}
C_m^{n+1}\leq&m!\Big[\left(\frac{n}{2}\right)^{\lfloor
\frac{m}{2}\rfloor}+\frac{1}{2!}\left(\frac{n}{2}\right)^{\lfloor
\frac{m-2}{2}\rfloor}\\
&+\frac{1}{3!}\left(\frac{n}{2}\right)^{\lfloor \frac{m-3}{2}\rfloor}+...
+\frac{1}{(m-2)!}\left(\frac{n}{2}\right)^{\lfloor \frac{2}{2}\rfloor}\Big]
+1.
\end{align*}
We now assume that $m$ is even. So,
\begin{align*}
C_m^{n+1}\leq&m!\left[ \left(\frac{n}{2}\right)^\frac{m}{2}+\frac{1}{2!}\left(\frac{n}{2}\right)^{\frac{m}{2}-1}+\frac{1}{3!}\left(\frac{n}{2}\right)^{\frac{m}{2}-2}+...+\frac{1}{(m-2)!}\left(\frac{n}{2}\right)\right]+1\\
=&m!\left[\left(\frac{n}{2}\right)^\frac{m}{2}+\frac{1}{2!}\left(\frac{n}{2}\right)^{\frac{m}{2}-1}+\sum_{j=2}^{\frac{m}{2}-1}\left(\frac{1}{(2j-1)!}+\frac{1}{\left(2j\right)!}\right)\left(\frac{n}{2}\right)^{\frac{m}{2}-j}\right]+1\\
\leq&m!\left[\left(\frac{n}{2}\right)^\frac{m}{2}+\frac{m}{4}\left(\frac{n}{2}\right)^{\frac{m}{2}-1}+\sum_{j=2}^{\frac{m}{2}-1}{\frac{m}{2}\choose j}\left(\frac{1}{2}\right)^j\left(\frac{n}{2}\right)^{\frac{m}{2}-j}+\left(\frac{1}{2}\right)^\frac{m}{2}\right]\\
=&m!\sum_{j=0}^{\frac{m}{2}}{\frac{m}{2}\choose j}\left(\frac{1}{2}\right)^j\left(\frac{n}{2}\right)^{\frac{m}{2}-j}=
m!\left( \frac{n+1}{2}\right)^{\lfloor \frac{m}{2}\rfloor}.
\end{align*}
This completes the proof for $m>2$ with $m$ even. We note finally, that for
odd $m>2$ we have $C_m^n<mC_{m-1}^n\leq m!\left(\frac{n}{2}\right)^{\lfloor \frac{m}{2}\rfloor}. $
\end{proof}
We now settle the bounded part of the problem. Bounded random variables
are in particular subexponential, so one could apply results from
\citep{Viens07} for
example. But for our purposes, a
direct treament as in the following seems to be more suitable.
\begin{lemma}
\label{lemma.helpwoe}
Let $l\in\mathbb{N}$, $p\geq 2$ and $A\geq 2$. Then, for the truncation level $K=\frac{ A}{2}+\sqrt{\frac{A^2}{4}-1}$,
\begin{equation*}
\mathbb{E} \left[\max_{1\leq j \leq
N}(\mathbb{P}_n -P)\underline{Z}(j)-AM\frac{\log(N)}{n}\right]_+^l\leq \left(\frac{M}{A}+\frac{lAM}{n}\right)^{l}.
\end{equation*}
\end{lemma}
\begin{pro}
We assume w.l.o.g. $M=1$ and observe that
\begin{equation*}
\mathbb{E} \left[\underline{Z}_i(j)-\mathbb{E}\underline{Z}_i(j)\right]^2\leq \mathbb{E}\underline{Z}_i(j)^2\leq 1.
\end{equation*}
Moreover, because of H\"older's and Chebyshev's Inequalities and $K\geq 1$
it holds that
\begin{equation*}
|\underline{Z}_i(j)-\mathbb{E}\underline{Z}_i(j)|\leq|\underline{Z}_i(j)|+|\mathbb{E}\underline{Z}_i(j)|\leq K+\frac{1}{K}=A.
\end{equation*}
These observations, the independence of the random variables and Lemma~\ref{lemma.helpfirst} yield then
\begin{align*}
&\mathbb{E} e^\frac{n(\mathbb{P}_n-P) \underline{Z}(j)}{A}\\
=&\mathbb{E} e^\frac{\sum_{i=1}^n(\underline{Z}_i(j)-\mathbb{E}\underline{Z}_i(j))}{A}\\
\leq&\left(1 +\frac{1}{A^2}\right)^n.
\end{align*}
Next, one checks easily, that the map $x\mapsto e^{x^\frac{1}{l}}$ is convex on the set $[(l-1)^{l},\infty)$. Hence, using Jensen's Inequality again, we obtain
\begin{align*}
&\mathbb{E} \left[\max_{1\leq j \leq
N}(\mathbb{P}_n-P) \underline{Z}(j)-A\frac{\log(N)}{n}\right]_+^{l}\\
\leq& \frac{A^{l}}{n^{l}}\mathbb{E}\left[ \left(\max_{1\leq j \leq
N}n(\mathbb{P}_n-P) \underline{Z}(j)/A-\log(N)\right)_+^{l}\vee (l-1)^{l}\right]\\
\leq& \frac{A^{l}}{n^{l}}\log^{l}\left(\mathbb{E} \exp\left(\left(\max_{1\leq j \leq
N}n(\mathbb{P}_n - P) \underline{Z}(j)/A-\log(N)\right)_+\vee (l-1)\right)\right)\\
=& \frac{A^{l}}{n^{l}}\log^{l}\left(\mathbb{E} \exp\left(\left(\max_{1\leq j \leq
N}n(\mathbb{P}_n - P) \underline{Z}(j)/A-\log(N)\right)\vee (l-1)\right)\right)\\
\leq& \frac{A^{l}}{n^{l}}\log^{l}\left(\max_{1\leq j \leq
N}\mathbb{E} \exp(n(\mathbb{P}_n - P) \underline{Z}(j)/A)+ e^{l-1}\right)\\
\leq& \frac{A^{l}}{n^{l}}\log^l\left(\left(1+\frac{1}{A^2}\right)^n+e^{l-1}\right).
\end{align*}
We finally note that $a+b< eab$ for all $a,b\geq 1$ and find
\begin{align*}
&\mathbb{E} \left[\max_{1\leq j \leq
N}(\mathbb{P}_n-P) \underline{Z}(j)-A\frac{\log(N)}{n}\right]_+^{l}\\
<& \frac{A^{l}}{n^{l}}\log^l\left(\left(1+\frac{1}{A^2}\right)^ne^{l}\right)\\
=& \frac{A^{l}}{n^{l}}\left(\log\left(1+\frac{1}{A^2}\right)^n+\log e^{l}\right)^l\\
\leq&\left(\frac{1}{A}+\frac{lA}{n}\right)^{l}.
\end{align*}
\end{pro}
The results above can now be used to derive a generalization of the main problem.
\begin{lemma}
\label{theorem.mainwoe}
Assume that the random variables $Z_i(j)$ are centered. Then, for
$l\in\mathbb{N}$, $p\geq 2$ , $p\geq l$ and $A\geq 2$,
\begin{equation*}
\label{eq.mainwoe}
\mathbb{E} \left[Z-AM\frac{\log(2N)}{n}\right]_+^l\leq \left(2\left(\frac{2}{A}\right)^{p-1}+(l!)^\frac{1}{l}\sqrt{\frac{2}{n}}+\frac{1}{A}+\frac{lA}{n}\right)^{l}M^l.
\end{equation*}
\end{lemma}
\begin{proof}
The idea is again to separate the bounded and the unbounded quantities. The part with the unbounded quantities is treated by elementary
means and Lemma~\ref{lemma.ConIn4.Combi}. For the bounded part, we use Lemma~\ref{lemma.helpwoe}.\\
First, we assume w.l.o.g. that $M=1$ and set $K=\frac{
A}{2}+\sqrt{\frac{A^2}{4}-1}$. Then, we deduce with the triangle inequality that
\begin{align}\nonumber
&\mathbb{E} \left[\max_{1\leq j \leq
N}\mathbb{P}_n Z(j)-A\frac{\log(N)}{
n}\right]_+^l\\\nonumber
=&\mathbb{E} \left[\max_{1\leq j \leq
N}(\mathbb{P}_n-P)Z(j)-A\frac{\log(N)}{n}\right]_+^l\\\nonumber
\leq&\mathbb{E} \left[\max_{1\leq j \leq
N}(\mathbb{P}_n-P)\overline{Z}(j)+\max_{1\leq j \leq
N}(\mathbb{P}_n-P)\underline{Z}(j)-A\frac{\log(N)}{
n}\right]_+^l\\\nonumber
\leq&\mathbb{E} \left[\left[\max_{1\leq j \leq
N}(\mathbb{P}_n-P)\overline{Z}(j)\right]_++\left[\max_{1\leq j \leq
N}(\mathbb{P}_n-P)\underline{Z}(j)-A\frac{\log(N)}{
n}\right]_+\right]^l\\
\leq&\left(\left(\mathbb{E} \mbox{\fontsize{9}{12}\selectfont $ \left[{\displaystyle \max_{1\leq j \leq
N}}(\mathbb{P}_n-P)\overline{Z}(j)\right]_+^l$}\right)^{\frac{1}{l}}+\left(\mathbb{E}\mbox{\fontsize{9}{12}\selectfont $\left[{\displaystyle \max_{1\leq j \leq
N}}(\mathbb{P}_n-P)\underline{Z}(j)-A\frac{\log(N)}{
n}\right]_+^{l}$}\right)^\frac{1}{l}\right)^l.\label{eq.ConIn4.splitbeg}
\end{align}
So, we are able to treat the unbounded and the bounded quantities separately. We begin with the unbounded quantities. We first note that
\begin{align*}
\left[(\mathbb{P}_n-P)\overline{Z}(j)\right]_+^l
\leq\left((\mathbb{P}_n+P)\overline{\mathcal{E}}\right)^l= \left((\mathbb{P}_n-P)\overline{\mathcal{E}}+2P\overline{\mathcal{E}}\right)^l.
\end{align*}
Hence,
\begin{equation}\label{eq.ConIn4.Splitun}
\mathbb{E} \left[\max_{1\leq j \leq
p}(\mathbb{P}_n-P)\overline{Z}(j)\right]_+^l
\leq \left(2P\overline{\mathcal{E}}+(\mathbb{E} \left[(\mathbb{P}_n-P)\overline{\mathcal{E}}\right]^l)^\frac{1}{l}\right)^{l}.
\end{equation}
H\"older's and Chebyshev's Inequalities are then used to find
\begin{align}
\label{eq.ConIn4.first}
&P\overline{\mathcal{E}}=\frac{1}{n}\sum_{i=1}^n\mathbb{E}\overline{\mathcal{E}}_i
\leq \frac{1}{n}\sum_{i=1}^n({\mathbb{E}\mathcal{E}_i^p})^\frac{1}{p}({\mathbb{E}
1_{\{\mathcal{E}_i>K\}}})^{1-\frac{1}{p}}
\leq \frac{1}{K^{p-1}}.
\end{align}
To bound the quantity left, we note that for all $i$ and $p\geq q\in\mathbb{N}$
\begin{align*}
\mathbb{E}\left[\overline{\mathcal{E}}_i-\mathbb{E}\overline{\mathcal{E}}_i\right]^q
\leq 2^q
\end{align*}
so that
\begin{align*}
\mathbb{E}\left[(\overline{\mathcal{E}}_{i_1} - \mathbb{E}\overline{\mathcal{E}}_{i_1})...(\overline{\mathcal{E}}_{i_l} - \mathbb{E}\overline{\mathcal{E}}_{i_l})\right]
\leq 2^l.
\end{align*}
Moreover, it holds that
\begin{align*}
\mathbb{E}\left[(\overline{\mathcal{E}}_{i_1} - \mathbb{E}\overline{\mathcal{E}}_{i_1})...(\overline{\mathcal{E}}_{i_l} - \mathbb{E}\overline{\mathcal{E}}_{i_l})\right]=0
\end{align*}
for all $i_1,...,i_l$ such that there is a $j$ with $i_j\neq i_{j'}$ for
all $j'\neq j$. With Lemma~\ref{lemma.ConIn4.Combi}, we then get for $n>1$
\begin{align}
&\mathbb{E} \left[(\mathbb{P}_n-P)\overline{\mathcal{E}}\right]^{l}\leq
\frac{2^lC_l^n}{n^l}\leq \frac{2^{l}
l!}{n^l}\left(\frac{n}{2}\right)^{\lfloor \frac{l}{2}\rfloor} \leq l!\sqrt{\frac{2}{n}}^l.\label{eq.ConIn4.second}
\end{align}
Clearly, this also holds for $n=1$ and $l=1$. For $n=1$ and $l>1$ we
note that
\begin{align*}
\mathbb{E} \left[(\mathbb{P}_n-P)\overline{\mathcal{E}}\right]^{l}\leq 2^l \leq l! \sqrt 2^l,
\end{align*}
so that Inequality \eqref{eq.ConIn4.second} holds for all $n$ and $l$ under
consideration. Inserting then Inequalities \eqref{eq.ConIn4.first} and \eqref{eq.ConIn4.second} in Inequality \eqref{eq.ConIn4.Splitun}, we obtain the result for the unbounded part
\begin{align}
\label{eq.ConIn4.boundunbound}
\mathbb{E} \left[\max_{1\leq j \leq p}(\mathbb{P}_n-P)\overline{Z}(j)\right]_+^l
\leq \left(\frac{2}{K^{p-1}}+(l!)^\frac{1}{l}\sqrt{\frac{2}{n}} \right)^l
\end{align}
Next, we plug the result of Lemma~\ref{lemma.helpwoe} and
Inequality~\eqref{eq.ConIn4.boundunbound} in Inequality~\eqref{eq.ConIn4.splitbeg} to derive
\begin{align*}
&\mathbb{E} \left[ \max_{1\leq j \leq
N}\mathbb{P}_n Z(j)-A\frac{\log(N)}{n}\right]_+^l\\
\leq&\left(2\left(\frac{2}{A}\right)^{p-1}+(l!)^\frac{1}{l}\sqrt{\frac{2}{n}}+\frac{1}{A}+\frac{lA}{n}\right)^{l}.
\end{align*}
Finally, we define $Z(j+N):=\text{-}Z(j)$ for $1\leq j \leq N$. We then get
\begin{align*}
\mathbb{E} \left[Z-A\frac{\log(2N)}{n}\right]_+^l=
&\mathbb{E} \left[ \max_{1\leq j \leq
2N}\mathbb{P}_n Z(j)-A\frac{\log(2N)}{n}\right]_+^l\\
\leq &\left(2\left(\frac{2}{A}\right)^{p-1}+(l!)^\frac{1}{l}\sqrt{\frac{2}{n}}+\frac{1}{A}+\frac{lA}{n}\right)^{l}
\end{align*}
replacing $N$ by $2N$ in the results above.
\end{proof}
Theorem~\ref{sec4.Corrollary1} is now a simple corollary.
\begin{proof}[Proof of Theorem \ref{sec4.Corrollary1}]
Set $A=2\sqrt n$ in Lemma~\ref{theorem.mainwoe}.
\end{proof}
\section{Main Result}
\label{sec.M1}
We are mainly concerned with concentration inequalities for unbounded empirical processes that only
fulfill weak moment conditions. Additionally, we want to incorporate empirical
processes with envelopes that may be much larger than the single
functions under consideration.\\
The following theorem is the main result of this paper:
\begin{theorem}
\label{lemma.ConIn2.FirstCor}
For $1\leq l\leq p$ and $\epsilon\in\mathbb{R}^+$ it holds that
\begin{equation*}
\mathbb{E} \left[Z-(1+\epsilon)\mathbb{E} {Z}\right]_+^l\leq
\left(\left(\frac{64}{\epsilon}+7+\epsilon\right)\frac{lM}{n^{1-\frac{l}{p}}}+\frac{4\sqrt
l{\sigma}}{\sqrt n} \right)^l
\end{equation*}
and, if additionally $\epsilon\in (0,1]$,
\begin{align*}
\mathbb{E} \left[(1-\epsilon)\mathbb{E} {Z}-Z\right]_+^l\leq
\left(\left(\frac{86.4}{\epsilon}+7-\epsilon\right)\frac{lM}{n^{1-\frac{l}{p}}}+\frac{4.7\sqrt l{\sigma}}{\sqrt n} \right)^l.
\end{align*}
\end{theorem}
\noindent As discussed in the preceding section, we state our results in terms of random vectors instead
of empirical processes. The connection can be made as
described. Furthermore, we relinquished slightly better constants to obtain
incisive bounds, however, a considerable improvement in $l$ does not seem
to be possible. We finally note that the expectation $\mathbb{E} Z$ can be
replaced by suitable approximations. Such approximations are usually found with chaining and
entropy (see for example \citep{Buhlmann11}, \citep{vdVaart00},
\citep{vdVaart11}) or generic chaining (see for example \citep{Talagrand96},
\citep{Talagrand05}).\\
Let us now have a closer look at the above result. In contrast to the known results given in the introduction, the single functions may be
unbounded and may only fulfill weak moment conditions. For the envelope, the moment restrictions are increasing with
increasing power $l$, as expected.\\
And what about large envelopes, that is $M\gg
{\sigma}$? Theorem~\ref{lemma.ConIn2.FirstCor}
separates the part including the size of the envelope (measured by $M$) from the part
including the size of the single random vectors (measured by
${\sigma}$). For $p>2l$ and $n
\gg 1$, a possibly large value of $M$ is counterbalanced by
$\frac{1}{n^{1-\frac{l}{p}}}\ll\frac{1}{\sqrt n} $ and thus, the influence
of large envelopes is tempered. In particular, the term
including the size of the envelope can be neglected for
$n\to \infty$ if $p$ is sufficiently large.\\
We conclude this section with two straightforward consequences of
Theorem~\ref{lemma.ConIn2.FirstCor}.
\begin{corollary} Theorem~\ref{lemma.ConIn2.FirstCor} directly implies
probability bounds via
Chebyshev's Inequality. Under the above assumptions it holds for $x > 0$
\begin{equation*}
\mathbb{P}\left(Z\geq (1+\epsilon)\mathbb{E} {Z} +x \right)\leq\min_{1\leq l \leq p}\mbox{\fontsize{12}{12}\selectfont $\frac{\left(\left(\frac{64}{\epsilon}+7+\epsilon\right)\frac{lM}{n^{1-\frac{l}{p}}}+\frac{4\sqrt
l{\sigma}}{\sqrt n}\right)^l}{x^l}$}
\end{equation*}
and similarly
\begin{equation*}
\mathbb{P}\left(Z\leq (1-\epsilon)\mathbb{E} {Z} -x \right)\leq\min_{1\leq l \leq p}\mbox{\fontsize{12}{12}\selectfont $\frac{\left(\left(\frac{86.4}{\epsilon}+7-\epsilon\right)\frac{lM}{n^{1-\frac{l}{p}}}+\frac{4.7\sqrt
l{\sigma}}{\sqrt n}\right)^l}{x^l}$}.
\end{equation*}
\end{corollary}
\begin{corollary} Concrete first order bounds under the above assumptions are for example
\label{lemma.ConIn.Bsp}
\begin{alignat*}{5}
\mathbb{E}& \left[Z-2\mathbb{E}{Z}\right]_+
&\leq&~72\frac{M}{n^{1-\frac{1}{p}}}&+&4\frac{\sigma}{\sqrt n}\\
\text{and~~~}\mathbb{E} &\left[\frac{1}{2}\mathbb{E}{Z}-Z\right]_+
&\leq&179.3\frac{M}{n^{1-\frac{1}{p}}}&+&4.7\frac{\sigma}{\sqrt n}.
\end{alignat*}
\end{corollary}
\section{Complementary Bounds}
\label{sec.M2}
In this section, we complement the main result
Theorem~\ref{lemma.ConIn2.FirstCor} with two additional bounds.
These
additional bounds can be of interest if $l$ is close to $p$.\\
The first result reads:
\begin{theorem}\label{theorem.Massart2} Assume that the random variables $Z_i(j)$ are centered. For $l\geq 1$ and $p\geq l$ it holds that
\begin{equation*}
\mathbb{E}\left[Z-4\mathbb{E} Z\right]_+^l\leq l~\Gamma\left(\frac{l}{2}\right)\left(\frac{32}{n}\right)^\frac{l}{2}M^l,
\end{equation*}
where ${\Gamma}$ is the usual Gamma function.
\end{theorem}
Let us compare Theorem~\ref{theorem.Massart2}
with Theorem~\ref{lemma.ConIn2.FirstCor}. On the one hand, the above
result does not possess the flexibility of the factor $(1+\epsilon)$ and is
a deviation inequality only. On the other hand, the term including the size
of the envelope $M$ is independent of $p$ and has a
different power of $n$ in the denominator compared to
the corresponding term in Theorem~\ref{lemma.ConIn2.FirstCor}. Comparing
these two terms in detail, we find that the bound of
Theorem~\ref{theorem.Massart2} may be sharper than the
corresponding bound in Theorem~\ref{theorem.Massart2} if $l\leq p <
2l$.\\
We finally give explicit deviation inequalities for $Z$ in the case of finitely many random
vectors. For $p\geq 2$, explicit bounds are found immediately by replacing $\mathbb{E} Z$ in
Theorem~\ref{lemma.ConIn2.FirstCor} or Theorem~\ref{theorem.Massart2} by the upper bound $\sqrt{\frac{8\log
(2N)}{n}}M$ (see \cite{Duembgen09}). Another bound is found by an
approach detailed in Section~\ref{sec.proofs}. The bound reads:
\begin{theorem}
\label{sec4.Corrollary1}
Let the random variables $Z_i(j)$ be centered. Then, for
$p\geq 2$, $l\in\mathbb{N}$, and $p\geq l$
\begin{equation*}
\mathbb{E} \left[Z-2M\frac{\log(2N)}{\sqrt n}\right]_+^l\leq
\left(\frac{35l^2}{n}\right)^{\frac{l}{2}}M^l.
\end{equation*}
\end{theorem}
\noindent This can supersede the bound in Theorem~\ref{theorem.Massart2}
for $2\sqrt{\log(2N)}\leq 8\sqrt{2}$.
|
\section{Introduction}
It remains a theoretical challenge
to explain the observed pattern of quark and lepton masses and mixings,
in particular the striking differences between the quark sector and the
neutrino sector.
Promising elements of a theory of flavour are grand unification (GUT) based on
the groups $\mathrm{SU(5)}$, $\mathrm{SO(10)}$ or $\mathrm{E_6}$,
supersymmetry, the seesaw mechanism and additional
flavour symmetries \cite{Raby:2008gh}. A successful example is the
Froggatt-Nielsen mechanism~\cite{Froggatt:1978nt} based on spontaneously
broken Abelian symmetries, which parametrizes quark and lepton mass ratios and mixings by powers of a small `hierarchy parameter'~$\eta$.
The resulting structure of mass matrices also arises in compactifications of
higher-dimensional field and string theories, where the parameter $\eta$
is related to the location of matter fields in the compact dimensions
or to vacuum expectation values of moduli fields
(cf.~\cite{extradimensions}).
In this article we consider a Froggatt-Nielsen symmetry which commutes with
the GUT group $\mathrm{SU(5)}$, and which naturally explains the large
$\nu_{\mu}-\nu_{\tau}$ mixing \cite{Froggatt-Nielsen+SU(5)}.
This symmetry implies a particular
hierarchy pattern in the Majorana mass matrix
for the light neutrinos,
\begin{equation}
\label{eq_mnu1}
m_{\nu} \propto
\begin{pmatrix} \eta^{2} & \eta & \eta \\ \eta & 1 & 1 \\
\eta & 1 & 1 \end{pmatrix} \ ,
\end{equation}
which can be regarded as a key element for our analysis.
The predicted Dirac and Majorana neutrino mass matrices are also consistent with leptogenesis~\cite{Buchmuller:1998zf}. Despite these successes,
the predictive power of the Froggatt-Nielsen mechanism is rather limited
due to unknown $\mathcal{O}(1)$ coefficients in all entries of the mass matrices.
For example, the considered model \cite{Buchmuller:1998zf} can
accommodate both a small as well as a large `solar' mixing angle
$\theta_{12}$~\cite{Froggatt-Nielsen+SU(5), Vissani:1998xg}. To get an
idea of the range of possible predictions for a given flavour structure,
it is instructive to
treat the $\mathcal{O}(1)$ parameters as random
variables~\cite{random_coeff}.
In the following we shall employ Monte-Carlo techniques to study
quantitatively the dependence of yet undetermined, but soon testable
parameters of the neutrino sector on the unknown ${\cal O}(1)$ factors
of the mass matrices. Using the already measured
neutrino masses and mixings as input, we find surprisingly sharp predictions
which indicate a large value for the smallest mixing angle $\theta_{13}$
in accordance with recent results from T2K~\cite{Abe:2011sj},
Minos~\cite{Adamson:2011qu} and Double Chooz~\cite{doublechooz}, a value for the lightest neutrino mass
of ${\cal O}(10^{-3})$~eV and one Majorana phase in the mixing
matrix peaked at $\alpha_{21} = \pi$.
\section{Masses and mixings in the lepton sector}
As far as orders of magnitude are concerned, the masses of quarks and charged leptons approximately satisfy the relations
\begin{equation}
\begin{split}
&m_t:m_c:m_u \sim 1:\eta^2 : \eta^4 \,, \\
&m_b : m_s : m_d \sim m_{\tau} : m_{\mu} : m_e \sim 1: \eta : \eta^3 \,,
\end{split}
\label{eq_masses}
\end{equation}
with $\eta^2 \simeq 1/300$ for masses defined at the GUT scale. This mass
hierarchy can be reproduced by a simple $\mathrm{U(1)}$ flavour symmetry.
Grouping the standard model
leptons and quarks into the $\mathrm{SU(5)}$ multiplets
$\textbf{10} = (q_L, u_R^c, e_R^c)$ and $\textbf{5}^* = (d_R^c, l_L)$,
the Yukawa interactions take the form
\begin{equation}
\label{eq_L}
{\cal L}_Y = h_{ij}^{(u)} \textbf{10}_i \textbf{10}_j H_u
+ h_{ij}^{(e)} \textbf{5}^*_i \textbf{10}_j H_d
+ h_{ij}^{(\nu)} \textbf{5}_i^* \textbf{1}_j H_u
+ \frac{1}{2} \, h_{i}^{(n)} \textbf{1}_i \textbf{1}_i S + \mathrm{c.c.}\ ,
\end{equation}
where $\textbf{1} = \nu_{R}^c$ denote the charge conjugates of right-handed neutrinos and
$i,j = 1\ldots 3$
are flavour indices.
Note that the Yukawa matrix $h^{(n)}$ for the right-handed neutrinos
can always be chosen to be real and diagonal.
$H_u$, $H_d$ and $S$ are the Higgs fields for
electroweak and ${B-L}$ symmetry breaking, i.e., their vacuum
expectation values
generate the Dirac masses of quarks and leptons and the Majorana masses for
the right-handed neutrinos, respectively.
In this setup, the Yukawa couplings are determined up to complex ${\cal O}(1)$ factors
by assigning $\mathrm{U(1)}$ charges to the fermion and Higgs fields in Eq.~\eqref{eq_L},
\begin{equation}
\label{eq_h}
h_{ij} \sim \eta^{Q_i + Q_j} \ .
\end{equation}
With the charge assignment given in Tab.~\ref{tab_fn-charges} the mass
relations in Eq.~\eqref{eq_masses} are reproduced. Additionally, perturbativity of the Yukawa couplings and constraints on $\tan \beta = \langle H_u \rangle / \langle H_d \rangle$ require $0 \leq a \leq 1$.
\begin{table}
\begin{center}
\begin{tabular}{c|cccccccccccc}
$\psi_i$ & $\textbf{10}_3$ & $\textbf{10}_2$ & $\textbf{10}_1$ &
$\textbf{5}^*_3$ & $\textbf{5}^*_2$ & $\textbf{5}^*_1$ & $\textbf{1}_3$ &
$\textbf{1}_2$ & $\textbf{1}_1$ & $H_u$ & $H_d$ & $S$ \\
\hline
$Q_i$ & 0& 1 & 2 & $a$ & $a$ & $a+1$ & $b$ & $c$ & $d$ & 0 & 0 & 0
\end{tabular}\end{center}
\caption{Froggatt-Nielsen charge assignments. From Ref. \cite{Buchmuller:1998zf}.}
\label{tab_fn-charges}
\end{table}
\medskip
\newpage \noindent \textit{Masses} \\
\noindent From Eq.~\eqref{eq_L} and Tab.~\ref{tab_fn-charges} one obtains for the Dirac
neutrino mass matrix $m_D$ and the Majorana mass matrix of the right-handed
neutrinos $M$,
\begin{equation}\
\frac{m_D}{v_{EW}\sin\beta} = h^{(\nu)}_{ij} \sim \eta^a
\begin{pmatrix} \eta^{d+1} & \eta^{c+1} & \eta^{b+1} \\
\eta^{d}& \eta^{c} & \eta^{b} \\ \eta^{d} & \eta^{c} & \eta^{b}\end{pmatrix}\ ,
\quad \frac{M}{v_{B-L}} = h_{ij}^{(n)} \sim
\begin{pmatrix} \eta^{2d} & 0 & 0 \\ 0 & \eta^{2c} & 0 \\
0 & 0 & \eta^{2b} \end{pmatrix} \ ,
\end{equation}
with the electroweak and ${B-L}$ symmetry breaking vacuum expectation
values $v_{EW}=\sqrt{\langle H_u \rangle^2 + \langle H_d \rangle^2}$
and $v_{B-L} = \langle S \rangle$,
respectively.
In the seesaw formula
\begin{equation}
\label{eq_seesaw}
m_{\nu} = - m_D \frac{1}{M} m_D^T\ ,
\end{equation}
the dependence on the right-handed neutrino charges drops out, and one finds
for the light neutrino mass matrix,
\begin{equation}
\label{eq_mnu}
m_{\nu} \sim
\frac{v_{EW}^2\sin^2\beta}{v_{B-L}} \ \eta^{2a} \
\begin{pmatrix} \eta^{2} & \eta & \eta \\ \eta & 1 & 1 \\
\eta & 1 & 1 \end{pmatrix} \ .
\end{equation}
The charged lepton mass matrix is given by
\begin{equation}
\frac{m_e}{v_{EW}\cos\beta} = h^{(e)}_{ij} \sim \eta^a
\begin{pmatrix} \eta^{3} & \eta^{2 } & \eta \\
\eta^{2}& \eta & 1 \\ \eta^{2} & \eta & 1\end{pmatrix}\ .
\end{equation}
Note that the second and third row of the matrix $m_e$ have the same hierarchy pattern. This is a consequence of the same flavour charge for the second and third generation of leptons, which is the origin of the large neutrino mixing. Hence, diagonalizing $m_e$ can a priori give a sizable contribution to the mixing in the lepton sector. \medskip
\noindent\textit{Mixing} \\ \noindent
The lepton mass matrices are diagonalized by bi-unitary and unitary
transformations, respectively,
\begin{equation}
V_L^{T} m_e V_R = m_e^{\mathrm{diag}}\ , \quad
U^T m_{\nu} U = m_{\nu}^{\mathrm{diag}} \ ,
\end{equation}
with $V_L^{\dagger}V_L = V_R^{\dagger}V_R = U^{\dagger}U = \mathbb{1}$.
From $V_L$ and $U$ one obtains the leptonic mixing matrix
$U_{\mathrm{PMNS}} = V_L^{\dagger}U$,
which is parametrized as \cite{Nakamura:2010zzi}
\begin{align}\label{pmns}
U_{\mathrm{PMNS}} =
\begin{pmatrix} c_{12}c_{13} & s_{12}c_{13}e^{i\frac{\alpha_{21}}{2}}
& s_{13}e^{i(\frac{\alpha_{31}}{2}-\delta)} \\
-s_{12}c_{23}-c_{12}s_{23}s_{13}e^{i\delta} &
\left(c_{12}c_{23}
-s_{12}s_{23}s_{13}e^{i\delta}\right)e^{i\frac{\alpha_{21}}{2}} &
s_{23}c_{13}e^{i\frac{\alpha_{31}}{2}} \\
s_{12}s_{23}-c_{12}c_{23}s_{13}e^{i\delta} &
\left(-c_{12}s_{23}
-s_{12}c_{23}s_{13}e^{i\delta}\right)e^{i\frac{\alpha_{21}}{2}} &
c_{23}c_{13}e^{i\frac{\alpha_{31}}{2}}\end{pmatrix}\ ,
\end{align}
with $c_{ij} = \cos \theta_{ij}$ and $s_{ij} = \sin \theta_{ij}$. Since the light neutrinos are Majorana fermions, all three phases are physical.
In the following we study the impact of the unspecified ${\cal O}(1)$ factors
in the lepton mass matrices on the various parameters of the neutrino sector
by using a Monte Carlo method, taking present knowledge on neutrino masses
and mixings into account.
Naively, one might expect large uncertainties in the predictions for the observables of the neutrino sector obtained in this setup. For instance, the neutrino mass matrix is calculated by multiplying three matrices, in which each entry comes with an unspecified ${\cal O}(1)$ factor, cf.\ Eq.~\eqref{eq_seesaw}. However, carrying out the analysis described below and calculating the $68\%$ confidence intervals, we find that in many cases our results are sharply peaked, yielding a higher precision than only an order-of-magnitude estimate.
\section{Random variables}
\noindent \textit{Monte-Carlo study}
\noindent The unknown ${\cal O}(1)$ coefficients of the Yukawa matrices $h^{(e)}$,
$h^{(\nu)}$ and $h^{(n)}$ are constrained by the experimental data on neutrino
masses and mixings, with the $3 \sigma$ confidence ranges given by~\cite{Nakamura:2010zzi}:
\begin{equation}
\label{eq_exp}
\begin{split}
& 2.07 \times 10^{-3} \, \text{eV}^2 \leq |\Delta m^2_{\text{atm}}| \leq 2.75 \times 10^{-3} \, \text{eV}^2 \,, \\
& 7.05 \times 10^{-5} \, \text{eV}^2 \leq \Delta m^2_{\text{sol}} \leq 8.34 \times 10^{-5} \, \text{eV}^2 \,, \\
& 0.75 \leq \sin^2(2 \theta_{12}) \leq 0.93 \,, \\
& 0.88 \leq \sin^2(2 \theta_{23}) \leq 1 \,.
\end{split}
\end{equation}
In the following we explicitly do not use the current bound on the smallest mixing angle ($\theta_{13} < 0.21$ at $3 \sigma$ \cite{Nakamura:2010zzi}). This allows us to demonstrate that nearly all values we obtain for $\theta_{13}$ automatically obey the experimental bound, cf.~Fig.~\ref{fig_mixing}.
In a numerical Monte-Carlo study we generate random numbers to model the
39 real parameters of the three mass matrices.\footnote{Nine complex ${\cal O}(1)$ factors in each $h^{(\nu)}$ and $h^{(e)}$, as well as three real ${\cal O}(1)$ factors in $h^{(n)}$. Note that here we are treating the low energy Yukawa couplings as random variables, which are related to the couplings at higher energy scales via renormalization group equations. However, we expect that the effect of this renormalization group running can essentially be absorbed into a redefinition of the effective scale $\bar{v}_{B-L}$, hence leaving the results presented in the following unchanged.} The absolute values are taken
to be uniformly distributed in $ [10^{-1/2}, 10^{1/2}]$ on a logarithmic
scale. The phases in $h^{(e)}$ and $h^{(\nu)}$ are chosen to be uniformly
distributed in $ [0, 2 \pi)$. In the following, we shall refer to those sets
of coefficients which are consistent with the
experimental constraints in Eq.~\eqref{eq_exp} as hits.
In a preliminary run, we consider the neutrino mixing matrix $U$,
with the effective scale $\bar{v}_{B-L} \equiv \eta^{-2a} v_{B-L}/\sin^2\beta$
treated as random variable in the interval
$[10^{-1/2}, 10^{1/2}]\times10^{15}$~GeV. We find that the percentage of hits
strongly peaks at $\bar{v}_{B-L} \simeq 1 \times 10^{15}$~GeV. This is
interesting for two reasons. Firstly, it implies that given $0 \leq a \leq 1$,
the high seesaw scale lies in the range
$3\times 10^{12}~\text{GeV}\lesssim v_{B-L}/\sin^2\beta\lesssim
1\times 10^{15}~\text{GeV}$.
Note that the upper part of this mass range is close the GUT scale, which is
important for recent work on the connection of leptogenesis, gravitino dark
matter and hybrid inflation
\cite{cosmo}.
Secondly, this result allows us to fix the parameter $\bar{v}_{B-L}$ in the
following computations without introducing a significant bias.
In the main run, for fixed $\bar{v}_{B-L}$, we include the mixing matrix $V_L$ of
the charged leptons to compute the full PMNS matrix. We require the mass ratios
of the charged leptons to fulfill the experimental constraints up to an accuracy
of $5\%$ and allow for ${1\leq\tan\beta \leq 60}$ to achieve the correct
normalization of the charged lepton mass spectrum. Finally, imposing the
$3\sigma$ constraints on the two large mixing angles of the full PMNS matrix, we find
parameter sets of ${\cal O}(1)$ factors which yield mass matrices
fulfilling the constraints in Eq.~\eqref{eq_exp}. Our final results are based on roughly
20\,000 such hits. For each hit we calculate the observables in the neutrino
sector as well as parameters relevant for leptogenesis. The resulting
distributions are discussed below.\medskip
\noindent \textit{Statistical analysis}
\noindent In our theoretical setup the relative frequency with which we encounter a
certain value for an observable might indicate the probability that this value
is actually realized within the large class of flavour models under study.
In the following we shall therefore treat the distributions for the various
observables as probability densities for continuous random variables.
That is, our predictions for the respective observables represent best-guess
estimates according to a probabilistic interpretation of the relative
frequencies.
For each observable we would like to deduce measures
for its central tendency and statistical dispersion from
the respective probability distribution.
Unfortunately, it is infeasible to fit all obtained distributions
with one common template distribution.
Such a procedure would lack a clear statistical justification, and it also
appears impractical as the distributions that we obtain differ substantially
in their shapes.
We therefore choose a different approach.
We consider the median of a distribution as its centre
and we use the $68\ \%$ `confidence' interval around it
as a measure for its spread.
Of course, this range of the confidence interval is reminiscent of
the $1\sigma$ range of a normal distribution.
More precisely, for an observable
$x$ with probability density $f$ we will
summarize its central tendency and variability in the following
form \cite{Cowan:1998ji},
\begin{align}
x = \hat{x}_{\Delta_-}^{\Delta_+}
\,,\quad
\Delta_\pm = x_\pm - \hat{x}
\,.
\end{align}
Here, $x_-$ and $x_+$ denote the $16\,\%$- and $84\,\%$-quantiles with respect to the density function
$f$.
The central value $\hat{x}$ is the median of $f$ and thus
corresponds to its $50\,\%$-quantile.
All three values of $x$ can be calculated from
the quantile function $Q$,
\begin{align}
Q(p) = \textrm{inf} \left\{x \in
\left[x_\textrm{min},x_\textrm{\text{max}}\right] : p \leq F(x)\right\}
\,,\quad
F(x) = \int_{x_\textrm{min}}^x dt \: f(t)
\,,
\end{align}
where $F$ stands for the cumulative distribution function of $x$.
We then have:
\begin{align}
x_- = Q(0.16)\,,\quad
\hat{x} = Q(0.50)\,,\quad
x_+ = Q(0.84)\,.
\end{align}
Intuitively, the intervals from $x_{\textrm{min}}$ to $x_-$, $\hat{x}$, and $x_+$ respectively correspond to
the $x$ ranges into which $16\,\%$, $50\,\%$ or $84\,\%$
of all hits fall.
This is also illustrated in the histogram
for $\sin^2 2\theta_{13}$ in Fig.~1.
Moreover, we have included vertical lines into each plot
to indicate the respective positions of $x_-$, $\hat{x}$, and $x_+$.
In our case the median is a particularly useful measure
of location.
First of all, it is resistant against outliers and hence an appropriate statistic for such skewed distributions as we observe them.
But more importantly, the average absolute deviation from
the median is minimal in comparison to any other reference point.
The median is thus the best guess for the outcome of a
measurement if one is interested in being as close as possible
to the actual result, irrespective of the sign of the error.
On the technical side the definition of the median fits nicely
together with our method of assessing statistical dispersion.
The $68\,\%$ confidence interval as introduced above
is just constructed in such a way that equal numbers of hits
lie in the intervals from $x_-$ to $\hat{x}$ and from $\hat{x}$
to $x_+$, respectively.
In this sense, our confidence interval represents a symmetric
error with respect to the median.
As a test of the robustness of our results, we checked the dependence of our distributions on the precise choice of the experimental error intervals. The results presented here proved insensitive to these variations. For definiteness, we therefore stick to the $3 \sigma$ intervals. We also checked the effect of taking the random ${\cal{O}}(1)$ factors to be distributed uniformly on a linear instead of a logarithmic scale. Again, the results proved to be robust.
\section{Observables and results}
\noindent \textit{Mass hierarchy}
\noindent An important open question which could help unravel the flavour structure of
the neutrino sector is the mass hierarchy. Since the sign of $\Delta m^2_{\text{atm}}$
is not yet known, we cannot differentiate with current experimental data between
a normal hierarchy with one heavy and two light neutrino mass eigenstates and
an inverted hierarchy, which has two heavy and one light neutrino mass eigenstate.
Measuring the Mikheyev-Smirnov-Wolfenstein (MSW) effect of the earth could
resolve this ambiguity.
With the procedure described above, all hits match the structure of the normal
hierarchy and there are no examples with inverted hierarchy. It is however
notable that imposing the structure of the neutrino mass matrix given by
Eq.~\eqref{eq_mnu} alone does not exclude the inverted mass hierarchy. Only
additionally imposing the measured bounds on the mixing angles rejects this
possibility.\medskip
\begin{figure}
\subfigure{
\includegraphics[width=0.48\textwidth]{theta13.eps}}
\subfigure{
\includegraphics[width=0.48\textwidth]{theta23.eps}}
\caption{Neutrino mixing angles $\theta_{13}$ and $\theta_{23}$. The vertical lines denote the position of the median (solid line) and the boundaries of the $68 \%$ confidence region (dashed lines) of the respective distribution.}
\label{fig_mixing}
\end{figure}
\noindent\textit{Mixing angles}
\noindent The mixing in the lepton sector is described by the matrix $U_{\mathrm{PMNS}}$
given in Eq.~\eqref{pmns}. Of the three angles, two are only bounded from one
side by experiment: for the largest mixing angle $\theta_{23}$ there exists
a lower bound, whereas the smallest mixing angle $\theta_{13}$ is so far only
bounded from above. Recent results from T2K~\cite{Abe:2011sj}, Minos~\cite{Adamson:2011qu} and the preliminary result of Double Chooz~\cite{doublechooz}
point to a value of
$\theta_{13}$ just below the current experimental bound. The respective best
fit points, assuming a normal hierarchy, are
$\sin^2 2\theta_{13} = 0.11$ (T2K),
$2 \sin^2 \theta_{23} \sin^2 2 \theta_{13} = 0.041$ (MINOS) and $\sin^2 2 \theta_{13} = 0.085$ (Double Chooz). The 90$\%$ and 68$\%$ confidence regions respectively
read
\begin{align}
&0.03 < \sin^2 2\theta_{13} < 0.28 &&\text{T2K, 90 $\%$ CL}, \, \delta_{CP} = 0, \nonumber \\
&2 \sin^2 \theta_{23} \sin^2 2 \theta_{13} < 0.12
\qquad &&\text{MINOS, 90 $\%$ CL}, \, \delta_{CP} = 0, \\
& 0.01 < \sin^2 2\theta_{13} < 0.16 &&\text{Double Chooz, 68 $\%$ CL}. \nonumber
\end{align}
With the procedure described above, we find sharp predictions for the smallest
and the largest mixing angle within the current experimental bounds,
\begin{equation}
\sin^2 2\theta_{13} = 0.07^{+0.11}_{-0.05} \ ,
\qquad \sin^2 2\theta_{23} = 0.97^{+0.03}_{-0.05} \ ;
\end{equation}
the corresponding distributions are shown in Fig.~\ref{fig_mixing}.
These results are quite remarkable: the atmospheric mixing angle points
to maximal mixing, while the rather large value for $\theta_{13}$ is consistent
with the recent T2K, Minos and Double Chooz results.
In our Monte-Carlo study we observe that the dominant contribution to the strong mixing in the
lepton sector is primarily due to the neutrino mass matrix $m_{\nu}$.
The numerical results are not much affected by including the charged lepton
mixing matrix $V_L$. The PMNS matrix is thus approximately given by the matrix
$U$ which diagonalizes the light neutrino mass matrix $m_{\nu}$. \medskip
\begin{figure}[t]
\subfigure{
\includegraphics[width=0.48\textwidth]{m1.eps}}
\subfigure{
\includegraphics[width=0.48\textwidth]{mtritium.eps}}
\caption{Lightest neutrino mass $m_1$ and effective neutrino mass in tritium decay $m_\beta$. Vertical lines and shadings as in Fig.~\ref{fig_mixing}.}
\label{fig_m1mtritium}
\end{figure}
\newpage
\noindent \textit{Absolute mass scale}
\noindent The absolute neutrino mass scale is a crucial ingredient for the study of neutrinoless double-beta decay and leptogenesis. Although inaccessible in neutrino oscillation experiments, different experimental setups have succeeded in constraining this mass scale. Cosmological observations of the fluctuations in the cosmic microwave background, of the density fluctuations in the galaxy distribution and of the Lyman-$\alpha$ forest yield a constraint for the sum of the light neutrino masses, weighted by the number of spin degrees of freedom per Majorana neutrino, $g_{\nu}=2$, \cite{Nakamura:2010zzi}
\begin{equation}
m_{\text{tot}} = \sum_{\nu} \frac{g_{\nu}}{2} m_{\nu} \lesssim 0.5 \, \text{eV} \,.
\end{equation}
\begin{figure}[t]
\subfigure{
\includegraphics[width=0.48\textwidth]{m0nubb.eps}}
\subfigure{
\includegraphics[width=0.48\textwidth]{alpha21.eps}}
\caption{Effective mass in neutrinoless double-beta decay $m_{0\nu\beta\beta}$ and Majorana phase $\alpha_{21}$. Vertical lines and shadings as in Fig.~\ref{fig_mixing}.}
\label{fig_masses}
\end{figure}
The Planck satellite is expected to be sensitive to values of $m_{\text{tot}}$ as low as roughly $0.1$~eV~\cite{:2006uk}. A further constraint arises from measuring the $\beta$-spectrum in tritium decay experiments. The current bound~\cite{Nakamura:2010zzi} is
\begin{equation}
m_{\beta}^2 = \sum_{i} |(U_{PMNS})_{ei}|^2 m_{i}^2 < 4~\text{eV}^2 \,.
\end{equation}
By comparison, the KATRIN experiment, which will start taking data soon, aims at reaching a sensitivity of $0.04 \,\text{eV}^2$~\cite{Beck:2010zzb}. Finally, the neutrino mass scale can also be probed by neutrinoless double-beta decay. The relevant effective mass is
\begin{equation}
m_{0\nu\beta\beta} = |\sum_i (U_{PMNS})^2_{ei} m_i | \,.
\label{eq_m0nubb}
\end{equation}
Here, Ref.~\cite{KlapdorKleingrothaus:2001ke} claims a value of $0.11 - 0.56$~eV. Dedicated experiments, such as GERDA~\cite{Meierhofer:2011zz} with a design sensitivity of $0.09 - 0.20$~eV, are on the way. Note that $m_{0\nu\beta\beta}$ does
not only depend on the absolute neutrino mass scale and the mixing angles, but also on the phases $(\alpha_{31} - 2 \delta)$ and $\alpha_{21}$ in the PMNS matrix.
We find sharp predictions for the neutrino mass parameters discussed above. The lightest neutrino, $\nu_1$, is found to be quite light, cf.\ Fig.~\ref{fig_m1mtritium},
\begin{equation}
m_1 = 2.2^{+1.7}_{-1.4} \times 10^{-3} \, \text{eV}\,,
\end{equation}
hence favouring a relatively low neutrino mass scale beyond the reach of current and upcoming experiments. More precisely, we find for the neutrino mass parameters discussed above:
\begin{equation}
m_{\text{tot}} = 6.0^{+0.3}_{-0.3} \times 10^{-2} \, \text{eV} , \quad m_{\beta} = 8.6^{+3.3}_{-2.2} \times 10^{-3} \, \text{eV} , \quad m_{0\nu\beta\beta} = 1.5^{+0.9}_{-0.8} \times 10^{-3} \, \text{eV} .
\end{equation}
\noindent \textit{CP-violating phases}
\noindent The small value of the mass parameter measured in neutrinoless double-beta decay, $m_{0\nu\beta\beta}$, is due to the relative minus sign between the $m_1$ and $m_2$ terms in Eq.~\eqref{eq_m0nubb}, caused by a strong peak of the value for the Majorana phase $\alpha_{21}$ at $\pi$,
\begin{equation}
\frac{\alpha_{21}}{\pi} = 1.0^{+0.2}_{-0.2} \,.
\end{equation}
This is depicted in Fig.~\ref{fig_masses}. An analytic analysis of how this phenomena arises from the structure of the neutrino mass matrix, cf.\ Eq.~\eqref{eq_mnu}, is presented in Appendix~\ref{app_alpha1}. For the other Majorana phase $\alpha_{31}$ and the Dirac phase $\delta$ we find no such distinct behaviour but approximately flat distributions. \medskip
\begin{figure}[t]
\subfigure{
\includegraphics[width=0.48\textwidth]{m1tilde.eps}}
\subfigure{
\includegraphics[width=0.48\textwidth]{epsilon1.eps}}
\caption{Effective neutrino mass of the first generation $\widetilde{m}_1$ and $CP$ violation parameter $\varepsilon_1$. Vertical lines and shadings as in Fig.~\ref{fig_mixing}.}
\label{fig_CP}
\end{figure}
\newpage
\noindent \textit{Leptogenesis parameters}
\noindent Finally, leptogenesis~\cite{Fukugita:1986hr} links the low energy neutrino physics to the high energy physics of the early universe. The parameters that capture this connection are the effective neutrino mass of the first generation $\widetilde{m}_1$ and the $CP$ violation parameter $\varepsilon_1$~\cite{leptogenesis},
\begin{equation}
\widetilde{m}_1 = \frac{(m^{\dagger}_D m_D)_{11}}{M_1}\,, \qquad
\varepsilon_1 = - \sum_{j=2,3}\frac{\text{Im} \left[ (h^{(\nu) \, \dagger} h^{(\nu)})_{1j} \right]^2 }{8 \pi (h^{(\nu) \, \dagger} h^{(\nu)})_{11}} F\left(\frac{M^2_j}{M^2_{1}} \right)\,,
\end{equation}
with $F(x) = \sqrt{x} \left(\text{ln} \frac{1+x}{x} + \frac{2}{x-1} \right)$ and $M_{j}$ denoting the masses of the heavy neutrinos. Here, $\widetilde{m}_1$ determines the coupling strength of the lightest of the heavy neutrinos to the thermal bath and thus controls the significance of wash-out effects. It is bounded from below by the lightest neutrino mass $m_1$. The absolute value of the $CP$ violation parameter $\varepsilon_1$ is bounded from above by~\cite{epsilon}
\begin{equation}
\varepsilon_{\text{max}} = \frac{3}{8 \pi}\frac{|\Delta m_{\text{atm}}^2|^{1/2} \, M_1 }{ v^2_{EW} \sin^2 \beta} \simeq 2.1 \times 10^{-6} \, \left( \frac{1}{\sin^2 \beta} \right) \left(\frac{M_1}{10^{10} \, \text{GeV}} \right).
\end{equation}
With the procedure described above, we find
\begin{equation}
\widetilde{m}_1 = 4.0^{+3.1}_{-2.0} \times 10^{-2} \, \text{eV} \,, \qquad \frac{\varepsilon_1}{\varepsilon_{\text{max}}} = 0.25^{+0.28}_{-0.18}\,,
\end{equation}
and hence a clear preference for the strong wash-out regime~\cite{leptogenesis}. Notice that there typically is a hierarchy between $\widetilde{m}_1$ and $m_1$ of about one order of magnitude. The relative frequency of the $CP$ violation parameter $\varepsilon_1$ peaks close to the upper bound $\varepsilon_{\text{max}}$, with the majority of the hits lying within one order of magnitude or less below $\varepsilon_{\text{max}}$, cf.\ Fig.~\ref{fig_CP}. This justifies the use of $\varepsilon_{\text{max}}$ when estimating the produced lepton asymmetry in leptogenesis. Here, in the discussion of $\varepsilon_1$, we assumed hierarchical heavy neutrinos, $M_{2,3} \gg M_1$. \medskip
\noindent \textit{Theoretical versus experimental input}\\
\noindent The results of this section are obtained by combining two conceptually different inputs, on the one hand the hierarchy structure of the neutrino mass matrix $m_{\nu}$ given by Eq.~\eqref{eq_mnu1} and on the other hand the experimentally measured constraints listed in Eq.~\eqref{eq_exp}. In general, the distributions presented above really arise from the interplay between both of these ingredients. For example, the hierarchy structure alone does not favour a large solar mixing angle $\theta_{12}$ and the ratio $\Delta m^2_{\text{sol}} / \Delta m^2_{\text{atm}}$ tends to be too large (cf.~\cite{masina,winter}). This discrepancy is eased by generating the random coefficients in Eq.~\eqref{eq_mnu1} via the seesaw mechanism. Imposing the experimental constraints finally singles out the subset of parameter sets used for the distributions presented above. As another example, consider the smallest mixing angle $\theta_{13}$ and the lightest neutrino mass eigenstate $m_1$. In these cases, the hierarchy structure of the neutrino mass matrix automatically implies small values, similar to those shown in the distributions above. However, the exact distributions including the precise position of the peaks only arise after implementing the experimental constraints. A notable exception to this scheme is the Majorana phase $\alpha_{21}$. Here the peak at $\alpha_{21} = \pi$ is a result of the hierarchy structure of the neutrino matrix $m_{\nu}$ alone, as demonstrated in Appendix~\ref{app_alpha1}.
\section{Discussion and outlook}
In summary, we find that starting from a flavour symmetry which accounts for
the measured quark and lepton mass hierarchies and large neutrino mixing, the present knowledge of
neutrino parameters strongly constrains the yet unknown observables, in particular the smallest mixing angle $\theta_{13}$,
the smallest neutrino mass $m_1$, and the Majorana phase $\alpha_{21}$.
This statement is based on a Monte-Carlo study:
Treating unspecified $\mathcal{O}(1)$ parameters of the considered Froggatt-Nielsen
model as random variables, the observables of interest are sharply peaked
around certain central values.
We expect that these results hold beyond Froggatt-Nielsen flavour models.
An obvious example are extradimensional models which lead to the same type
of light neutrino mass matrix (cf.~\cite{Asaka:2003iy}). On the other hand, quark-lepton
mass hierarchies and the presently known neutrino observables cannot determine
the remaining observables in a model-independent way. This is illustrated
by the fact that our present knowledge about quark and lepton masses and
mixings is still consistent with $\theta_{13}\simeq0$ as well as with an inverted neutrino mass hierarchy (cf.~\cite{theta13}). As a
consequence, further measurements of neutrino parameters will be able to falsify
certain patterns of flavour mixing and thereby provide valuable guidance for
the theoretical origin of quark and lepton mass matrices.
\bigskip\bigskip \noindent \textbf{Acknowledgements}
The authors thank G.~Altarelli, F.~Br\"ummer, G.~Ross, D.~Wark, W.~Winter and T.~Yanagida for helpful discussions and comments. This
work has been supported by the German Science Foundation (DFG) within the Collaborative
Research Center 676 ``Particles, Strings and the Early Universe''.
|
\section{Introduction}
Cobalt Ferrite is under examination since more than sixty years, but still there is quite a lot of open problems. In 1988 Guillot observed a jump in the field dependence of the magnetization in pure Cobalt Ferrite with the composition Co$_{1.04}$Fe$_{1.96}$O$_{4}$ and also in Cd substituted Cobalt Ferrite \cite{guillot_high_1988}. This jump was explained as a spin-flip. Because of the rather low critical field this explanation was not conclusive. Therefore within the present work this transition was studied in more detail.
\section{Sample Preparation}
The single crystal with the composition Co$_{0.8}$Fe$_{2.2}$O$_{4}$ was grown by flux method. The starting materials are 18~g Na$_{2}$B$_{4}$O$_{7}$~$\cdot$~10H$_{2}$O (Borax), 2.3~g CoO (99.99\%), and 6.7~g Fe$_{2}$O$_{3}$ (99.99\%). After sufficiently mixing, the materials were put in a tightly closed Pt crucible and heated from room temperature to 1370~$^{\circ}$C at a rate of 100~$^{\circ}$C/h, and then held for a period of 6~h; slowly cooled from 1370 to 990~$^{\circ}$C at 2~$^{\circ}$C/h, followed by a furnace cooling by switching off the power supply \cite{wang_flux_2006}. The composition of the single crystal was checked by XRD and SEM investigation.
\section{Experimental Procedures}
Magnetization was measured from 5 to 400~K in a vibrating sample magnetometer with a superconducting 9~T coil. The single crystal was oriented by XRD in a Laue setup and then transfered to the VSM sample holder. The error from transferring the single crystal from one sample holder to the other was usually smaller than 1.5$^{\circ}$.\\
The magnetostriction was measured with a miniature capacitive dilatometer described in \cite{rotter_miniature_1998} using a cryostat with a variable temperature insert (VTI) and a superconducting 9~T coil.\\
\section{Results and Discussion}
\subsection{Magnetization}
The degree of inversion of the cation distribution of the inverse spinel (A$^{2+}$)[B$_{2}^{3+}$]O$_{4}$ was calculated to $i = 0.625$ with (Co$_{0.8-i}^{2+}$ Fe$_{0.2+i}^{3+}$)[Co$_{i}^{2+}$Fe$_{2-i}^{3+}$]O$_{4}$ and the magnetic moment of $\mu = \mu_{B-Sites} - \mu_{A-Sites} = 4.1 \mu_{B}$ at $T = 5$~K. The value of saturation magnetization at 5 and 10~K is practically the same due to the the very high ordering temperature.\\
\begin{figure}
\includegraphics{G10Kall_lin_Fit.eps
\caption{\label{fig:MH10K}Normalised Magnetization of Co$_{0.8}$Fe$_{2.2}$O$_{4}$ single crystal at $T = 10$~K}
\end{figure}
The magnetization measurements revealed that the [100] axis is the easy axis of magnetization of Co-ferrite over the whole temperature range. Below 150~K a jump in the magnetization at fields occurs as it was also found in a similar material (Co$_{1.04}$Fe$_{1.96}$O$_{4}$) in Ref.~\cite{guillot_high_1988}. In figure~\ref{fig:MH10K} the normalized magnetization at $T = 10$~K is plotted versus the external magnetic field. The jump is clearly visible at roughly 7~T in the [111] axis. The critical field differs from that published by Guillot which can be understood regarding the different sample composition. A linear fit between 1 and 6~T was performed showing that the extrapolation to 0~T leads to a value of $\frac{1}{\sqrt[]{3}}$, which indicates that the magnetization vector lies in the [100] axis and only the projection ($cos(\alpha)$) of the [100] magnetization vector into the [111] axis is measured. By increasing the magnetic field the vector starts to rotate towards the [111] axis. However as soon as the measured moment in the [111] axis achieves the value of $\frac{1}{\sqrt[]{2}}$, the measured moment becomes equal to the [110] value measured at 0~T. At this point the system needs more energy (field) to overcome the magnetic anisotropy energy corresponding to the [110] axis. As a consequence the magnetization vector rotates in the [110] axis to a local energy minimum in order to reduce energy. In figure~\ref{fig:Geo} the offset of the linear fits assuming such a geometrical model below the transition are shown. The error due to misalignment is found to be 1.5$^{\circ}$ for the [111] axis and 0.3$^{\circ}$ for the [110] axis. Obviously our geometric model works well from 5~K up to 250~K. Above 250~K the anisotropy along the [110] and [111] axis decreases and the magnetization vector can rotate smoothly (without a discontinuity) out of the easy axis.\\
\begin{figure}
\includegraphics{Geometric_Factor.eps
\caption{\label{fig:Geo}Offset of the linear fits of normalized magnetization of Co$_{0.8}$Fe$_{2.2}$O$_{4}$ as a function of temperature. The straight lines indicate the geometrical fit}
\end{figure}
By this geometrical considerations one can see that the origin of the magnetization jump is not a spin-flip, but a rotation in the magnetization due to the higher order anisotropy constants which cause a more complicated shape of the anisotropy energy $E_{A}$.\\
Such anisotropy driven magnetization jumps are generally called FOMP (First Order Magnetization Process) and were also found in other complex systems and alloys (as e.g. PrCo5 - see \cite{asti_abstract:_1979}, RE$_{2}$Fe$_{17}$C \cite{grossinger_anisotropy_1991}, REFe$_{11}$Ti \cite{kou_magnetic_1993} or Nd$_{2}$Fe$_{14}$B \cite{kou_spin-reorientation_1997}).
\subsection{Magnetic Anisotropy}
The magnetic anisotropy was determined with the integral method $E_{A} = \int_{0}^{M_{S}} H dM$ along the measured crystallographic axis of the single crystal. Applying the formula $E_{A} = K_{0} + K_{1} (\alpha_{1}^{2}\alpha_{2}^{2} + \alpha_{2}^{2}\alpha_{3}^{2} + \alpha_{3}^{2}\alpha_{1}^{2}) + K_{2} (\alpha_{1}^{2}\alpha_{2}^{2}\alpha_{3}^{2})$, where the $\alpha_{i}$ are the direction cosines in polar coordinates, the anisotropy constants $K_{0}$, $K_{1}$ and $K_{2}$ were determined. For using the integral method it is pre-requisite to saturate the sample fully. It is reported that the saturation along the intermediate and hard axis is reached at around 18~T at $T = 4$~K \cite{guillot_high_1988}. Our measurements were only performed up to 9~T, therefore extrapolating the M(H) curves to saturation causes a rather large error ($\pm$ 20\% for $K_{1}$ and $\pm$ 30\% $K_{2}$), but they are still in the range of reported values of $K_{1}$ in literature.\\
In figure~\ref{fig:EaConst} the anisotropy constants $K_{0}$, $K_{1}$ and $K_{2}$ of the single crystal and $K_{1}$ of a polycrystalline sample are plotted against the temperature. For obtaining $K_{1}$ for the polycrystalline sample we used the law of approach to saturation. Measuring the polycrystalline sample the maximum magnetic field was only 9~T and the sample not fully saturated. Accordingly all reported values in literature using the law of approach to saturation are not giving the correct value of $K_{1}$, because they estimate only $K_{1}$ and not higher order anisotropy constants.\\
It is interesting to note that no values for $K_{2}$ are reported in literature. Above $T = 150$~K the value of $K_{2}$ is $\sim$~6 times higher than $K_{1}$ and below the FOMP the factor is even much higher, yielding in an increased anisotropy along the [111] and [110] axes.\\
\begin{figure}
\includegraphics[width=0.47\textwidth]{anisotropy_constants.eps
\caption{\label{fig:EaConst}Anisotropy constants of Co$_{0.8}$Fe$_{2.2}$O$_{4}$ single crystal and one polycrystalline specimen as a function of temperature}
\end{figure}
\subsection{Magnetostriction}
We are presenting the first magnetostriction measurement at $T = 4.2$~K demonstrating also the effect of such a field induced transition in the magnetoelastic behaviour as shown in figure~\ref{fig:MS4K}.
\begin{figure}
\includegraphics{MS_4K.eps
\caption{\label{fig:MS4K}Magnetostriction measurements of Co$_{0.8}$Fe$_{2.2}$O$_{4}$ single crystal at $T = 4.2$K}
\end{figure}
Due to hysteresis effects (remanence) the magnetostriction measurements at low temperatures are very difficult and need a special measuring procedure as will be published elsewhere. But when increasing the magnetic field above the critical field, a huge jump in the magnetostriction occurs. At higher fields due to rotational magnetization process no hysteresis effects are observed.\\
\section{Conclusions}
We have proved that the jump in the magnetization can be explained as an anisotropy driven transition which is called ``FOMP''. The transition is caused by a rotation of the magnetization vector jumping over an energy barrier. At low temperatures the second anisotropy constant $K_{2}$ is increasing which strengthens the [100] axis as easy axis and underlines the [110] as intermediate and the [111] as hard axis. These assumptions are also supported by geometric considerations explaining the values of the $M(H)$ curves as measured in the different crystallographic directions.
Additionally we present an accurate measurement of $\lambda_{111}$ at a temperature of 4.2~K showing also this critical transition in the magnetoelastic behaviour.\\
\begin{acknowledgments}
The financial support by the FWF under the NFN-project numbers S~10406 and S~10403 is gratefully acknowledged.
\end{acknowledgments}
|
\section{Introduction}
A transition-edge sensor (TES) is a thin superconducting film that can be used as a sensitive thermometer when voltage biased within the normal metal - superconducting transition region and read out with superconducting SQUID sensors \cite{irwin}. TES based devices are used as extremely sensitive bolometers and calorimeters to detect radiation in a wide energy range from gamma-rays to sub-millimeter radiation \cite{enss}, and typically the thermal conductance to the bath is controlled by mounting the TES on a thin insulating SiN membrane. Although the performance of these detectors is already excellent,
the most sensitive TES devices have not yet reached the theoretical limits in energy resolution.
This is mostly due to excess noise that has been shown to be present in
many devices \cite{ullom,NASA,hoevers,IM_LTD12,KK_LTD12}. Several candidates for the noise sources have been proposed, such as thermal fluctuations within the TES \cite{hoevers}, fluctuations in the Cooper-pair density \cite{FSN,secondFSN} or phase-slips \cite{slipsold,slipsnew}, but a definitive answer is still missing.
However, before resorting to more exotic noise sources to explain the data, one should be sure the known noise mechanisms are fully understood. These are the thermal fluctuations of the electrical degrees of freedom, or Johnson noise, and the thermal fluctuations of the energy degrees of freedom, usually called thermal fluctuation noise or phonon noise \cite{enss}. Both of these noise mechanisms are unavoidable in bolometric TES detectors; moreover, their magnitude depends on the details of the device in question. It is therefore of utmost importance to characterize the detector accurately, both electrically and thermally, before one can understand the origin of noise.
A useful tool in characterizing TES detectors is the measurement of their frequency-dependent complex electrical impedance \cite{linde1}. Importantly, within the transition this impedance depends not only on the electrical, but also on the thermal circuit of the device, through the electrothermal feedback effect \cite{irwin,enss}. With the help of this technique, a more accurate picture of the electrical and thermal properties of TES sensors has emerged: First of all, it was realized that the dependence of the detector resistance on current, and not only on temperature, is critical for the detector response, as well \cite{linde1,enss}. This was later shown to influence the Johnson noise directly \cite{irwinnoise}, and some of the excess noise could then be explained as non-equilibrium Johnson noise. In addition, it has become clear that for many detectors, the simplest thermal circuit of one heat capacity connected to heat bath through one thermal conductance is not adequate \cite{NIST,NASASaab,nasaIEEE,cambridge,SRON,KK_LTD12,mikko}. A more complex thermal circuit then adds new components to the thermal fluctuation part of the noise spectrum \cite{hoevers,galeazzi,enectali}.
Here, we present a study of the noise and complex impedance of several different designs of TES devices. Many of the devices are based on the so-called Corbino-geometry TES (CorTES)\cite{FSN, IM_LTD12}, where current spreads out radially from a central contact with radius $r_i$ into the outer contact at $r_o$, instead of flowing linearly [Fig. 1 (a)]. Although it is a bit more complicated to fabricate, it offers advantages in modelling, as the superconducting and normal regions separate due to the non-uniform current density profile. This means that we can determine the phase boundary radius $ r_b $ from the measured resistance $R$, which depends logarithmically on $r_b$ in Corbino geometry:
\begin{equation}
\label{eq:rb}
R \ln\left ( \frac{r_o}{r_i} \right )= R_{N} \ln\left ( \frac{r_b}{r_i} \right ) \Leftrightarrow \frac{r_b}{r_i} = \left ( \frac{r_o}{r_i} \right )^{R/R_{N}},
\end{equation}
where $R_{N}$ is the normal state resistance of the device. Once we know the size of the N and S phases, their expected theoretical heat capacities can be calculated as well. This level of theoretical description is not possible in more common square shaped TES devices.
CorTES has also shown an excess noise component, if compared to the simplest thermal model. This was originally explained by the fluctuations of the N-S boundary, or fluctuation superconductivity noise (FSN) \cite{FSN}. However, here we show that most of the "excess" noise is simply internal thermal fluctuation noise originating from a more complex thermal circuit within the device. By analyzing the data thoroughly, we suggest that most of the "excess" noise is generated by the thermal decoupling of the spatially separated superconducting and normal regions of the devices. Data from typical square shaped devices also supports this picture: Due to non-uniformities of real devices and/or the lateral proximity effect \cite{latp}, phase separation can still take place, although not in such a controllable manner as in CorTES devices.
\begin{figure}
\includegraphics{Fig1.eps}
\caption{(Color online) (a) Diagram illustrating the radial current distribution and separation into normal (N) and superconducting (S) phases in a CorTES. $ r_i $ and $ r_o $ are the radii of the inner and outer superconducting contacts and $ r_b $ is the phase boundary radius. (b) A schematic side view of a CorTES with an absorber. The arrows indicate the path of the bias current. \label{CorTES}}
\end{figure}
\section{Thermal modelling}
The simplest thermal model of a TES device is shown in Fig. \ref{block}(a), consisting of a single heat capacity connected to the heat bath. The complex impedance $Z$ of a device with this one-body thermal circuit always traces a semicircle in the complex plane as a function of frequency \cite{enss}. Thus, it is experimentally straightforward to determine whether a more complex thermal circuit is required by studying the shape of the $Z$ curve. Whenever additional thermal blocks are added to the system, $Z$ develops bulges outwards from the ideal case ( see Fig. \ref{MvsITFN} (a)). The shape and size of the new features in $Z$ depend on the heat capacities $C_i$ of the added bodies and the thermal conductance links $g_i$ between them, in such a way that the features grow in size with both $C_i$ and $1/g_i$.
\begin{figure}
\includegraphics{Fig2.eps}
\caption{(a) Basic calorimeter thermal model. (b) Three-body model used in this work. \label{block}}
\end{figure}
Fig. \ref{MvsITFN} (a) shows a typical measured impedance curve of a CorTES device, together with a one-, two- and three-body theoretical model fits. Although simpler one and two-body models seem to fit some device designs by other groups \cite{iyomoto, NASASaab, SRON, Eckart}, for our devices a three-body model is the simplest model that fits the data well, as is clear from Fig. \ref{MvsITFN} (a). Recently, simple three-body models have also been used by other groups for more accurate modelling of their devices \cite{NIST,nasaIEEE}.
\begin{figure}
\includegraphics{Fig3.eps}
\caption{(Color online) (a) A comparison of one- two- and three-body fits to a typical impedance data. (b) An example of a current noise spectrum from a CorTES device, showing the minor difference between fitting the high frequency noise with ITFN or with excess Johnson (M) noise. \label{MvsITFN}}
\end{figure}
In this work, we have therefore used the thermal model shown in Fig. \ref{block}(b), which is the simplest three-body model that is physically justifiable for our devices. In addition to the TES film with heat capacity $C_{TES}$ where the bias power $P_{bias}$ is dissipated, there is a hanging thermal body with heat capacity $C_1$, and an intermediate heat capacity $C_2$ between the TES and the heat bath.
$ C_2 $ could be associated with the supporting SiN membrane, and $C_1$ could represent the absorber on top of a TES film [Fig. 1 (b)], but in general it could be some other part of the device also. Most devices studied here do not actually have absorbers. We call this the IH (intermediate + hanging) model, and use analytical expressions for the impedance and noise of a TES with this thermal circuit presented in \cite{mikko}. Their derivation and full theoretical discussion will be published elsewhere \cite{TPB}. We simply note that with each new thermal block, a new thermal noise source is also introduced to the system, arising from fluctuations in energy between the bodies through the thermal link. We call the noise due to any additional block internal thermal fluctuation noise (ITFN) to distinguish it from the thermal noise between the system and the heat bath (phonon noise).
\subsection{Intermediate body}
The intermediate body is characterized by its temperature $T_2$ and heat capacity $C_2$. We emphasize that we do not make the simplifying assumptions that $ T_2 = T_{TES} $ or that $ T_2 = T_{bath}$, when the TES is biased. Associated with the intermediate block, we thus have four different values of dynamic thermal conductance, two for each physical link connected to $C_2$, evaluated at each temperature end of the link \cite{TPB}.
The parameters $ G_{dyn}=dP/dT_{TES} $ and $ T_{TES} $ can be calculated from a careful analysis of the measured I-V characteristics, as described in Ref \cite{linde2}.
In our fitting procedure we use $T_2$, $C_2$ and $g_{TES,2}(T_{TES})$ as the free parameters. The other unknown conductances $g_{TES,2}(T_2)$, $g_{2,b}(T_{bath})$ and $g_{2,b}(T_2)$ are then calculated using the free parameters and the known values of $T_{TES}$ and $G_{dyn}$,
as \cite{TPB}
\begin{equation}
G_{dyn}=\frac{g_{TES,2}(T_{TES})g_{2,b}(T_2)}{g_{TES,2}(T_{2})+g_{2,b}(T_2)}.
\end{equation}
To simplify the model and to minimize fitting parameters, it is also assumed that the links on both sides of the intermediate block have the same thermal exponent $n$, i.e.
$g_{TES,2}(T_j) = AT_{j}^{n-1}$ and $g_{2,b}(T_j) = BT_{j}^{n-1}$. This is physically reasonable if all the conductances are dominated by the phononic transport properties of the SiN membrane.
Later, we will show that the intermediate body fitting parameters are nearly constant throughout the transition, and only change when $ T_{bath} $ is changed. Also, $C_2$ appears to depend on both the area and thickness of the membrane. This behavior is thus consistent with the assumption that the intermediate body represents the SiN membrane phonons. Furthermore, $ T_2 $ is usually always slightly below $ T_{TES} $, and not close to the bath temperature.
\subsection{Hanging body}
In contrast to the intermediate body, the second added heat capacity $C_1$ in our model is hanging [Fig. \ref{block} (b)], which means that no steady state power flows through it, and the average values of $T_{TES}$ and $T_1$ are equal. This simplifies the description in comparison to the intermediate body, as only one thermal conductance $g_{TES,1}$ is required. The Joule power of the TES bias current is all assumed to be dissipated inside $C_{TES}$.
In real X-ray and $\gamma$-ray devices with thick absorbers on top of the TES film \cite{iyomoto,NIST,mikko}, $C_1$ could well be the absorber, in which case $g_{TES,1}$ describes the thermal conduction within the TES and the absorber. In devices without an absorber (most samples in this work), the idea is to model electronic degrees of freedom, thus $g_{TES,1}$ models the thermal conduction within the TES film.
The hanging block is naturally still a simplification, in reality there can be also a thermal coupling from $C_{1}$ directly to the bath or to $C_{2}$. However, the effect of the missing link, which would have a magnitude of the order of $g_{TES,2}$, is believed to be minor for the devices studied here, where $g_{TES,1}$ is much larger than $g_{TES,2}$, as will be seen later.
Any effects of temperature gradients within the TES film are also not captured in this simple model. Numerical estimates of the gradients within a CorTES device \cite{AL} have shown that the isothermal simplification is a reasonable assumption.
For the CorTES devices with phase separated N and S regions, one may at this point wonder, if the blocks $C_{TES}$ and $C_{1}$ could actually directly represent the normal and superconducting regions of the TES film, respectively. The great benefit of the CorTES geometry is that we can a priori calculate the values of $C_{TES}$ and $C_{1}$ as a function of bias, which is not possible for the more standard square shaped devices. As we later show, the fitted values of the heat capacities and their behavior as a function of bias point do indeed follow a trend predicted by this interpretation. $g_{TES,1}$ is thus still the electronic thermal conductance, but now it can be affected by the location of the NS boundary, for example. Understanding of how $g_{TES,1}$ should behave as a function of the bias (phase boundary location) is quite sketchy at the moment, and it needs to be studied further theoretically.
The hanging body $C_{1}$ produces internal thermal fluctuation noise (ITFN) that appears in current noise spectra at the high-frequency side of the effective thermal time constant, just like Johnson noise \cite{mikko,TPB}, see Fig. \ref{MvsITFN} (b). In addition, when $ g_{TES,1} $ is large enough, the roll-off for the ITFN noise occurs at very high frequencies, even above the electrical cut-off frequency of the read-out circuit. Thus, the ITFN noise produced by the hanging body can look very similar to the Johnson noise. Quite often in the past, excess noise in TES devices has been quantified by the parameter $M$ \cite{ullom,IM_LTD12,SRON}, which is a measure of excess Johnson noise. Thus, one has to be very careful about drawing conclusions based on just the noise data: If the thermal circuit is not adequately characterized, one can misinterpret ITFN noise as excess Johnson noise. In our devices the time constant for ITFN roll-off is indeed near the electric cut-off of our read-out circuit, so that fitting the spectra either with the three-body ITFN model, or two-body + excess Johnson noise ($M$ parameter) give almost identical results, as illustrated in Fig. \ref{MvsITFN} (b). The only difference seen is that with the ITFN model, the roll-off has a steeper slope, as it is a higher order roll-off produced by the combination of the thermal and electrical circuits. However, the point is that in many cases the impedance can only be fitted with the three block model [see Fig. \ref{MvsITFN} (a)], which supports the choice of ITFN noise over other noise sources.
\section{ Experimental setup and methods}
\label{setup}
The experiments were performed in a compact homemade dilution refrigerator with a base temperature $T= 40$ mK. The TES devices were voltage biased by a shunt resistor (8.9 m$\Omega$) at the sample stage, and a NIST two-stage SQUID \cite{cherv} mounted on the 1 K stage (equivalent input current noise density $\sim 4$ pA/$\sqrt{\mathrm{Hz}}$ and geometric input inductance of 300 nH) served as the current amplifier for the readout. The SQUID was operated in the flux locked loop, with dedicated room temperature electronics designed by SRON. More details on the setup can be found in \cite{Kimmothesis}.
All studied TES devices were first characterized by measuring their resistance vs. temperature transitions with a four-probe lock-in measurement, to find the normal state resistance $ R_N $ and critical temperature $ T_c$. Next, each detector was connected to the SQUID readout, and a series of current-voltage (I-V) curves was measured at several different bath temperatures. The thermal exponent $n$ was extracted from the $T_{bath}$ dependence of the I-V data, as explained in Ref. \cite{linde2}. Then, using the known $n$, the TES temperature $T_{TES}$ and dynamic thermal conductance $G_{dyn}$ at any point in the transition were determined from the I-V curves \cite{Kimmothesis}. We also calculated the transition steepness parameter $ \alpha_{TOT,IV} = (T/R)dR/dT $, which is a dimensionless measure of the sensitivity of the TES \cite{irwin}.
Finally, a set of noise and complex impedance measurements were performed at a desired $ T_{bath} $ and at various bias points within the transition. The TES impedance was extracted from the measured circuit impedance by dividing out the effect of the read-out circuit transfer function, as described in ref. \cite{transfer}. From the TES impedance curves, the low and high frequency limits $ Z_0 $ and $ Z_{\infty} $ can be estimated, which can then be used to calculate the important transition parameters $ \alpha = (T/R)\partial R/\partial T|_{I_0} $ and $ \beta = (I/R)\partial R/\partial I|_{T_0}$ as
\begin{equation}
\label{eq:alpha}
\alpha = \frac{n}{\phi}\frac{Z_0 - Z_{\infty}}{Z_0 + R_0}, \; \beta = \frac{Z_{\infty}}{R_0} - 1
\end{equation}
where $ \phi = 1 - (T_{bath}/T_{TES})^n $ and $ R_0 $ is the TES resistance. Additionally, $ \alpha _{TOT,Z} = (n/\phi)(Z_0 - R_0)/(Z_0 + R_0)=[2\alpha+(n/\phi)\beta]/(2+\beta)$ from the impedance data should give \cite{AIP} the same value as $ \alpha _{TOT,IV} $, so that we can check for consistency between the impedance and I-V measurements. Notice how $\alpha _{TOT}$ depends both on $\alpha$ and $\beta$.
We have performed impedance measurements between 4 Hz and 100 kHz both by the white noise method \cite{linde1}, where a white noise excitation is used and the broadband response is measured, and by the lock-in method where a sinusoidal excitation is used at certain frequency points \cite{transfer}.
Both methods can in general be used, however, we have seen that the sine-wave method is more reliable, as the total heating power generated by the excitation is typically less in the sine-wave method. Data with heating problems could be identified by the consistency check mentioned above: With excess heating $ \alpha _{TOT,Z} $ did not agree with $ \alpha _{TOT,IV} $, indicating a heating induced shift in the bias point.
The measured impedance and noise data are finally fitted {\em simultaneously} to the IH model equations. We emphasize that the fitting is done by eye, and free fit parameters are varied manually, as high-dimensional non-linear least-squares fitting would be demanding to implement. In some cases we found that the three-body equations still did not explain all the observed noise, even if the impedance fit was good. In those cases we have quantified the remaining truly excess noise as excess Johnson noise using the M-parameter. Note also that although here we only show the complex plane plots of $Z$, we also make sure that the real and imaginary parts fit separately as a function of frequency. Included in the fits, but not shown in the plots because of their small values, are the Johnson noise of the known shunt resistor, and the equivalent white input noise of the SQUID. We have also used the equation for the (lowest order) non-equilibrium TES Johnson noise \cite{irwinnoise} $V_n^2=4k_BTR_0(1+2\beta)$ in the analysis for all the devices. Even though $Z$ is measured only up to 100 kHz, the theory curves are always calculated up to 2 MHz, in order to see the high-frequency differences between the fits, as demonstrated in section \ref{sec:slice1}.
\section{Experimental results}
In this section we present the measured data and fit results from several different detectors. Table \ref{TEStable} lists some of the key parameters for each device studied. Most of them have radially spreading current distribution as in the fully circulary symmetric CorTES device shown in Fig. \ref{CorTES}. However, the device labelled STES is divided into four equal size parallel slices (see inset, Fig. \ref{STESparam}), and the detectors labelled Slice 1 and Slice 2 are individual slices (Fig. \ref{111fits}). Also, data from one traditional square-shaped device is also presented. All measurements were done at a regulated bath temperature of 60 mK unless stated otherwise. The detectors were not shielded against external magnetic fields, nor was earth's field compensated for.
\begin{table}[h]
\caption{\label{TEStable} Device parameters for the TESs measured in this work.}
\begin{tabular}{r c c c c}
TES & Ti/Au [nm] & $T_c$ [mK] & $R_N$ [m$\mathrm{\Omega}$] & SiN [$\mathrm{\mu m}$] \\ \hline
CorTES & 40/55 & 98 & 200 & 0.30$\times$800$\times$750 \\
STES & 40/55 & 99 & 220 & 0.30$\times$800$\times$750 \\
Slice 1 & 71/105 & 126 & 166 & 0.30$\times$830$\times$730 \\
Slice 2 & 58/83 & 162 & 220 & 0.75$\times$830$\times$730 \\
Square & 48/70 & 156 & 425 & 0.75$\times$460$\times$410 \\
\end{tabular}
\end{table}
\subsection{CorTES}
\label{sec:cortesdata}
The data we have measured on a full CorTES turned out to be quite difficult to fit accurately even with the three-body IH model, as seen from the examples shown in Fig. \ref{OTESdata}. The general trends are, nevertheless quite well reproduced, including the very strong deviations of $Z$ from the simplest circular shape and all the trends in the noise spectra. Typical values for the most important fit parameters were $C_{1} \sim 0.2$ pJ/K, $C_{2} \sim 0.35$ pJ/K, $C_{TES} \sim 0.05$ pJ/K and $g_{TES,1} \sim 20-30$ nW/K throughout the transition for the bare devices without an absorber. Fig. \ref{OTESdata} also shows the data after the deposition of a 2 $\mu$m thick Bi absorber on top, as shown schematically in Fig. \ref{CorTES}(b). In that case, all other parameters stayed about the same, except for $C_{1}$, which, quite reasonably, doubled to $\sim 0.4$ pJ/K. We should note that the obtained $C$ values are reasonable if compared to estimates from the detector size, except for $C_{2}$ which is surprisingly large. This could be because of the 120 nm thick AlOx insulator layer separating the Nb bias lines [\ref{CorTES}(b)] in the CorTES device, creating unwanted thermal links and heat capacity inside, as shown in Ref. \cite{KK_LTD14} for a different device.
\begin{figure}
\includegraphics{Fig4.eps}
\caption{(Color online) Comparison of (a) the impedance (measured with the noise method) and (b) the noise of a full CorTES with (black crosses) and without (red open circles) an absorber, at two different bias points. The lines are best fits to the data using the IH model. No $M$-parameter was used. \label{OTESdata}}
\end{figure}
Very accurate determination of how the addition of the absorber affected the detector characteristics is unfortunately complicated, because the absorber processing changed also the transition properties, by lowering the $ T_c $ from 98 mK to 85 mK, and by broadening the transition. For the bias values shown in Fig. \ref{OTESdata}, $\alpha _{TOT,IV}$ decreased from $\sim 300$ to $\sim 50$ at $R/R_{N}=0.6$, whereas at $R/R_{N}=0.1$ it stayed constant at $\sim 150$. These changes influence the detector responsivity, which means that noise and $Z$ are affected not only through the increase of $C_{1}$ but also directly. The large difference of noise at $R/R_{N}=0.6$ can be attributed mostly to the change in $\alpha$, whereas at $R/R_{N} = 0.1$ the lowering of noise is consistent with the lower $T_c$ of the device after the addition of the absorber.
\subsection{STES}
We have also measured a set of data from another absorberless CorTES device, shown in Fig. \ref{STESfits}. It was otherwise nominally identical to the full device discussed above, but was cut into four parallel slices (dubbed STES), with an SEM image shown in the inset of Fig. \ref{STESparam}(b). In this case, the impedance fits to the IH model look nearly perfect, and the noise data agrees also quite well without the need of an empirical $M$ parameter, except at the very highest frequencies near the cut-off (reason unknown at the moment). Achieving better fits than in the full CorTES case was easier, most likely because the STES device has a lower responsivity (lower maximum $\alpha \sim 200$ ). In Fig. \ref{STESparam} we plot the obtained fit parameters of interest as a function of the bias point $R/R_{N}$. The heat capacities, Fig. \ref{STESparam} (a), do not vary strongly, and their values are generally consistent with the CorTES results. $C_{2} \sim 0.55$ pJ/K is again high, even higher than in the CorTES device. To compare the data with the suggestion that $C_{TES}$ and $C_{1}$ are associated with the normal and superconducting regions, respectively, we also show theoretical curves corresponding to the calculated heat capacities of the normal and superconducting phase regions, using the location of the phase boundary calculated from Eq. \ref{eq:rb}, the known film thicknesses and Sommerfeld constants $\gamma_{Ti}= 330$ J/K$^2$m$^3$ and $\gamma_{Au}= 65$ J/K$^2$m$^3$ from literature \cite{constants}. In addition, the heat capacity of the S region was calculated using two different assumptions: (a) the jump at $T_c$ is given by simple BCS theory as $1.43 C_{N}$ (dash-dotted line), and (b) the jump is slightly suppressed due to proximity effect according to Ref. \cite{kozorezov}. We see that $C_{1}$ is clearly consistent with it being associated with the S region, however $C_{TES}$ is a bit elevated compared to the simplest theory. Moreover, lower in the transition at $R/R_{N} < 0.3$, there seems to be a trend that $ C_{1}$ increases and $ C_{TES} $ decreases. This trend is very clear in the data, as forcing the values of $C_{1}$ and $C_{TES}$ for $R/R_{N} < 0.3$ to equal the values at $R/R_{N} = 0.3$ produces very poor fits, as shown in Figs. \ref{STESfits} and \ref{STESparam}.
\begin{figure}[h]
\includegraphics{Fig5v3.eps}
\caption{(Color online) Measured (symbols) and fitted (lines, IH model) (a) impedance (measured using the lock-in method up to 100 kHz) and (b) noise in the STES device. Decreasing bias makes the impedance curve smaller and noise level higher. Bias values shown are $R/R_{N}=0.1$, 0.2, 0.3, 0.4, 0.5, 0.8, rest are omitted for clarity. Dashed lines show fits with alternative parameter values shown in Fig. \ref{STESparam} as crosses. \label{STESfits}}
\end{figure}
\begin{figure}[h]
\includegraphics{Fig6v3.eps}
\caption{(Color online) Parameters obtained from the fits to the data in Fig. \ref{STESfits} as a function of the bias point.(a) The heat capacities, and (b) the thermal conductances. $G_{dyn}$ is fixed by the I-V data and is not a fit parameter. Theory curves show the expected dependence of the normal region (dashed line) and the superconducting region with two different models for the jump at $T_c$: full jump (dash-dotted) and suppressed according to \cite{kozorezov} (solid line). Crosses show alternate parameter values used for "bad fits" shown in Fig. \ref{STESfits}. Inset in (b) is an SEM image of the STES. \label{STESparam}}
\end{figure}
\subsection{Slice TES}
The usual CorTES devices have an AlOx insulator layer between the bias leads. Furthermore, from the high values of $C_{2}$ obtained above, and from previous measurements in Ref. \cite{KK_LTD14}, we suspect that the AlOx layer could be responsible for the high value of $C_{2}$. Therefore, we also studied devices where only a quarter "slice" of the full CorTES disk is retained: this way the insulator layer is not required, as the center contact lead can come from the opposite side (see inset Fig. \ref{111fits}). However, we still retain a geometry that promotes a phase separation. Here we report on measurements on two such samples, with different Au/Ti thicknesses, leading to different $T_c$ and $R_{N}$ values (Table \ref{TEStable}). The Au/Ti layer thicknesses were increased compared to the full CorTES devices, in order to keep $ R_N $ approximately the same. Also, the thickness of the SiN membrane (300 nm) was chosen to be the same as in the usual CorTES devices for sample 1, whereas sample 2 had much thicker SiN (750 nm).
\subsubsection{Sample 1}
\label{sec:slice1}
The first slice TES has relatively low values of $ \alpha \sim 60 - 80 $, with the result that the measured impedance data, shown in Fig. \ref{112fits}, lacks any striking features. The data is therefore too easy to fit: That is, we can find several ways to fit the data (both impedance and noise), and choosing the best fit is not simple based on this data alone. To illustrate this, we have fitted the same data in three different ways, with the resulting fitted curves almost identical up to 100 kHz, but with significant differences in the obtained parameter values. The three cases we consider are: Case 1: we force $ C_{TES} $ to the theoretically calculated value and let the other parameters vary freely. In case 2, $C_{TES} $ is not fixed but also free (as in all the fits for the CorTES and STES devices above). Finally, in case 3, we set $ C_{1} = 0 $, reducing the thermal model to two bodies: the TES and the intermediate.
\begin{figure}[h]
\includegraphics{Fig7.eps}
\caption{(Color online) (a) Impedance and (b) noise data for the slice TES, Sample 1 (symbols), for $R/R_{N}$=0.1, 0.2, 0.4, 0.5, 0.6, 0.8, rest omitted for clarity. We also show the best fit for case 1 (solid black line) and case 2 (dashed red line). The dotted green line in (a) is the impedance according to the thermal model of Fig. \ref{block}(a), illustrating how much our data deviates from the simple one-body case, and it was calculated using $Z_{0}$ and $Z_{\infty}$ obtained from case 1 at $R/R_N$ = 0.4. \label{112fits}}
\end{figure}
From the CorTES sample results above, we still expect that the three-block model is required to describe the sample physics, thus we first discuss the comparison between cases 1 and 2, and only later comment on the case 3 fits. Fig. \ref{112fits} shows a comparison of cases 1 and 2 with the data. It is clear that both cases fit the data well, and deviate from each other only at high frequencies close to 100 kHz, with case 1 making a more pronounced kink in the calculated impedance and slightly weaker roll-off for the noise.
\begin{figure}[h]
\includegraphics{Fig8.eps}
\caption{(Color online) Parameters obtained from the fits in Fig. \ref{112fits}. Case 1 is shown by filled symbols, and case 2 with open symbols. (a) Squares: $ C_{TES}$, triangles: $ C_{2} $ and circles: $ C_{1}$. Lines are the theoretical calculations as discussed in the text. (b) Thermal conductances of interest. (c) $ \alpha $ and $\alpha_{TOT}$ parameters calculated from the I-V (line) and impedance data (points). \label{112param}}
\end{figure}
Looking at the obtained parameters plotted in Fig. \ref{112param}, we find case 1 quite interesting. With $ C_{TES} $ forced to the theoretical values, $ C_{1} $ follows the calculated theoretical value of the reduced BCS heat capacity \cite{kozorezov}, supporting again the proposed picture of thermal decoupling between the N and S phases. As expected, due to the lack of AlOx, $C_{2} \sim 0.1$ pJ/K is much reduced compared to the CorTES and STES samples, even though the SiN membrane dimensions are the same. Case 2 reproduces all the trends, but now $C_{TES}$ is higher than the theory (as in the STES device) and $ C_{1} $ lower, but with their sum approximately constant between the two cases. We also note, how for both cases $ C_{1} $ grows fast below $R/R_{N} < 0.2$, similar to the STES device.
On the other hand, comparing the two cases in terms of $ g_{TES,1} $ fit values produces a striking difference: The trends are quite clear, but opposite for the two cases. Thus, it is hard to draw solid conclusions, except that one can see that the highest values obtained $\sim 100$ nW/K are consistent with the increased TES bilayer film thickness, if compared to the STES results. The trend in case 1 of decreasing $ g_{TES,1} $ with decreasing $R$ is of course consistent with the fact that thermal conductance starts to decrease in the superconducting phase, and that the size of the S phase increases when going down in bias.
In cases 1 and 2 we did not need to use the $M$-parameter (excess noise) in the upper part of the transition, but had to include it in the lower part, in case 1 below 30 \% bias and in case 2 already at 50 \%. This is in contrast to the CorTEs and STES fits, where $M$-parameter was not used at all (those fits were not as accurate, though). The fitted values of M are shown in Fig. \ref{MbetaCtot}(a). For this sample, we also found that in a narrow bias range $ R/R_N$ = 0.3 to 0.35 there was an enhancement of the mid-frequency (1 kHz - 30 kHz) noise that could be explained neither by the thermal circuit nor an $M$-noise component. Curiously, this bias range
corresponds to the sudden drop in the $\alpha$ parameters, as can be seen in Fig. \ref{112param}(c)).
\begin{figure}[h]
\includegraphics{Fig9.eps}
\caption{(Color online) Impedance fits in cases 3 (solid black line) and 1 (blue squares) compared. Red circles are the measured data. For comparison with the simple calorimeter model of Fig. \ref{block}(a), the dashed lines show the calculated one-block impedances using $Z_{\infty}$ of case 1 (short dash) and case 3 (long dash).
\label{112Cabs0}}
\end{figure}
To study Case 3, we show in Fig. \ref{112Cabs0} representative impedance fits at two bias points (solid lines), with case 1 fits included for comparison. Again, if we only look at frequencies up to the measurement maximum, we do not see any difference in the fits. The difference in the noise fits (not shown) is also limited to a small variation in the roll-off. In case 3 we naturally do not have the high frequency ITFN component present because $ C_{1} $ is missing, and thus a large part of the noise at high freqencies has to be accounted for by excess noise with $M$ parameter, at all bias points. The effect of the analysis case on the obtained $M$-parameter values vs the bias point is shown in Fig. \ref{MbetaCtot} (a). In addition, the chosen fitting case has a big effect also on $ \beta $, as it is proportional to $ Z_{\infty} $ [Fig. \ref{MbetaCtot} (b)].
The plots remind us that many parameters strongly depend on the chosen thermal model and how it is interpreted. If we do not know with certainty which of the possible thermal models is correct, we should not jump to conclusions about the nature of the excess noise based on the dependence of $M$ parameter on bias, magnetic field etc. In other words, the $Z$ data may look like it fits a simpler one- or two-block model, but the underlying true model could still be different. For example, if one is developing a detector where $ \beta $ is an important parameter, mistakes can be made in TES design if the decisions are based on $ \beta $ from a wrong fit. For the case of this particular device, we believe that Cases 1 and 2 are more accurate because of the evidence from the other samples studied here, where two-block model fits are impossible to the impedance data.
\begin{figure}
\includegraphics{Fig10.eps}
\caption{(Color online) Comparison of (a) M, (b) $ \beta $ and (c) the sum of $ C_{TES} $ and $ C_{1} $ between different fits cases. Symbols: values from fits. Line: BCS heat capacity with the bilayer correction, dashed line, full BCS theory. \label{MbetaCtot}}
\end{figure}
Another interesting point to notice is that even though we have three different sets of values for $ C_{TES} $ and $ C_{1} $, the sums $ C_{TES}+C_{1} $ are essentially constant between the different fitting cases, as shown in Fig. \ref{MbetaCtot}(c). The values are also consistent with the modified BCS theory \cite{kozorezov}, except at the lowest bias points, where the rapid increase mentioned before is clearly seen.
At this point we wish to speculate about the observations at low bias points. A possible scenario for the effects seen (increasing $C$, increasing $M$-noise) could well be related to vortex physics. If vortices are not pinned, they could contribute to heat capacity and also generate excess noise by the so called phase-slip shot noise mechanism \cite{slipsold,slipsnew}. The sudden drop in $ \alpha_{TOT,IV} $ is also correlated with the onset of the need for $M$- parameter in Case 1. The lower $\alpha$ would then correspond to extra resistance in the transition, caused by the vortex motion. In this light we can perhaps also understand why the full CorTES devices do not require $M$ to fit their noise, as the extra Nb top layer over the TES film may pin vortices more strongly, and prevent the generation of excess phase slip noise. Another option is that some of the excess $M$-noise could be generated by the FSN mechanism \cite{FSN}. However, FSN theory in its current state cannot explain the increased heat capacity or lowered $\alpha$.
\subsubsection{Sample 2}
The geometrical design of slice TES sample 2 was identical to sample 1, however the SiN and TES layer thickness are different (Table \ref{TEStable}), resulting also to a higher $T_c$. In contrast to sample 1, the second slice TES featured a more complicated transition with very large peaks in $ \alpha _{TOT,IV} $ as shown in the inset of Fig. \ref{111fits}(b). The high values and large variation in $ \alpha _{TOT,IV} $ produces complicated impedance features and high noise levels. Sample 2 was also measured at four different bath temperatures: 140 mK, 110 mK, 85 mK and 60 mK. Fig. \ref{111fits} shows the data for a few representative bias points measured at a bath temperature of 60 mK. The fits were done with the full IH model using Case 2, keeping $C_{TES}$ free. Now Case 3 is again out of the question due to the complex shapes of the $Z$-data. Case 1 was tried but did not produce as good fits as Case 2.
\begin{figure}
\includegraphics{Fig11.eps}
\caption{(Color online) Measured data (symbols) and fits (lines) for slice TES 2 at $T_{bath}$ = 60 mK. Insets: (a) SEM image of the TES, (b) Total $ \alpha $ calculated from I-V. \label{111fits}}
\end{figure}
Again, the main features of both $Z$ and noise data are captured with the IH model. However, at bias points where $\alpha$ is very large (for example $R/R_{N} =0.47 $) it was more difficult to find good fits, and determination of $ Z_{\infty} $ is more uncertain. The inevitable result of this is more scatter in the fitted parameters. In Fig. \ref{111param} we have plotted all the results from all four bath temperature runs in one graph, as $C_{TES}$, $ C_{1} $ and $ g_{TES,1} $ should not depend on $ T_{bath} $ at all according to the thermal model ($C_{2}$ depends on $T_{bath}$ through its influence on $T_{2}$, thus we show only $T_{bath}=$ 60 mK fits). Comparing to Sample 1, we see that $C_{TES}$ and $C_{1}$ have the same trends, and only slightly higher values, consistent with the increase of $T_c$. On the other hand, $C_{2}$ is three times larger. This again supports the picture that $C_{2}$ originates from the SiN membrane, as it is 2.5 times thicker for Sample 2. $ g_{TES,1} $ is of the same order of magnitude, and does not follow a monotonous trend throughout the whole transition. Most of the noise data did not require any additional $M$-noise, all the fits shown in Fig. \ref{111fits} are without the $M$ parameter.
This, we speculate, could be because of the much higher $\alpha$ compared to Sample 1: The themal noise grows so much that is swamps any possible additional noise sources.
\begin{figure}
\includegraphics{Fig12.eps}
\caption{(Color online) Parameters obtained from the fits for sample 2 at all values of $T_{bath}$, except $C_{2}$ is shown only at $T_{bath}=$ 60 mK. (a) Squares: $ C_{TES}$, circles: $C_{1} $ and triangles: $ C_{2} $. Dashed black and solid red lines are theoretical calculations for $ C_{TES}$ and $ C_{1}$, respectively, whereas the dash-dotted line is their sum. The reduced BCS value \cite{kozorezov} was used for $ C_{1} $. \label{111param}}
\end{figure}
The data for $ g_{TES,2} $ was not shown because it depends on the bath temperature. For comparison with sample 1, at 60 mK sample 2 had $ g_{TES,2}(T_{TES}) $ = 1.1 nW/K and $ G_{dyn} $ = 0.97 nW/K at 0.5 $ R/R_N $.
As can be seen for bias point $R/R_{N}=0.2$ in Fig. \ref{111fits}, the measured noise sometimes develops a peak near the high-frequency roll-off, when the TES is biased low in the transition. This indicates electrothermal oscillations and it is predicted by our thermal model, as shown by the fit. It arises due to a resonant-like interaction between the electrical and thermal circuits: When $ C_{1} $ starts to decouple from a TES that has a very small heat capacity and large enough $ \alpha $, the TES response becomes oscillatory.
The increased responsivity means that all noise sources develop a peak, including the Johnson noise, for example.
\subsection{Square TES}
Finally, we show also results for a simple square TES. The measured detector was a bare $\mathrm{300 \mu m} \times \mathrm{300 \mu m}$ TES with Nb bias lines and no extra features added (more sample parameter details in Table \ref{TEStable}). It has a smoothly changing $\alpha_{TOT,IV}$ with a fairly low maximum value of 50. In Fig. \ref{TTESfits} we plot the measured and fitted impedance and noise data. Just as for all the other geometries, our square TES clearly has a large high frequency noise component. Again, the shapes of the impedance curves shows that the model without a hanging block (case 3 discussed in section \ref{sec:slice1}) will not work here either. The IH model again produces a very good fit for both $Z$ and noise, which gives us confidence that the noise in the square TES can be explained by the same mechanism as in the other geometries. Note that no $M$ parameter was needed in these fits.
\begin{figure}
\includegraphics{Fig13.eps}
\caption{(Color online) Measured square TES data (symbols) and fits with IH model (lines). \label{TTESfits}}
\end{figure}
Fig. \ref{TTESparam} shows the relevant parameters from the fits. We again observe a decreasing $C_{TES}$ and an increasing $C_{1}$, consistent with our model of thermal decoupling of the N and S regions, and comparable values to the Corbino-like samples. To calculate some estimate for the N and S phase heat capacities, we take a crude model of a linear transition, where the ratio of the volumes is given directly by $R/R_N$, so for example in the middle of the transition we would have exactly half of the device in the superconducting state. This model produces a correct order of magnitude for $C_{TES}$ and $C_{1}$, but naturally no real agreement, although the parameter values do seem to change linearly with $R$, as our overly simple model predicts. The values of thermal conductances are again consistent with the Corbino devices, and $g_{TES,1}$ does not show any simple dependence on $R$.
\begin{figure}
\includegraphics{Fig14.eps}
\caption{(Color online) Parameters from the fits in Fig. \ref{TTESfits}. Lines show the simple estimates for $C_{TES}$ (dashed) and $C_{1}$ (solid) discussed in the text. \label{TTESparam}}
\end{figure}
\section{Conclusions and discussion}
We have fabricated several different geometrical designs of TES devices, and studied their impedance and noise properties. All data are consistent with the picture that a three block thermal circuit is required to accurately describe the data. In all devices, the fitted heat capacity values follow trends which suggest that one of the extra heat capacity blocks arises from the insulating films such as the SiN membrane and, if present, the AlOx layer used in some devices. The other two blocks seem to represent the normal and superconducting regions of the TES film itself, with the observation that the finite value of thermal conductance within the TES film (or a possible thermal boundary resistance between the N and S phases), in combination with the large value of the heat capacity in the superconducting phase near $T_c$, leads to fairly large observable features in the impedance and noise data at high frequencies. In most cases, this internal thermal fluctuation noise component explains all of the "excess noise" in the devices. However, in some devices indications of additional noise sources remained, especially low in the transition.
In our Ti/Au devices the resistivity of the TES film is such that $ g_{TES,1} $ is "low enough". By this we mean that the resulting cut-off frequency of the ITFN noise due to the hanging block is near the electrical cut-off of the readout, so that we can observe its effect in the measured data. Our data, therefore, agrees with the previous suggestion that a large ITFN noise exists in Ti/Au TESs \cite{hoevers,SRON}.
In contrast, the lower resistivity Mo/Au or Mo/Cu devices usually have a lower excess noise level that is white voltage noise, in other words noise that looks like excess Johnson noise \cite{ullom,iyomoto}.
Our model could perhaps explain part of this difference between Ti based and Mo based devices through the higher thermal conductance $g_{TES,1}$ of the Mo/Au and Mo/Cu TES films. If $g_{TES,1}$ grows, the ITFN noise level will fall, and the cut-off is pushed to a higher frequency. Thus if the ITFN roll-off happens much after the readout cut-off, ITFN noise will look like excess Johnson noise in the frequency range of the measurement. More measurements on Mo devices are required to test this hypothesis.
The results reported here indicate that in an optimal TES design, the internal thermal conductances within the metallic parts of the device (TES film, absorber) should be maximized. As the actual phase separation between the superconducting and normal regions may be impossible to prevent, the effect of the superconducting heat capacity is minimized if the heat capacity of the normal part of the TES is larger. We note that the leading designs in terms of reported energy resolution \cite{ullom,iyomoto} employ added normal metal features, and thus we speculate that the reduction of their noise could originate, at least partly, due to the increased normal region heat capacity.
\begin{acknowledgments}
We acknowledge helpful discussions about data analysis with M. Lindeman, J. van der Kuur, J. Ullom and D. Swetz. This work was supported by the Finnish Funding Agency for Technology and Innovation TEKES and EU through the regional funds, and the Finnish Academy project no. 128532. M. P. would like to thank the National Graduate School in Materials Physics for funding.
\end{acknowledgments}
|
\section{#1}}
\renewcommand{\theequation}{\arabic{section}.\arabic{equation}}
\begin{document}
\title{\bf Anisotropic Dark Energy and the Generalized Second Law of Thermodynamics}
\author{M. Sharif \thanks {<EMAIL>} and Farida
Khanum \thanks {<EMAIL>}\\
Department of Mathematics, University of the Punjab,\\
Quaid-e-Azam Campus, Lahore-54590, Pakistan.}
\date{}
\maketitle
\begin{abstract}
We consider a Bianchi type $I$ model in which anisotropic dark
energy is interacting with dark matter and anisotropic radiation.
With this scenario, we investigate the validity of the generalized
second law of thermodynamics. It is concluded that the validity of
this law depends on different parameters like shear, skewness and
equation of state.
\end{abstract}
\textbf{Keywords:} Bianchi Type $I$ Model; Generalized Second Law of
Thermodynamics.\\
\textbf{PACS:} 95.36.+x, 98.80.-k
\section{Introduction}
The observational evidence that our universe is making a transition
from a decelerating phase to an accelerating is the major
development in cosmology. Supernova $Ia$ data \cite{1,2} gave the
first indication of the accelerated expansion of the universe. This
was confirmed by the observations of anisotropies in the cosmic
microwave background (CMB) radiations as seen in the data from
satellites such as Wilkinson Microwave Anisotropy Probe (WMAP)
\cite{3} and large scale structure \cite{4}. Astrophysical
observations indicate that accelerated expansion of the universe is
driven by an exotic energy with large negative pressure which is
known as \emph{dark energy} (DE).
The data indicates that the universe is spatially flat and is
dominated by 76\% DE and 24\% other matter (20\% dark matter and 4\%
other cosmic matter). Despite all lines of observational evidence,
the nature of DE is still a challenging problem in theoretical
physics. Several models have been proposed such as, quintessence
\cite{5}, phantom field \cite{6}, tachyon field \cite{7}, quintom
\cite{8} and the interacting DE models, Chaplygin gas \cite{9},
holographic models \cite{10} and braneworld models \cite{11}, etc.
However, none of these models can be regarded as being entirely
convincing so far.
In black hole physics, the temperature and entropy are proportional
to the surface gravity at the horizon and the area of the horizon
\cite{12,13}, respectively. Also, the temperature, entropy and mass
of the black hole satisfy the first law of thermodynamics \cite{14}.
This leads to the relationship between the black hole thermodynamics
and the Einstein field equations. Jacobson \cite{15} derived the
field equations from the first law of thermodynamics for all local
Rindler causal horizons. Padmanavan \cite{16} formulated the first
law of thermodynamics on the horizon using the field equations for a
general static spherically symmetric spacetime. Later, this
relationship was developed in the cosmological context by taking
universe as a thermodynamical system bounded by the apparent
horizon. Cai and Kim \cite{17} showed the equivalence of the first
law of thermodynamics to the Friedmann equations with the
generalized second law of thermodynamics (GSLT) at the horizon.
Cosmological models in which DE interacts with dark matter and other
cosmic matter are well-known in the literature \cite{18}. The
discovery of black hole thermodynamics has led to the thermodynamics
of cosmological models. Bekenstein \cite{13} proved that there is a
relation between an event horizon and thermodynamics of black hole.
The event horizon of black hole is in fact the measure of entropy
which is generalized to the cosmological models so that each horizon
corresponds to an entropy. The GSLT is generalized in such a way
that the sum of the time derivative of each entropy must be
increasing. People studied the thermodynamics of different models in
which different cosmic constituents interact.
Wang et al. \cite{20} found that both the first and the second law
of thermodynamics break down at the event horizon. The first law of
thermodynamics is restricted to nearby states of local thermodynamic
equilibrium while the event horizon reflects global features of the
spacetime. Also, due to the existence of the cosmological event
horizon, the universe should be non-static in nature and, as a
result, the usual definition of the thermodynamical quantities on
the event horizon may not be as simple as in the static spacetime.
Finally, they have argued that as the event horizon is larger than
the apparent horizon so the universe bounded by the event horizon is
not a Bekenstein system.
There is a large body of literature \cite{21}-\cite{28e} dealing
with the thermodynamics of the universe bounded by the apparent
horizon and the validity of GSLT. Mazumder et al. \cite{29},
\cite{30} have explored the validity of GSLT using the first law of
thermodynamics and found some conditions on its validity. Debnath
\cite{31} investigated the validity of GSLT in the scenario of
Friedmann Robertson Walker (FRW) model in which DE interacts with
dark matter and radiation. Mubasher et al. \cite{32} proved the
validity of GSLT for all time, independent of geometry and equation
of state (EoS) parameter. In a recent paper \cite{32a}, the validity
of GSLT is explored in the Kaluza-Klein cosmology with modified
holographic DE.
Jacobs \cite{33} studied the spatially homogeneous and anisotropic
Bianchi type $I$ (BI) cosmological model with expansion and shear
but without rotation. He discussed anisotropy in the temperature of
CMB and cosmic expansion both with and without magnetic field.
Akarsu and Kilinc \cite{34} investigated anisotropic BI models in
the presence of perfect fluid and minimally interacting DE with
anisotropic EoS parameter. They found that anisotropy of the DE did
not always promote the anisotropy in the expansion.
In this paper, we take the BI model in which anisotropic DE
interacts with isotropic dark matter as well as anisotropic
radiation and investigate the validity of GSLT. The plan of this
paper is the following: in next section, we take the BI model as the
representation of the universe and evaluate shear parameters
$(R,S)$, skew parameters $(\delta,~\gamma,~\zeta,~\xi)$. Also, we
formulate the corresponding field equations, the generalized
Friedmann equation and equations of continuity in the effective
field theory. Section \textbf{3} presents the analysis of the
validity of GSLT. The last section concludes the paper results.
\section{Anisotropic DE Interacting with Two Fluids}
The metric representation of the BI universe is given as
\begin{equation}\label{1}
ds^{2}=-dt^{2}+A^2(t)dx^{2}+B^2(t)dy^{2}+C^2(t)dz^{2},
\end{equation}
where $A,~B,~C$ are scale factors along the $x,~y,~z$ axes
respectively. When $A=B=C$, the BI model reduces to the flat FRW
model. Thus BI is the generalization of the flat FRW model. The
energy-momentum tensor is defined as
\begin{equation}\label{2}
T^{\mu}_{\nu}=T^{\mu}_{(m)\nu}+T^{\mu}_{(\Lambda)\nu}+T^{\mu}_{(\chi)\nu},
\end{equation}
where the subscripts $m,~\Lambda,~\chi$ denote dark matter,
anisotropic DE and anisotropic radiation, respectively, and
corresponding energy momentum tensors are given as follows:
\begin{eqnarray}\label{3}
T^{\mu}_{(m)\nu}&=&(-\rho_{m},P_{m},P_{m},P_{m}),\\\label{4}
T^{\mu}_{(\Lambda)\nu}&=&(-\rho_{\Lambda},P_{(\Lambda)x},P_{(\Lambda)y},P_{(\Lambda)z}),\\\label{5}
T^{\mu}_{(\chi)\nu}&=&(-\rho_{\chi},P_{(\chi)x},P_{(\chi)y},P_{(\chi)z}).
\end{eqnarray}
The corresponding EoSs for these components are
\begin{eqnarray}\label{6}
P_{(\Lambda)x}&=&\rho_{\Lambda}\omega_{(\Lambda)x},\quad
P_{(\Lambda)y}=\rho_{\Lambda}\omega_{(\Lambda)y},\quad
P_{(\Lambda)z}=\rho_{\Lambda}\omega_{(\Lambda)z},\\\label{7}
P_{(m)x}&=&\rho_{m}\omega_{(m)x},\quad
P_{(m)y}=\rho_{m}\omega_{(m)y},\quad
P_{(m)z}=\rho_{m}\omega_{(m)z},\\\label{8}
P_{(\chi)x}&=&\rho_{\chi}\omega_{(\chi)x},\quad
P_{(\chi)y}=\rho_{\chi}\omega_{(\chi)y},\quad
P_{(\chi)z}=\rho_{\chi}\omega_{(\chi)z}.
\end{eqnarray}
To parameterize the anisotropy of pressure, we define skewness
parameters. These parameters actually measure the deviation in
pressure along the $y$ and $z$-axis from the $x$-axis. The skewness
parameters for anisotropic DE are defined as \cite{35}
\begin{equation}\label{9}
3\delta=\frac{P_{(\Lambda)x}-P_{(\Lambda)y}}{\rho_{\Lambda}},\quad
3\gamma=\frac{P_{(\Lambda)z}-P_{(\Lambda)x}}{\rho_{\Lambda}}.
\end{equation}
Using Eqs.(\ref{6}) and (\ref{9}), we can write Eq.(\ref{4}) as
follows:
\begin{equation}\label{10}
T^{\mu}_{(\Lambda)\nu}=[-\rho_{\Lambda},\omega_{\Lambda}\rho_{\Lambda},
(\omega_{\Lambda}+3\delta)\rho_{\Lambda},(\omega_{\Lambda}+3\gamma)\rho_{\Lambda}].
\end{equation}
The skewness parameters for anisotropic radiation are
\begin{equation}\label{11}
3\zeta=\frac{P_{(\chi)x}-P_{(\chi) y}}{\rho_{\chi}},\quad
3\xi=\frac{P_{(\chi)z}-P_{(\chi) x}}{\rho_{\chi}}.
\end{equation}
With the help of Eqs.(\ref{8}) and (\ref{11}), Eq.(\ref{5}) turns
out to be
\begin{eqnarray} \label{12}
T^{\mu}_{(\chi)\nu}=[-\rho_{\chi},\omega_{\chi}\rho_{\chi},
(\omega_{\chi}+3\zeta)\rho_\chi,(\omega_{\chi}+3\xi)\rho_{\chi}].
\end{eqnarray}
Making use of Eqs.(\ref{10}) and (\ref{12}) in Eq.(\ref{2}), we have
\begin{eqnarray}\nonumber
T^{\mu}_{\nu}&=&[-\rho_{m},\omega_{m}\rho_{m},\omega_{m}\rho_{m},\omega_{m}\rho_{m}]
+[-\rho_{\Lambda},\omega_{\Lambda}\rho_{\Lambda},(\omega_{\Lambda}
+3\delta)\rho_{\Lambda},(\omega_{\Lambda}+3\gamma)\rho_{\Lambda}]\\\label{2(a)}
&+&[-\rho_{\chi},\omega_{\chi}\rho_{\chi},(\omega_{\chi}
+3\zeta)\rho_{\chi},(\omega_{\chi}+3\xi)\rho_{\chi}].
\end{eqnarray}
The corresponding field equations take the following form:
\begin{eqnarray}\label{13}
\frac{\dot{A}\dot{B}}{AB}+\frac{\dot{B}\dot{C}}{BC}+\frac{\dot{C}\dot{A}}{CA}
&=&[\rho_{m}+\rho_{\Lambda}+\rho_{\chi}],\\\label{14}
\frac{\ddot{B}}{B}+\frac{\ddot{C}}{C}+\frac{\dot{B}\dot{C}}{BC}
&=&-[\omega_{m}\rho_{m}+\omega_{\Lambda}\rho_{\Lambda}+\omega_{\chi}\rho_{\chi}],\\\label{15}
\frac{\ddot{A}}{A}+\frac{\ddot{C}}{C}+\frac{\dot{A}\dot{C}}{AC}
&=&-[\omega_{m}\rho_{m}+(\omega_{\Lambda}+3\delta)\rho_{\Lambda}
+(\omega_{\chi}+3\zeta)\rho_{\chi}],\\\label{16}
\frac{\ddot{A}}{A}+\frac{\ddot{B}}{B}+\frac{\dot{A}\dot{B}}{AB}
&=&-[\omega_{m}\rho_{m}+(\omega_{\Lambda}+3\gamma)\rho_{\Lambda}+(\omega_{\chi}+3\xi)\rho_{\chi}].
\end{eqnarray}
The continuity equations, ${T^{\mu\nu}}_{;\nu}=0$, are
\begin{eqnarray}\label{17}
\dot{\rho_{\Lambda}}+H\rho_{\Lambda}[3(1+\omega_{\Lambda})
+\delta(3-2R+S)+\gamma(3-2S+R)]=-Q',\\\label{18}
\dot{\rho_{m}}+3H\rho_{m}(1+\omega_{m})=Q.
\end{eqnarray}
Here $Q$ and $Q'$ are introduced because two principal components of
the universe are mutually interacting, which lead to some loss in
other cosmic component. We take this component as radiation for
which the continuity equation becomes
\begin{equation}\label{19}
\dot{\rho_{\chi}}+H\rho_{\chi}[3(1+\omega_{\chi})+\zeta(3-2R+S)+\xi(3-2S+R)]=Q'-Q.
\end{equation}
Here, we define the mean expansion rate as an average Hubble rate
$H$ by
\begin{equation}\label{26}
H=\frac{1}{3}(\frac{\dot{A}}{A}+\frac{\dot{B}}{B}+\frac{\dot{C}}{C}).
\end{equation}
The difference of expansion rates as the Hubble normalized shear
parameters $R$ and $S$ are given as \cite{37}
\begin{equation}\label{27}
R=\frac{1}{H}(\frac{\dot{A}}{A}-\frac{\dot{B}}{B}),\quad
S=\frac{1}{H}(\frac{\dot{A}}{A}-\frac{\dot{C}}{C}).
\end{equation}
Shear parameters are also deviation measuring parameters which
parameterize anisotropy in scale factors by taking expansion along
$x$-axis as standard. Here, all expansion rates are positive and
$R>-3,~S<3$. This is also true either for $R=0$ or $S=0$. The
generalized Friedmann equation takes the form \cite{35}
\begin{equation}\label{28}
H^{2}=\frac{\rho_{m}+\rho_{\Lambda}+\rho_{\chi}}{3(1-\frac{1}{9}(R^{2}+S^{2}-RS))}.
\end{equation}
We take \cite{36}
$Q'=\Gamma_{\Lambda}\rho_{\Lambda},~Q=\Gamma_{m}\rho_{\Lambda}$.
The ratios for energy densities as
\begin{equation}\nonumber
r_{1}=\frac{\rho_{m}}{\rho_{\Lambda}},\quad
r_{2}=\frac{\rho_{\chi}}{\rho_{\Lambda}}.
\end{equation}
The corresponding EoS parameters in the effective field theory will
become
\begin{eqnarray}\label{20}
\omega^{eff}_{\Lambda}&=&\omega_{\Lambda}+\frac{\Gamma_{\Lambda}}{3H}
+\frac{\delta(3-2R+S)+\gamma(3-2S+R)}{3},\\\label{21}
\omega^{eff}_{m}&=&\omega_{m}-\frac{\Gamma_{m}}{3Hr_{1}},\\\label{22}
\omega^{eff}_{\chi}&=&\omega_{\chi}+\frac{\Gamma_{m}-\Gamma_{\Lambda}}
{3Hr_{2}}+\frac{\zeta(3-2R+S)+\xi(3-2S+R)}{3}.
\end{eqnarray}
Consequently, the continuity equations (\ref{17})-(\ref{19}) turn
out to be
\begin{eqnarray}\label{23}
\dot{\rho_{\Lambda}}+3H\rho_{\Lambda}(1+\omega^{eff}_{\Lambda})&=&0,\\\label{24}
\dot{\rho_{m}}+3H\rho_{m}(1+\omega^{eff}_{m})&=&0,\\\label{25}
\dot{\rho_{\chi}}+3H\rho_{\chi}(1+\omega^{eff}_{\chi})&=&0.
\end{eqnarray}
\section{Generalized Second Law Of Thermodynamics}
Now we investigate the validity of the GSLT in BI universe bounded
by apparent horizon with size $L$ which coincides with Hubble
horizon in the case of flat geometry, i.e., $L=\frac{1}{H}$. The
first law of thermodynamics gives
\begin{equation*}
TdS=PdV+dE,\quad dS=\frac{PdV+dE}{T},
\end{equation*}
where $T,~S,~E$ and $P$ are the temperature, entropy, internal
energy and pressure of the system respectively. The corresponding
entropies will become
\begin{equation}\label{29}
dS_{\Lambda}=\frac{P_{\Lambda}dV+dE_{\Lambda}}{T},\quad
dS_{m}=\frac{P_{m}dV +dE_{m}}{T},\quad
dS_{\chi}=\frac{P_{\chi}dV+dE_{\chi}}{T},
\end{equation}
where $P_{\Lambda},~P_{\chi},~P_{m},~E_{\Lambda},~E_{\chi}$ and
$E_{m}$ are the pressures and internal energies of anisotropic DE,
anisotropic radiation and dark matter, respectively. We assume that
the system is in equilibrium, which implies that all the components
of the system have the same temperature given by \cite{13}
\begin{equation}\label{30}
T=\frac{K}{2\pi },
\end{equation}
where $K$ is the surface gravity of the black hole. At equilibrium,
the horizon and other components of the system have same
temperature. Thermodynamical quantities are related to the
cosmological quantities by the following relations:
\begin{eqnarray}\label{31}
P_{\Lambda}&=&\omega^{eff}_{\Lambda}\rho_{\Lambda},\quad
P_{\chi}=\omega^{eff}_{\chi}\rho_{\chi},\quad
P_{m}=\omega^{eff}_{m}\rho_{m},\\\label{32}
E_{\Lambda}&=&ABC\rho_{\Lambda},\quad E_{\chi}=ABC \rho_{\chi},\quad
E_{m}=ABC\rho_{m},
\end{eqnarray}
where $V=ABC$ is the volume of the system containing all the
matter.
The entropy of the horizon is $S_{h}=\frac{kA^*}{4}$, where $A^*$ is
the surface area of the black hole and $k$ is the Boltzmann's
constant. Here $A^*={4\pi L^{2}}$ so that we have $S_{h}=k \pi
L^{2}$, which leads to
\begin{equation}\label{34}
\dot{S_{h}}=2k\pi L \dot{L}.
\end{equation}
Also, the time derivative of Eq.(\ref{29}) yields
\begin{equation}\label{35}
\dot{S_{\Lambda}}=\frac{P_{\Lambda}\dot{V}+\dot{E_{\Lambda}}}{T},\quad
\dot{S_{\chi}}=\frac{P_{\chi}\dot{V}+\dot{E_{\chi}}}{T},\quad
\dot{S_{m}}=\frac{P_{m}\dot{V}+\dot{E_{m}}}{T}.
\end{equation}
Using Eqs.(\ref{23})-(\ref{25}), (\ref{34}) and (\ref{35}), it
follows that
\begin{equation}\label{36}
\dot{S}_{total}=2k\pi L\dot{L},
\end{equation}
where $\dot{S}_{total}$ is the sum of the matter entropy and horizon
entropy. Now making use of Eqs.(\ref{23})-(\ref{25}) and (\ref{28}),
we obtain
\begin{equation}\label{37}
\dot{L}=-\frac{\dot{H}}{H^2}=\frac{L^{3}H\alpha}{2\beta}-\frac{L^{3}H^{2}[\dot{R}(2R-S)
+\dot{S}(2S-R)]}{18\beta},
\end{equation}
where
\begin{eqnarray}\label{38}
&&\alpha=(1+\omega_{\Lambda})\rho_{\Lambda}+(1+\omega_{m})\rho_{m}
+(1+\omega_{\chi})\rho_{\chi}\nonumber\\
&&+\frac{1}{3}[(\rho_{\Lambda}\delta+\rho_{\chi}\zeta)(3-2R+S)
+(\rho_{\Lambda}\gamma+\rho_{\chi}\xi)(3-2S+R)],
\end{eqnarray}
and
\begin{equation}\label{39}
\beta={1}-\frac{R^{2}+S^{2}-R S}{9}\geq0.
\end{equation}
Inserting this value of $\dot{L}$ in Eq.(\ref{36}), it follows
that
\begin{equation}\label{40}
\dot{S}_{total}=\frac{ k \pi H L^{4}\alpha}{\beta}-\frac{ k \pi
H^{2} L^{4} [\dot{R}(2R-S)+\dot{S}(2S-R)]}{9\beta}.
\end{equation}
This implies that if $\dot{S}_{total}\geqslant0$, then GSLT holds.
Thus the validity of GSLT depends upon the skewness, shear and
state parameters. From Eq.(\ref{40}), $\dot{S}_{total}\geqslant0$
implies that
\begin{equation*}
\frac{ k \pi H L^{4}\alpha}{\beta}-\frac{ k \pi H^{2} L^{4}
[\dot{R}(2R-S)+\dot{S}(2S-R)]}{9\beta}\geq0.
\end{equation*}
After substituting the values of $\alpha$ and $\beta$ in the above
equation, it follows that
\begin{eqnarray}\label{41}
&&9(\rho+p)\geq[\dot{R}(2R-S)+\dot{S}(2S-R)]-3[(\rho_{\Lambda}\delta+
\rho_{\chi}\zeta)(3-2R+S)\nonumber\\
&&+(\rho_{\Lambda}\gamma+\rho_{\chi}\xi)(3-2S+R)].
\end{eqnarray}
Now, we consider the following three interesting
cases involving the condition on different parameters:\\
\textbf{I.} $\beta=0$, \textbf{II.} $\beta>0$, \textbf{III.}
$R=S=\delta=\gamma=\xi=\zeta=0$.
\subsection*{Case I}
Here $H$ and $\dot{S}_{total}$ tend to infinity for $\beta=0$ from
Eqs.(\ref{28}) and (\ref{40}) respectively. It may happen for very
large time $(t\rightarrow\infty)$, i.e., when the expansion rate is
very high. In this case, all the useable energy in the universe will
be converted into another form of energy which is not useable. This
stage is also known as the heat death of the system. The heat death
of the universe is said to be a suggested fate of the universe. At
this time, all the thermodynamic free energy will be diminished from
the universe and motion or life cannot sustain any more. In the
language of physics, the entropy in the universe will reach to its
maximum value.
\subsection*{Case II}
In this case, by takeing $\beta>0$ in Eq.(\ref{40}), we observe that
the validity of this law depends on the skewness, shear and EoS
parameters. Further, if we remove anisotropy in the expansion, its
validity depends on skewness as well as equation of state
parameters, and if we remove anisotropy in fluid, then it depends on
shear and equation of state parameters. Hence, GSLT is conditionally
valid.
\subsection*{Case III}
When we take $R=S=\delta=\gamma=\xi=\zeta=0$, BI reduces to the flat
FRW model. Using these values in Eq.(\ref{41}), we have
$(\rho+p)\geq0$, which is the null energy condition in the FRW
universe. Consequently, Eq.(\ref{40}) leads to
$\dot{S}_{total}=k\pi(\rho+p)$ and hence $\dot{S}_{total}\geq0$ for
$(\rho+p)\geq0$. Thus we can say that GSLT holds for all time in
this case, if the null energy condition for the considered matter is
satisfied in the FRW universe. This case implies that the validity
of GSLT depends on the null energy condition which is not discussed
in \cite{32}. Therefore using results produced in this case, we can
say that the validity of GSLT depends on null energy condition and
background geometry.
\section{Conclusion}
This paper is devoted to study the validity of GSLT with the BI
universe model. We assume that anisotropic DE is interacting with
dark matter and anisotropic radiation. Using this scenario, we have
formulated the conditions under which GSLT is valid. We have
discussed three cases for its validity. For a particular value of
$\beta=0$, the first case provides the general validity of GSLT. In
this case, we see that at infinite expansion the heat death of
universe will take place. At this stage all types of motion and life
will be finished. The second case $\beta>0$ gives conditional
validity of this law. In the third case, BI reduces to the flat FRW
metric and the general validity of GSLT depends upon the condition
$(\rho+p)\geq0$, which is the null energy condition. We would like
to mention here that the validity of this law for the FRW universe
with the same scenario has been proved \cite{32} independent of
background geometry, EoS parameter and interacting scenario. This
analysis indicates that the validity of GLST for FRW depends on
geometry as well as null energy condition. We conclude that GSLT is
conditionally valid in the BI type universe with anisotropic perfect
fluid.
|
\section{Introductions}
Quantization of string theory in $AdS_5\times S^5$ background is very important for the study of $AdS_5/CFT_4$ correspondence
\cite{Mal97, Gubser:1998bc, Witten:1998qj}. In spite of many efforts \cite{Berkovits:2000fe, Berkovits:2007rj, Bonelli:2008rv}, this is still very challenging even for free strings, due to the existence of the Ramond-Ramond fluxes. This difficulty led people to study Penrose limit of this background \cite{BMN} and semi-classical string solutions \cite{GKP} inside it. Largely inspired by these studies, people found
remarkable integrable structure in the planar limit \cite{Minahan:2002ve, Bena:2003wd} (for a collection of reviews, see \cite{Beisert:2010jr}). With integrability at hand, people can compute many non-trivial quantities. Cusp anomalous dimension at arbitrary coupling \cite{Freyhult:2010kc} and scattering amplitudes at strong coupling \cite{Alday:2010kn} are two important examples of these quantities. Later this integrable structure was also found \cite{Minahan:2008hf, Bak:2008cp, Arutyunov:2008if, Stefanski:2008ik, arXiv:0811.1566, arXiv:1009.3498, arXiv:1101.3777} in $AdS_4/CFT_3$ correspondence \cite{Aharony:2008ug}. The latter correspondence states that type IIA string theory in $AdS_4\times \cp$ is dual to certain three-dimensional ${\cal N}=6$ Chern-Simons-matter theory with gauge group $U(N)\times U(N)$ and Chern-Simons levels $(k, -k)$. For reviews on integrability in this correspondence, see \cite{Klose:2010ki, Lipstein:2011}.
Planar approximation in the gauge theory is dual to free approximation in the string theory side. Studying the interaction of strings in these nontrivial backgrounds is quite attracting and difficult. A single closed string can split into two because of interactions. One necessary condition for the splitting is that there are two points in the string which have the same positions and velocities. Folded strings satisfy this condition in a very simple manner.
The splitting in flat spacetime can be studied at the full quantum level \cite{Iengo:2003ct}. However in backgrounds like $AdS_5\times S^5$, currently we can only study the splitting at the classical level \cite{Peeters:2004pt, Murchikova:2011ea, Vicedo:2011vn}. The splitting of strings in $AdS_5\times S^5$ was studied in the gauge theory side in \cite{Peeters:2004pt, Casteill:2007td}. People hope that studies of the splitting will help us to improve our understanding about the interactions of strings in these non-trivial backgrounds and the behaviors of non-planar corrections in the gauge theory side \cite{Beisert:2002bb, Beisert:2003tq} ( for related review, please see \cite{Kristjansen:2010kg}) \footnote{Some discussions on non-planar sector via other approaches appeared in \cite{deMelloKoch:2009zm, Carlson:2011hy, Koch:2011jk, deMelloKoch:2011ci}.}.
Among other classical string solutions, two folded strings in $AdS_4\times \cp$ background were found in \cite{Chen:2008qq}. In this paper, we will study the splitting of these two strings. We compute their energy and the conserved charges from the isometry group $SU(4)$ of the $\cp$ part of the background geometry, both before \footnote{The energy and the Noether charges from the Cartan generators of $SU(4)$ before splitting have already been computed in \cite{Chen:2008qq}.} and after the splitting. Some quite interesting patterns for these charges emerge for both folded strings. For the first folded string, we find that the
generators giving nonzero charges all belong to an $SU(3)$ subgroup of $SU(4)$. For the second folded string, we find that the conserved has the same form as the charges given in \cite{Peeters:2004pt} for splitting string in $AdS_5\times S^5$. The latter pattern can be thought as the extension of the similarity between the charges before the splitting already noticed in \cite{Chen:2008qq}. These similarities are quite interesting since there are many differences between $AdS_5/CFT_4$ and $AdS_4/CFT_3$ correspondences on both string theory side and gauge theory side. The study here on the concrete examples should be complementary to more abstract approach in \cite{Vicedo:2011vn}. Some aspects on non-planar corrections in the gauge theory side was discussed in \cite{Kristjansen:2008ib}.
This paper is organized as follows: after discussing the Noether charges from the isometry group of $\cp$ in the next section, we study the splitting of two kinds of folded strings one by one in the following two sections. The last section is devoted to conclusions and discussions. We list the generators of the isometry group $SU(4)$ of $\cp$ in the fundamental representation in the appendix.
\section{Conserved charges}
In this paper, we focus on type IIA string theory in $AdS_4\times {\bf CP}^3$ background. The background metric is:
\be ds^2=R^2(\frac14 ds^2_{AdS_4}+ds^2_{{\bf CP}^3}),\ee
where $ds^2_{AdS_4}$ is metric of $AdS_4$ with unit radius
\be ds^2_{AdS_4}=- \cosh^2\rho dt^2+d\rho^2+\sinh^2\rho(d\theta^2+\sin^2\theta d\phi^2),\ee and the $ds^2_{{\bf} CP^3}$ is Fubini-Study metric of $\cp$ whose
details will be given below. The relation between $R$ and 't Hooft coupling $\lambda=N/k$ in the gauge theory side is:
\be R=2^{5/4}\pi^{1/2}\alpha^{\prime 1/2}\lambda^{1/4}, \label{lambda}\ee
The dilaton field is a constant in this background:
\be e^{2\phi}=2^{5/2}\pi N^{1/2}k^{-5/2}.\ee There are also Ramond-Ramond background fields, but they do not play any roles in the studies here.
For the study of splitting classical string inside ${\bf CP}^3$, we would like to compute the conserved charges of these strings.
We start with the $\sigma$-model action:
\be S=\frac{1}{4\pi\alpha^\prime}\int \sqrt{-g}g^{\a\b}\p_\a X^\m \p_\b X^\n G_{\m\n}.\ee
The Killing vector $v^\mu$ of the target space gives a symmetry of the above action. And the corresponding
Noether current is:
\be j^\a=\frac1{2\pi\alpha^\prime} \sqrt{-g}g^{\a\b}\p_\b X^\n v^\m G_{\m\n}.\ee
We take the conformal gauge in which $g_{\a\b}$ is proportional to $\eta_{\a\b}=diag(-1, 1)$. Then we have
\be j^\a=\frac1{2\pi\alpha^\prime} \eta^{\a\b}\p_\b X^\n v^\m G_{\m\n}.\ee
${\bf CP}^3$ has isometry group $SU(4)$. The corresponding Killing vectors were given in \cite{Hoxha:2000jf} and more
explicitly in the coordinates we will use in \cite{Huang:2010qy}.
Let us begin with an unit $S^7$ \be\sum_{i=1}^4 |Z_i|^2=1,\ee inside ${\bf C}^4$. First define
\be\Omega_i=\sum_{I, J=1}^{4}(T_i)_I^{\,\,J}Z^IZ^\dagger_J \ee
with $T_i$ generators of $SU(4)$ in fundamental representation \footnote{They are listed in Appendix A.}.
Then the Killing vectors of ${\bf CP}^3$ are given by
\be v^\mu_i=J^{\mu\nu}\p_\n \Omega_i,\ee
where $J$ is the K\"ahler form of $\cp$.
From the above, we can get that
\be j^0_i=-\frac1{2\pi\alpha^\prime}\p_0 X_\mu J^{\mu\nu}\p_\nu \Omega_i.\ee
Now we choose a explicit system of coordinates for
$\cp$.
We begin with the following parameterization of $S^7$:
\bea
Z_1&=&\cos\xi\cos\frac{\theta_1}2\exp[i(y+\frac{\psi+\varphi_1}2)],\\
Z_2&=&\cos\xi\sin\frac{\theta_1}2\exp[i(y+\frac{\psi-\varphi_1}2)],\\
Z_3&=&\sin\xi\cos\frac{\theta_2}2\exp[i(y+\frac{-\psi+\varphi_2}2)],\\
Z_4&=&\sin\xi\sin\frac{\theta_2}2\exp[i(y+\frac{-\psi-\varphi_2}2)],\eea
where $0\le\xi<\frac\pi2, 0\le y<2\pi, -2\pi<\psi<2\pi, 0\le\theta_i\le\pi,
0\le\varphi_i<2\pi$.
Now the
induced metric on $S^7$ can be written as a $U(1)$ fiber over
$CP^3$: \be ds^2_{S^7}=ds^2_{CP^3}+(dy+A)^2. \ee Here $A$ is a one-form,
and the metric on ${\bf CP}^3$ is:
\bea ds^2_{{\bf CP}^3}&=&d\xi^2+\cos^2\xi\sin^2\xi
(d\psi+\frac12\cos\theta_1d\varphi_1-\frac12\cos\theta_2d\varphi_2)^2+\nn\\
&&\frac14\cos^2\xi(d\theta_1^2+\sin^2\theta_1
d\varphi_1^2)+\frac14\sin^2\xi(d\theta_2^2+\sin^2\theta_2 d\varphi_2^2). \eea
The K\"ahler form in this coordinate is \footnote{We choose the normalization of $J$ such that $J^\m_{\,\,\n} J^\n_{\,\,\rho}=-\delta^\m_\rho$.
}:
\bea J&=&-\frac12\sin\xi\cos\xi d\xi\wedge (2d\psi+\cos\theta_1d\varphi_1-\cos\theta_2d\varphi_2)-\frac14\cos^2\xi\sin\theta_1
d\theta_1\wedge d\varphi_1\nn\\
&& -\frac14 \sin^2\xi \sin\theta_2d\theta_2\wedge d\varphi_2.\eea
For the solutions consider in this paper, the $\tau$-dependent coordinates of ${\bf CP}^3$ are:
\be \psi=\omega_1\tau, \varphi_1=\omega_2\tau, \varphi_2=\omega_3\tau, \label{tau} \ee
then \bea j^0_i&=&-\frac{R^2}{2\pi\alpha^\prime}(\omega_1 J_{\psi}^{\,\,\,\,\nu}\p_\nu\Omega_i
+\omega_2 J_{\varphi_1}^{\,\,\,\,\nu}\p_\nu\Omega_i+\omega_3 J_{\varphi_2}^{\,\,\,\,\nu}\p_\nu\Omega_i)\\
&=&-\frac{\sqrt{\tilde\lambda}}{2\pi}(\omega_1 J_{\psi}^{\,\,\,\,\nu}\p_\nu\Omega_i
+\omega_2 J_{\varphi_1}^{\,\,\,\,\nu}\p_\nu\Omega_i+\omega_3 J_{\varphi_2}^{\,\,\,\,\nu}\p_\nu\Omega_i),\eea
where in the last line eq.~(\ref{lambda}) is used and we have defined that \be \tilde{\lambda}\equiv 32\pi^2\lambda. \ee
\section{Splitting of Folded String I}\label{folded1}
The first folded string solution we will study has the following configuration \cite{Chen:2008qq}:
\bea & & t=\kappa\tau, \rho=0, \theta_1=\theta_2=0,\\
& & \psi=\omega_1\tau, \varphi_1=\omega_2\tau, \varphi_2=\omega_3\tau,\eea
with $\xi=\xi(\sigma)$ being the only nontrivial function of $\sigma$.
When $\sigma$ increases from $0$ to $\pi$, $\xi$ increases from $-\xi_0$ to $\xi_0$ \footnote{Here when $\xi<0$, we change the coordinates $\xi, \varphi_2$ into $-\xi, \varphi_2+2\pi$.\label{footnote}}. While when $\sigma$ increases
from $\pi$ to $2\pi$, $\xi$ decreases from $\xi_0$ to $-\xi_0$. The Virasoro constraints gives:
\be \frac{\kappa^2}{4}=\xi^{\prime
2}+\frac{\sin^22\xi}{4}\tilde{\omega}^2. \ee
where $\tilde{\omega}=\omega_1+(\omega_2-\omega_3)/2$
(we assume $\tilde{\omega}$ is positive in the remaining part
of this section). When $\xi=\xi_0$, we have $\xi^\prime=0$, then we get $\kappa^2=\sin^22\xi_0\tilde\omega^2$.
From this we have \be \xi^{\prime2}=\frac{\tilde \omega^2}4 (\sin^22\xi_0-\sin^22\xi).\label{xi}\ee
Then we have \be
2\pi=4\int_0^{\xi_0}\frac{2d\xi}{\tilde{\omega}\sqrt{\sin^22\xi_0-\sin^22\xi}},\ee
which gives \footnote{We use $E(q), K(q)$ to denote the elliptic integrals
of first kind and second kind, respectively. And we use $E(q, x)$ and $F(q, x)$ to denote the corresponding incomplete elliptic integral. Our
convention is that $E(q, \pi/2)=E(q), F(q, \pi/2)=K(q)$.} \be \tilde{\omega}=\frac2{\pi}K(q) \ee with $q=\sin^22\xi_0$.
Now we compute the conserved charges of this string before splitting. The energy is related
to translation along $t$ direction insider $AdS_4$ and is give by:
\be
=\frac14\cosh^2\rho\sqrt{\tilde{\lambda}}\kappa.\ee
For this solution, the energy becomes:
\be E=\frac{\sqrt{\tilde{\lambda}}}{4}\kappa=\frac{\sqrt{\tilde{\lambda}}}{4}\tilde{\omega}
\sin2\xi_0=\frac{\sqrt{\tilde{\lambda}}}{2\pi}\sin2\xi_0K(q). \ee
The other Noether charges can be computed from the currents given in the previous section
\be Q_i=\int_0^{2\pi}j^0_i d\sigma.\ee We find that all
charges corresponding to non-Cartan generators vanish. This means that the solution is in highest weight representation of the isometry
group of the background \cite{Arutyunov:2003za}.
For Killing vectors corresponding to Cartan generators, we have \be v_3=-4\p_{\varphi_1}, v_8=\frac1{\sqrt{3}} (-4\p_\psi+4\p_{\varphi_2}),
v_{15}=-\frac1{\sqrt{6}} (4\p_\psi+8\p_{\varphi_2}),\ee
The corresponding Noether charges are:\bea Q_3&=&\frac{\sqrt{\tilde{\lambda}}}{\pi}(K(q)-E(q)),\\
Q_8&=&\sqrt{3}Q_3,\\ Q_{15}&=&0. \eea
Now assume that this folded string splits at $\sigma=a\in(0, \pi)$ and becomes two closed string. The first fragment
corresponds to $\sigma\in [0, a]\cup [2\pi-a, 2\pi]$ while the second fragment corresponds to $\sigma\in[a, 2\pi-a]$.
We denote $\xi(a)$ by $\xi_1$. From eq.~(\ref{xi}) The relation between $\xi_1$ and $a$ is:
\be \frac{\tilde\omega a}2=\int^{\xi_1}_{-\xi_0}\frac{d\xi}{\sqrt{\sin^22\xi_0-\sin^22\xi}},\ee
which gives:
\be a=\frac{1}{\tilde\omega}(K(q)+F(q, x_1)), \label{a}\ee
with \be x_1=\sin^{-1}\left(\frac{\sin2\xi_1}{\sin2\xi_0}\right)\ee
The energy of the first fragment is: \be E^I=\frac{a}\pi E=\frac{a\sqrt{\tilde{\lambda}}}{2\pi^2}\sin2\xi_0K(q).\label{ei}\ee
The other Noether charges for the first fragment is:
\be Q^I_{i}=2\int_0^a j^0_id\sigma. \ee
By some computations, we find that the Noether charges corresponding to the Cartan generators are:
\bea Q_3^I&=&\frac{\sqrt{\tilde{\lambda}}}{2\pi}(K(q)-E(q)+F(q, x_1)-E(q, x_1)),\label{q3i}\\
Q_8^I&=&\sqrt{3}Q_3^I,\\ Q_{15}^I&=&0. \eea
The non-vanishing Noether charges corresponding to the non-Cartan generators are:
\bea Q_4^I&=&\frac{\tilde\omega\sqrt{\tilde\lambda}}{2\pi}\cos\left(\frac12 (\omega_2-\omega_3+2\omega_1)\tau\right)\sqrt{q-\sin^22\xi_1} \\
Q_5^I&=&\frac{\tilde\omega\sqrt{\tilde\lambda}}{2\pi}\sin\left(\frac12 (\omega_2-\omega_3+2\omega_1)\tau\right)\sqrt{q-\sin^22\xi_1}. \eea
Without loss of generality, we can choose the time of splitting to be $\tau=0$, then using eq.~(\ref{tau}), we get:
\bea Q_4^I&=&\frac{\tilde\omega\sqrt{\tilde\lambda}}{2\pi} \sqrt{q-\sin^22\xi_1}, \label{q4i}\\
Q_5^I&=&0. \eea
For the charges of the second fragment we have $E^{II}=E-E^I$ and $Q^{II}_i=Q_i-Q^{I}_i, i=1, \cdots, 15$.
From eqs.~(\ref{a}, \ref{ei}, \ref{q3i}, \ref{q4i}) we can get relations among $E^I, Q^I_3$ and $Q^I_4$.
We notice that the non-vanishing charges belong to the set $Q^I_i, i=1, \cdots, 8$, with the corresponding
generators generate a $SU(3)$ subgroup of $SU(4)$.
\section{Splitting of folded string II}
The configuration of another folded string in \cite{Chen:2008qq} is:
\bea & & t=\kappa\tau, \xi=\pi/4, \theta_1(\sigma)=\pm\theta_2(\sigma),\\
& & \psi=\omega_1\tau, \varphi_1=\omega_2\tau, \varphi_2=\omega_3\tau,\eea
with $\omega_2=-\omega_3, \omega_1\neq 0$. Without losing generality, we assume that $\omega_1>0, \omega_2<0$, then the folded
string is around $\theta_1=0$.
When $\sigma$ increases from $0$ to $\pi$, $\theta_1$ increases from $-\theta_1^0$ to $\theta_1^{0}$ \footnote{Here, as in section~\ref{folded1}, when $\theta_1<0$, we change the coordinates $\theta_1, \theta_2, \psi, \varphi_1, \varphi_2$ into $-\theta_1, -\theta_2, \psi+\pi, \varphi_1-\pi, \varphi_2+\pi$. We thank the referee for suggestions on this footnote and footnote~\ref{footnote}.}. While when $\sigma$ increases
from $\pi$ to $2\pi$, $\theta_1$ decreases from $\theta_1^{0}$ to $-\theta_1^{0}$.
From the Virasoro constraint, we get:
\be \theta_1^{\prime2}=2\omega_1\omega_2(\cos\theta_1^0-\cos\theta_1), \ee
and \be \kappa^2=\omega_1^2+\omega_2^2+2\omega_1\omega_2\cos\theta_1^0.\ee
From \bea
2\pi=\int_0^{2\pi}d\sigma=\frac{2}{\sqrt{-2\omega_1\omega_2}}\int_0^{\theta_1(0)}\frac{d\theta_1}
{\sqrt{\cos\theta_1-\cos\theta_1^0}}, \eea
we get \be \sqrt{-\omega_1\omega_2}=\,\frac{2}{\pi}K(x),\ee
with $x=\sin^2\frac{\theta_1^0}{2}$.
The energy of this folded string is:
\be E=\frac14 \sqrt{\tilde\lambda}\kappa. \ee
As in last section, the conserved charges corresponding to the non-Cartan generators vanish before the splitting of the string.
As for the ones for the Cartan generators we have:
\bea Q_3&=&-4Q_{\varphi_1},\\
Q_8&=&\frac1{\sqrt{3}}(-4Q_{\psi}+4Q_{\varphi_2}),\\
Q_{15}&=&-\frac1{\sqrt{6}}(4Q_{\psi}+8Q_{\varphi_2}).\eea
Now for this solution, we have \cite{Chen:2008qq}:
\bea
Q_\psi
&=&-\frac{\sqrt{\tilde{\lambda}}}{2\pi\sqrt{-\omega_1\omega_2}}\left((\omega_1-\omega_2)K(x)+2\omega_2E(x)
\right),\\
Q_{\varphi_1}
&=&-\frac{\sqrt{\tilde{\lambda}}}{4\pi\sqrt{-\omega_1\omega_2}}\left((\omega_2-\omega_1)K(x)+2\omega_1E(x)\right),\\
Q_{\varphi_2}&=&-Q_{\varphi_1}. \eea
These charges satisfy the following relations:
\be\frac{\omega_1}{2}Q_{\psi}-\omega_2Q_{\varphi_1}=-\frac{\sqrt{\tilde{\lambda}}}{8}(\omega_1^2-\omega_2^2),
\ee
\bea\label{relation}
\left(\frac{E}{K(x)}\right)^2-\left(\frac{Q_\psi+2Q_{\varphi_1}}{E(x)}\right)^2&=&\frac{4\tilde\lambda x}{\pi^2},
\\
\left(\frac{Q_\psi-2Q_{\varphi_1}}{E(x)-K(x)}\right)^2-\left(\frac{Q_\psi+2Q_{\varphi_1}}{E(x)}\right)^2&=&\frac{4\tilde\lambda}{\pi^2}.
\eea
As already pointed out in \cite{Chen:2008qq}, these relations are similar to the ones in \cite{Beisert:2003ea}. We will see that such kind of similarity appears after splitting as well.
As in previous section, we assume that this folded string splits at $\sigma=a\in(0, \pi)$ and becomes two closed string. The first fragment
corresponds to $\sigma\in [0, a]\cup [2\pi-a, 2\pi]$ while the second fragment corresponds to $\sigma\in[a, 2\pi-a]$.
And we denote $\theta_1(a)$ by $\theta_1^1$. We can get the relation between $a$ and $\theta_1^1$ as:
\be a=\frac{K(x)+F(x, \tilde x)}{\sqrt{-\omega_1\omega_2}},\label{a2} \ee
with \be\tilde x=\sin^{-1}\frac{\sin\frac{\theta_1^1}2}{\sin\frac{\theta_1^0}2}.\ee
The energy of the first fragment is: \be E^I=\frac{a}{\pi}E.\ee As for the charges corresponding to the Cartan generators, we have:
\bea Q_3^I&=&-4Q_{\varphi_1}^I,\\
Q_8^I&=&\frac1{\sqrt{3}}(-4Q_{\psi}^I+4Q_{\varphi_2}^I),\\
Q_{15}^I&=&-\frac1{\sqrt{6}}(4Q_{\psi}^I+8Q_{\varphi_2}^I),\eea with
\bea
Q_\psi^I
&=&-\frac{\sqrt{\tilde{\lambda}}}{4\pi\sqrt{-\omega_1\omega_2}}\left((\omega_1-\omega_2)(K(x)+F(x, \tilde x))\right.\nn\\
&&\left. +2\omega_2(E(x)+E(x, \tilde x)
\right),\\
Q_{\varphi_1}^I
&=&-\frac{\sqrt{\tilde{\lambda}}}{8\pi\sqrt{-\omega_1\omega_2}}\left((\omega_2-\omega_1)(K(x)+F(x, \tilde x))\right. \nn\\
&& \left.+2\omega_1(E(x)+E(x, \tilde x))\right),\\
Q_{\varphi_2}^I&=&-Q_{\varphi_1}^I. \eea
The nonzero charges for non-Cartan generators are:
\bea Q_1^I&=&-\frac{2\sqrt{\tilde\lambda}}{\pi}\frac{\omega_1}{\sqrt{-\omega_1\omega_2}}\cos(\omega_2\tau)\sqrt{\sin^2\frac{\theta^0_1}{2}-\sin^2\frac{\theta^1_1}{2}},\\
Q_2^I&=&-\frac{2\sqrt{\tilde\lambda}}{\pi}\frac{\omega_1}{\sqrt{-\omega_1\omega_2}}\sin(\omega_2\tau)\sqrt{\sin^2\frac{\theta^0_1}{2}-\sin^2\frac{\theta^1_1}{2}},\\
Q_6^I&=&\frac{2\sqrt{\tilde\lambda}}{\pi}\frac{\omega_2}{\sqrt{-\omega_1\omega_2}}\cos\left(\frac12 (\omega_2+\omega_3-2\omega_1)\tau\right)\sqrt{\sin^2\frac{\theta^0_1}{2}-\sin^2\frac{\theta^1_1}{2}},\\
Q_7^I&=&-\frac{2\sqrt{\tilde\lambda}}{\pi}\frac{\omega_2}{\sqrt{-\omega_1\omega_2}}\sin\left(\frac12 (\omega_2+\omega_3-2\omega_1)\tau\right)\sqrt{\sin^2\frac{\theta^0_1}{2}-\sin^2\frac{\theta^1_1}{2}},\\
Q_9^I&=&\pm\frac{2\sqrt{\tilde\lambda}}{\pi} \frac{\omega_2}{\sqrt{-\omega_1\omega_2}}\cos\left(\frac12 (\omega_2+\omega_3+2\omega_1)\tau\right)\sqrt{\sin^2\frac{\theta^0_1}{2}-\sin^2\frac{\theta^1_1}{2}},\\
Q_{10}^I&=&\pm \frac{2\sqrt{\tilde\lambda}}{\pi}\frac{\omega_2}{\sqrt{-\omega_1\omega_2}}\sin\left(\frac12 (\omega_2+\omega_3+2\omega_1)\tau\right)\sqrt{\sin^2\frac{\theta^0_1}{2}-\sin^2\frac{\theta^1_1}{2}},
\eea
the signs in the last two equations correlate with the sign in $\theta_1=\pm \theta_2$.
If we set the time of the splitting to be $\tau=0$, we get:
\bea Q_1^I&=&-\frac{2\sqrt{\tilde\lambda}}{\pi}\frac{\omega_1}{\sqrt{-\omega_1\omega_2}}\sqrt{\sin^2\frac{\theta^0_1}{2}-\sin^2\frac{\theta^1_1}{2}},\\
Q_2^I&=&0,\\
Q_6^I&=&\frac{2\sqrt{\tilde\lambda}}{\pi} \frac{\omega_2}{\sqrt{-\omega_1\omega_2}}\sqrt{\sin^2\frac{\theta^0_1}{2}-\sin^2\frac{\theta^1_1}{2}},\\
Q_7^I&=&0,\\
Q_9^I&=&\pm Q_6^I,\\
Q_{10}^I&=&0.
\eea
For the charges of the second fragment we still have $E^{II}=E-E^I$ and $Q^{II}_i=Q_i-Q^{I}_i, i=1, \cdots, 15$.
Now we notice that, interestingly, the results in eq.~(\ref{a2}) and the results for $E^I, Q_\psi^I, Q_{\varphi_1}^I, Q^I_1, Q^I_6$
have the same forms as the results for conversed charges in \cite{Peeters:2004pt} after suitably identifying some quantities here with some linear combinations of quantities there. Because of this relation, one can use the discussions in \cite{Peeters:2004pt} to get similar relations among these conserved charges before and after the splitting for the current case.
\section{Conclusions}
We study the splitting of one closed folded string in $AdS_4\times \cp$ into two for two cases at the classical level. We pay main attention to
the conserved charges after the splitting. In each cases, we find interesting pattern among these
conserved charges. For the first case, we find that the generators of the isometry group $SU(4)$ which lead to the non-vanishing Noether charges are in an $SU(3)$
subgroup of $SU(4)$. For the second case, the charges have the same forms as the ones in \cite{Peeters:2004pt} for splitting of strings in $AdS_5\times S^5$. It will be quite interesting to study reasons behind these patterns from both string theory and gauge theory sides.
The study here can be generalized into folded strings rotating in $AdS_4$ and $\cp$ at the same time following \cite{Murchikova:2011ea}.
It will be also interesting to study the higher conserved charges from the integrable structure of the strings,
using the methods in \cite{Eichenherr:1979mx, Scheler:1980yv, Chou:1980ym, Chou:1980zy}. We hope to address these issues in the future.
\section*{Acknowledgments}
JW would like to thank Bin~Chen and San-Min~Ke for discussions and encouragement.
The work of JW is partly supported by NSFC under Grant No. 11105154 by Youth Innovation Promotion Association, CAS.
|
\section{Introduction}
Slowly pulsating B (SPB) stars are a pulsator of low frequency modes excited by
the $\kappa$ mechanism associated with the iron opacity bump at $T\sim 2\times10^5$K
in the interior (e.g., Dziembowski, Moskalik, Pamyatnykh 1993; Gautschy \& Saio 1993).
The observed amplitudes of the oscillations range from $\sim 0.1$ to $\sim1$ mmag (e.g., Huat et al 2009; Diago et al 2009;
Neiner et al 2009; Cameron et al 2008; Balona et al 2011).
Although SPB stars are not necessarily a rapid rotator,
the effects of rotation on the low frequency modes can be significant, particularly for
$|2\Omega/\omega|>\kern-1.2em\lower1.1ex\hbox{$\sim$} 1$, where $\Omega$ denotes the rotation frequency and $\omega$
is the oscillation frequency observed in the co-rotating frame of the stars.
Here, we are interested in theoretically determining the amplitudes of low frequency modes in SPB stars,
taking account of the effects of rotation on the modes.
Fully hydrodynamical calculation of radial pulsation has a long history, but
that of non-radial pulsation is not always feasible to carry out,
since it usually requires a huge amount of numerical
resources, particularly for low frequency modes in a rotating star.
Instead of hydrodynamical calculation,
we may apply weakly non-linear theory of
oscillation to non-radial pulsation, expecting the amplitudes of oscillation modes are
limited by weak non-linear coupling between them.
As the simplest case,
we may consider non-linear coupling between three modes, whose
amplitudes are expected to reach an equilibrium state as a result of
parametric instability between one unstable mode
and two stable modes (e.g., Dziembowski 1982; Kumar \& Goldreich 1989;
Wu \& Goldreich 2001; see also Craik 1985).
For the weakly non-linear theory of oscillation, it is essential to calculate the coupling coefficient
between oscillation modes as well as their excitation and damping rates.
To calculate the non-linear coupling coefficient between three oscillation modes,
we employ the formulation by Schenk et al (2002), who extended the theory to the case of rotating stars.
To compute oscillation modes in a rotating star,
we use the method of calculation by Townsend (2005), who applied the traditional approximation
to calculate both adiabatic and non-adiabatic oscillations of the star.
The traditional approximation is quite helpful to largely reduce the computing time spent for calculating a large number of coupling coefficients
between various non-radial modes in a rotating star.
We briefly describe the method of solution in \S 2, and \S 3 is for numerical results,
and \S 4 is for conclusions.
\section{method of solution}
\subsection{Linear Oscillation Equation}
Adiabatic oscillation in a uniformly rotating star may be governed by the linear differential equation:
\begin{equation}
-\omega^2\pmb{\xi}+i\omega\pmbmt{B}\left(\pmb{\xi}\right)+\pmbmt{C}\left(\pmb{\xi}\right)=0,
\end{equation}
where $\omega$ is the oscillation frequency in the co-rotating frame, $\pmb{\xi}$ is the displacement vector,
$
\pmbmt{B}\left({\pmb{\xi}}\right)=2\pmb{\Omega}\times{\pmb{\xi}}
$
with $\pmb{\Omega}$ being the angular velocity vector of rotation, and $\pmbmt{C}$ is the differential operation on $\pmb{\xi}$ and its expression as well as its
derivation may be found, for example, in Schenk et al (2002).
Assuming the time dependence of the perturbations is given by the factor $e^{{\rm i}\omega t}$, we may write the displacement vector as
\begin{equation}
\pmb{\xi}(\pmb{x},t)=e^{{\rm i}\omega t}\pmb{\xi}(\pmb{x})=e^{{\rm i}\omega t}\left(\xi^r\pmb{e}_r+\xi^\theta\pmb{e}_\theta+\xi^\phi\pmb{e}_\phi\right),
\end{equation}
where $\pmb{e}_r$, $\pmb{e}_\theta$, and $\pmb{e}_\phi$ are the orthonormal base vectors in spherical polar coordinates $(r,\theta,\phi)$.
We assume that the equilibrium of the rotating star is axisymmetric about the rotation axis,
and that the dependence of the oscillation modes on the azimuthal angle $\phi$ is given by
the factor $e^{{\rm i} m\phi}$ with $m$ being an integer representing the azimuthal wave number.
The oscillation frequency $\omega$ may be given as $\omega=\sigma+m\Omega$ with $\sigma$ being the oscillation
frequency observed in an inertial frame.
Because of the dependence given by $e^{{\rm i}(m\phi+\omega t)}$, the oscillation mode with
$m\omega<0$ ($m\omega>0$) is a prograde (retrograde) mode.
Although a separation of variables is in general impossible for oscillations in a rotating star,
it becomes possible under the traditional approximation, in which the term $-\sin\theta\Omega\pmb{e}_\theta$
in $\pmb{\Omega}=\cos\theta\Omega\pmb{e}_r-\sin\theta\Omega\pmb{e}_\theta$ is ignored (e.g., Lee \& Saio 1997).
In the traditional approximation, the components of $\pmb{\xi}(\pmb{x})$ are given by
\begin{equation}
\xi^r=\xi^r(r)\Theta_{km}(\mu;\nu)e^{{\rm i} m\phi},
\end{equation}
\begin{equation}
\xi^\theta={1\over r\omega^2}{p^\prime(r)\over \rho(r)}\Theta^\theta_{km}(\mu;\nu)e^{{\rm i} m\phi},
\end{equation}
\begin{equation}
\xi^\phi={1\over r\omega^2}{p^\prime(r)\over \rho(r)}{\rm i}\Theta^\phi_{km}(\mu;\nu)e^{{\rm i} m\phi},
\end{equation}
and $p^\prime$ and $\rho^\prime$, which respectively stand for the Eulerian perturbation of the pressure and the density,
are given by
\begin{equation}
p^\prime=p^\prime(r)\Theta_{km}(\mu;\nu)e^{{\rm i} m\phi}e^{{\rm i}\omega t},
\end{equation}
\begin{equation}
\rho^\prime=\rho^\prime(r)\Theta_{km}(\mu;\nu)e^{{\rm i} m\phi}e^{{\rm i}\omega t},
\end{equation}
where $\mu=\cos\theta$, $\nu=2\Omega/\omega$, $k$ is an integer used as a modal index, and
\begin{equation}
\Theta^\theta_{km}(\mu;\nu)={1\over (1-\nu^2\mu^2)\sqrt{1-\mu^2}}\left[-(1-\mu^2){d\over d\mu}+m\nu\mu\right]\Theta_{km}(\mu;\nu),
\end{equation}
\begin{equation}
\Theta^\phi_{km}(\mu;\nu)={1\over (1-\nu^2\mu^2)\sqrt{1-\mu^2}}\left[-\nu\mu(1-\mu^2){d\over d\mu}+m\right]\Theta_{km}(\mu;\nu).
\end{equation}
The function $\Theta_{km}(\mu;\nu)$, called the Hough function (e.g., Lindzen \& Holton 1968), is the eigenfunction, associated with the eigenvalue $\lambda_{km}$,
of Laplace tidal equation given by
\begin{equation}
{\cal L}_\nu\left[\Theta_{km}(\mu;\nu)\right]=-\lambda_{km}\Theta_{km}(\mu;\nu),
\end{equation}
where the definition of the differential operator $\cal L_\nu$ may be found, e.g., in Lee \& Saio (1997).
Note that the oscillation modes in a rotating star are separated into even modes and odd modes,
depending on symmetry of the eigenfunctions about the equator of the star.
For example, the angular dependence of $p^\prime(r,\theta,\phi,t)$ is symmetric (antisymmetric) about the equator
for even (odd) modes.
In this paper, we normalize the function $\Theta_{km}$ as
\begin{equation}
\int_0^\pi d\theta\int_0^{2\pi}d\phi \sin\theta \left|\tilde\Theta_{km}\right|^2=2\pi\int_{-1}^1d\mu\left|\Theta_{km}\right|^2=1,
\end{equation}
where
\begin{equation}
\tilde\Theta_{km}=\Theta_{km}e^{{\rm i} m\phi}.
\end{equation}
For a given azimuthal wavenumber $m$, $\lambda_{km}$
depends on the parameter $\nu$
and tends to $l_k(l_k+1)$ with $l_k=|m|+k$ as $\nu\rightarrow 0$ for $k\ge0$,
which corresponds to $\tilde\Theta_{km}\rightarrow Y_{l_k}^m$ as $\nu\rightarrow 0$.
The quantity $\sqrt{\lambda_{km}}$ represents a kind of surface wave number.
Except for the prograde sectoral modes ($k=0$), $\lambda_{km}$ increases
as $\nu$ increases.
The prograde sectoral modes (associated with $\lambda_{0m}$ for modes with $m\omega<0$)
are special modes in rapidly rotating stars whose
surface wavenumber is lower than the value at $\Omega=0$, and hardly changes with $\Omega$.
Note that $g$-modes belong to $\lambda_{km}$ with positive $k$.
On the other hand, $r$-modes, which are a retrograde mode, belong to $\lambda_{km}$ with
negative $k$ (Lee \& Saio 1997).
In the limit of $\Omega\rightarrow 0$, we have
$
\omega\rightarrow {2m\Omega/ l_k^\prime(l_k^\prime+1)},
$
where
$l_k^\prime=|m|+|k+1|$ for negative integer $k$.
Note that $\lambda_{km}\rightarrow 0$ as $\omega\rightarrow 2m\Omega/l_k^\prime(l_k^\prime+1)$.
In this paper, we employ the labeling $(l_k,m)$ with $l_k= |m| + k$ for $g$-modes
and the labeling $(l_k^\prime,m)$ with $l_k^\prime=|m|+|k+1|$ for $r$-modes where $k$ is non-negative integer for the former and negative integer for the latter, extending
the familiar notation for non-radial pulsations of a non-rotating star.
Note that in our convention we have ${\rm mod}(l_k-|m|,2)=0$ and ${\rm mod}(l^\prime_k-|m|,2)=1$
for even modes, while ${\rm mod}(l_k-|m|,2)=1$ and ${\rm mod}(l^\prime_k-|m|,2)=0$ for odd modes.
\subsection{Weakly Nonlinear Oscillation Equation and Parametric Instability}
Nonlinear evolution of small amplitude oscillation modes in a uniformly rotating star is governed by
the oscillation equation with nonlinear terms:
\begin{equation}
\ddot{\pmb{\xi}}+\pmbmt{B}\left(\dot{\pmb{\xi}}\right)+\pmbmt{C}\left(\pmb{\xi}\right)=
\pmb{a}^{(2)}\left(\pmb{\xi},\pmb{\xi}\right),
\end{equation}
where $\dot{\pmb{\xi}}=d\pmb{\xi}/dt$ and $\ddot{\pmb{\xi}}=d^2\pmb{\xi}/dt^2$,
and $\pmb{a}^{(2)}\left(\pmb{\xi},\pmb{\xi}\right)$ represents a collection of nonlinear terms of second order in $\pmb{\xi}$, and the $i$th component of $\pmb{a}^{(2)}$ is given by (Schenk et al 2002)
\begin{equation}
a_i^{(2)}\left(\pmb{\xi},\pmb{\xi}\right)=-{\rho^{-1}}\nabla_j\left\{p\left[\left(\Gamma_1-1\right)\Pi_i^j+\Xi_i^j
+\Psi\delta_i^j\right]\right\}-({1/ 2})\xi^k\xi^l\nabla_k\nabla_l\nabla_i\Phi,
\end{equation}
where $\nabla_j$ denotes the covariant derivative with respect to the coordinate $x^j$,
\begin{equation}
\Pi_i^j=(\nabla_i\xi^j)\nabla\cdot\pmb{\xi},
\end{equation}
\begin{equation}
\Xi_i^j=(\nabla_i\xi^k)(\nabla_k\xi^j),
\end{equation}
\begin{equation}
\Psi=\left({1/ 2}\right)\Pi\left[\left(\Gamma_1-1\right)^2+{\partial\Gamma_1/\partial\ln\rho}\right]+\left({1/2}\right)\left(\Gamma_1-1\right)\Xi,
\end{equation}
$
\Pi=\delta^i_j\Pi^j_i=\left(\nabla\cdot\pmb{\xi}\right)^2,
$
$
\Xi=\delta^i_j\Xi^j_i=\left(\nabla_j\xi^k\right)\left(\nabla_k\xi^j\right),
$
$\delta^i_j$ is the Kronecker delta, $\Phi$ is the gravitational potential,
and the repeated indices imply
the summation over the indices from 1 to 3, and $\Gamma_1=\left(\partial\ln p/\partial\ln \rho\right)_{\rm ad}$.
Note that we have applied the Cowling approximation, neglecting the Eulerian perturbation of
the gravitational potential.
Following Schenk et al (2002), we use eigenvalues $\omega$ and eigenfunctions $\pmb{\xi}$ of the linear oscillation equation (1) to
expand the displacement vector $\pmb{\xi}(\pmb{x},t)$ and its time derivative $\dot{\pmb{\xi}}(\pmb{x},t)$ in the nonlinear equation (13):
\begin{equation}
\left[\matrix{\pmb{\xi}(\pmb{x},t)\cr\dot{\pmb{\xi}}(\pmb{x},t)\cr}\right]=\sum_Ac_A(t)\left[\matrix{\pmb{\xi}_A(\pmb{x})\cr {\rm i}\omega_A\pmb{\xi}_A(\pmb{x})\cr}\right],
\end{equation}
for which
\begin{equation}
\sum_A\left(\dot c_A-{\rm i}\omega_A c_A\right)\pmb{\xi}_A(\pmb{x})=0,
\end{equation}
where the subscript $A$ stands for a collection of numbers such as harmonic degree $l$, azimuthal order $m$, and
radial order $n$ used to identify a linear mode.
Note that if a mode with with $(m_A,\omega_A)$ satisfies the linear oscillation equation
the mode with $(-m_A,-\omega_A)$ also satisfies the same equation, and both modes are included in the expansion given above.
Substituting the expansion (18) into the governing equation (13), and
making a scaler product with $\pmb{\xi}_A^*$ and integrating over the volume of the star, we obtain
\begin{equation}
\dot c_A(t)-{\rm i}\omega_Ac_A(t)=-{{\rm i}}\left<\pmb{\xi}_A,\pmb{a}^{(2)}\left(\pmb{\xi},\pmb{\xi}\right)\right>/b_A,
\end{equation}
where
\begin{equation}
c_A(t)=\left<\pmb{\xi}_A,\omega_A\pmb{\xi}(t)-{\rm i}\dot{\pmb{\xi}}(t)-{\rm i}\pmbmt{B}\left(\pmb{\xi}(t)\right)\right>/b_A,
\end{equation}
\begin{equation}
b_A=-\left<\pmb{\xi}_A,{\rm i}\pmbmt{B}\left(\pmb{\xi}_A\right)\right>+2\omega_A\left<\pmb{\xi}_A,\pmb{\xi}_A\right>,
\end{equation}
and for $A\not= B$ we have used a modified type of orthogonality relation given by
\begin{equation}
\left<\pmb{\xi}_A,{\rm i}\pmbmt{B}\left(\pmb{\xi}_B\right)\right>-\left(\omega_A+\omega_B\right)\left<\pmb{\xi}_A,
\pmb{\xi}_B\right>=0,
\end{equation}
where
\begin{equation}
\left<\pmb{\xi}_A,\pmb{\xi}_B\right>=\int d^3\pmb{x}\rho(\pmb{x})\pmb{\xi}^*_A(\pmb{x})\cdot\pmb{\xi}_B(\pmb{x}),
\end{equation}
and the asterisk ${}^*$ in the superscript implies the complex conjugation.
If we substitute the expansion
$
\pmb{\xi}\left(\pmb{x},t\right)=\sum_Bc_B^*(t)
\pmb{\xi}_B^*\left(\pmb{x}\right)
$
into $\pmb{a}^{(2)}\left(\pmb{\xi},\pmb{\xi}\right)$, we may obtain
\begin{equation}
\dot c_A(t)-{\rm i}\omega_Ac_A(t)=-{\rm i}\omega_A\sum_{B,C}\left({\kappa^*_{ABC}/\epsilon_A}\right)c_B^*(t)c_C^*(t),
\end{equation}
where
\begin{equation}
\kappa_{ABC}=\left<\pmb{\xi}_A^*,\pmb{a}^{(2)}\left(\pmb{\xi}_B,\pmb{\xi}_C\right)\right>,
\end{equation}
and $\epsilon_A\equiv \omega_A b_A$ corresponds to the total energy of oscillation in the co-rotating frame of the star (e.g., Lee \& Saio 1990).
If we introduce
$
{\hat c}_A=c_A\exp(-{\rm i} \omega_A t),
$
the equation (25) reduces to
\begin{equation}
\dot{\hat c}_A=-{\rm i} \omega_A
\sum_{B,C}\left({\kappa^*_{ABC}/ \epsilon_A}\right){\hat c}_B^*(t){\hat c}_C^*(t)e^{-{\rm i}\Delta\omega t},
\end{equation}
where
$
\Delta\omega=\omega_A+\omega_B+\omega_C.
$
If the driving rate of an unstable mode is smaller than the damping rates of stable modes nonlinearly coupled with the unstable one,
the growth of the unstable mode can be saturated by a transfer of energy to a small number of the damped modes.
Here, we consider parametric instability between three modes, one unstable mode and two stable modes and
we call the former the parent mode and the latter the daughter modes, and we expect the amplitude of the parent mode
is saturated by energy transfer to the daughter modes.
In the following, for convenience, we call mode $A$ the parent and modes $B$ and $C$ the daughter modes.
If we consider nonlinear mode coupling between three modes $A$, $B$, and $C$, we obtain
\begin{equation}
\dot{\hat c}_A=-\gamma_A\hat c_A-{\rm i} \omega_A
{\eta^*_{ABC}}{\hat c}_B^*(t){\hat c}_C^*(t)e^{-{\rm i}\Delta\omega t},
\end{equation}
and two similar equations for $\dot{\hat c}_B$ and $\dot{\hat c}_C$,
where we have included the effects of linear destabilization ($\gamma<0$) and stabilization ($\gamma>0$) of the modes, and
we have normalized the eigenfunctions $\pmb{\xi}_A$, $\pmb{\xi}_B$, and $\pmb{\xi}_C$ such that
$\epsilon_A=\epsilon_B=\epsilon_C=GM^2/R$, which leads to
$
{\kappa_{ABC}/\epsilon_A}={\kappa_{BCA}/\epsilon_B}={\kappa_{CAB}/\epsilon_C}\equiv{\eta_{ABC}/ 2}.
$
Parametric instability may occur when the amplitude of the parent mode, $|c_A|$, exceeds
the critical amplitude given by (e.g., Dziembowski 1982; Arras et al 2003)
\begin{equation}
|c_{A:c}|^2={1\over |\eta_{ABC}|^2Q_BQ_C}\left[1+\left({\Delta\omega\over\gamma_B+\gamma_C}\right)^2\right],
\end{equation}
where $Q_j=-\omega_j/\gamma_j$.
The equilibrium amplitude of the parent mode is then given by
\begin{equation}
|c_{A:e}|^2={1\over|\eta_{ABC}|^2Q_BQ_C}\left[1+\left({\Delta \omega\over\Delta\gamma}\right)^2\right],
\end{equation}
and those of the daughter modes are by
\begin{equation}
\left|{c_{B:e}}\right|^2=\left|c_{A:e}\right|^2{Q_B/ Q_A}, \quad
\left|{c_{C:e}}\right|^2=\left|c_{A:e}\right|^2{Q_C/ Q_A},
\end{equation}
where $\Delta\gamma=\gamma_A+\gamma_B+\gamma_C$.
Here, we have assumed $Q_BQ_C>0$, $Q_CQ_A>0$, and $Q_AQ_B>0$, which is equivalent to the relation given by
$Q_A>0,~Q_B>0,~Q_C>0$ or by $Q_A<0,~Q_B<0,~Q_C<0$, that is, the signs of
$\omega_B$ and $\omega_C$ are the same to each other but are different from that of $\omega_A$, because
the parent mode $A$ is assumed unstable ($\gamma_A<0$) and the daughter modes $B$ and $C$
stable ($\gamma_B>0$ and $\gamma_C>0$).
Since $\omega_B$ and $\omega_C$ have the same sign, to obtain a resonant coupling satisfying
$\Delta\omega\sim0$, we have
$|\omega_B|<\kern-1.2em\lower1.1ex\hbox{$\sim$} ~|\omega_A|$ and $|\omega_C|<\kern-1.2em\lower1.1ex\hbox{$\sim$} ~|\omega_A|$.
We use the condition $\Delta\gamma>0$ as the criteria for effectively stable equilibrium state
of three mode coupling (e.g., Wu \& Goldreich 2001; Arras et al 2003).
One of the selection rules giving non-zero coupling coefficient $\eta_{ABC}\not=0$ is
\begin{equation}
m_A+m_B+m_C=0,
\end{equation}
and another selection rule may be simply stated that
the coupling coefficient $\eta_{ABC}$ is non-zero only when the mode triad is composed of
three even modes or of one even mode and two odd modes (e.g., Schenk et al 2002).
For $m_A<0$, for example, we have two cases because of the selection rule (32), that is, both $m_B$ and $m_C$ are positive or one of $m_B$ and $m_C$ is negative so that $m_Bm_C<0$.
In the former case, if the parent mode is a prograde (retrograde) mode, the two daughter modes are
prograde (retrograde) modes.
In the latter case, however, if the parent mode is a prograde mode having $\omega_A>0$,
the daughter mode with $m<0$ is a retrograde mode since $\omega_B<0$ and $\omega_C<0$.
On the other hand, if the parent mode is a retrograde mode having $\omega_A<0$,
the daughter mode with $m<0$ is a prograde mode since $\omega_B>0$ and $\omega_C>0$.
\section{Numerical Results}
As a background model for mode calculation, we use a $4M_\odot$ main sequence model
computed by a standard stellar evolution code,
where we have used OPAL opacity (Iglesias \& Rogers 1996).
The physical parameters of the model are $\log T_{\rm eff}=4.142$, $\log (L/L_\odot)=2.470$, $R/R_\odot=2.980$, and $X_c=0.4602$,
and the initial abundance is given by $X=0.7$ and $Z=0.02$, where $T_{\rm eff}$, $L$, $R$, and $X_c$ respectively denote
the effective temperature, the surface luminosity, the radius of the model, and
the hydrogen mass fraction at the center.
Since we use the main sequence model slightly evolved from ZAMS, the model have a thin $\mu$-gradient
zone above the convective core where $\mu$ denotes the mean molecular weight.
We employ the method of calculation given by Townsend (2005) to calculate in the traditional approximation
both adiabatic and non-adiabatic modes of a uniformly rotating star, where
no effects of rotational deformation are considered.
We also employ the Cowling approximation, which is good enough
for high radial order $g$-modes of low degree $l$.
For this model, as well as pulsationally stable low frequency modes,
we obtain many unstable $g$-modes and $r$-modes, which are excited by the $\kappa$ mechanism
associated with the iron opacity bump located at $T\sim 2\times10^5$K.
It is our main concern here how non-linear three mode coupling determines the amplitudes
of the low frequency modes in a rotating B-type star.
We use the eigenfrequencies and eigenfunctions of adiabatic modes to compute
the nonlinear coupling coefficient $\eta_{ABC}$.
The excitation and damping rates
$\gamma$ are given by the imaginary part of
the complex eigenfrequency, $\omega_{\rm I}={\rm Im}(\omega)$, which is obtained by non-adiabatic mode calculation.
In this paper, to prepare a set of daughter modes used to calculate the coupling coefficient $\eta_{ABC}$ for a given low $|m|$ parent mode,
we consider low frequency modes of $|m|$ ranging from $|m|=0$ to 5 and of
$l$ in the limited range of $|m|\le l\le |m|+1$
for $|m|\not=0$ and $l=1$ and 2 for $m=0$, that is, we compute
even and odd $g$-modes with $(l,m)=(|m|,m)$ and $(|m|+1,m)$, and
odd and even $r$-modes with $(l^\prime,m)=(|m|,m)$ and $(|m|+1,m)$
for $|m|=1$ to 5, and even and odd $g$-modes with $(l,m)=(2,0)$ and $(1,0)$ for $m=0$, in the frequency range
$0.05\le\bar\omega\le 2$, where $\bar\omega\equiv\omega/\sqrt{GM/R^3}$, and
$M$ is the mass of the star and $G$ is the gravitational constant.
Note that the $r$-modes are in the frequency range of $|\omega|<2|m|\Omega/l^\prime(l^\prime+1)$.
For a given combination of $(m_A,m_B,m_C)$ for mode triad,
the even/odd mode combinations giving non-zero $\eta_{ABC}$ are
$(A_e,B_e,C_e)$, $(A_e,B_o,C_o)$, $(A_o,B_e,C_o)$, and
$(A_o,B_o,C_e)$, where the subscripts $e$ and $o$ stand for even and odd modes, respectively.
For a given unstable parent mode, there are numerous combinations of a pair of stable daughter modes,
satisfying the selection rules for non-zero coupling coefficient $\eta_{ABC}$.
For each of the combinations we
compute the critical amplitude $c_{A:c}$ and equilibrium amplitude $c_{A:e}$ for the parent mode using equations (29) and (30),
where the excitation and damping rates are obtained by non-adiabatic mode calculation.
Among the critical amplitudes calculated for various combinations of a pair of daughter modes for a given parent mode, we chose the smallest one,
considering that the parent mode reaches the smallest critical amplitude first to be in an equilibrium state.
Since we search for the combination giving the smallest $c_{A:c}$ from a limited set of daughter modes for a parent mode,
the critical amplitude $c_{A:c}$ thus determined should be regarded as an upper limit for the parent mode.
As indicated by equation (29),
the critical amplitude $|c_{A:c}|$ becomes smaller for larger values of $|\eta_{ABC}|$ and $\sqrt{Q_BQ_C}$
and has a dip at $\Delta \omega=0$ because of the factor $1+[\Delta\omega/(\gamma_B+\gamma_C)]^2$ in equation (29).
Since normalized damping rates $\bar\gamma\equiv\gamma/\sqrt{GM/R^3}$ of high radial order $g$-modes of the model
are of order $10^{-5}\sim 10^{-4}$, which are much larger than the growth rates ranging from
$|\bar\gamma|\sim10^{-8}$ to $\sim 10^{-5}$ for the $g$-modes,
the dip at $\Delta\omega=0$ will be very sharp for triads of $g$-modes having frequencies $|\bar\omega|>\kern-1.2em\lower1.1ex\hbox{$\sim$} 0.05$.
In this paper,
we search for the smallest critical amplitude $c_{A:c}$ for a parent mode
among mode triads satisfying $\Delta\omega\sim 0$, that is, nearly in frequency resonance.
This procedure is helpful to
reduce the number of combinations of daughter modes we have to try in order to find the smallest $|c_{A:c}|$.
Because of the assumption $\Delta\omega\sim0$, the factor
$1+[\Delta\omega/(\gamma_B+\gamma_C)]^2$ in equation (29) takes values between $\sim 1$ and $\sim10$ for most of the triads examined,
and $|c_{A:c}|$ is practically dependent on the two quantities $|\eta_{ABC}|$ and $\sqrt{Q_BQ_C}$, that is,
the smallest critical amplitude is likely to take place when either $|\eta_{ABC}|$ or $\sqrt{Q_BQ_C}$ is very large or when both of them are large.
Since $\eta_{ABC}$, which is calculated by using the eigenfunctions of adiabatic modes,
is the sum of products of the three eigenfunctions of
modes $A$, $B$, and $C$, and since the eigenfunction of $g$-modes has an asymptotic form proportional to
$\cos\left(\int^r k_{r}dr\right)$ with $k_r$ being the wave number in the radial direction and
$\int k_{r}dr=n_{g}\pi$ with $n_g$ being the number of $g$-nodes of the eigenfunction
when integrated over the entire propagation zone of the $g$-modes,
the sum of terms proportional to $I_{ABC}=\int_0^R dr f(r)\cos\left(\int^r k_{rA}dr\right)\cos\left(\int^r k_{rB}dr\right)\cos\left(\int^r k_{rC}dr\right)$ with $f(r)$ being a spatially slowly varying weighting function may be maximized when
$n_{gA}\sim|n_{gB}-n_{gC}|$, which is
numerically confirmed (see also Wu \& Goldreich 2001).
The quantity $\sqrt{Q_BQ_C}$ can be large when
a low radial order $g$-mode or $r$-mode with a very small damping rate is in the mode triad, and in this case
we do not necessarily have the property $n_{gA}\sim|n_{gB}-n_{gC}|$.
\subsection{Slow Rotation}
\begin{figure}
\resizebox{0.5\columnwidth}{!}{
\includegraphics{f1a.epsi}}
\resizebox{0.5\columnwidth}{!}{
\includegraphics{f1b.epsi}}
\resizebox{0.5\columnwidth}{!}{
\includegraphics{f1c.epsi}}
\resizebox{0.5\columnwidth}{!}{
\includegraphics{f1d.epsi}}
\caption{Mode triad quantities $|\eta_{ABC}|$, $\sqrt{Q_BQ_C}$, $|c_{A:e}|$, and $|\delta L_{\rm rad}/L_{\rm rad}|$ are plotted
versus the frequency $\omega_A/\sqrt{GM/R^3}$ of the parent mode for a $4M_\odot$ main sequence
model
for $\Omega/\sqrt{GM/R^3}=0.01$, where
$\delta L_{\rm rad}$ denotes the Lagrangian variation of
the surface luminosity $L_{\rm rad}$ caused by the parent mode, $Q_BQ_C =\omega_B\omega_C/\gamma_B\gamma_C$, and
the red, blue, cyan dots and black open circle stand for the parent modes of
$(l_A,m_A)=(1,-1)$, $(2,-1)$, $(2,-2)$, and $(3,-2)$, respectively.
Note that for a given parent mode a combination of a pair of daughter modes has been chosen so that the critical amplitude $|c_{A:c}|$ of the parent mode
be smallest, and that the thus determined combination of daughter modes makes the mode triad associated with the parent mode.
Here, the parent modes with positive (negative) $\omega_A$ are prograde (retrograde) modes.
}
\end{figure}
\begin{figure}
\resizebox{0.33\columnwidth}{!}{
\includegraphics{f2a.epsi}}
\resizebox{0.33\columnwidth}{!}{
\includegraphics{f2b.epsi}}
\resizebox{0.33\columnwidth}{!}{
\includegraphics{f2c.epsi}}
\caption{Coupling coefficient $\eta_{ABC}(r)\equiv\kappa_{ABC}(r)/\epsilon_A$ and eigenfunctions $xz_1$ and $xz_2/(c_1\bar\omega^2)$
versus $x\equiv r/R$ for a mode triad
composed of a parent mode of $l_A=-m_A=1$ and daughter modes of $l_B=-m_B=4$ and $l_C=m_C=5$ for $\bar\Omega=0.01$,
where the eigenfunctions $z_1$ and $z_2$ are normalized so that
the oscillation energy $\epsilon_A$ in the co-rotating frame be equal to $GM^2/R$ with $G$, $M$ and $R$ being the gravitational constant,
and the mass and radius of the star, respectively.
See the Appendix B for the definition of the quantities $c_1$, $z_1$ and $z_2$.
Here, $\bar\omega_A=0.1680$,
$\bar\omega_B=-0.07082$, and $\bar\omega_C=-0.09769$.
In panels (b) and (c), the black, red, and blue lines indicate the mode A, B, and C, respectively.}
\end{figure}
In Figure 1, the quantities $|\eta_{ABC}|$, $\sqrt{Q_BQ_C}$, $|c_{A:e}|$,
and $|\delta L_{\rm rad}/L_{\rm rad}|$ computed for the mode triad composed of low frequency $g$-modes and
corresponding to the smallest $|c_{A:e}|$ are plotted versus the frequency $\omega_A/\sqrt{GM/R^3}$ of the parent mode
for the case of $\bar\Omega\equiv\Omega/\sqrt{GM/R^3}=0.01$, where
$\delta L_{\rm rad}$ is the Lagrange variation of the radiative luminosity at the surface caused by the parent mode, and
we have considered only $g$-modes for the mode triads, ignoring $r$-modes because of the slow rotation.
Note that we have $|c_{A:e}|\sim|c_{A:c}|$ for the mode triads plotted in the figure.
Since the ratio $|2\Omega/\omega|$ is much smaller than unity for unstable $g$-modes (parent modes) for the slow rotation,
the distributions of the points in the figure are almost
symmetric between prograde and retrograde modes, although there exists slight deviation from
the symmetry.
As suggested by the panels (a) to (c), $|c_{A:c}|$ for the
parent modes can be small either when $|\eta_{ABC}|$ or $\sqrt{Q_BQ_C}$ is very large or when both of them
are large.
For the mode triads in the figure,
$|\eta_{ABC}|$ ($|c_{A:e}|$) tends to increase (decrease) with decreasing $|\bar\omega|$, but
the dependence of $|\delta L_{\rm rad}/L_{\rm rad}|$ on $\bar\omega$
is not simple because the normalizing amplitudes, defined to make the energy of oscillation equal to $GM^2/R$, increase as $|\bar\omega|$ decreases.
The fractional luminosity amplitudes $|\delta L_{\rm rad}/L_{\rm rad}|$ at the surface
take values ranging from $\sim 10^{-6}$ to $\sim 10^{-4}$, and those of the parent $l=|m|=1$ $g$-modes tend to be smaller than the others.
As shown by Figure 1, the coupling coefficient $|\eta_{ABC}|$ can be as large as $\sim 10^6$ for the very low frequency parent modes for $\bar\Omega=0.01$.
In Figure 2, the coupling coefficient $\eta_{ABC}(r)=\kappa_{ABC}(r)/\epsilon_A$ (see Appendix B) and the eigenfunctions $x z_1$ and $x z_2/(c_1\bar\omega^2)$
are plotted
versus $x\equiv r/R$ for a mode triad composed of a parent mode of $l_A=-m_A=1$ and daughter modes of $l_B=-m_B=4$ and $l_C=m_C=5$
for $\bar\Omega=0.01$, where the eigenfunctions are normalized so that the oscillation energy be equal to $GM^2/R$.
The panel (a) shows that the mean magnitude of $|\eta_{ABC}(r)|$ increases rapidly with increasing $r$ in the $\mu$-gradient region above the convective core,
and the panel (c) indicates that this rapid increase is caused by the amplitude trapping of the eigenfunctions into the $\mu$-gradient zone.
Note that $\eta_{ABC}(r)$ stays almost constant outside the $\mu$-gradient region as shown by panel (a).
For this mode triad, the modes $B$ and $C$ are very high radial order $g$-modes, for which
we have $n_{gA}\sim |n_{gB}-n_{gC}|$.
The panel (a) suggests that for the mode triad the non-linear coupling between the modes preferentially occurs in the thin $\mu$-gradient zone
having a high Brunt-V\"ais\"al\"a frequency.
Note that for the ZAMS model with no $\mu$-gradient zone, there occurs no rapid increase in the mean magnitude of $|\eta_{ABC}(r)|$ immediately above the
convective core.
\begin{figure}
\resizebox{0.5\columnwidth}{!}{
\includegraphics{f3a.epsi}}
\resizebox{0.5\columnwidth}{!}{
\includegraphics{f3b.epsi}}
\resizebox{0.5\columnwidth}{!}{
\includegraphics{f3c.epsi}}
\resizebox{0.5\columnwidth}{!}{
\includegraphics{f3d.epsi}}
\caption{Same as Figure 1 but for the case of $\bar\Omega=0.2$. Here, we have included $r$-modes as a possible member in the mode triads
to calculate the coupling coefficient $\eta_{ABC}$.}
\end{figure}
\subsection{Rapid Rotation}
For weakly non-linear coupling of oscillations in a rapidly rotating star, $r$-modes may come into play as an important member
in mode triads for parametric instability.
For a rotation rate $\bar\Omega=0.2$, for example,
the maximum frequency $2|m|\bar\Omega/l^\prime(l^\prime+1)$ for even (odd) $r$-modes in the co-rotating frame
is 1/15 (0.2), 1/15 (2/15), 0.06 (0.1), 4/75 (0.08), 1/21 (1/15) for $|m|=1$ to $|m|=5$, and the
corresponding inertial frame frequency $|\bar\sigma|$ is
2/15 (0), 1/3 (4/15), 0.54 (0.5), 56/75 (0.72), and 20/21 (14/15), respectively.
In Figure 3, the mode triad quantities $|\eta_{ABC}|$, $\sqrt{Q_BQ_C}$, $|c_{A:e}|$,
and $|\delta L_{\rm rad}/L_{\rm rad}|$ are plotted versus the frequency $\bar\omega_A$ of the low $|m|$ parent modes
for $\bar\Omega=0.2$.
The distribution of the points of the parent modes of $l=|m|=1$, for example,
is not symmetric any more between prograde and retrograde modes, because
the frequency spectra of low frequency $g$-modes themselves largely deviate from the symmetry for $|2\Omega/\omega|>\kern-1.2em\lower1.1ex\hbox{$\sim$} 1$,
and $r$-modes appear only as a retrograde mode.
Since the damping rates $\bar\gamma$ of low radial order $r$-modes can be smaller than $10^{-10}$
and those of low radial order $g$-modes are of order $10^{-8}$ to $10^{-9}$, it is likely that
the mode triads giving
the smallest critical amplitude $|c_{A:c}|$ for the parent modes are those containing a low radial order $r$-mode or $g$-mode.
Figure 4 shows mode triad quantities $|\eta_{ABC}|$, $\sqrt{Q_BQ_C}$,
and $|\delta L_{\rm rad}/L_{\rm rad}|$ the same as those plotted in Figure 3, but here
the red, blue, and cyan dots respectively stand for the mode triads containing no $r$-modes, one $r$-mode,
and two $r$-modes.
The low frequency parent modes tend to be coupled to one $r$-mode or two.
We find cases where the parent modes are an unstable $r$-mode coupled with
a stable $r$-mode.
Since both $|\eta_{ABC}|$ and $\sqrt{Q_BQ_C}$ are large, the amplitude $|\delta L_{\rm rad}/L_{\rm rad}|$
becomes very small.
We also find cases in which the parent modes are a low frequency even $g$-mode coupled with
two stable odd $r$-modes, one low radial order and the other high radial order $r$-modes.
If no $r$-modes are in a mode triad, on the other hand, the amplitude $|\delta L_{\rm rad}/L_{\rm rad}|$ for
the parent mode is in general larger than those of the mode triads that contain one $r$-mode or two.
\begin{figure}
\resizebox{0.33\columnwidth}{!}{
\includegraphics{f4a.epsi}}
\resizebox{0.33\columnwidth}{!}{
\includegraphics{f4b.epsi}}
\resizebox{0.33\columnwidth}{!}{
\includegraphics{f4c.epsi}}
\caption{Mode triad quantities $|\eta_{ABC}|$, $\sqrt{Q_BQ_C}$, and $|\delta L_{\rm rad}/L_{\rm rad}|$ are plotted
versus the frequency $\omega_A/\sqrt{GM/R^3}$ of the parent mode for a $4M_\odot$ main sequence
model for the case of $\bar\Omega=0.2$, where
$\delta L_{\rm rad}$ denotes the Lagrangian variation of
the surface luminosity $L_{\rm rad}$ caused by the parent mode, $Q_BQ_C =\omega_B\omega_C/\gamma_B\gamma_C$, and
the red, blue, cyan dots indicate the mode triads containing
no $r$-modes, one $r$-mode, and two $r$-modes, respectively.
Here, the parent modes with positive (negative) $\omega_A$ are prograde (retrograde) modes.}
\end{figure}
In Figure 5, the quantities $\eta_{ABC}(r)$, $x z_1$, and $x z_2/(c_1\bar\omega^2)$ are plotted
versus $x=r/R$ for a mode triad composed of a parent mode of $l_A=-m_A=2$ and daughter modes of $l_B-1=m_B=1$ and $l_C-1=m_C=1$ for $\bar\Omega=0.2$,
where the eigenfunctions are normalized so that the oscillation energy be equal to $GM^2/R$.
Here, the daughter mode C is an $r$-mode, and the eigenfunction $z_1$ of the $r$-mode has an amplitude much smaller than those of
the $g$-modes A and B, although the amplitude $z_2/(c_1\bar\omega^2)$ of the $r$-mode is
comparable to the $g$-modes.
The figure suggests that the terms containing $z_2/(c_1\bar\omega^2)$ and/or $d[z_2/(c_1\bar\omega^2)]/dx$ make
dominating contributions to $\eta_{ABC}$ (see Appendix B).
We also note that the amplitudes of the modes in the triad are not strongly trapped in the $\mu$-gradient zone.
Although the coefficient $\eta_{ABC}(r)$ is spatially oscillatory
in the region $0.1<\kern-1.2em\lower1.1ex\hbox{$\sim$} x<\kern-1.2em\lower1.1ex\hbox{$\sim$} 0.3$, it takes almost a constant value in $x>\kern-1.2em\lower1.1ex\hbox{$\sim$} 0.5$.
\begin{figure}
\resizebox{0.33\columnwidth}{!}{
\includegraphics{f5a.epsi}}
\resizebox{0.33\columnwidth}{!}{
\includegraphics{f5b.epsi}}
\resizebox{0.33\columnwidth}{!}{
\includegraphics{f5c.epsi}}
\caption{Coupling coefficient $\eta_{ABC}(r)\equiv\kappa_{ABC}(r)/\epsilon_A$ and eigenfunctions $xz_1$ and $xz_2/(c_1\bar\omega^2)$
versus $x\equiv r/R$ for a mode triad
composed of a parent mode of $l_A=-m_A=2$ and daughter modes of $l_B-1=m_B=1$ and $l_C-1=m_C=1$ for $\bar\Omega=0.2$,
where the eigenfunctions $z_1$ and $z_2$ are normalized so that
the oscillation energy $\epsilon_A$ in the co-rotating frame be equal to $GM^2/R$ with $G$, $M$ and $R$ being the gravitational constant,
and the mass and radius of the star.
See the Appendix B for the definition of the quantities $c_1$, $z_1$ and $z_2$.
Here, $\bar\omega_A=-0.5332$,
$\bar\omega_B=0.3329$, and $\bar\omega_C=0.1999$, and the mode C is an $r$-mode.
In panels (b) and (c), the black, red, and blue lines indicate the mode A, B, and C, respectively.}
\end{figure}
\subsection{In an Inertial Frame}
\begin{figure}
\resizebox{0.5\columnwidth}{!}{
\includegraphics{f6a.epsi}}
\resizebox{0.5\columnwidth}{!}{
\includegraphics{f6b.epsi}}
\caption{Fractional amplitude $|\delta L_{\rm rad}/L_{\rm rad}|$ of the surface luminosity
versus the oscillation frequency $|\sigma|/\sqrt{GM/R^3}$ observed in an inertial frame for
$\bar\Omega=0.01$ in panel (a) and for $\bar\Omega=0.2$ in panel (b), where the filled (open)
symbols stand for the parent (daughter) modes, and the legend of the symbols for both panels are given in
the lower right corner in panel (a), and
only the modes of $1\le l\le3$ are plotted.
For the parent modes, the combination $(l,|m|)$ is given by $(1,1)$, $(2,1)$, $(2,2)$, and $(3,2)$, while
for the daughter modes the combination $(l,|m|)$ is given by $(1,0)$, $(1,1)$, $(2,0)$, $(2,1)$,$(2,2)$, $(3,2)$, and $(3,3)$.
}
\end{figure}
It may be useful to plot the fractional amplitude of both the parent and daughter modes
as a function of the oscillation frequency
observed in an inertial frame, where the daughter modes are regarded as being non-linearly excited
by the parametric instability.
Figure 6 shows $|\delta L_{\rm rad}/L_{\rm rad}|$
versus the inertial frame oscillation frequency $|\bar\sigma|$ for the case of
$\bar\Omega=0.01$ (panel a) and $\bar\Omega=0.2$ (panel b), where $\bar\sigma=\bar\omega-m\bar\Omega$,
and the filled (open) symbols stand for the parent (daughter) modes.
Here, only the daughter modes having $l\le 3$ are plotted in the figure.
If one of the daughter modes in a mode triad has a damping rate much smaller that the excitation rate of the parent mode
such that $|Q_B|\gg |Q_A|$ or $|Q_C|\gg |Q_A|$ (see equation (31)),
the amplitude $|\delta L_{\rm rad}/L_{\rm rad}|$ of the daughter mode can be comparable to or even larger than
that of the parent mode.
In fact, although the upper limit of $|\delta L_{\rm rad}/L_{\rm rad}|$
for the parent modes is of order $\sim 10^{-4}$,
there are many daughter modes whose amplitude is as large as $|\delta L_{\rm rad}/L_{\rm rad}|>\kern-1.2em\lower1.1ex\hbox{$\sim$} 10^{-3}$.
Since $|\bar\omega_B|+|\bar\omega_C|<\kern-1.2em\lower1.1ex\hbox{$\sim$} |\bar\omega_A|$, the daughter modes are likely to be in the low frequency domain of $|\bar\sigma|$
for the case of $\bar\Omega=0.01$ since the term $m\bar\Omega$ in $\bar\sigma=\bar\omega-m\bar\Omega$ is small compared to $\bar\omega$
and hence $\bar\sigma\sim\bar\omega$.
For the case of $\bar\Omega=0.2$, however, the term $m\bar\Omega$ can be comparable to $\bar\omega$ and even the daughter modes
can have the inertial frame oscillation frequencies comparable to those of the parent modes.
The range of the fractional amplitude $|\delta L_{\rm rad}/L_{\rm rad}|$ for $\bar\Omega=0.2$ is much wider than
that for $\bar\Omega=0.01$, although
the upper limits are almost the same.
It may be interesting to note that for the case of $\bar\Omega=0.2$,
low radial order odd $l^\prime=m=1$ $r$-modes, which are linearly stable but non-linearly excited,
have very low oscillation frequency $\bar\sigma\sim 0$ in the inertial frame, although the amplitudes
are not very high.
As indicated by the existence of vertical sequences of open symbols (daughter modes) in the panels (a) and (b),
there arise some cases in which a stable low radial order $r$-mode or $g$-mode, which has a very small damping rate $\bar\gamma$,
is shared by several parent modes, indicating that one stable daughter mode has different amplitudes depending on the mode triads
it belongs to.
This may suggest that the set of daughter modes we use for the computation of $\eta_{ABC}$ is not large enough, or
that the three mode non-linear coupling theory is
too simplified to be applied to the case where modes having an extremely small damping rate exist
in dense frequency spectra of oscillation modes.
\section{conclusions}
Using the weakly non-linear theory of oscillation,
we have estimated the amplitudes of low $m$ $g$-modes and $r$-modes in a slightly evolved $4M_\odot$ main sequence model,
assuming the mode amplitudes are limited by parametric instability between one unstable mode and two stable modes.
Here, the unstable low frequency modes are assumed destabilized by the $\kappa$-mechanism associated with the iron opacity bump,
and we have taken account of
the effects of rotation on low frequency modes in the traditional approximation.
For a given unstable mode (parent mode A), we compute three mode non-linear coupling coefficient $\eta_{ABC}$
for various combinations of two stable modes (daughter modes B and C), and we choose, among the numerous combinations, the one that gives the smallest critical amplitude.
The traditional approximation is employed in order to reduce the amount of computing time
necessary to find the optimal combination of daughter modes.
It is important to note that since we can use only a limited set of daughter modes the critical amplitude thus determined for a parent mode
should be regarded as an upper limit.
The critical amplitude essentially depends on $|\eta_{ABC}|$ and $\sqrt{Q_BQ_C}$ if we assume resonant mode coupling satisfying $\Delta\omega\sim0$,
and the smallest critical amplitude may take place when either $|\eta_{ABC}|$ or $\sqrt{Q_BQ_C}$ is very large or when both of them are large.
If the damping rate of a parametrically excited daughter mode in a mode triad is less than the growth rate of
the parent mode, the equilibrium amplitude of the daughter mode can be
larger than the parent mode.
It is therefore likely that parametrically destabilized daughter modes like low radial order $g$-modes and $r$-modes
are among the periodicities observed in a rapidly rotating B star.
The fractional amplitudes $|\delta L_{\rm rad}/L_{\rm rad}|$ of the parent and daughter modes can be of order
$\sim10^{-4}$ to $\sim 10^{-3}$ for the main sequence model, the magnitudes of which may be
consistent with those observed in B type variable stars (e.g., Huat et al 2009; Diago et al 2009;
Neiner et al 2009; Cameron et al 2008; Balona et al 2011).
We also find that the amplitudes $|\delta L_{\rm rad}/L_{\rm rad}|$ of the parent mode
tend to be large for high $l$ values,
although the visibility of the modes decreases with increasing $l$.
We find that $r$-modes significantly affects the amplitude determination of low frequency modes in a rapidly rotating star.
Since low radial order $r$-modes of the model
have damping rates $\bar\gamma$ as small as or even smaller than $\sim 10^{-10}$,
the mode triads giving the smallest critical amplitude for the low frequency parent modes are likely to have
a low radial order $r$-mode as a member.
When a mode triad has a low radial order $r$-mode, the equilibrium amplitude of the parent mode tends to be smaller
than those for the mode triads without $r$-modes.
We find some cases in which a low radial order $r$-mode is shared by several parent modes.
We think the degeneracy of the daughter mode in mode triads is a problem in the weak non-linear coupling theory we use,
since the equilibrium amplitude of the daughter mode depends on the parent modes it is coupled to.
This degeneracy might be removed if we use an enlarged set of daughter modes to determine the optimal combination, or
this degeneracy may suggest that the three mode coupling theory we use is too simplified to be applied to the problem we are considering, that is,
we have to consider higher order non-linear mode coupling to lift the degeneracy.
In spite of the problem,
the weakly non-linear theory could be useful when we try to compare theoretical mode calculations to observations.
Generally, the number of observationally detected low frequency modes for SPB stars
is much smaller than that of theoretically calculated linearly unstable modes
(e.g., Walker et al 2005, Saio et al 2007), and we may use
the weakly non-linear theory to decide which linearly unstable modes can have amplitudes
large enough to be detected observationally.
We may suggest that, if a low radial order $l^\prime=m=1$ $r$-mode, which may be linearly stable,
is parametrically excited, the $r$-mode can produce
very long period variations observed in an inertial frame.
We may attribute very slow pulsations detected
in SPBe stars (e.g., Walker et al 2005, Saio et al 2007) to low radial order $l^\prime=m=1$ $r$-modes, although the amplitudes would not be very high as indicated by Figure 6.
We have carried out an additional calculation to obtain the optimal critical amplitudes $|c_{A:c}|$
for the parent modes by extending the set of daughter modes from $l_{\rm max}=|m|+1=6$ to
$l_{\rm max}=8$ for $\bar\Omega=0.01$, and we obtained almost the same result for their amplitudes $|\delta L_{\rm rad}/L_{\rm rad}|$.
However, it is extremely time consuming to carry out similar calculations for $l_{\rm max}$ much larger than $l_{\rm max}\sim 10$, and
from the numerical results we currently have it would be fair to say
we are not able to correctly specify what the most likely degrees of the daughter modes are for the parent modes.
To construct a set of daughter modes extended for a very large $l_{\rm max}$,
asymptotic methods would be useful to represent the eigenfunctions and eigenfrequencies for high $l$ and high radial order $g$-modes and to
calculate the coupling coefficient $\eta_{ABC}$ (e.g., Dziembowski 1982), and we may use an asymptotic treatment
by Lee \& Saio (1989) for low frequency modes in uniformly rotating stars.
Extending our weakly non-linear analysis to large values of $l$,
we can also consider non-linear
couplings not only between a parent mode and many pairs of daughter modes having similar frequencies but also between a daughter mode and granddaughter
modes with frequencies still lower than that of the daughter mode (e.g., Kumar \& Goodman 1996).
The problems of these highly multiple mode couplings could be important for the amplitude determination for both the parent modes and daughter modes,
and we may have to include these mechanisms in our analysis to obtain definite answers for the oscillation amplitudes in the stars.
We have assumed uniform rotation in the present analysis.
It is, however, quite likely that differential rotation is a rule
in reality in a rotating star, and that
even a weak differential rotation would affect the frequency spectrum of the
low frequency modes.
We need to understand the property of low frequency modes in a differentially
rotating star as well as how a differential rotation law is established in a star.
The problem of differential rotation in a rotating star is quite difficult to find answer and is beyond the scope of this paper.
Since we used the traditional approximation to compute low frequency modes in a rotating star,
we could not correctly take account of the effects of linear coupling between the modes associated with different $\lambda_{km}$s.
As discussed by Aprilia et al (2011), the linear coupling between low frequency modes
tends to preferentially stabilize retrograde $g$-modes for rapidly rotating B stars, particularly
for those having lower effective temperatures.
It is therefore desirable to
use the expansion method (e.g., Lee \& Saio 1987; Lee \& Baraffe 1995) to compute low frequency modes in a rotating star
for the weakly non-linear coupling calculation (as well as for the analysis of the pulsational stability),
although the calculation using the expansion method would be much more time consuming than that using the traditional approximation to find the optimal
combination of daughter modes for a given parent mode.
In a B type main sequence star,
a $\mu$-gradient zone with a high Brunt V\"ais\"al\"a frequency forms above the convective core as it evolves from the ZAMS,
and the $\mu$-gradient zone has the effect of enhancing the coupling coefficient $|\eta_{ABC}|$
compared to the case of the ZAMS model with no $\mu$-gradient zone.
This enhancement of $|\eta_{ABC}|$ would affect the equilibrium amplitudes $|c_{A:e}|$ of the low frequency modes.
It is therefore important to examine the effects of stellar evolution on the quantities $|\eta_{ABC}|$ and $|c_{A:e}|$ and hence on the amplitude determination
of the low frequency modes in SPB stars in the weakly non-linear coupling theory.
\begin{appendix}
\section{Oscillation Equation in rotating stars in the traditional approximation}
The oscillation equations for uniformly rotating stars in the traditional approximation may be given by (e.g., Lee \& Saio 1990)
\begin{equation}
r{dz_1\over dr}=\left({V\over\Gamma_1}-3\right)z_1+\left({\lambda_{km}\over c_1\bar\omega^2}-{V\over\Gamma_1}\right) z_2,
\end{equation}
\begin{equation}
r{dz_2\over dr}=\left(c_1\bar\omega^2+rA\right)z_1+\left(1-U-rA\right)z_2,
\end{equation}
where
\begin{equation}
z_1={\xi^r(r)\over r}, \quad z_2={p^\prime(r)\over \rho g r},
\end{equation}
and
\begin{equation}
V=-{d\ln p\over d\ln r}, \quad U={d\ln M_r\over d\ln r}, \quad rA={d\ln\rho\over d\ln r}-{1\over\Gamma_1}{d\ln p\over d\ln r},
\end{equation}
$
c_1={(r/R)^3/ (M_r/M)},
$
$M_r=\int_0^R4\pi r^2\rho dr$, $g=GM_r/r^2$, $\Gamma_1=\left(\partial\ln p/\partial\ln\rho\right)_{\rm ad}$,
and $M$ and $R$ are the mass and radius of the star, and $G$ is the gravitational constant.
With appropriate boundary conditions imposed at the centre and surface of the star, we solve the above set of
differential equations as a boundary-eigenvalue problem for the eigenfrequency $\omega$.
Since $\xi^r=rz_1\tilde\Theta_{km}$, $\xi^\theta=rz_2\tilde\Theta^\theta_{km}/c_1\bar\omega^2$, and
$\xi^\phi=r z_2\tilde\Theta^\phi_{km}/c_1\bar\omega^2$, the derivatives $\partial\xi^r/\partial r$,
$\partial\xi^\theta/\partial r$, and
$\partial \xi^\phi/\partial r$ may be calculated by making use of equations (A1) and (A2).
Using the eigenfunctions $\pmb{\xi}$, the oscillation energy $\epsilon$ observed in the corotating frame
of the star is given by
(Lee \& Saio 1990)
\begin{equation}
\epsilon\equiv\omega b=\omega^2\int_0^R\pmb{\xi}^*\cdot\pmb{\xi}\rho r^2dr,
\end{equation}
and it is interesting to note that for positive $\lambda_{km}$
\begin{equation}
\left(\lambda_{km}+\nu{\partial\lambda_{km}\over\partial \nu}\right)\int_{-1}^1d\mu\left|\Theta_{km}\right|^2=
\int_{-1}^1d\mu\left(\left|\Theta^\theta_{km}\right|^2+\left|\Theta^\phi_{km}\right|^2\right).
\end{equation}
\section{Calculation of Coupling Coefficient $\kappa_{ABC}$}
If we neglect the boundary terms using the pressure zero surface boundary condition,
by use of partial integrations
we can rewrite the expression for the coupling coefficient $\kappa_{ABC}$ as (Schenk et al 2002)
\begin{equation}
\kappa_{ABC}=\kappa_{ABC}^{(1)}+\kappa_{ABC}^{(2)}+\kappa_{ABC}^{(3)}+\kappa_{ABC}^{(4)},
\end{equation}
where
\begin{equation}
\kappa_{ABC}^{(1)}={1\over 2}\int d^3\pmb{x}p\left(\Gamma_1-1\right)\left(\Xi_{AB}\nabla\cdot\pmb{\xi}_C
+\Xi_{BC}\nabla\cdot\pmb{\xi}_A+\Xi_{CA}\nabla\cdot\pmb{\xi}_B\right),
\end{equation}
\begin{equation}
\kappa_{ABC}^{(2)}=
{1\over 2}\int d^3\pmb{x}p\left[\left(\Gamma_1-1\right)^2+{\partial\Gamma_1\over\partial\ln\rho}\right]\nabla\cdot\pmb{\xi}_A\nabla\cdot\pmb{\xi}_B\nabla\cdot\pmb{\xi}_C,
\end{equation}
\begin{equation}
\kappa_{ABC}^{(3)}
={1\over 2}\int d^3\pmb{x}p\left(\chi_{ABC}+\chi_{ACB}\right),
\end{equation}
\begin{equation}
\kappa_{ABC}^{(4)}=
-{1\over 2}\int d^3\pmb{x}\rho\xi^i_A\xi^j_B\xi^k_C\nabla_i\nabla_j\nabla_k\Phi,
\end{equation}
where
$
\Xi_{AB}=\delta^i_j\Xi_i^j\left(\pmb{\xi}_A,\pmb{\xi}_B\right),
$
$
\chi_{ABC}=\delta_i^j\chi^i_j\left(\pmb{\xi}_A,\pmb{\xi}_B,\pmb{\xi}_C\right),
$
$
\chi^i_j\left(\pmb{\xi}_A,\pmb{\xi}_B,\pmb{\xi}_C\right)=(\nabla_l\xi^i_A)(\nabla_k\xi^l_B)(\nabla_j\xi^k_C),
$
and the repeated indices imply the summation over the indices from 1 to 3.
We note that since $\chi_{ABC}=\chi_{BCA}=\chi_{CAB}$ and $\chi_{ACB}=\chi_{BAC}=\chi_{CBA}$,
the coefficient $\kappa_{ABC}$ does not depend on the order of the indices $A$, $B$, and $C$.
Note that if we consider $\kappa_{ABC}(r)=\int_0^r\left(d\kappa_{ABC}/dr\right)dr$, we have $\kappa_{ABC}=\kappa_{ABC}(R)$.
In spherical polar coordinates $(r,\theta,\phi)$, the covariant derivatives of the displacement vectors are
\begin{equation}
\xi^r_{;r}={\partial \xi^r\over\partial r}={\partial (rz_1)\over\partial r}\tilde\Theta,
\end{equation}
\begin{equation}
\xi^r_{;\theta}={\partial\xi^r\over\partial\theta}-\xi^\theta=r\left(z_1\partial\tilde\Theta-{z_2\over c_1\bar\omega^2}\tilde\Theta^\theta\right),
\end{equation}
\begin{equation}
\xi^r_{;\phi}={\partial\xi^r\over\partial\phi}-\sin\theta\xi^\phi={\rm i} r\left(z_1m\tilde\Theta
-{z_2\over c_1\bar\omega^2}\sin\theta\tilde\Theta^\phi\right),
\end{equation}
\begin{equation}
\xi^\theta_{;r}={1\over r}{\partial\xi^\theta\over \partial r}={1\over r}{\partial\over \partial r}\left(r{z_2\over c_1\bar\omega^2}\right)\tilde\Theta^\theta,
\end{equation}
\begin{equation}
\xi^\theta_{;\theta}={1\over r}{\partial\xi^\theta\over\partial\theta}+{\xi^r\over r}=
z_1\tilde\Theta+{z_2\over c_1\bar\omega^2}\partial\tilde\Theta^\theta,
\end{equation}
\begin{equation}
\xi^\theta_{;\phi}={1\over r}{\partial\xi^\theta\over\partial\phi}-{1\over r}\cos\theta\xi^\phi
={\rm i} {z_2\over c_1\bar\omega^2}\tilde\Theta^\theta_\phi,
\end{equation}
\begin{equation}
\xi^\phi_{;r}={1\over r\sin\theta}{\partial\xi^\phi\over\partial r}
={\rm i} r{\partial\over \partial r}\left(r{z_2\over c_1\bar\omega^2}\right){\tilde\Theta^\phi\over \sin\theta},
\end{equation}
\begin{equation}
\xi^\phi_{;\theta}={1\over r\sin\theta}{\partial\xi^\phi\over\partial\theta}
={\rm i} {z_2\over c_1\bar\omega^2}{\partial\tilde\Theta^\phi\over\sin\theta},
\end{equation}
\begin{equation}
\xi^\phi_{;\phi}={1\over r\sin\theta}{\partial\xi^\phi\over\partial\phi}+{\xi^r\over r}
+{1\over r}{\cos\theta\over\sin\theta}\xi^\theta=z_1\tilde\Theta+{z_2\over c_1\bar\omega^2}\tilde\Theta^\phi_\phi,
\end{equation}
where we have written $\xi^i_{;j}$, instead of $\nabla_j\xi^i$, for the covariant derivatives, and
\begin{equation}
\partial\tilde\Theta={\partial\tilde\Theta\over\partial\theta}, \quad
\partial\tilde\Theta^\theta={\partial\tilde\Theta^\theta\over\partial\theta}, \quad
\partial\tilde\Theta^\phi={\partial\tilde\Theta^\phi\over\partial\theta},
\end{equation}
and
\begin{equation}
\tilde\Theta^\theta_\phi=m\tilde\Theta^\theta-\cos\theta\tilde\Theta^\phi, \quad
\tilde\Theta^\phi_\phi=-{m\over\sin\theta}\tilde\Theta^\phi+{\cos\theta\over\sin\theta}\tilde\Theta^\theta.
\end{equation}
Note that we have omitted the subscript $km$ attached to the functions $\tilde\Theta_{km}$, $\tilde\Theta^\theta_{km}$, $\tilde\Theta^\phi_{km}$, etc., for simplicity.
The integrands of $\kappa$ may be given by
\begin{equation}
\begin{array}{l}
\displaystyle
\Xi_{AB}\nabla\cdot\pmb{\xi}_C
+\Xi_{BC}\nabla\cdot\pmb{\xi}_A+\Xi_{CA}\nabla\cdot\pmb{\xi}_B=
{1\over 2}S\left({\partial(rz_1)\over \partial r}{\partial(rz_1)\over \partial r}H:\tilde\Theta\tilde\Theta\tilde\Theta\right)
\\
\displaystyle
+{1\over 2}S\left({z_2\over c_1\bar\omega^2}{z_2\over c_1\bar\omega^2}H:\left[\partial\tilde\Theta^\theta
\partial\tilde\Theta^\theta+\tilde\Theta^\phi_\phi\tilde\Theta^\phi_\phi
-2{\tilde\Theta^\theta_\phi\partial\tilde\Theta^\phi\over\sin\theta}\right]\tilde\Theta\right)
+S\left(z_1z_1H:\tilde\Theta\tilde\Theta\tilde\Theta\right)
+S\left({z_2\over c_1\bar\omega^2}z_1H:\left[\partial\tilde\Theta^\theta\tilde\Theta+\tilde\Theta^\phi_\phi\tilde\Theta\right]\tilde\Theta\right)
\\
\displaystyle
+S\left(z_1\left({\partial\over\partial r}r{z_2\over c_1\bar\omega^2}\right)H:\left[\partial\tilde\Theta\tilde\Theta^\theta-{m\tilde\Theta\tilde\Theta^\phi\over\sin\theta}\right]\tilde\Theta\right)
-S\left(H{z_2\over c_1\bar\omega^2}\left({\partial\over\partial r}r{z_2\over c_1\bar\omega^2}\right):
\tilde\Theta\left[\tilde\Theta^\theta\tilde\Theta^\theta-\tilde\Theta^\phi\tilde\Theta^\phi\right]\right),\\
\end{array}
\end{equation}
\begin{equation}
\nabla\cdot\pmb{\xi}_A\nabla\cdot\pmb{\xi}_B\nabla\cdot\pmb{\xi}_C=H_AH_BH_C\tilde\Theta_A\tilde\Theta_B\tilde\Theta_C,
\end{equation}
\begin{equation}
\begin{array}{l}
\displaystyle
\chi_{ABC}+\chi_{ACB}=S\left({\partial(rz_1)\over\partial r}\left(z_1-{z_2\over c_1\bar\omega^2}\right)
\left({\partial\over\partial r}r{z_2\over c_1\bar\omega^2}\right):\tilde\Theta\left[\tilde\Theta^\theta\tilde\Theta^\theta-\tilde\Theta^\phi\tilde\Theta^\phi\right]\right)\\
\displaystyle
+S\left({\partial(rz_1)\over\partial r}z_1
\left({\partial\over\partial r}r{z_2\over c_1\bar\omega^2}\right):\tilde\Theta\left(\partial\tilde\Theta\tilde\Theta^\theta-{m\tilde\Theta\tilde\Theta^\phi\over\sin\theta}\right)-\tilde\Theta\left[\tilde\Theta^\theta\tilde\Theta^\theta-\tilde\Theta^\phi\tilde\Theta^\phi\right]\right)\\
\displaystyle
+S\left(z_1{z_2\over c_1\bar\omega^2}\left({\partial\over\partial r}r{z_2\over c_1\bar\omega^2}\right):
\partial\tilde\Theta\partial\tilde\Theta^\theta\tilde\Theta^\theta-{m\tilde\Theta\tilde\Theta^\phi_\phi
\tilde\Theta^\phi+m\tilde\Theta\partial\tilde\Theta^\phi\tilde\Theta^\theta+\partial\tilde\Theta\tilde\Theta^\theta_\phi\tilde\Theta^\phi\over\sin\theta}
-\tilde\Theta\left[\tilde\Theta^\theta\tilde\Theta^\theta-\tilde\Theta^\phi\tilde\Theta^\phi\right]\right)\\
\displaystyle
+S\left({z_2\over c_1\bar\omega^2}{z_2\over c_1\bar\omega^2}\left({\partial\over\partial r}r{z_2\over c_1\bar\omega^2}\right):\tilde\Theta^\phi\tilde\Theta^\phi_\phi\tilde\Theta^\phi-\tilde\Theta^\theta\partial\tilde\Theta^\theta\tilde\Theta^\theta+\tilde\Theta^\phi\partial\tilde\Theta^\phi\tilde\Theta^\theta+{\tilde\Theta^\theta_\phi\tilde\Theta^\theta\tilde\Theta^\phi\over\sin\theta}\right)\\
\displaystyle
+S\left(z_1z_1\left({\partial\over\partial r}r{z_2\over c_1\bar\omega^2}\right):\partial\tilde\Theta\tilde\Theta\tilde\Theta^\theta-{m\tilde\Theta\tilde\Theta\tilde\Theta^\phi\over\sin\theta}\right)\\
\displaystyle
-S\left({z_2\over c_1\bar\omega^2}{z_2\over c_1\bar\omega^2}{z_2\over c_1\bar\omega^2}:
{\partial\tilde\Theta^\theta\tilde\Theta^\theta_\phi\partial\tilde\Theta^\phi+\partial\tilde\Theta^\phi\tilde\Theta^\theta_\phi\tilde\Theta^\phi_\phi\over\sin\theta}\right)
-2S\left(z_1{z_2\over c_1\bar\omega^2}{z_2\over c_1\bar\omega^2}:{\tilde\Theta\tilde\Theta^\theta_\phi\partial\tilde\Theta^\phi\over\sin\theta}\right)\\
\displaystyle
+{1\over 2}S\left(z_1{z_2\over c_1\bar\omega^2}{z_2\over c_1\bar\omega^2}:\tilde\Theta\tilde\Theta^\phi_\phi\tilde\Theta^\phi_\phi+\tilde\Theta\partial\tilde\Theta^\theta\partial\tilde\Theta^\theta\right)
+{1\over 2}S\left(z_1z_1{z_2\over c_1\bar\omega^2}:\tilde\Theta\tilde\Theta\partial\tilde\Theta^\theta+
\tilde\Theta\tilde\Theta\tilde\Theta^\phi_\phi\right)\\
\displaystyle
+\left({\partial(rz_1)_A\over\partial r}{\partial(rz_1)_B\over\partial r}{\partial(rz_1)_C\over\partial r}
+2(z_{1})_A(z_{1})_B(z_{1})_C\right)\tilde\Theta_A\tilde\Theta_B\tilde\Theta_C\\
\displaystyle
+\left({z_2\over c_1\bar\omega^2}\right)_A\left({z_2\over c_1\bar\omega^2}\right)_B\left({z_2\over c_1\bar\omega^2}\right)_C
\left(\tilde\Theta^\phi_{\phi A}\tilde\Theta^\phi_{\phi B}\tilde\Theta^\phi_{\phi C}
+\partial\tilde\Theta^\theta_A\partial\tilde\Theta^\theta_B\partial\tilde\Theta^\theta_C
\right),
\end{array}
\end{equation}
and
\begin{equation}
r^{-3}\xi^i_A\xi^j_B\xi^k_C\Phi_{;ijk}={1\over 2}S\left(z_1{z_2\over c_1\bar\omega_2}{z_2\over c_1\bar\omega_2}:
\tilde\Theta\left[\tilde\Theta^\theta\tilde\Theta^\theta-\tilde\Theta^\phi\tilde\Theta^\phi\right]\right){\partial\over\partial r}\left({1\over r}{\partial\Phi\over\partial r}\right)
+(z_1)_A(z_1)_B(z_1)_C\tilde\Theta_A\tilde\Theta_B\tilde\Theta_C{\partial^3\Phi\over\partial r^3},
\end{equation}
where the function $H(r)$ is defined by
\begin{equation}
\nabla\cdot\pmb{\xi}=H(r)\tilde\Theta(\theta,\phi)=-{V\over\Gamma_1}\left(z_2-z_1\right)\tilde\Theta(\theta,\phi),
\end{equation}
and
\begin{equation}
\begin{array}{r}
\displaystyle
S\left(f^1f^2f^3:p^1p^2p^3\right)=
f^1_Af^2_Bf^3_Cp^1_Ap^2_Bp^3_C
+f^1_Af^2_Cf^3_Bp^1_Ap^2_Cp^3_B
+f^1_Bf^2_Cf^3_Ap^1_Bp^2_Cp^3_A\\
\displaystyle
+f^1_Bf^2_Af^3_Cp^1_Bp^2_Ap^3_C
+f^1_Cf^2_Af^3_Bp^1_Cp^2_Ap^3_B
+f^1_Cf^2_Bf^3_Ap^1_Cp^2_Bp^3_A,
\end{array}
\end{equation}
where the functions $f^j$s depend only on $r$ and the functions $p^j$s only on $\theta$ and $\phi$.
Note that
\begin{equation}
S\left(f^1f^2f^3:p^1p^2p^3\right)=S\left(f^2f^1f^3:p^2p^1p^3\right)=S\left(f^1f^3f^2:p^1p^3p^2\right)=S\left(f^3f^2f^1:p^3p^2p^1\right)=\cdots.
\end{equation}
Integrating $S\left(f^1f^2f^3:p^1p^2p^3\right)$ over a sphere of radius $r$, we obtain
\begin{equation}
\begin{array}{r}
\displaystyle
\int S\left(f^1f^2f^3:p^1p^2p^3\right)d\Omega=f^1_Af^2_Bf^3_CZ^{123}_{ABC}
+f^1_Af^2_Cf^3_BZ^{123}_{ACB}+f^1_Bf^2_Cf^3_AZ^{123}_{BCA}~\\
+f^1_Bf^2_Af^3_CZ^{123}_{BAC}
+f^1_Cf^2_Af^3_BZ^{123}_{CAB}+f^1_Cf^2_Bf^3_AZ^{123}_{CBA},
\end{array}
\end{equation}
where
\begin{equation}
Z^{123}_{ABC}=\int p^1_Ap^2_Bp^3_C d\Omega.
\end{equation}
\end{appendix}
|
\section{Introduction}
\label{Introduction}
Metamaterials are new class of engineered materials that exhibit electromagnetic properties not readily found in nature. The novelty is that unconventional electromagnetic properties can be created by carefully chosen sub-wavelength configurations of conventional materials. The distinctive properties of metamaterials are derived from geometrically induced resonances localized to specific frequencies. These resonances are used to control propagating modes with wavelengths longer than the characteristic length scale of the material. Metamaterials are envisaged for several application areas ranging from telecommunication and solar energy harvesting to the electromagnetic cloaking
of material objects.
A generic metamaterial comes most often in the form of a crystal made from a periodic array of scatterers embedded within a host medium. The physical notions of frequency dependent effective magnetic permeability and dielectric permittivity are used to describe the behavior of propagating modes at wavelengths larger than the length scale of the metamaterial crystal. The past decade has witnessed the development and identification of new sub-wavelength geometries for novel metamaterial properties. These include the simultaneous appearance of negative effective dielectric permittivity and magnetic permeability. Such ``left handed media'' are predicted to exhibit negative group velocity, inverse Doppler effect, and an inverted Snell's law \cite{Veselago}. The first metamaterial configurations imparted electromagnetic properties consistent with the appearance of a negative bulk dielectric constant \cite{Pendry1998}. Subsequently electromagnetic behavior associated with negative effective magnetic permeability
at microwave frequencies were derived from periodic arrays of non-magnetic metallic split ring resonators \cite{PendryHolden}.
Double negative or left handed metamaterials with simultaneous negative bulk permeability and permittivity at microwave frequencies have been verified for arrays of metallic posts and split ring resonators \cite{Smith}. Subsequent work has delivered several new designs using
different configurations of metallic resonators for double negative behavior \cite{23,2A,17,21,18,20,22}.
Current state of the art metallic resonators do not perform well at optical frequencies and
alternate strategies are contemplated employing the use of both metals and dielectric materials for optical frequencies \cite{Shalaev}.
For higher frequencies in the infrared and optical range new strategies for generating double negative response rely on Mie resonances
generated inside coated rods consisting of a high dielectric core coated with a dielectric exhibiting plasmonic or Drude type frequency response at optical frequencies \cite{11,Yannopappas,Yanno2}. A second strategy for generating double negative response employs dielectric resonances
associated with small rods or particles made from dielectric materials with large permittivity, \cite{Plasmon,LPeng,VinkFelbacq}. Alternate strategies for generating negative permeability at infrared and optical frequencies use special configurations of plasmonic nanoparticles \cite{6A,7A,shevts}. The list of metamaterial systems is rapidly growing and comprehensive reviews of the subject can be found in \cite{Service} and \cite{Shalaev}.
Despite the large number of physically based strategies for generating unconventional properties the theory lacks mathematical frameworks that:
\begin{enumerate}
\item Provide the explicit relationship connecting the leading order influence of double negative behavior to the existence of Bloch wave modes inside metamaterials.
\item Provide a systematic identification of the underlying spectral problems related to the crystal geometry that control the location of stop bands and propagation bands for metamaterial crystals.
\end{enumerate}
In this article we provide such a framework for a generic class of metamaterial crystals made from non-magnetic constituents. The crystal is given by a periodic array of two aligned non-magnetic rods; one of which possesses a large frequency independent dielectric constant while the other is characterized by a frequency dependent dielectric response. In this treatment the frequency dependent dielectric response $\epsilon_P$ is associated with plasmonic or Drude behavior at optical frequencies given by \cite{11,Yannopappas}
\begin{eqnarray}
\epsilon_P(\omega^2)=1-\frac{\omega_p^2}{\omega^2},
\label{singleosc}
\end{eqnarray}
where $\omega$ is the frequency and $\omega_p$ is the plasma frequency \cite{bohren}.
Here we develop a rigorous method for calculating the frequency intervals where either double negative or double positive bulk properties appear and show how these intervals imply the existence of Bloch wave modes in the dynamic regime away from the quasi static limit, see Theorems \ref{summable} and \ref{bands}. It is shown that
these frequency intervals are explicitly determined by two distinct spectra. These are the Dirichlet spectrum of the Laplacian associated with the high dielectric rod and the electrostatic spectrum of a three phase high contrast medium obtained by sending the dielectric constant inside the high dielectric rod to $\infty$. The electrostatic spectra is introduced in Theorem \ref{completeeigen} and discussed in section \ref{genrealizedelectrostaticspectra}. The methods are illustrated for $\epsilon_P$ given by \eqref{singleosc}, however they apply to dielectrics characterized by single oscillator or multiple oscillator models that include dissipation and are of the form
\begin{eqnarray}
\epsilon_P(\omega)=1+\sum_{j=1}^N\frac{\omega_{p}^2}{\omega^2_j-\omega^2-i\gamma_j\omega},
\label{multleosc}
\end{eqnarray}
where $\omega_j$ are resonant frequencies, and $\gamma_j$ are damping factors.
We start with a metamaterial crystal characterized by a period cell containing two parallel infinitely long cylindrical rods. The rods are parallel to the $x_3$ axis and are periodically arranged within a square lattice over the transverse $\mathbf{x}=(x_1,x_2)$ plane. The period of the lattice is denoted by $d$. There is no constraint placed on the shape of the rod cross sections other than they have smooth boundaries and are simply connected.
The objective is to characterize the branches of the dispersion relation for for H-polarized Bloch-waves inside the crystal. For this case the magnetic field is aligned with the rods and the electric field lies in the transverse plane. The direction of propagation is described by the unit vector $\hat{\kappa}=(\kappa_1,\kappa_2)$ and $k=2\pi/\lambda$ is the wave number for a wave of length~$\lambda$ and the fields are of the form
\begin{eqnarray}
H_3=H_3(\mathbf{x})e^{i(k\hat{\kappa}\cdot\mathbf{x}-t\omega )},\,\,E_1=E_1(\mathbf{x})e^{i(k\hat{\kappa}\cdot\mathbf{x}-t\omega)},\,\,
E_2=E_2(\mathbf{x})e^{i(k\hat{\kappa}\cdot\mathbf{x}-t\omega)} \label{em3}
\end{eqnarray}
where $H_3(\mathbf{x})$, $E_1(\mathbf{x})$, and $E_2(\mathbf{x})$ are $d$-periodic for $\mathbf{x}$ in $\mathbb{R}^2$.
In the sequel $c$ will denote the speed of light in free space.
We denote the unit vector pointing along the $x_3$ direction by ${\bf e}_3$, and the periodic dielectric permittivity and magnetic permeability are denoted by $a_d$ and $\mu$ respectively. The electric field component $\mathbf{E}=(E_1,E_2)$ of the wave is determined by $${\bf E}=-\frac{ic}{\omega a_d}{\bf e}_3\times \nabla H_3.$$
The materials are assumed non-magnetic hence the magnetic permeability $\mu$ is set to unity inside the rods and host.
The oscillating dielectric permittivity for the crystal
is a $d$ periodic function in the transverse plane and
is described by $a_d=a_d(\mathbf{x}/d)$ where $a_d(\mathbf{y})$ is
the unit periodic dielectric function taking the values
\begin{equation}
a_d(\textbf{y})=
\begin{cases}
\epsilon_H &\text{ in the host material},\\
\epsilon_P=\epsilon_P(\omega) & \text{ in the frequency dependent ``plasmonic'' rod},\\
\epsilon_R=\epsilon_r/d^2 &\text{ in the high dielectric rod} .
\end{cases}
\end{equation}
\par
This choice of high dielectric constant $\epsilon_R$ follows that of \cite{felbacqbouchette} where $\epsilon_r$ has dimensions of area.
Setting $h^d(\mathbf{x})=H_3(\mathbf{x})e^{i(k\hat{\kappa}\cdot\mathbf{x})}$ the Maxwell equations take the form of the Helmholtz equation given by
\begin{equation}
-\nabla_{\mathbf{x}} \cdot \left(a_d^{-1}(\frac{\mathbf{x}}{d})\nabla_{\mathbf{x}} h^d(\mathbf{x})\right)=\frac{\omega^2}{c^2}h^d ~~~\text{ in } \mathbb{R}^2.
\label{Helmholtzr2}
\end{equation}
The band structure is given by the Bloch eigenvalues $\frac{\omega^2}{c^2}$ which is a subset of the parameter space $\{\frac{\omega^2}{c^2} , -2\pi\leq k_1\leq 2\pi , -2\pi\leq k_2\leq 2\pi\}$, with $k=(k_1^2+k_2^2)^{1/2}$ and $\hat{\kappa}_i=k_i/k$. This constitutes the first Brillouin Zone for this problem.
We set $\textbf{x}=d\textbf{y}$ for $\textbf{y}$ inside the unit period $Y=[0,1]^2$, put $\beta=dk\hat{\kappa}$ and write $u(\textbf{y})=H_3(d\textbf{y})$. The dependent variable is written $u^d(\textbf{y})=h^d(d\textbf{y})=u(\textbf{y})\exp^{i\beta\cdot\textbf{y}}$, and we recover the equivalent problem over the unit period cell given by
\begin{equation}
-\nabla_{\textbf{y}} \cdot \left(a_d^{-1}(\textbf{y})\nabla_{\textbf{y}} u^d\right)=\frac{d^2\omega^2}{c^2}u^d ~~~\text{ in } Y.
\label{Helmholtzd}
\end{equation}
To proceed we work with the dimensionless ratio $\rho=d/\sqrt{\epsilon_r}$, wave number $\tau=\sqrt{\epsilon_r}k$ and square frequency $\xi=\epsilon_r\frac{\omega^2}{c^2}$. The dimensionless parameter measuring the departure away from the quasi static regime is given by the ratio of period size to wavelength $\eta=dk=\rho\tau\geq 0$. The regime $\eta>0$ describes dynamic wave propagation while the infinite wavelength or quasi static limit is recovered for $\eta=0$. Metamaterials by definition are structured materials operating in the sub-wavelength regime $0<\eta<1$ away from the quasistatic limit \cite{PendryHolden}. For these parameters the dielectric permittivity takes the values $\epsilon_P=1-\frac{\epsilon_r\omega_p^2/c^2}{\xi}$, $\epsilon_R=\frac{1}{\rho^2}$, $\epsilon_H=1$, and is denoted by $a_\rho(\textbf{y})$ for $\textbf{y}$ in $Y$ and \eqref{Helmholtzd} is given by
\begin{equation}
-\nabla_{\textbf{y}} \cdot \left(a_\rho^{-1}(\textbf{y})\nabla_{\textbf{y}} u^d(\textbf{y})\right)=\rho^2\xi u^d(\textbf{y}) ~~~\text{ in } Y.
\label{Helmholtz}
\end{equation}
The unit period cell for the generic metamaterial system is represented in Fig. \ref{unitcell}. In what follows $R$ represents the rod cross section containing high dielectric material, $P$ the cross section containing the plasmonic material and $H$ denotes the connected host material region.
\par
\begin{figure}[h]
\centering
\psfrag{n1}{$\textbf{n}$}
\psfrag{H}{$H$}
\psfrag{P}{$P$}
\psfrag{R}{$R$}
\epsfig{figure=fignoncoating.eps , width=2in}
\caption{Cross section of unit cell.}
\label{unitcell}
\end{figure}
It is shown in Theorem \ref{convergenceofxi} that the band structure for the metamaterial is characterized by a power series in $\eta$ and is governed by two distinct types of spectra determined by the shape and configuration of the rods inside the period cells. The series delivers the explicit relationship connecting the leading order influence of quasi static behavior, as mediated by {\em effective magnetic permeability and dielectric permittivity}, to the propagation of Bloch wave solutions inside metamaterial crystals made from sub-wavelength $1>\eta>0$ structures, see Theorems \ref{convergenceofxi}, \ref{summable}, and \ref{bands}.
The relevant spectra for this problem is found to be given by the Dirichlet spectra for the Laplacian over the rod cross sections $R$ together with a generalized electrostatic spectra associated with the infinite connected region exterior to the rods. These spectra provide two distinct criteria that taken together are sufficient for the existence of power series solutions see, Theorem \ref{summable}. In what follows the power series is developed in terms of a hierarchy of boundary value problems posed separately over the domain $R$ and the domain exterior to the high dielectric rod $Y\setminus R$. The existence of solutions for the boundary value problems inside the high dielectric rod $R$ is controlled by the Dirichlet spectra see, section \ref{higherorder}. Existence of solutions for boundary value problems exterior to $R$ are determined by the generalized electrostatic spectrum see, Theorems \ref{completeeigen} and \ref{existencetheoremforoutsideR}.
The electrostatic spectra is identified and is shown to be given by the eigenvalues of a compact operator acting on an appropriate Sobolev space of periodic functions see, the discussion in section \ref{genrealizedelectrostaticspectra} and Theorem \ref{T}.
The generic class of double negative metamaterials introduced here appears to be new. The motivation behind their construction draws from earlier investigations. The influence of electrostatic resonances on the effective dielectric tensor associated with crystals made from a single frequency dependent dielectric inclusion is developed in the pioneering work of \cite{kantorbergman,McPhedranMilton,MiltonBook} see also the more recent work \cite{shevts} in the context of matematerials. These resonances are responsible for negative effective dielectric permittivity. In this context we point out that the sub-wavelength geometry introduced here is a three phase medium and the categorization of all possible electrostatic resonances requires a different approach see section \ref{genrealizedelectrostaticspectra}. On the other hand Dirichlet resonances generate negative effective permeability inside high contrast non-dispersive dielectric inclusions, this phenomena is discovered in \cite{felbacqbouchette,bouchettefelbacq,9A,3A,ObrienPendry}, see also the mathematically related investigations of \cite{Cherdansev,HempelandLenau,Smyshalaev,zhikov}. Motivated by these observations we have constructed a hybrid composite crystal that combines both high dielectric inclusions and frequency dependent inclusions for generating double negative response from non-magnetic materials.
The power series approach to sub-wavelength $\eta<1$ analysis has been developed in \cite{FLS} for characterizing the dynamic dispersion relations for Bloch waves inside plasmonic crystals. It has also been applied to assess the influence of effective negative permeability on the propagation of Bloch waves inside high contrast dielectrics \cite{FLS2}, the generation of negative permeability inside metallic - dielectric resonators \cite{shipman}, and for concentric coated cylinder assemblages generating a double negative media \cite{chenlipton}.
We conclude noting that earlier related work introduces the use of high contrast structures for opening band gaps in photonic crystals, this is developed in \cite{FK1,FK2,FK3}. For two phase high contrast media integral equation methods are applied to recover dispersion relations about frequencies corresponding to Dirichlet eigenvalues \cite{amari1} and \cite{amari2}. The connection between high contrast interfaces and negative effective magnetic permeability for time harmonic waves is made in \cite{KohnShipman}. More recently
two-scale homogenization theory has been developed for three dimensional split ring structures that deliver negative effective magnetic permeability \cite{bouchetteschwizer,4A}. For periodic arrays made from metal fibers a homogenization theory delivering negative effective dielectric constant \cite{bouchettebourel} is established. A novel method for creating metamaterials with prescribed effective dielectric permittivity and effective magnetic permeability at a fixed frequency
is developed in \cite{Milton2}.
\section{Background, basic theory, and generalized electrostatic resonances}
\label{Background}
In this section we outline the steps in the power series development of Bloch waves and establish the mathematical foundations for establishing existence of power series solutions.
We introduce the function space $H^1_{per}(Y)$ defined to be all $Y$-periodic, complex valued square integrable functions with square integrable derivatives with the usual inner product and norm given by
\begin{eqnarray}
( u, v)_{Y}=\int_Y\left(\nabla u\cdot\nabla \overline{v}+u\overline{v}\right)\,dy \hbox{ and }\Vert u\Vert_Y=(u,u)_Y^{1/2}.
\label{innerY}
\end{eqnarray}
We also introduce the space $H^1_{per}(Y\setminus R)$ given by all $Y$-periodic, complex valued square integrable functions with square integrable derivatives and the inner product and seminorm
\begin{eqnarray}
( u,v)=\int_{Y\setminus R}\nabla u\cdot\nabla \overline{v}\,dy \hbox{ and }\Vert u\Vert=(u,u)^{1/2}.
\label{innerproduct}
\end{eqnarray}
The variational form of (\ref{Helmholtz}) is given by
\begin{equation}
\int_Y a_\rho^{-1}\nabla u^d\cdot \nabla \bar{\tilde{v}}= \int_Y \frac{\rho^2\xi}{c^2}u^d\bar{\tilde{v}}
\label{variational}
\end{equation}
for any $\tilde{v}=v(\textbf{y})e^{i\hat{\kappa}\cdot \tau\rho \textbf{y}}$ , where $v \in H ^1_{per}(Y)$ . On writing $u^d=u(\textbf{y})e^{i\hat{\kappa}\cdot \tau\rho \textbf{y}}$ and setting $\eta=\tau\rho$ we transform (\ref{variational}) into
\begin{eqnarray}
&&\int_H\tau^2(\xi-\epsilon_r\frac{\omega_p^2}{c^2})(\nabla+i\eta \hat{\kappa})u\cdot \overline{(\nabla+i\eta \hat{\kappa})v }+\int_P\tau^2\xi(\nabla+i\eta \hat{\kappa})u\cdot \overline{(\nabla+i\eta \hat{\kappa})v }\nonumber
\\&&+\int_R\eta^2(\xi-\epsilon_r\frac{\omega_p^2}{c^2})(\nabla+i\eta \hat{\kappa})u\cdot \overline{(\nabla+i\eta \hat{\kappa})v }=\int_Y\eta^2\xi(\xi-\epsilon_r\frac{\omega_p^2}{c^2})u\overline{v}
\label{variational2}
\end{eqnarray}
We introduce the power series
\begin{eqnarray}
&&u=\sum_{m=0}^\infty \eta^m u_m \label{upower}
\\&&\xi=\sum_{m=0}^\infty\eta^m\xi_m \label{xipowerexpansion}
\end{eqnarray}
where $u_m$ belongs to $H^1_{per}(Y)$. In view of the algebra it is convenient to write $u_m=i^m\underline{u}_0\psi_m$ where $\underline{u}_0$ is an arbitrary constant factor.
We now describe the underlying variational structure associated with the power series solution. Set $$z=\epsilon_P^{-1}(\xi_0)=\left(1-\frac{\epsilon_r\omega_p^2/c^2}{\xi_0}\right)^{-1}$$
and for $u$, $v$ belonging to $H^1_{per}(Y)$ we introduce the sesquilinear form
\begin{eqnarray}
B_z(u,v)=\int_{H}\nabla u\cdot\nabla\overline{v}\,d\textbf{y}+\int_P\,z\,\nabla u\cdot\nabla\overline{v}\,d\textbf{y}.
\label{bilinear}
\end{eqnarray}
Here $Y\setminus R=H\cup P$ and the form $B_z(u,v)$ is well defined for functions in $H^1_{per}(Y\setminus R)$.
Substitution of the series
into the system (\ref{variational2}) and equating like powers of $\eta$ produce the infinite set of coupled equations for $m=0,1,2\ldots$ given by
\begin{eqnarray}
&&\tau^2 B_z(\psi_m,v)+\xi_0^{-1}\epsilon_p^{-1}(\xi_0)\tau^2\int_{Y\setminus R}\big[\sum^{m-1}_{l=1}(-i)^l\xi_l\nabla \psi_{m-l}\cdot \nabla \overline{ v}\nonumber \\&&+\hat{\kappa}\cdot\sum_{l=0}^{m-1}(-i)^l\xi_l(\psi_{m-1-l}\nabla \overline{v}-\nabla \psi_{m-1-l}\overline{v})-\sum_{l=0}^{m-2}(-i)^l\xi_l\psi_{m-2-l}\overline{v}\big]\nonumber \\&&-\xi_0^{-1}\epsilon_p^{-1}(\xi_0)\tau^2\epsilon_r\frac{\omega_p^2}{c^2}\int_H\big[\hat{\kappa}\cdot(\psi_{m-1}\nabla \overline{v}-\nabla \psi_{m-1}\overline{v})-\psi_{m-2}\overline{v}\big]\nonumber
\\&&-\xi_0^{-1}\epsilon_p^{-1}(\xi_0)\int_R\big[\sum^{m-2}_{l=0}(-i)^l\xi_l\nabla \psi_{m-2-l}\cdot \nabla \overline{ v}\nonumber \\&&+\hat{\kappa}\sum_{l=0}^{m-3}(-i)^l\xi_l(\psi_{m-3-l}\nabla \overline{v}-\nabla \psi_{m-3-l}\overline{v})
-\sum_{l=0}^{m-4}(-i)^l\xi_l\psi_{m-4-l}\overline{v}\big]\nonumber
\\&&+\xi_0^{-1}\epsilon_p^{-1}(\xi_0)\int_R\epsilon_r\frac{\omega_p^2}{c^2}\big[\nabla \psi_{m-2}\cdot \nabla \overline{ v}+\hat{\kappa}(\psi_{m-3}\nabla \overline{v}-\nabla \psi_{m-3}\overline{v})+\psi_{m-4}\overline{v}\big]\nonumber
\\&&-\xi_0^{-1}\epsilon_p^{-1}(\xi_0)\int_Y\big[\sum_{l=0}^{m-2}\sum_{n=0}^l \xi_{m-2-l}\xi_n \psi_{l-n} i^{l-n-m}\overline{ v}\nonumber
\\&&+\epsilon_r\frac{\omega_p^2}{c^2}\sum_{l=0}^{m-2}(-i)^l\xi_l \psi_{m-2-l}\overline{ v}\big]
=0,\quad\quad\hbox{ for all $v$ in $H^1_{per}(Y)$}.
\label{summation2}
\end{eqnarray}
Here the convention is $\psi_m=0$ for $m<0$.
The determination of $\{\psi_m\}_{m=0}^\infty$ proceeds iteratively. We start by determining $\psi_0$ on $Y\setminus R$, this function is used as boundary data to determine $\psi_0$ in $R$ from which we determine $\psi_1$ on $Y\setminus R$ and the full sequence is determined on iterating this cycle. The elements $\xi_m$ are recovered from solvability conditions obtained by setting $v=1$ in \eqref{summation2} and proceeding iteratively. The complete algorithm together with explicit boundary value problems necessary for the determination of the sequences $\{\psi_m\}_{m=0}^\infty$, $\{\xi_m\}_{m=0}^\infty$ is described in section \ref{higherorder} and Theorem \ref{Existencesequence}.
The existence theory for the solution of the sequence of boundary value problems is based on the sesquilinear form $B_z(u,v)$ defined for functions $u$ and $v$ belonging to $H^1_{per}(Y\setminus R)/\mathbb{C}$. Here $H^1_{per}(Y\setminus R)/\mathbb{C}$ is the subspace of functions $u$ belonging to $H_{per}^1(Y\setminus R)$ with zero mean $\int_{Y\setminus R} u\,dy=0$. This space is a Hilbert space with inner product \eqref{innerproduct}. Although in this treatment $\epsilon_P$ is given by \eqref{singleosc} we are motivated by the general case \eqref{multleosc} and proceed in full generality allowing for the possibility that $z$ can lie anywhere on the complex plane $\mathbb{C}$ including the negative real axis. Thus for each $z$ in $\mathbb{C}$ we are required to characterize the range of the map $u \mapsto B_z(u,\cdot)$ viewed as a linear transformation $T_z$ mapping $u$ into the space of bounded skew linear functionals on $H^1_{per}(Y\setminus R)/\mathbb{C}$. This is linked to the following eigenvalue problem characterizing all pairs $\lambda$ in $\mathbb{C}$, $\psi$ in $H^1_{per}(Y\setminus R)/\mathbb{C}$ that solve
\begin{eqnarray}
-\frac{1}{2}\int_P\nabla\psi\cdot\nabla \overline{v}\,d\textbf{y}+\frac{1}{2}\int_H\nabla\psi\cdot\nabla\overline{v}\,d\textbf{y}=(\lambda\psi,v),
\label{resonance}
\end{eqnarray}
for every $v$ in $H^1_{per}(Y\setminus R)/\mathbb{C}$.
Inspection shows that for $z=(\lambda+1/2)/(\lambda-1/2)$ that $B_z(\psi,v)=0$, for every $v$ in $H^1_{per}(Y\setminus R)/\mathbb{C}$. In other words the kernel of the operator $T_z$ is nonempty for $z=(\lambda+1/2)/(\lambda-1/2)$.
The eigenvalues $\lambda$ will be referred to as generalized electrostatic resonances. In what follows we show that $T_z$ is a one to one and onto map from $H_{per}^1(Y\setminus R)/\mathbb{C}$ into the dual space provided that $z\not=(\lambda+1/2)/(\lambda-1/2)$.
The generalized electrostatic resonances are characterized by introducing a suitable orthogonal decomposition of $H^1_{per}(Y\setminus R)/\mathbb{C}$.
We introduce the subspace $H^1_0(P)$ given by the closure in the $H^1$ norm of smooth functions $v$ with compact support on $P$ and the subspace $H^1_{0,per}(H)$ given by the closure in the $H^1$ norm of all periodic continuously differentiable functions with support outside $P$. Extending elements of $H^1_{0,per}(H)$ by zero to $Y\setminus R$ delivers $W_1\subset H^1_{per}(Y\setminus R)$, extending elements of $H^1_0(P)$ by zero to $Y\setminus R$ delvers $W_2\subset H^1_{per}(Y\setminus R)$. Define $W_3$ to be all functions $w$ in $H^1_{per}(Y\setminus R)$ for which the boundary integral $\int_{\partial P}\, w\,dS$ vanishes and that belong to the orthogonal complement of $W_1\cup W_2$ with respect to the inner product \eqref{innerproduct}. Its easily verified that $W_1$, $W_2$, and $W_3$ are pairwise orthogonal with respect to the inner product \eqref{innerproduct} and
\begin{eqnarray}
H^1_{per}(Y\setminus R)/\mathbb{C}=W_1\oplus W_2\oplus W_3\oplus \mathbb{C},
\label{decomppp}
\end{eqnarray}
where the constant part of a function $u$ belonging to this space is uniquely determined by the condition $\int_{Y\setminus R} u\,dy=0$.
The following theorem
describing all eigenvalue eigenfunction pairs is established in section \ref{genrealizedelectrostaticspectra}.
\begin{theorem}
The eigenvalues for \eqref{resonance} are real and constitute a denumerable set contained inside $[-1/2,1/2]$ with the only accumulation point being zero. The eigenspaces associated with $\lambda=1/2$ and $\lambda=-1/2$ are $W_1$ and $W_2$ respectively. Eigenspaces associated with distinct eigenvalues in $(-1/2,1/2)$ are pairwise orthogonal, and their union spans the subspace $W_3$.
\label{completeeigen}
\end{theorem}
We denote the denumerable set of eigenvalues for \eqref{resonance} by the sequence $\{\lambda_i\}_{i=1}^\infty$. Here we put $\lambda_1=0$. The orthogonal projections onto $W_1$ and $W_2$ are denoted by $\mathcal{P}_1$ and $\mathcal{P}_2$. The orthogonal projections associated with $\lambda_i$ in $(-1/2,1/2)$ are denoted by $\mathcal{P}_{\lambda_i}$.
For $z\not=(\lambda_i+1/2)/(\lambda_i-1/2)$ the following existence theorem holds.
\begin{theorem}
Suppose $z\neq (\lambda_i+1/2)/(\lambda_i-1/2)$ for $\lambda_i \in [-\frac{1}{2},\frac{1}{2}]$ then
\begin{itemize}
\item
For any $F\in \left[H^1_{per}(Y\setminus R)/\mathbb{C}\right]^*$ such that $F(v)=0 $ for constant $v$, there exists a unique solution $u\in H^1_{per}(Y\setminus R)/\mathbb{C} $ of the variational problem $B_z(u,v)=\overline{F(v)}$ for all $v\in H^1_{per}(Y\setminus R)/\mathbb{C}$.
\item The transformation $T_z$ from $H^1_{per}(Y\setminus R)/\mathbb{C}$ onto itself has the representation formula given by
\begin{eqnarray}
T_z =\mathcal{P}_1 +z \mathcal{P}_2 +\sum_{-\frac{1}{2}<\lambda_n<\frac{1}{2}} (1+(z-1)(\frac{1}{2}-\lambda_n))\mathcal{P}_{\lambda_n},
\label{Tz1}
\end{eqnarray}
with inverse
\begin{eqnarray}
T_z^{-1}=\mathcal{P}_1 +z^{-1} \mathcal{P}_2 +\sum_{-\frac{1}{2}<\lambda_n<\frac{1}{2}} (1+(z-1)(\frac{1}{2}-\lambda_n))^{-1}\mathcal{P}_{\lambda_n},\label{Tzinus1}
\end{eqnarray}
and $B_z(u,v)=(T_z u,v)$ for all $u$, $v$ in $H^1_{per}(Y\setminus R)/\mathbb{C}$.
\end{itemize}
\label{existencetheoremforoutsideR}
\end{theorem}
\noindent This theorem is proved in section \ref{exist}.
In the following sections we present the sequence of boundary value problems for determining $\psi_m$ in $Y\setminus R$ and $R$ together with the sequence of solvability conditions characterizing $\xi_m$. In this section we use the complete orthonormal systems of eigenfunctions associated with electrostatic resonances and Dirichlet eigenvalues to explicitly solve for fields $\psi_0$ in $Y$ and $\psi_1$ in $Y\setminus R$ and provide an explicit formula for $\xi_0$.
In the following section the boundary value problems used to determine $\psi_m$ in $Y\setminus R$, $m\geq 2$ and $\psi_m$ in $R$ for $m\geq 1$
together with solvability conditions for $\xi_m$ for $m\geq 1$ are shown to be a well posed infinite system of equations see Theorem \ref{Existencesequence}.
Applying the convention $\psi_m=0$ for $m<0$ in \eqref{summation2} shows that $\psi_0$ is the solution of
\begin{eqnarray}
B_z(\psi_0,v)=0,\hbox{ for all $v$ in $H^1(Y\setminus R)$}.
\label{zero}
\end{eqnarray}
From Theorem \ref{completeeigen} and \eqref{decomppp} we have the dichotomy:
\begin{enumerate}
\item $\xi_0$ satisfies $\epsilon_P^{-1}(\xi_0)=(\lambda_i+1/2)/(\lambda_i-1/2)$ and $\psi_0$ is an eigenfunction for \eqref{resonance}.
\item $\xi_0$ satisfies $\epsilon_P^{-1}(\xi_0)\not=(\lambda_i+1/2)/(\lambda_i-1/2)$, $i=1,2,\ldots$ and $\psi_0=constant$.
\end{enumerate}
In this article we assume the second alternative. Subsequent work will investigate the case when the first alternative is applied.
For future reference the condition $\epsilon_P^{-1}(\xi_0)\not=(\lambda_i+1/2)/(\lambda_i-1/2)$ is equivalent to
\begin{eqnarray}
\xi_0\not=\zeta_i,&\zeta_i\equiv(\lambda_i+\frac{1}{2})\frac{\epsilon_r\omega_p^2}{c^2}&,0\leq\zeta_i\leq\frac{\epsilon_r\omega_p^2}{c^2} \hbox{ $i=1,2,\ldots$.}
\label{notonresonance}
\end{eqnarray}
Restricting to test functions $v$ with support in $R$ in \eqref{summation2} we get
\begin{eqnarray}
\int_R\left(\nabla \psi_0\cdot\nabla \overline{v}-\xi_0\psi_0 \overline{v}\right)\,d\textbf{y}=0.
\label{weakformofu0inR}
\end{eqnarray}
From continuity we have the boundary condition for $\psi_0$ on $R$ given by $\psi_0=const.$
We denote the Dirichlet eigenvalues for $R$ by $\nu_j$, $j=1,2,\ldots$. Here we have the alternative:
\begin{enumerate}
\item If $\xi_0$ is a Dirichlet eigenvalue $\nu_i$ of $-\Delta$ in $R$ then $\psi_0(\textbf{y})=0$ for $\textbf{y}$ in $Y\setminus R$.
\item If $\xi_0\not=\nu_i$, $i=1,2,\ldots$ then $\psi_0$ is the unique solution of the Helmholtz equation \eqref{weakformofu0inR} and $\psi_0=const.$ in $Y\setminus R$.
\label{altdir}
\end{enumerate}
In this treatment we will choose the second alternative $\xi_0\not=\nu_i$. The case when the first alternative is chosen will be taken up in future investigation. Since $u_0=\underline{u}_0\psi_0$ where $\underline{u}_0$ is an arbitrary constant we can without loss of generality make the choice $\psi_0=1$ for $\textbf{y}$ in $Y\setminus R$. Since $\xi_0\not=\nu_i$, $i=1,\dots$ and $\psi_0=1$ in $Y\setminus R$ a straight forward calculation gives $\psi_0$ in $R$ in terms of the complete set of Dirichlet eigenfunctions and eigenvalues:
\begin{eqnarray}
&&\psi_0=\sum_{n=1}^{\infty}\frac{\mu_n<\phi_n>_R}{\mu_n-\xi_0}\phi_n , \hbox{ in $R$}.
\label{explicitformofpsi0}
\end{eqnarray}
Note here that $\mu_n$ denote the Dirichlet eigenvalues of $-\Delta$ in $R$ whose eigenfunctions $\phi_n$ have nonzero mean, $<\phi_n>_R=\int_R\phi_n(y)dy\neq 0$. The Dirichlet eigenvalues associated with zero mean eigenfunctions are denoted by $\mu'_n$ and $\{\nu_n\}_{n=1}^\infty=\{\mu_n\}_{n=1}^\infty\cup\{\mu'_n\}_{n=1}^\infty$.
\par
To find $\psi_1$ in $Y\setminus R$, we appeal to (\ref{summation2}) with $\psi_0=1$ in $Y\setminus R$ to discover
\begin{eqnarray}
B_z(\psi_1,v)=-\int_H \hat\kappa \cdot \nabla\overline{v}-\int_P\epsilon_P^{-1}(\xi_0)\hat\kappa\cdot \nabla\overline{v} ~~~\forall ~~ v\in H^1_{per}(Y)
\end{eqnarray}
It follows from Theorem \ref{existencetheoremforoutsideR} that the problem has a unique solution subject to the mean-zero condition: $\int_{Y\setminus R}\psi_1=0$ provided that $\xi_0\not=\zeta_i$, $i=1,2,\ldots$.
We apply the decomposition of $H^1_{per}(Y\setminus R)/\mathbb{C}$ given by \eqref{decomppp} and represent $\psi_1$ in terms of the complete set of orthonormal eigenfunctions $\{\psi_{\lambda_n}\}\subset W_3$ associated with $-1/2<\lambda_n <1/2$ together with
the complete orthonormal sets of functions for $W_1$ and $W_2$, denoted by $\{\psi^1_n\}_{n=1}^\infty$ and $\{\psi^2_n\}_{n=1}^\infty$ respectively.
A straight forward calculation gives the representation for $\psi_1$ in $Y\setminus R$
{\footnotesize\begin{eqnarray}
\psi_1=-\sum_{-\frac{1}{2}<\lambda_n<\frac{1}{2}}\left(\frac{(\alpha^1_{\lambda_n}+\epsilon_P^{-1}(\xi_0)\alpha^2_{\lambda_n})}{1+(\epsilon_P^{-1}(\xi_0)-1)(1-\lambda_n)}\right)\psi_{\lambda_n}+\sum_{n=1}^\infty \alpha_{1,n}\psi^1_n,\hbox{ in $Y\setminus R$}
\label{expansionpsi1}
\end{eqnarray}}
with
{\footnotesize\begin{eqnarray}
\alpha^1_{\lambda_n}=\hat{\kappa}\cdot\int_H\nabla\psi_{\lambda_n}\,d\textbf{y}, & \alpha^2_{\lambda_n}=\hat{\kappa}\cdot\int_P\nabla\psi_{\lambda_n}\,d\textbf{y}, \hbox{~and}&
\alpha_{1,n}=\hat{\kappa}\cdot\int_H\nabla\psi^1_n\,d\textbf{y}.
\label{coefficients}
\end{eqnarray}}
Setting $v=1$ and $m=2$ in \eqref{summation2} we recover the solvability condition given by
\begin{eqnarray}
&& \tau^2\int_{H\cup P}\big[ -\hat\kappa\cdot\xi_0\nabla \psi_1+\xi_0 \big]-\tau^2\epsilon_r\frac{\omega_p^2}{c^2}\int_H(-\hat\kappa\nabla \psi_1+1)\label{solvablety}
\\&&=\int_Y(\xi_0^2\psi_0-\epsilon_r\frac{\omega_p^2}{c^2}\xi_0 \psi_0)\nonumber
\end{eqnarray}
Substitution of the spectral representations for $\psi_1$ and $\psi_0$ given by \eqref{explicitformofpsi0} and \eqref{expansionpsi1} into \eqref{solvablety} delivers the quasistatic dispersion relation
\begin{eqnarray}
\xi_0=\tau^2 n_{eff}^{-2}(\xi_0),
\label{subwavelengthdispersion1}
\end{eqnarray}
where the effective index of diffraction $n_{eff}^2$ depends upon the direction of propagation $\hat{\kappa}$ and is written
\begin{eqnarray}
n_{eff}^2(\xi_0)=\mu_{eff}(\xi_0)/\epsilon_{eff}^{-1}(\xi_0)\hat\kappa\cdot\hat\kappa.
\label{indexofrefract}
\end{eqnarray}
The frequency dependent effective magnetic permeability $\mu_{eff}$ and effective dielectric permittivity $\epsilon_{eff}$ are given by
\begin{eqnarray}
\mu_{eff}(\xi_0)=\int_Y\psi_0=\theta_H+\theta_P+\sum_{n=1}^\infty\frac{\mu_n<\phi_n>^2_R}{\mu_n-\xi_0}\label{effperm}
\end{eqnarray}
and
{\footnotesize\begin{eqnarray}
&&\epsilon^{-1}_{eff}(\xi_0)\hat\kappa\cdot\hat\kappa =\int_{H}(I-\mathbb{P})\hat{\kappa}\cdot\hat{\kappa}\,d\textbf{y}+\frac{\xi_0}{\xi_0-\frac{\epsilon_r\omega_p^2}{c^2}}\theta_P\nonumber\\
&&-\sum_{-\frac{1}{2}<\lambda_h <\frac{1}{2}}\left( \frac{\left(\xi_0-\frac{\epsilon_r\omega_p^2}{c^2} \right)|\alpha_{\lambda_h}^{(1)}|^2+2\frac{\epsilon_r\omega_p^2}{c^2}\alpha_{\lambda _h}^{(1)}\alpha_{\lambda_h}^{(2)}+\frac{\left(\frac{\epsilon_r\omega_p^2}{c^2}\right)^2}{\xi_0-\frac{\epsilon_r\omega_p^2}{c^2}}|\alpha_{\lambda_h}^{(2)}|^2}{\xi_0-s_h} \right),\label{effdielectricconst}
\end{eqnarray}}
where $\theta_H$ and $\theta_P$ are the areas occupied by regions $H$ and $P$ respectively. The first term on the right hand side of \eqref{effdielectricconst} is positive and frequency independent. It is written in terms of
the spectral projection $\mathbb{P}$ of square integrable vector fields over $H$ onto the subspace of gradients of potentials $\psi$ in $H_{0,per}^1(H)$.
Here $\int_{H}\mathbb{P}\hat{\kappa}\cdot\hat{\kappa}\,d\textbf{y}<\theta_H$ and $\mathbb{P}\sigma=\sum_{n=1}^\infty\left(\int_H\nabla\psi_n^1\cdot\sigma\,d\textbf{y}\right)\nabla\psi_n^1$ with $\sum_{n=1}^\infty|\alpha_{1,n}|^2=\int_{H}\mathbb{P}\hat{\kappa}\cdot\hat{\kappa}\,d\textbf{y}$. The poles $s_h$ are the subset of values $\zeta_h$ for which at least one of the weights $\alpha_{\lambda_h}^{(1)}$ and $\alpha_{\lambda_h}^{(2)}$ are non zero. The values of $\zeta_h$ for which both weights vanish are denoted by
$s'_h$ and $\{\zeta_h\}_{h=1}^\infty=\{s_h\}_{h=1}^\infty\cup\{s'_h\}_{h=1}^\infty$. The graphs
of $\mu_{eff}$ and $\epsilon^{-1}_{eff}\hat\kappa\cdot\hat\kappa$ as functions of $\xi_0$ are displayed in Fig. \ref{ximueffrelation} and Fig. \ref{xiepsiloneffrelation}. Here the intervals $a'<\xi_0 <b'$ and $a''<\xi_0 <b''$ are the same in all graphs.
\par
\begin{figure}[h!]
\begin{center}
\begin{psfrags}
\psfrag{mu}{$\mu_{eff}$}
\psfrag{wc}{$\xi_0$}
\psfrag{k1}{$\mu_1$}
\psfrag{k2}{$\mu_2$}
\psfrag{k3}{$\mu_3$}
\psfrag{c}{$\cdots$}
\psfrag{a1}{$a'$}
\psfrag{b1}{$b'$}
\psfrag{a2}{$a''$}
\psfrag{b2}{$b''$}
\includegraphics[width=0.4\textwidth]{mueff.eps}
\end{psfrags}
\caption{The relation between $\mu_{eff}$ and $\xi_0$.}
\label{ximueffrelation}
\end{center}
\end{figure}
\par
\begin{figure}[h!]
\begin{center}
\begin{psfrags}
\psfrag{e}{$\epsilon^{-1}_{eff}\hat\kappa\cdot\hat\kappa$}
\psfrag{wc}{$\xi_0$}
\psfrag{k3}{$\epsilon_r\frac{\omega_p^2}{c^2}$}
\psfrag{k2}{$s_1$}
\psfrag{k1}{$s_2$}
\psfrag{c}{$\cdots$}
\psfrag{a1}{$a'$}
\psfrag{b1}{$b'$}
\psfrag{a2}{$a''$}
\psfrag{b2}{$b''$}
\includegraphics[width=0.45\textwidth]{epsiloneff.eps}
\end{psfrags}
\caption{The relation between $\epsilon^{-1}_{eff}\hat\kappa\cdot\hat\kappa$ and $\xi_0$.}
\label{xiepsiloneffrelation}
\end{center}
\end{figure}
\par
For future reference is is convenient to write the dispersion relation explicitly in terms of $\mu_{eff}$ and $\epsilon_{eff}$ and
\begin{eqnarray}
\mu_{eff}(\xi_0)\xi_0=\tau^2\epsilon_{eff}^{-1}(\xi_0)\hat\kappa\cdot\hat\kappa.
\label{subwavelengthdispersion1alternate}
\end{eqnarray}
The branches of the quasistatic dispersion relation $(\tau,\hat{\kappa})\mapsto\xi_0$ are controlled by the poles and zeros of $\mu_{eff}$ and $\epsilon_{eff}^{-1}$ explicitly
determined through the Dirichlet spectra and generalized electrostatic resonances.
\section{Solution of higher order problems}
\label{higherorder}
Having determined $\psi_0$ in $Y$, $\psi_1$ in $Y\setminus R$, and $\xi_0$ we now
provide the algorithm for determining the rest of the elements in the sequences $\{\psi_m\}_{m=0}^\infty$, $\{\xi_m\}_{m=0}^\infty$.
The algorithm is summarized in Theorem \ref{Existencesequence}.
\noindent Step I. {\em Solution of $\psi_m$ in $R$ for $m\geq 1$.}
We restrict the trial space to test functions $v$ with support in $R$
and put $m\mapsto m+2$ in \eqref{summation2}. We decompose $\psi_m$ according to
\begin{eqnarray}
\psi_m=\tilde{\psi}_m+(-i)^{m}\xi_{m}\psi_*
\label{decomp}
\end{eqnarray}
and substitute into \eqref{summation2}.
This decomposition is chosen such that $\tilde{\psi}_m$ depends on $\xi_n$ and $\psi_n$ inside $R$ for $n\leq m-1$. The function $\psi_*$ depends only upon $\xi_0$ and $\psi_0$ in $R$.
The function $\tilde{\psi}_m$ solves the Dirichlet boundary value problem with $\tilde{\psi}_m|_{\partial R-}=\psi_m|_{\partial R+}$ and
\begin{eqnarray}
\int_R\left(\nabla\tilde{\psi}_m\cdot\nabla\overline{v}-\xi_0\tilde{\psi}_m\overline{v}\right)\,d\textbf{y}=\int_R F\cdot\nabla\overline{v}\,d\textbf{y}+\int_R G\overline{v}\,d\textbf{y},
\label{Rproblem}
\end{eqnarray}
where
\begin{eqnarray}
F&=&-\xi_0^{-1}\epsilon_p^{-1}(\xi_0)[\sum^{m-1}_{l=1}(-i)^l\xi_l\nabla \psi_{m-l}+\hat{\kappa}(\sum_{l=0}^{m-1}(-i)^l\xi_l\psi_{m-1-l})]
\nonumber
\\&&+\xi_0^{-1}\epsilon_p^{-1}(\xi_0)\epsilon_r\frac{\omega_p^2}{c^2}\hat{\kappa}\psi_{m-1},
\label{F}
\end{eqnarray}
and
{\small
\begin{eqnarray}
G&=&\xi_0^{-1}\epsilon_p^{-1}(\xi_0)(\hat{\kappa}\cdot \sum_{l=0}^{m-1}(-i)^l\xi_l\nabla \psi_{m-1-l}+\sum_{l=0}^{m-2}(-i)^l\xi_l\psi_{m-2-l})\nonumber
\\&&-\xi_0^{-1}\epsilon_p^{-1}(\xi_0)(\hat{\kappa}\cdot\nabla \psi_{m-1}+\psi_{m-2})\label{G}
\\&&+\xi_0^{-1}\epsilon_p^{-1}(\xi_0)[\sum_{l=1}^{m-1}\sum_{n=0}^l \xi_{m-l}\xi_n\psi_{l-n}i^{l-n-m}-\epsilon_r\frac{\omega_p^2}{c^2}\sum_{l=1}^{m-1}(-i)^l\xi_l\psi_{m-l}\nonumber
\\&&+\xi_0\sum_{n=1}^{m-1}\xi_n\psi_{m-n}i^{-n}].
\nonumber
\end{eqnarray}}
This boundary value problem has a unique solution for $\xi_0\not=\nu_i$, $i=1,\ldots$.
The function $\psi_*$ is the solution of
\begin{equation}
\begin{cases}
\int_R(-\nabla\psi_*\cdot \nabla \overline{v}+ \xi_0\psi_*\overline{v})+\int_R\psi_0\overline{v}=0\\
\psi_*|_{\partial R-}=0\label{inproblem}
\end{cases}
\end{equation}
The explicit representation for $\psi_*$ is obtained using the Dirichlet eigenfunctions on $R$ and is given by,
\begin{equation}
\psi_*= \sum_{n=1}^{\infty}\frac{\mu_n<\phi_n>_R}{(\mu_n-\xi_0)^2}\phi_n ~~\text{ in } R.
\label{psi*}
\end{equation}
\noindent Step II. {\em Solution of $\psi_m$ in $Y\setminus R$ for $m\geq 2$.}
We decompose $\psi_m$
\begin{eqnarray}
\psi_m=\psi_m'+(-i)^{m-1}\xi_{m-1}\hat{\psi}
\label{decompp}
\end{eqnarray}
and substitute into \eqref{summation2}.
This decomposition is chosen such that $\psi'_m$ depends on $\xi_n$ for $0\leq n \leq m-2$, the functions $\psi_n$, $n\leq m-1$ in $Y\setminus R$ and the functions $\psi_n$, $n\leq m-2$ in $R$. The function $\hat{\psi}$ depends only on $\xi_0$ and $\psi_1$ in $Y\setminus R$.
For $m\geq 2$, $\psi_m'$ in $Y\setminus R$ subject to the mean zero condition $\int_{Y\setminus R} \psi_m'\,dy=0$ is the solution of
{\footnotesize
\begin{eqnarray}
\tau^2B_z(\psi_m',v)&=&\int_{Y\setminus R} \left(F_1\cdot\nabla \overline{v}+G_1\overline{v}\right)\,d\textbf{y}\nonumber
\\&&+\int_R \left(F_2\cdot\nabla\overline{v}+G_2\overline{v}\right)\,d\textbf{y}, \quad \hbox{ for all $v$ in $H^1_{per}(Y)$},
\label{varformpsimprime}
\end{eqnarray}}
where
\begin{eqnarray}
F_1&=&\xi_0^{-1}\epsilon_p^{-1}(\xi_0)\tau^2\big[\sum^{m-2}_{l=1}(-i)^l\xi_l\nabla \psi_{m-l} +(\sum_{l=0}^{m-2}(-i)^l\xi_l\psi_{m-1-l})\hat{\kappa}\big]\nonumber\\
&&-\xi_0^{-1}\epsilon_p^{-1}(\xi_0)\tau^2\epsilon_r\frac{\omega_p^2}{c^2}\chi_H\hat{\kappa}\psi_{m-1},
\label{F1}
\end{eqnarray}
{\footnotesize
\begin{eqnarray}
G_1&=&\xi_0^{-1}\epsilon_p^{-1}(\xi_0)\tau^2\big[ \hat{\kappa}\cdot\sum_{l=0}^{m-2}(-i)^l\xi_l(-\nabla \psi_{m-1-l})-\sum_{l=0}^{m-2}(-i)^l\xi_l\psi_{m-2-l}\big]\nonumber\\ &&-\xi_0^{-1}\epsilon_p^{-1}(\xi_0)\tau^2\epsilon_r\frac{\omega_p^2}{c^2}\chi_H\big[\hat{\kappa}\cdot(-\nabla \psi_{m-1})-\psi_{m-2}\big]\nonumber
\\&&+\xi_0^{-1}\epsilon_p^{-1}(\xi_0)\big[\sum_{l=0}^{m-2}\sum_{n=0}^l \xi_{m-2-l}\xi_n \psi_{l-n} i^{l-n-m}\label{G1}
\\&&-\epsilon_r\frac{\omega_p^2}{c^2}\sum_{l=0}^{m-2}(-i)^l\xi_l \psi_{m-2-l}\big],\nonumber
\end{eqnarray}}
\begin{eqnarray}
F_2&=&-\xi_0^{-1}\epsilon_p^{-1}(\xi_0)\big[\sum^{m-2}_{l=0}(-i)^l\xi_l\nabla \psi_{m-2-l} +\hat{\kappa}(\sum_{l=0}^{m-3}(-i)^l\xi_l\psi_{m-3-l})
\big] \nonumber\\
&&+\xi_0^{-1}\epsilon_p^{-1}(\xi_0)\epsilon_r\frac{\omega_p^2}{c^2}\big[\nabla \psi_{m-2}+\hat{\kappa}(\psi_{m-3})\big],
\label{F2}
\end{eqnarray}
and
{\footnotesize
\begin{eqnarray}
G_2&=&-\xi_0^{-1}\epsilon_p^{-1}(\xi_0)\big[ +\hat{\kappa}(\sum_{l=0}^{m-3}(-i)^l\xi_l(-\nabla \psi_{m-3-l}))\nonumber
\\&&-\sum_{l=0}^{m-4}(-i)^l\xi_l\psi_{m-4-l}\big] +\xi_0^{-1}\epsilon_p^{-1}(\xi_0)\epsilon_r\frac{\omega_p^2}{c^2}\big[\hat{\kappa}\cdot(-\nabla \psi_{m-3})+\psi_{m-4}\big]\nonumber
\\&&+\xi_0^{-1}\epsilon_p^{-1}(\xi_0)\big[\sum_{l=0}^{m-2}\sum_{n=0}^l \xi_{m-2-l}\xi_n \psi_{l-n} i^{l-n-m}\overline{v}\label{G2}
\\&&-\epsilon_r\frac{\omega_p^2}{c^2}\sum_{l=0}^{m-2}(-i)^l\xi_l \psi_{m-2-l}\overline{ v}\big],\nonumber
\end{eqnarray}}
where $\chi_H$ is the indicator function of the host domain $H$ taking $1$ inside and zero outside. Here we have used the identity $\psi_0(\textbf{y})=1$, for $\textbf{y}\in Y\setminus R$ in determining the formulas for the right hand side of \eqref{varformpsimprime}.
Setting $v=1$ in \eqref{varformpsimprime} delivers the solvability condition
\begin{eqnarray}
\int_{Y\setminus R}\,G_1\,d\textbf{y}+\int_{R}\,G_2\,d\textbf{y}=0.
\label{solvabilitycondt}
\end{eqnarray}
The condition
\begin{eqnarray}
-\nabla\cdot F_2+G_2=0\hbox{ for $\textbf{y}$ in $R$}\label{divzero}
\end{eqnarray}
follows from the solution of Step I see, \eqref{decomp}, \eqref{Rproblem} and \eqref{inproblem}.
Integration by parts on the right hand side of \eqref{varformpsimprime} together with \eqref{divzero} transforms \eqref{varformpsimprime}
into the equivalent Neumann boundary value problem for $\psi'_m$, $m\geq 2$, given by
{\footnotesize
\begin{eqnarray}
\tau^2B_z(\psi_m',v)&=&\int_{Y\setminus R} \left(F_1\cdot\nabla \overline{v}+G_1\overline{v}\right)\,d\textbf{y}\label{varformpsimprimeYminusR}
\\&&+\int_{\partial R} F_2\cdot n\overline{v}\,ds, \quad\hbox{ for all $v$ in $H^1_{per}(Y\setminus R)$},\nonumber
\end{eqnarray}}
where $n$ is the outward directed unit normal to $\partial R$. Finally we observe that the solvability condition for the Neumann problem \eqref{varformpsimprimeYminusR} given by
\begin{eqnarray}
\int_{Y\setminus R}\,G_1\,d\textbf{y}+\int_{\partial R}\,F_2\cdot n\,ds=0,
\label{solvabilitycondtneuman}
\end{eqnarray}
follows immediately from \eqref{solvabilitycondt} and \eqref{divzero}. This Neumann problem satisfies the hypotheses of Theorem \ref{existencetheoremforoutsideR} and we assert the existence of a solution $\psi'_m$ for \eqref{varformpsimprimeYminusR}, uniquely determined by the condition $\int_{Y\setminus R} \psi_m'\,dy=0$, provided that $\xi_0\not=\zeta_i$, $0\leq \zeta_i\leq\frac{\epsilon_r\omega_P^2}{c^2}$, $i=1,2,\ldots$.
The field $\hat{\psi}$
solves
\begin{equation}
\begin{cases}
B_z(\hat{\psi},v)=-\xi_0^{-1}\epsilon_p^{-1}(\xi_0)\int_{Y\setminus R}(\nabla \psi_1+\hat\kappa)\cdot\nabla\overline{v}, \\
\\
\int_{Y\setminus R}\hat\psi=0 ,
\end{cases}
\label{hatpsisolution}
\end{equation}
for all trials $v\in H^1_{per}(Y\setminus R)$ .
The solution $\hat{\psi}$ is represented explicitly in terms of the eigenvectors associated with the generalized electrostatic resonances.
A straight forward calculation gives
{\footnotesize
\begin{eqnarray}
\hat{\psi}&=&\sum_{-\frac{1}{2}<\lambda_n<\frac{1}{2}}\frac{\psi_{\lambda_n}}{\xi_0-(\frac{1}{2}+\lambda_n)\frac{\epsilon_r\omega_p^2}{c^2}}\nonumber
\\&&\left(-(\alpha_{\lambda_n}^1+\alpha_{\lambda_n}^2)+\frac{(\xi_0-\frac{\epsilon_r\omega_p^2}{c^2})\left(\alpha_{\lambda_n}^1+\alpha_{\lambda_n}^2\left(\frac{\xi_0}{\xi_0-\frac{\epsilon_r\omega_p^2}{c^2}}\right)\right)}{\xi_0-s_n}\right).
\label{psihatformula}
\end{eqnarray}}
\noindent Step III. {\em Solution of $\xi_m$ for $m\geq 1$.}
We apply the decomposition \eqref{decomp} to $\psi_{m-2}$ in $R$ and \eqref{decompp} to $\psi_{m-1}$ in $Y\setminus R$. Substitution of the decompositions into the solvability condition \eqref{solvabilitycondt} and applying the explicit spectral representations \eqref{psi*}, \eqref{psihatformula} for $\psi_*$ and $\hat{\psi}$ we obtain the recursion relation for determining $\xi_1,\xi_2,\ldots$ given by
{\footnotesize
\begin{eqnarray}
\xi_{m-2}\mathcal{G}(\xi_0) &=&\tau^2\int_{Y\setminus R}\big[\hat{\kappa}\cdot (\sum_{l=1}^{m-3}(-i)^l\xi_l\nabla\psi_{m-1-l}+\xi_0\nabla \psi_{m-1}')+\sum_{l=0}^{m-3}(-i)^l\xi_l\psi_{m-2-l}\big]\nonumber
\\&&-\tau^2\int_H\epsilon_r\frac{\omega_p^2}{c^2}(\hat\kappa\cdot\nabla\psi_{m-1}'+\psi_{m-2})\nonumber
\\&&-\int_R\big[\hat{\kappa}\cdot \sum_{l=0}^{m-3}(-i)^l\xi_l\nabla\psi_{m-3-l}+\sum_{l=0}^{m-4}(-i)^l\xi_l\psi_{m-4-l}\big]\label{iterationforxim1}
\\&&-\int_Y
\big[\sum_{l=1}^{m-3}\sum_{n=0}^l i^{l-n-m}\xi_{m-2-l}\xi_n\psi_{l-n}-\sum_{l=1}^{m-3}(-i)^l\xi_0\xi_l\psi_{m-2-l}\big]\nonumber
\\&&-\int_R\xi_0^2\tilde\psi_{m-2}-\int_{Y\setminus R}\xi_0^2\psi_{m-2} +\int_Y \epsilon_r\frac{\omega_p^2}{c^2}\sum_{l=1}^{m-3}(-i)^l\xi_l\psi_{m-2-l}\nonumber
\\&&+\int_R \epsilon_r\frac{\omega_p^2}{c^2}\xi_0\tilde\psi_{m-2}+\int_{Y\setminus R} \epsilon_r\frac{\omega_p^2}{c^2}\xi_0\psi_{m-2}.\nonumber
\end{eqnarray} }
Noting that \eqref{decomp} and \eqref{decompp} can also be applied to the lower order terms on the right hand side of \eqref{iterationforxim1} we see that \eqref{iterationforxim1} determines $\xi_{m-2}$ from $\xi_n$, $\psi_n$ in $Y\setminus R$, and $\psi_n$ in $R$ with $0\leq n\leq m-3$. Here the ``solvability matrix,'' $\mathcal{G}$ depends explicitly on $\xi_0$ and is given by
\begin{eqnarray}
\mathcal{G}(\xi_0) &=&\theta_H+\theta_P\nonumber
\\&&+\tau^2\big[\int_H\mathbb{P}\hat{\kappa}\cdot\hat{\kappa}\,d\textbf{y}-\sum_{n=1}^\infty (g_n^2(\xi_0)|\alpha_{\lambda_n}^1+\epsilon^{-1}_P(\xi_0)\alpha^2_{\lambda_n}|^2)\nonumber
\\&&+2g_n(\xi_0)(\alpha_{\lambda_n}^1+\alpha^2_{\lambda_n})(\alpha_{\lambda_n}^1+\epsilon^{-1}_P(\xi_0)\alpha^2_{\lambda_n})\big]\label{B}
\\&& -2(\xi_0+\frac{\epsilon_r\omega_p^2}{c^2})\mu_{eff}(\xi_0)+\xi_0(\xi_0-\frac{\epsilon_r\omega_p^2}{c^2})\sum_{n=1}^\infty\frac{\mu_n\langle\phi_n\rangle_R^2}{(\mu_n-\xi_0)^2},\nonumber
\end{eqnarray}
where $g_n(\xi_0)=(\xi_0-\frac{\epsilon_r\omega_p^2}{c^2})/(\xi_0-s_n)$ and $\epsilon_P^{-1}(\xi_0)=\xi_0/(\xi_0-\frac{\epsilon_r\omega_p^2}{c^2})$.
The set of poles and zeros for $\mathcal{G}(\xi_0)$ are explicitly controlled by the generalized electrostatic resonances and Dirichlet spectra. The zeros of $\mathcal{G}$ are denoted by $\gamma_n$, $n=1,\ldots$. It is now evident that the recursion relation \eqref{iterationforxim1} holds provided that $\xi_0\neq \nu_n$, $\xi_0\neq \zeta_n$, and $\xi_0\not=\gamma_n$.
Denote the union of the Dirichlet spectra, the electrostatic resonances, and zeros of $\mathcal{G}$ by
\begin{eqnarray}
U\equiv\{\nu_j\}_{j=1}^\infty\cup\{\zeta_j\}_{j=1}^\infty\cup\{\gamma_j\}_{j=1}^\infty
\label{unionofspectra}
\end{eqnarray}
where $\{\nu_j\}_{j=1}^\infty\subset \mathbb{R}^+$, $\{\gamma_j\}_{j=1}^\infty\subset \mathbb{R}^+$, and $\{\zeta_j\}_{j=1}^\infty\subset [0,\frac{\epsilon_r\omega_p^2}{c^2}]$.
We collect results and state the following theorem.
\begin{theorem}
If $\xi_0$ belongs to $\mathbb{R}^+\setminus U$ then the sequences $\{\psi_m\}_{m=0}^\infty\subset H^1_{per}(Y)$, $\{\xi_m\}_{m=0}^\infty\subset \mathbb{C}$ exist and are uniquely determined.
\label{Existencesequence}
\end{theorem}
\begin{proof}
The functions $\psi_0$ in $H^1(Y)$ and $\psi_1$ in $H^1(Y\setminus R)$ with $\int_{Y\setminus R}\psi_1\,dy=0$ exist as does $\xi_0$. Application of Step I uniquely determines $\tilde{\psi_1}$ in $R$. Application of Step II uniquely determines $\psi'_2$ in $Y\setminus R$. Step III uniquely determines $\xi_1$ and we apply \eqref{decomp} and \eqref{decompp} to recover $\psi_1$ in $H^1(Y)$ and $\psi_2$ in $H^1(Y\setminus R)$ with $\int_{Y\setminus R}\psi_2\,dy=0$.
We now adopt the induction hypotheses $H_n$, $n\geq 2$:\\
There exist\\
1) $\psi_n\in H^1_{per}(Y\setminus R)$ with $\int_{Y\setminus R}\psi_n=0$ ;\\
2) $\psi_{j}\in H^1_{per}(Y) $ for $0\leq j\leq n-1$ with $\int_{Y\setminus R}\psi_{j}=0$, $1\leq j\leq n-1$;\\
3) $\xi_{n-1} \in \mathbb{C}$.\\
Application of Step I uniquely determines $\tilde{\psi_n}$ in $R$. Application of Step II uniquely determines $\psi'_{n+1}$ in $Y\setminus R$. Step III uniquely determines $\xi_n$ and we apply \eqref{decomp} and \eqref{decompp} to recover $\psi_n$ in $H^1(Y)$ and $\psi_{n+1}$ in $H^1(Y\setminus R)$ with $\int_{Y\setminus R}\psi_{n+1}\,dy=0$.
\end{proof}
\section{Band structure for the metamaterial crystal}
\label{bandstructure}
In this section we present
explicit formulas describing the band structure for the metamaterial crystal in the dynamic, sub-wavelength regime $1>\eta>0$.
The formulas show that the frequency intervals associated with pass bands and stop bands are governed by the poles and zeros of the effective magnetic permittivity and dielectric permittivity tensors. The poles and zeros are explicitly determined by the Dirichlet spectra of $R$ and the generalized electrostatic resonances of $Y\setminus R$, see \eqref{effperm} and \eqref{effdielectricconst}. For the general class of parallel rod configurations treated here the pass bands and stop bands can exhibit anisotropy, i.e., dependence on the direction of propagation described by $\hat{\kappa}$. The anisotropy is governed
by the projection of $\hat{\kappa}$ onto the eigenfunctions associated with the generalized electrostatic resonances given by \eqref{coefficients}.
In what follows pass bands are explicitly linked to frequency intervals and propagation directions for which both the effective magnetic permittivity and dielectric permeability have the same sign. This includes frequency intervals where effective tensors are either simultaneously positive or negative.
\par
\begin{figure}[h!]
\begin{center}
\begin{psfrags}
\psfrag{a1}{$a'$}
\psfrag{b1}{$b'$}
\psfrag{a2}{$a''$}
\psfrag{b2}{$b''$}
\psfrag{tao}{$\tau^2$}
\psfrag{xi}{$\xi_0$}
\psfrag{I1}{$I_{n'}$}
\psfrag{I2}{$I_{n''}$}
\includegraphics[width=0.4\textwidth]{figdispersion.eps}
\end{psfrags}
\caption{The leading order dispersion relation over two selected intervals $I_{n'}$ and $I_{n''}$.}
\label{tauxirelation}
\end{center}
\end{figure}
\par
We begin by identifying the locations of the branches for the dispersion relation associated with the metamaterial crystal. We introduce the union of open intervals $\bigcup_n O_n$ on the positive real axis $\mathbb{R}^+$ obtained by removing the points $\{\zeta_j\}_{j=1}^\infty$, $\{\nu_j\}_{j=1}^\infty$, $\{\gamma_j\}_{j=1}^\infty$, $\{s_j^*\}_{j=1}^\infty$ and $\{\mu_j^*\}_{j=1}^\infty$. Here $s_j^*$ is the zero of $\epsilon_{eff}^{-1}(\xi_0)\hat\kappa\cdot\hat\kappa$ between $s_j$ and $s_{j+1}$ and $\mu_j^*$ is the zero of $\mu_{eff}(\xi_0)$ between $\mu_j$ and $\mu_{j+1}$. The leading order dispersion relation \eqref{subwavelengthdispersion1alternate}
implicitly defines the map $(\tau,\hat\kappa)\mapsto\xi_0$. We denote the branch associated with $O_n$ by $\xi_0^n=\xi_0^{(n)}(\tau,\hat\kappa)$ . Let $I_n$ be an open interval strictly contained inside $O_n$. For this choice $I_n\subset O_n$ does not intersect the union of the Dirichlet spectra, generalized electrostatic spectra and the zeros of the solvability matrix $\mathcal{G}$, see \eqref{unionofspectra}.
The band structure for the metamaterial in the dynamic sub-wavelength regime $1>\eta>0$ is given by the following theorem.
\par
\begin{theorem}
The dispersion relation for the metamaterial crystal contains an infinite sequence of branches with leading order behavior given by the
functions $\xi_0^n=\xi_0^n(\tau,\hat{\kappa})$. For $\tau$ such that $\xi_0^n$ belongs to $I_n$ there exists a constant $R$ depending only on $I_n$ such that for $0<\rho< R$, the branch of the dispersion relation for the metamaterial crystal is given by
\begin{eqnarray}
\xi=\xi_0^n(\tau,\hat\kappa)+\sum_{l=1}^{\infty}(\rho\tau)^{l}\xi_{l}^n,
\label{xipowerseries}
\end{eqnarray}
for
\begin{eqnarray}
\{-2\pi\leq \tau\rho\hat{\kappa}_1\leq 2\pi , -2\pi\leq \tau\rho\hat{\kappa}_2\leq 2\pi\}.
\label{brullionfortau}
\end{eqnarray}
Here the higher order terms $\xi_{l}^n$ are real and are uniquely determined by $\xi_0^n$ according to Theorem \ref{Existencesequence}.
\label{convergenceofxi}
\end{theorem}
The power series representation for the transverse magnetic Bloch wave \eqref{em3} is given by the following theorem.
\begin{theorem}
For $I_n$ and $R$ as in Theorem \ref{convergenceofxi} and for $\tau$ such that $\xi_0^n$ belongs to $I_n$ there exist transverse magnetic Bloch waves given by the expansion
\begin{eqnarray}
H_3=\underline u_0\left(\psi_0(\textbf{x}/d)+\sum_{l=1}^{\infty}(\rho\tau)^li^l\psi_l(\textbf{x}/d)\right)exp\left\{i\left(k\hat\kappa\cdot \textbf{x}-t\omega\right)\right\},
\end{eqnarray}
where the series
\begin{eqnarray}
\psi_0(\textbf{y})+\sum_{l=1}^{\infty}(\rho\tau)^li^l\psi_l(\textbf{y})
\label{convgncee}
\end{eqnarray}
is summable in $H^1(Y)$ for $0\leq\rho<R$ and $\frac{\omega}{c}=\sqrt{\frac{\xi}{\epsilon_r}}$.
\label{summable}
\end{theorem}
Theorem \ref{convergenceofxi} explicitly shows that the leading order behavior determines the existence of pass bands or stop bands when $\rho=d/\sqrt{\epsilon_r}$ is sufficiently small. With this in mind we can appeal to \eqref{subwavelengthdispersion1alternate} and state the following theorem.
\begin{theorem}
\label{bands}
For $\rho>0$ sufficiently small:
\begin{itemize}
\item There exist propagating Bloch wave solutions along directions $\hat{\kappa}$ over intervals $I_n$ for which $\mu_{eff}(\xi_0^n)$ and $\epsilon_{eff}^{-1}(\xi_0^n)\hat{\kappa}\cdot\hat{\kappa}$ have the same sign.
\item
Intervals $I_n$ for which $\mu_{eff}(\xi_0^n)$ and $\epsilon_{eff}^{-1}(\xi_0^n)\hat{\kappa}\cdot\hat{\kappa}$ have the opposite sign lie within stop bands.
\end{itemize}
\end{theorem}
Leading order behavior for pass bands associated with double negative and double positive behavior are illustrated in Fig. \ref{tauxirelation} for two intervals $I_{n'}$ and $I_{n''}$. Here the effective properties are both negative over the interval $I_{n'}$ while they are both positive over $I_{n''}$.
\section {Generalized electrostatic spectra}
\label{genrealizedelectrostaticspectra}
In this section we establish Theorem \ref{completeeigen} and characterize all pairs $(\lambda,\psi)$ in $\mathbb{C}\times H_{per}^1(Y\setminus R)$ satisfying \eqref{resonance}. We start by recalling the bilinear form in \eqref{resonance} and forming the quotient
\begin{eqnarray}
Q(u)=\frac{-\frac{1}{2}\int_P|\nabla u|^2\,dx+\frac{1}{2}\int_H|\nabla u|^2\,dx}{(u,u)}.
\label{quotient}
\end{eqnarray}
From \eqref{quotient} it is evident that $\lambda$ lies inside $[-1/2,1/2]$.
It is easily seen that the solutions $(\lambda,\psi)$ of \eqref{resonance} for the choices $\lambda=\frac{1}{2}$ and $\lambda=-\frac{1}{2}$ correspond to $\psi$ in $W_1$ and $W_2$ respectively. Only the $0$ element of $W_3$ satisfies \eqref{resonance} for the choices $\lambda=\frac{1}{2}$ and $\lambda=-\frac{1}{2}$.
We now investigate solutions $(\lambda,\psi)$ for $\psi$ belonging to $W_3$.
The bilinear form \eqref{resonance} defines a map $T$ from $W_3$ into the space of skew linear functionals on $W_3$. Here $W_3$ is a Hilbert space and
\begin{eqnarray}
(Tu,v)=-\frac{1}{2}\int_P\nabla u\cdot\nabla \overline{v}\,d\textbf{y}+\frac{1}{2}\int_H\nabla u\cdot\nabla\overline{v}\,d\textbf{y},
\label{resonanceT}
\end{eqnarray}
for $u$ and $v$ in $W_3$.
In what follows we show that $T$ is a compact map from $W_3$ onto $W_3$ with eigenvalues contained inside the open interval $(-1/2,1/2)$.
We do this by providing an explicit representation for $T$ given in terms of a composition of single and double layer potentials.
Set $\mathbb{D}=\cup_{n\in\mathbb{Z}^2}((Y\setminus R)+n)$ and introduce the Green's function $G(\textbf{x},\textbf{y})$ on $\mathbb{D}$ such that:
$G$ is separately periodic in $\textbf{x}$ and $\textbf{y}$ with period $Y$, $G$ is $C^2$ in each of the variables $\textbf{x}$ and $\textbf{y}$ for $\textbf{x}\not=\textbf{y}$, and
{\footnotesize
\begin{eqnarray}
\Delta_\textbf{x} G(\textbf{x},\textbf{y})&=&\sum_{n\in \mathbb{Z}^2}\delta(\textbf{x}-\textbf{y}+n) -1\quad\text{ in } \mathbb{D},\nonumber
\\\partial_{n_\textbf{x}}G(\textbf{x},\textbf{y})&=&\frac{|R|}{|\partial R|},\quad \text{ for } \textbf{x}\in \cup_{n\in\mathbb{Z}^2} (\partial R+n),
\label{greensfunction}
\end{eqnarray}}
where $\delta_\textbf{x}$ is the Dirac delta function located at $\textbf{x}$ and $|R|$ and $|\partial R|$ are the area and arc-length of $R$ and $\partial R$ respectively.
We can find $G(\textbf{x},\textbf{y})$ as a sum of the periodic free space Green's function $F$ for the Laplacian and a corrector $\phi^*$: $G(\textbf{x},\textbf{y})=F(\textbf{x},\textbf{y})+\phi^*(\textbf{x},\textbf{y})$ , where
\begin{eqnarray}
F(x,y)=-\sum_{n\in \mathbb Z^2\setminus\{0\}}\frac{e^{i2\pi n\cdot(x-y)}}{4\pi^2|n|^2},
\label{freespacegreens}
\end{eqnarray}
$\phi^*(\textbf{x},\textbf{y})$ is periodic in $\textbf{x}$ and $\textbf{y}$ with period $Y$, $C^2$ in $\textbf{x}$, $\textbf{y}$, and solves
\begin{eqnarray}
&&
\Delta_\textbf{x} \phi^*(\textbf{x},\textbf{y})=0 ~\hbox{ for $\textbf{x}$, $\textbf{y}$ in } \mathbb{D} ,\nonumber\\
&& ~~\partial_{n_\textbf{x}}\phi^*(\textbf{x},\textbf{y})_{{|_{\partial R}^+}}=-\partial_{n_\textbf{x}}F(\textbf{x},\textbf{y})_{{|_{\partial R}^+}} +\frac{|R|}{|\partial R|}, \hbox{for $\textbf{x}$ on $\partial R$} ,\label{pdeforphi}\nonumber\\
&&\hbox{and }\int_{\partial R}\phi^*(\textbf{x},\textbf{y})dS_\textbf{x}=0.
\label{correctorproblem}
\end{eqnarray}
Here the subscript ${\partial R}^+$, indicates traces of functions on $\partial R$ taken from the outside of $R$.
The space of mean zero square integrable functions defined on $\partial P$ is denoted by $L^2_0(\partial P)$
and for $\phi$ in $L^2_0(\partial P)$ we introduce the single layer potential $\tilde{S}_P$ given by
\begin{eqnarray}
\tilde{S}_P(\phi)=\int_{\partial P} G(\textbf{x},\textbf{y})\phi(\textbf{y})dS_\textbf{y},
\label{singlelayer}
\end{eqnarray}
and the modified single layer potential $S_P$ mapping $L^2_0(\partial P)$ into $W_3$
given by
\begin{eqnarray}
{S}_P(\phi)=\tilde{S}_P(\phi)-|\partial P|^{-1}\int_{\partial P} \tilde{S}_P(\phi)(\textbf{y})dS_\textbf{y}.
\label{singlelayerzeroavg}
\end{eqnarray}
On restricting $S_P(\phi)(\textbf{x})$ to $\textbf{x}$ on $\partial P$, one readily verifies that $S_P$ is self-adjoint with respect to the $L^2(\partial P)$ inner product and is a bounded one to one linear map from $L_0^2(\partial P)$ onto
$H^{1/2}(\partial P)\cap L^2_0(\partial P)$. The inverse denoted by $S_{\partial P}^{-1}:H^{1/2}(\partial P)\cap L_0^2(\partial P)\rightarrow L_0^2(\partial P)$ is linear and continuous. We introduce the trace operator $\gamma$ mapping $W_3$ into $H^{1/2}(\partial P)\cap L_0^2(\partial P)$, this map is bounded and onto \cite{Adams}.
Collecting results we have the following
\begin{lemma}
$S_p$ is a one to one, bounded linear transformation from $L^2_0(\partial P)$ onto $W_3$ with $S^{-1}_P: W_3\rightarrow L_0^2(\partial P)$ linear and continuous given by $S_P^{-1}=S_{\partial P}^{-1}\gamma$.
\end{lemma}
\begin{proof}
Given $u$ in $W_3$ consider its trace on $\partial P$ given by $\gamma u=u_{{|_{\partial P}}}$ and $u_{{|_{\partial P}}}$ belongs to $H^{1/2}(\partial P)\cap L^2_0(\partial P)$. For $\textbf{x}$ in $Y\setminus R$ set $w(\textbf{x})=S_P(S_{\partial P}^{-1}( u_{{|_{\partial P}}}))(\textbf{x})$. Since $w$ belongs to $W_3$ the difference $w-u$ also belongs to $W_3$. Noting further that the traces of $w$ and $u$ agree on $\partial P$ we conclude that $w-u$ also belongs to $W_1\oplus W_2$. However since $W_3=(W_1\oplus W_2)^\perp$ it is evident that $w-u=0$ and we conclude that $S_P^{-1}=S_{\partial P}^{-1}\gamma$.
\end{proof}
The outward pointing normal derivative of a quantity $q$ on $\partial P$ is denoted by $\frac{\partial q}{\partial{n_\textbf{x}}}$ and the jump relations satisfied by the single layer potential are given by
\begin{eqnarray}
S_P(\phi)_{{|_{\partial P}^+}}=S_P(\phi)_{{|_{\partial P}^-}}, \hbox{ on $\partial P$},
\label{cont}
\frac{\partial S_P(\phi)}{\partial{n_\textbf{x}}}{{|_{{\partial P}\stackrel{+}{-}}}}=\pm\frac{1}{2}\phi+K_P^*(\phi), \hbox{ on $\partial P$}
\label{jump}
\end{eqnarray}
where the double layer $K_P^*$ is a bounded linear operator mapping $L_0^2(\partial P)$ into $L_0^2(\partial P)$ defined by
\begin{eqnarray}
K^*_P(\phi)=\int_{\partial P} \frac{\partial G(\textbf{x},\textbf{y})}{\partial n_\textbf{x}}\phi(\textbf{y})dS_\textbf{y},
\label{double}
\end{eqnarray}
and the subscripts $+$ and $-$ indicate traces from the outside and inside of $P$ respectively. Here $\frac{\partial G}{\partial n_\textbf{x}}$ is a continuous kernel of order zero and it follows \cite{Folland,ColtonKress} that $K^*_P$ is a compact operator mapping $L_0^2(\partial P)$ into $L_0^2(\partial P)$.
Now we identify the transform $T$ defined by \eqref{resonanceT}.
\begin{theorem}
\label{T}
The linear map $T:W_3\rightarrow W_3$ defined by the sesquilinear form \eqref{resonanceT} is given by
\begin{eqnarray}
T=S_P K_P^*S_P^{-1},
\label{ident}
\end{eqnarray}
and is a compact bounded self-adjoint operator on $W_3$ with eigenvalues lying inside $(-1/2,1/2)$.
\end{theorem}
\noindent Since $T$ is self-adjoint, bounded, and compact: the spectrum is discrete with only zero as an accumulation point, the eigenspaces associated with distinct eignevalues are pairwise orthogonal, finite dimensional, and their union together with the null space of $T$ is $W_3$.
Theorem \ref{completeeigen} follows immediately from Theorem \ref{T} noting that the choices $\lambda=1/2$, $\lambda=-1/2$ in \eqref{resonance} correspond to $\phi$ belonging to $W_1$ and $W_2$ respectively.
We now prove Theorem \ref{T}.
\begin{proof}
\noindent Step I.
We establish \eqref{ident}. For $u$, $v$, consider
\begin{eqnarray}
(S_P K_p^*S_P^{-1}(u),v)=\int_{Y\setminus R}\nabla S_PK_p^*S_P^{-1}(u)\cdot\nabla \overline{v}\,dx.
\label{startup}
\end{eqnarray}
Integration by parts gives
\begin{eqnarray}
(S_PK_p^*S_P^{-1}(u),v)=\int_{\partial P}\left[\frac{\partial(S_PK_p^*S_P^{-1}(u))}{\partial n_\textbf{x}}\right]_+^-\overline{v}\,dS,
\label{startup1}
\end{eqnarray}
where $[q]_+^-=q|_{{\partial P}^-}-q|_{{\partial P}^+}$. Applying the jump condition \eqref{jump} we get
\begin{eqnarray}
(S_PK_p^*S_P^{-1}(u),v)=-\int_{\partial P}K_p^*S_P^{-1}(u)\overline{v}\,dS,
\label{startup2}
\end{eqnarray}
applying \eqref{jump} again gives
\begin{eqnarray}
K_p^*S_P^{-1}(u)&=&\frac{1}{2}\frac{\partial S_PS_P^{-1}(u)}{\partial n_\textbf{x}}|_{{\partial P}^-}+\frac{1}{2}\frac{\partial S_PS_P^{-1}(u)}{\partial n_\textbf{x}}|_{{\partial P}^+}\nonumber\\
&=&\frac{1}{2}\frac{\partial u}{\partial n_\textbf{x}}|_{{\partial P}^-}+\frac{1}{2}\frac{\partial u}{\partial n_\textbf{x}}|_{{\partial P}^+}.
\label{startup3}
\end{eqnarray}
Step I now follows on substitution of \eqref{startup3} into \eqref{startup2} and integration by parts.
\noindent Step II.
The remaining properties of the operator $T$ follow directly from the representation $T=S_PK_P^*S_P^{-1}$ and the properties of $S_P$ and $K_P^*$.
\end{proof}
We conclude noting that the spectrum of $T$ is the same as the spectrum of $K_P^*$. This is stated in the following Lemma.
\begin{lemma}
\label{samespectra}
$Tu=\lambda u$ if and only if $\lambda$ corresponds to an eigenvalue of $K_P^*$.
\end{lemma}
\begin{proof}
If a pair $(\lambda,u)$ belonging to $(-1/2,1/2)\times W_3$ satisfies $Tu=\lambda u$ then $S_PK_P^*S_P^{-1}u=\lambda u$. Multiplication of both sides by $S_P^{-1}$ shows that $S_P^{-1}u$ is an eigenfunction for $K_P^*$ associated with $\lambda$.
Suppose the pair $(\lambda,w)$ belongs to $(-1/2,1/2)\times L_0^2(\partial P)$ and satisfies $K_P^*w=\lambda w$. Since the trace map from $W_3$ to $H^{1/2}(\partial P)\cap L_0^2(\partial P)$ is onto then there is a $u$ in $W_3$ for which $w=S_P^{-1}u$ and $K_P^*S_P^{-1}u=\lambda S_P^{-1}u$. Multiplication of this identity by $S_P$ shows that $u$ is an eigenfunction for $T$ associated with $\lambda$.
\end{proof}
\section{Existence for exterior problems with a dielectric permitivity in $\mathbb{C}$}
\label{exist}
In this section we establish Theorem \ref{existencetheoremforoutsideR}.
Recall that any element $u$ of $H_{per}^1(Y\setminus R)/\mathbb{C}$ can be written as
\begin{eqnarray}
u=\mathcal{P}_1 u + \mathcal{P}_2 u +\sum_{-\frac{1}{2}<\lambda_n<\frac{1}{2}} \mathcal{P}_{\lambda_n} u +D.
\label{expandHper}
\end{eqnarray}
where $D$ is chosen such that $\int_{Y\setminus R} u \, dy=0$. Here $\mathcal{P}_i$ are the orthgonal projections onto $W_i$, $i=1,2$ and $\mathcal{P}_{\lambda_n}$ are the orthogonal projections associated with $\lambda_n$ in $(-1/2,1/2)$.
The orthogonal decomposition \eqref{expandHper} is used in the proof of Theorem \ref{existencetheoremforoutsideR} given below.
\begin{proof}
For $u$, $v$ in $H_{per}^1(Y\setminus R)/\mathbb{C}$ we apply \eqref{expandHper} to see that
{\footnotesize
\begin{eqnarray}
B_z(u,v) &=&\sum_{i=1}^2 B_z(\mathcal{P}_i u,\mathcal{P}_i v)+\sum_{-\frac{1}{2}<\lambda_n<\frac{1}{2}} B_z(\mathcal{P}_{\lambda_n} u, \mathcal{P}_{\lambda_n} v)\nonumber\\
&=&(\mathcal{P}_1 u,v)+z(\mathcal{P}_2 u,v)+\sum_{-\frac{1}{2}<\lambda_n<\frac{1}{2}} (1+(z-1)(\frac{1}{2}-\lambda_n))(\mathcal{P}_{\lambda_n} u,v)\nonumber\\
&=&(T_z u,v).
\label{bilinearform}
\end{eqnarray}}
From \eqref{bilinearform} we conclude that
\begin{eqnarray}
T_z =\mathcal{P}_1 +z \mathcal{P}_2 +\sum_{-\frac{1}{2}<\lambda_n<\frac{1}{2}} (1+(z-1)(\frac{1}{2}-\lambda_n))\mathcal{P}_{\lambda_n}.
\label{Tz2}
\end{eqnarray}
It is evident from \eqref{Tz2} that for $z\not=(\lambda_i+1/2)/(\lambda_i-1/2)$ that $T_z$ is a bounded one to one and onto map in $H_{per}^1(Y\setminus R)/\mathbb{C}$. The formula for $T_z^{-1}$ is given by
\begin{eqnarray}
T_z^{-1}=\mathcal{P}_1 +z^{-1} \mathcal{P}_2 +\sum_{-\frac{1}{2}<\lambda_n<\frac{1}{2}} (1+(z-1)(\frac{1}{2}-\lambda_n))^{-1}\mathcal{P}_{\lambda_n}.
\label{tzinverse}
\end{eqnarray}
Taking conjugates on both sides of \eqref{bilinearform} gives $\overline{B_z(u,v)}=(v,T_z u)$ and choosing $u=T_z^{-1}q$ for $q$ in $H_{per}^1(Y\setminus R)/\mathbb{C}$ delivers the identity
\begin{eqnarray}
\label{conjidenty}
\overline{B_z(T_z^{-1}q,v)}=(v,q).
\end{eqnarray}
To complete the proof consider any linear functional $F$ in $[H_{per}^1(Y\setminus R)/\mathbb{C}]^*$ with $F(v)=0$ for $v=const.$
Applying the Reisz representation theorem shows that there exists a unique solution $u$ in $H_{per}^1(Y\setminus R)/\mathbb{C}$ of
\begin{eqnarray}
B_z(u,v)=\overline{F(v)}, \hbox{ for all $v$ in $H_{per}^1(Y\setminus R)/\mathbb{C}$},
\label{finishproof}
\end{eqnarray}
\end{proof}
\noindent and Theorem \ref{existencetheoremforoutsideR} is proved.
\section{Convergence of the Power Series}
\label{convergence}
In this section we show that the series \eqref{xipowerseries}, \eqref{convgncee} identified in Theorems \ref{convergenceofxi} and \ref{summable} are convergent. Here the convergence radius $R$ depends upon the branch of the quasistatic dispersion relation $\xi_0^n=\xi_0^n(\tau,\hat{\kappa})$. The methodology applied here uses generating functions and follows the approach presented in \cite{FLS2}.
\noindent{Step I. {\em A priori estimates.}}
We establish a priori bounds for the solutions of \eqref{Rproblem} and \eqref{varformpsimprimeYminusR}. Before proceeding recall the Friderich's inequality and trace estimate satisfied by elements $\psi$ of
$H^1_{per}(Y\setminus R)/\mathbb{C}$ given by
\begin{eqnarray}
\Vert\psi\Vert_{Y\setminus R}=\left(\int_{Y\setminus R} |\nabla \psi|^2+|\psi|^2\,dy\right)^{1/2}\leq \Omega\Vert\psi\Vert
\label{Fredrichs}
\end{eqnarray}
and
\begin{eqnarray}
\Vert \psi\Vert_{H^{1/2}(\partial R)}\leq A \Vert\psi\Vert
\label{trace}
\end{eqnarray}
where $\Omega$ and $A$ are positive constants depending on $R$ and $\Vert\psi\Vert$ is the norm on $H_{per}^1(Y\setminus R)/\mathbb{C}$ defined by \eqref{innerproduct}.
The solutions $\tilde{\psi}$ of \eqref{Rproblem} solve boundary value problems of the following generic form. Given $\psi\in H^1_{per}(Y\setminus R)$, $G \in L^2(R) $ and $F\in [L^2(R)]^2$ the function $\tilde{\psi}$ is the $H^1(R)$ solution of
\begin{equation}
\begin{cases}
\int_R (\nabla\tilde{\psi}\cdot\nabla \overline v-\xi_0 \psi\overline v)=\int_R G\overline v+\int_R F\cdot\nabla \overline v ~~~~~~ \forall v\in H_0^1(R)\\
\tilde{\psi}|_{\partial R-}=\psi|_{\partial R+} .
\end{cases}
\end{equation}
Decompose $\tilde{\psi}$ according to $\tilde{\psi}=\psi_0+\psi_1+\psi_2$ such that
{\footnotesize
\begin{equation}
\begin{cases}
\int_R (\nabla\psi_0\cdot\nabla \overline v-\xi_0 \psi_0\overline v)=\int_R \left [G+(1+\xi_0)\psi_1
+(1+\xi_0)\psi_2\right]\overline v ~~\forall v\in H_0^1(R)\\
\\
\psi_0|_{\partial R-}=0,
\end{cases}
\label{psi_0}
\end{equation}}
\begin{equation}
\begin{cases}
\int_R (\nabla\psi_1\cdot\nabla \overline v +\psi_1\overline v)=0 ~~~\forall v\in H_0^1(R)\\
\\
\psi_1|_{\partial R-}=\psi|_{\partial R+},
\end{cases}
\label{psi_1}
\end{equation}
and
\begin{equation}
\begin{cases}
\int_R (\nabla\psi_2\cdot\nabla \overline v + \psi_2\overline v)=\int_R F\cdot\nabla \overline v ~~~~~~ \forall v\in H_0^1(R)\\
\\
\psi_2|_{\partial R-}=0.
\end{cases}
\label{psi_2}
\end{equation}
Let $J=G+(1+\xi_0)\psi_1+(1+\xi_0)\psi_2$ denote the right hand side for (\ref{psi_0}).
Expanding $\psi_0$ and $J$ with respect to the complete orthonormal set of Dirichlet eigenfunctions $\{\phi_j\}_{j=1}^\infty$, $\{\nu_j\}_{j=1}^\infty$ gives
\begin{eqnarray}
\Vert \psi_0\Vert_{H^1(R)}^2=\sum_{j=1}^\infty(1+\nu_j)\left|\frac{\langle J \psi_j\rangle_R}{(\xi_0-\nu_j)}\right|^2\leq C_{1,n}^2\Vert J\Vert_{L^2(R)}^2,
\label{expr}
\end{eqnarray}
where
\begin{eqnarray}
C_{1,n}=\max_{\xi_0\in I_n}\left\{max_j \left\{\frac{(1+\nu_j)^{1/2}}{|\xi_0-\nu_j|}\right\}\right\}.
\label{intest}
\end{eqnarray}
Collecting results gives
\begin{eqnarray}
||\psi_1||_{H^1(R)}&=& ||\psi_2||_{H^{1/2}(\partial R)}\leq A ||\psi||\nonumber
\\||\psi_2||_{H^1(R)}&\leq &||F||_{L^2(R)} \nonumber
\\||\psi_0||_{H^1(R)}&\leq & C_{1,n}||G+(1+\xi)\psi_1+(1+\xi)\psi_2||_{L^2(R)}
\end{eqnarray}
and
\begin{eqnarray}
||\tilde{\psi}||_{H^1(R)}\leq B_n( ||G||_{L^2(R)}+||F||_{L^2(R)}+||\psi||),
\label{boundnessforpsiinR}
\end{eqnarray}
where
\begin{eqnarray}
B_n=\max_{\xi_0\in I_n}\{ C_{1,n} ,C_{1,n} (1+\xi_0), A(C_{1,n}(1+\xi_0)+1)\}.
\label{Kn}
\end{eqnarray}
Now we estimate the field $\psi_*$. The explicit expression (\ref{psi*}) for $\psi_*$ gives
\begin{eqnarray}
||\psi_*||^2_{H^1(R)}=\sum_{n=1}^\infty\frac{(1+\mu_n)\mu_n^2<\phi_n>^2_R}{(\mu_n-\xi_0)^4},
\label{psi*norm}
\end{eqnarray}
and
\begin{eqnarray}
||\psi_*||_{H^1(R)}\leq L_n.
\label{boundnessforpsistarinR}
\end{eqnarray}
where $L_{n}$ is the maximum of the right hand side of \eqref{psi*norm} for $\xi_0(\tau,\hat\kappa )\in I_n$.
Next we provide an upper bound for the solutions of $\psi_m'$ of \eqref{varformpsimprimeYminusR}. We may continuously extend elements $v$ in $H_{per}^1(Y\setminus R)/\mathbb{C}$ onto $R$ as $H_{per}^1(Y)$ functions such that the extension $v^{ext}$ satisfies
\begin{eqnarray}
\Vert v^{ext}\Vert_Y\leq C\Vert v\Vert.
\label{extchion}
\end{eqnarray}
In what follows we continue to denote these extensions by $v$ and \eqref{varformpsimprimeYminusR} takes the equivalent form \eqref{varformpsimprime}. Application of H\"olders inequality to the right hand side of \eqref{varformpsimprime} gives
{\footnotesize
\begin{eqnarray}
&&\tau^2|B_z(\psi_m',v)|\nonumber
\\&&\leq C\left(\Vert F_1\Vert_{L^2(Y\setminus R)} +\Vert G_1\Vert_{L^2(Y\setminus R)} +\Vert F_2\Vert_{L^2(R)} +\Vert G_2\Vert_{L^2(R)}\right)\Vert v\Vert.
\label{upest1}
\end{eqnarray}}
\par
It is evident that $B_z(\psi_m',v)=(T_z \psi_m',v)$ and choosing $v=T_{\overline{z}}^{-1}\psi_m'$ gives $B_z(\psi_m',v)=\Vert \psi_m'\Vert^2, \Vert v\Vert\leq G_z\Vert \psi_m'\Vert$ and delivers the estimate
{\footnotesize
\begin{eqnarray}
\tau^2\Vert\psi_m'\Vert^2&=&
\tau^2|B_z(\psi_m',v)|\nonumber\\
&\leq& C\times G_z(\Vert F_1\Vert_{L^2(Y\setminus R)} +\Vert G_1\Vert_{L^2(Y\setminus R)} \label{Blowerestimate}
\\&&+\Vert F_2\Vert_{L^2(R)} +\Vert G_2\Vert_{L^2(R)})\Vert \psi_m'\Vert\nonumber
\end{eqnarray}}
with
\begin{eqnarray}
G_z=\max\left\{1,|z|^{-1},\sup_{-\frac{1}{2}<\lambda_n<\frac{1}{2}}\{|1+(z-1)(1/2-\lambda_n)|\}^{-1}\right\}.
\label{constzlowB}
\end{eqnarray}
For $z=\epsilon_p(\xi_0)$ we maximize \eqref{constzlowB} for $\xi_0\in I_n$ to obtain
{\footnotesize
\begin{eqnarray}
\tau^2\Vert\psi_m'\Vert\leq G_n\left(\Vert F_1\Vert_{L^2(Y\setminus R)} +\Vert G_1\Vert_{L^2(Y\setminus R)} +\Vert F_2\Vert_{L^2(R)} +\Vert G_2\Vert_{L^2(R)}\right).
\label{upperprimeest1}
\end{eqnarray}}
Applying Freidrich's inequality we arrive at the desired upper bound given by
{\footnotesize
\begin{eqnarray}
\tau^2\Vert\psi_m'\Vert_{Y\setminus R}\leq E_n\left(\Vert F_1\Vert_{L^2(Y\setminus R)} +\Vert G_1\Vert_{L^2(Y\setminus R)} +\Vert F_2\Vert_{L^2(R)} +\Vert G_2\Vert_{L^2(R)}\right),
\label{boundnessforpsiooutsideR}
\end{eqnarray}}
where $E_n$ denotes a generic constant depending only on $I_n$.
The a priori estimate for $\hat{\psi}$ follows from the representation formula \eqref{psihatformula}
and
\begin{eqnarray}
\Vert \hat{\psi}\Vert_{Y\setminus R}\leq H_n,
\label{hatpsiest}
\end{eqnarray}
where the constant $H_n$ depends only on $I_n$.
\noindent{Step II. \em{System of inequalities.}}
We apply the a priori estimates developed in Step I to the system of equations \eqref{Rproblem}, \eqref{varformpsimprimeYminusR}, and solvability conditions \eqref{iterationforxim1} derived in section \ref{higherorder}.
From \eqref{varformpsimprime} and (\ref{boundnessforpsiooutsideR}), it follows that
{\footnotesize
\begin{eqnarray}
\lefteqn{\tau^2\Vert\psi_m'\Vert_{Y\setminus R}\leq}\nonumber
\\ &&E_{n}\big[\tau^2(\sum_{l=1}^{m-2}|\xi_l|\Vert\psi_{m-l}\Vert_{Y\setminus R}+\sum_{l=1}^{m-2}|\xi_l|\Vert\psi_{m-1-l}\Vert_{Y\setminus R}+\sum_{l=1}^{m-2}|\xi_l|\Vert\psi_{m-2-l}\Vert_{Y\setminus R}\nonumber
\\&&+\Vert\psi_{m-1}\Vert_{Y\setminus R}+\Vert\psi_{m-2}\Vert_{Y\setminus R})+\sum_{l=0}^{m-2}|\xi_l|\Vert\psi_{m-2-l}\Vert_{H^1(R)}+\sum_{l=0}^{m-3}|\xi_l|\Vert\psi_{m-3-l}\Vert_{H^1(R)}\nonumber
\\&&+\sum_{l=1}^{m-4}|\xi_l|\Vert\psi_{m-4-l}\Vert_{H^1(R)}+\Vert\psi_{m-2}\Vert_{H^1(R)}+\Vert\psi_{m-3}\Vert_{H^1(R)}+\Vert\psi_{m-4}\Vert_{H^1(R)} \nonumber
\\&&+\sum_{l=0}^{m-2}\sum_{n=0}^l|\xi_{m-2-l}||\xi_n|(\Vert\psi_{l-n}\Vert_{Y\setminus R}+\Vert\psi_{l-n}\Vert_{H^1(R)})\nonumber
\\&&+\sum_{l=0}^{m-2}|\xi_l|(\Vert\psi_{m-2-1}\Vert_{Y\setminus R}+\Vert\psi_{m-2-1}\Vert_{H^1(R)})\big].
\label{bound1}
\end{eqnarray}}
From \eqref{Rproblem} and (\ref{boundnessforpsiinR}), it follows that
{\footnotesize
\begin{eqnarray}
\lefteqn{||\tilde\psi_{m}||_{H^1(R)}\leq }\nonumber
\\&&{B_n}\big[\sum_{l=1}^{m-1}\xi_l||\psi_{m-l}||_{H^1(R)}+\sum_{l=0}^{m-1}\xi_l||\psi_{m-1-l}||_{H^1(R)}\nonumber
\\&&+\sum_{l=0}^{m-2}\xi_l||\psi_{m-2-l}||_{H^1(R)}+||\psi_{m-1}||_{H^1(R)}+||\psi_{m-2}||_{H^1(R)}\nonumber
\\&&+\sum_{l=1}^{m-1}\sum_{n=0}^l\xi_{m-l}\xi_n||\psi_{l-n}||_{H^1(R)}+\sum_{l=1}^{m-1}\xi_l||\psi_{m-l}||_{H^1(R)}+||\psi_{m}||_{Y\setminus R}\big ]
\label{bound2}
\end{eqnarray}}
The solvability constant $\mathcal{G}$ in \eqref{iterationforxim1} is bounded away from zero and $\infty$ for $\xi_0\in I_n$
Put $m-2\mapsto m$ in \eqref{iterationforxim1} to obtain the inequality for $|\xi_m|$. The inequality is in terms of a constant $C_n$ depending only on $I_n$ and is given by
{\footnotesize
\begin{eqnarray}
\lefteqn{|\xi_m|\leq}\nonumber
\\&& C_n \big [\tau^2\big(\sum_{l=0}^{m-1}|\xi_l|\Vert\psi_{m+1-l}\Vert_{Y\setminus R}+\sum_{l=0}^{m-1}|\xi_l|\Vert\psi_{m-l}\Vert_{Y\setminus R}+2\Vert\psi_{m+1}'\Vert_{Y\setminus R}+\Vert\psi_{m}\Vert_{Y\setminus R}\big)\nonumber
\\&&+\sum_{l=0}^{m-1}|\xi_l|\Vert\psi_{m-1-l}\Vert_{H^1(R)}+\sum_{l=0}^{m-2}|\xi_l|\Vert\psi_{m-2-l}\Vert_{H^1(R)}\nonumber
\\&&+\sum_{l=1}^{m-1}\sum_{n=0}^l|\xi_{m-l}||\xi_n|(\Vert\psi_{l-n}\Vert_{H^1(R)}+\Vert\psi_{l-n}\Vert_{Y\setminus R})\label{bound3}
\\&&+2\sum_{l=1}^{m-1}|\xi_l|(\Vert\psi_{m-l}\Vert_{H^1(R)}+\Vert\psi_{m-l}\Vert_{Y\setminus R})\nonumber
\\&&+(\Vert\tilde\psi_{m}\Vert_{H^1(R)}+\Vert\psi_{m}\Vert_{Y\setminus R})+(\Vert\tilde\psi_{m}\Vert_{H^1(R)}+\Vert\psi_{m}\Vert_{Y\setminus R})
\big]\nonumber
\end{eqnarray}}
The decompositions (\ref{decomp}), \eqref{decompp} and bounds (\ref{boundnessforpsistarinR}), \eqref{hatpsiest} give
\begin{eqnarray}
\Vert\psi_m\Vert_{H^1(R)}&\leq & \Vert\tilde\psi_{m}\Vert_{H^1(R)}+L_n |\xi_m| ~~\hbox{ and}
\label{bound4}\\
\Vert \psi_m\Vert_{Y\setminus R}&\leq& \Vert\psi_m'\Vert_{Y\setminus R}+ H_n|\xi_{m-1}|.
\label{bound55}
\end{eqnarray}
\par
Here the upper bounds \eqref{bound1}, \eqref{bound2} \eqref{bound3}, \eqref{bound4}, and \eqref{bound55} hold for all values of $\tau$ for which the branch of the quasistatic dispersion relation $\xi_n^0$ lies in interval $I_n$.
\noindent{Step III. {\em Majorizing sequence and analyticity of generating functions.}
We simplify the exposition by introducing
\begin{eqnarray}
\overline p_m &=& ||\tau^m\psi_m||_{Y\setminus R}\\
p_m &=& ||\tau^m\psi_m||_{H^1(R)}\\
\tilde p_m &=& ||\tau^m\tilde\psi_m||_{H^1(R)}\\
p_m' &=& ||\tau^m\psi_m'||_{Y\setminus R}\\
s_m &=& |\tau^m\xi_m|,
\end{eqnarray}
and show that the corresponding series $\sum \rho^m\overline p_m$, $\sum \rho^m p_m$, $\sum \rho^m \tilde p_m$, $\sum \rho^m p_m'$, and $\sum \rho^m s_m$
converge. This is sufficient to establish summability and convergence for the series \eqref{convgncee} and \eqref{xipowerseries}.
Writing \eqref{bound1}, \eqref{bound2} \eqref{bound3}, \eqref{bound4}, and \eqref{bound55} in terms of the new notation gives the following system of inequalities.
For $m\geq 2$ we have:
{\footnotesize
\begin{eqnarray}
\lefteqn{ p_{m}'\leq} \nonumber
\\&& E_n \big[ \sum_{l=1}^{m-2}s_l\overline p_{m-l}+\tau\sum_{l=0}^{m-2}s_l\overline p_{m-1-l}+ +\tau^2\sum_{l=0}^{m-2}s_l\overline p_{m-2-l}\nonumber
\\&&+\tau\overline p_{m-1}+\tau^2\overline p_{m-2}+\sum_{l=0}^{m-2}s_l p_{m-2-l}+\tau\sum_{l=0}^{m-3}s_l p_{m-3-l}+\tau^2\sum_{l=0}^{m-4}s_lp_{m-4-l}\nonumber
\\&&+p_{m-2}+\tau p_{m-3}+\tau^2 p_{m-4}+\sum_{l=0}^{m-2}\sum_{n=0}^ls_{m-2-l}s_n(\overline p_{l-n}+p_{l-n})\nonumber
\\&&+\sum_{l=0}^{m-2}s_l(\overline p_{m-2-l}+p_{m-2-l})
\big],
\end{eqnarray}}
and for $m\geq 1$,
{\footnotesize
\begin{eqnarray}
\lefteqn{\tilde p_m\leq }\nonumber
\\&&B_n\big[\sum_{l=1}^{m-1}s_lp_{m-l}+\tau\sum_{l=0}^{m-1}s_lp_{m-1-l}+\tau^2\sum_{l=0}^{m-2}s_lp_{m-2-l}+\tau p_{m-1}\nonumber
\\&&+\tau^2 p_{m-2}+\sum_{l=1}^{m-1}\sum_{n=0}^ls_{m-l}s_n p_{l-n}+\sum_{l=1}^{m-1}s_lp_{m-l}+\overline p_m
\big].
\end{eqnarray}}
The bounds (\ref{bound3}), (\ref{bound4}), and \eqref{bound55} yield the following bounds for $m\geq 1$:
{\footnotesize
\begin{eqnarray}
\lefteqn{s_m\leq }\nonumber
\\&&C_n \big[ 2\tau p_{m+1}'+\tau \sum_{l=1}^{m-1}s_l\overline p_{m+1-l}+\tau^2\sum_{l=0}^{m-1}s_l\overline p_{m-l}\nonumber
\\&&+\tau \sum_{l=0}^{m-1}s_l p_{m-1-l}+\tau^2\sum_{l=0}^{m-2}s_l p_{m-2-l}+\sum_{l=0}^{m-3}\sum_{n=0}^ls_{m-l}s_n(\overline p_{l-n}+p_{l-n})\nonumber
\\&&+2\sum_{l=1}^{m}s_l(\overline p_{m-l}+p_{m-l})+2\tilde p_m+ (2+\tau^2)\overline p_{m}\big],
\end{eqnarray}}
\begin{eqnarray}
p_m&\leq& \tilde p_m+L_n s_m,\label{pmbd}\\
\overline{p}_m&\leq& p_m'+\tau H_n s_{m-1}.\label{pmoverlinebd}
\end{eqnarray}
The inequalities presented above hold for all $\tau$ for which $\xi_0^n$ belongs to $I_n$. In what follows it is convenient to increase the upper bounds by choosing the maximum value of $\tau$ for $\xi_0^n$ in $I_n$ and absorbing this value into the constants $B_n,C_n,E_n$
Additionally any constant factors other than unity that multiply terms in these upper bounds are absorbed into these constants. With these choices the system is written:
{\footnotesize
\begin{eqnarray}
\lefteqn{ p_{m}'\leq} \label{pprimebd}
\\&& E_n \big[ \sum_{l=1}^{m-2}s_l\overline p_{m-l}+\sum_{l=0}^{m-2}s_l\overline p_{m-1-l}+ +\sum_{l=0}^{m-2}s_l\overline p_{m-2-l}\nonumber
\\&&+\overline p_{m-1}+\overline p_{m-2}+\sum_{l=0}^{m-2}s_l p_{m-2-l}+\sum_{l=0}^{m-3}s_l p_{m-3-l}+\sum_{l=0}^{m-4}s_lp_{m-4-l}\nonumber
\\&&+p_{m-2}+ p_{m-3}+ p_{m-4}+\sum_{l=0}^{m-2}\sum_{n=0}^ls_{m-2-l}s_n(\overline p_{l-n}+p_{l-n})\nonumber
\\&&+\sum_{l=0}^{m-2}s_l(\overline p_{m-2-l}+p_{m-2-l})
\big],\nonumber
\end{eqnarray}}
{\footnotesize
\begin{eqnarray}
\lefteqn{\tilde p_m\leq }\label{tpmbd}
\\&&B_n\big[\sum_{l=1}^{m-1}s_lp_{m-l}+\sum_{l=0}^{m-1}s_lp_{m-1-l}+\sum_{l=0}^{m-2}s_lp_{m-2-l}+ p_{m-1}+ p_{m-2}\nonumber
\\&&+\sum_{l=1}^{m-1}\sum_{n=0}^ls_{m-l}s_n p_{l-n}+\sum_{l=1}^{m-1}s_lp_{m-l}+\overline p_m
\big],\nonumber
\end{eqnarray}}
{\footnotesize
\begin{eqnarray}
\lefteqn{s_m\leq }\label{smbd}
\\&&C_n \big[ p_{m+1}'+ \sum_{l=1}^{m-1}s_l\overline p_{m+1-l}+\sum_{l=0}^{m-1}s_l\overline p_{m-l}\nonumber
\\&&+ \sum_{l=0}^{m-1}s_l p_{m-1-l}+\sum_{l=0}^{m-2}s_l p_{m-2-l}+\sum_{l=0}^{m-3}\sum_{n=0}^ls_{m-l}s_n(\overline p_{l-n}+p_{l-n})\nonumber
\\&&+\sum_{l=1}^{m}s_l(\overline p_{m-l}+p_{m-l})+\tilde p_m+ \overline p_{m}\big],\nonumber
\end{eqnarray}}
\begin{eqnarray}
p_m&\leq& C_4(\tilde p_m+ s_m),\label{pmbd2}\\
\overline{p}_m&\leq& C_5(p_m'+ s_{m-1}),\label{pmoverlinebd2}
\end{eqnarray}
where $C_4$, $C_5$ are positive constants depending only on $I_n$ and $m\geq 1$.
We now introduce a majorizing sequence $\{(a_m,b_m,c_m,d_m,e_m)\}_{m=0}^\infty$ for which
\begin{eqnarray}
\sum_{n=0}^\infty a_n\rho^n=&\rho a(\rho)+\overline{p}_0&\geq\sum_{n=0}^\infty \overline{p}_n\rho^n\nonumber\\
\sum_{n=1}^\infty b_n\rho^n=&b(\rho)&\geq\sum_{n=1}^\infty \tilde{p}_n\rho^n\nonumber\\
\sum_{n=0}^\infty c_n\rho^n=&c(\rho)&\geq\sum_{n=0}^\infty s_n\rho^n\nonumber\\
\sum_{n=0}^\infty d_n\rho^n=&d(\rho)&\geq\sum_{n=0}^\infty p_n\rho^n\nonumber\\
\sum_{n=2}^\infty e_n\rho^n=&\rho e(\rho)&\geq\sum_{n=2}^\infty p_n'\rho^n, \label{majorant}
\end{eqnarray}
to show that the generating functions $a(\rho),\ldots,e(\rho)$ are analytic in a neighborhood of the origin $\rho=0$.
The majorizing sequence is chosen so that the system of inequalities \eqref{pprimebd}, \eqref{tpmbd}, \eqref{smbd}, \eqref{pmbd2}, \eqref{pmoverlinebd2}
hold with equality for $a_m=\overline{p}_m$, $b_m=\tilde{p}_m$, $c_m=s_m$, $p_m=d_m$, and $e_m=p_m'$.
Indeed, for this choice one observes that $\overline p_m\leq a_m$, $\tilde p_m\leq b_m$ , $s_m\leq c_m$, $p_m\leq d_m$, and $p_m'\leq e_n$.
Enforcing equality in \eqref{pprimebd}, \eqref{tpmbd}, \eqref{smbd}, \eqref{pmbd2}, \eqref{pmoverlinebd2}, multiplying each by the appropriate power of $\rho$ and summation delivers the equivalent system for the generating functions $a(\rho), b(\rho), c(\rho), d(\rho), e(\rho)$ given by:
\begin{eqnarray}
A_i(a,b,c,d,e,\rho)=0, \hbox{ $i=1,\ldots 5$},
\end{eqnarray}
where:
\begin{eqnarray}
A_1(a,b,c,d,e,\rho)&=&\overline{p}_1-a+C_5(e+(c-s_0)),\label{A1}
\end{eqnarray}
\begin{eqnarray}
\lefteqn{A_2(a,b,c,d,e,\rho)=}\label{A2}
\\&& -b+B_n\big[(c-s_0)(d-p_0)+(cd+d)(\rho+\rho^2)+c^2 d-s_0^2 p_0-\rho s_0 s_1 (p_0+p_1)\nonumber
\\&&-s_0 p_0(c-s_0)-s_0 c d+s_0(\rho(s_0 p_1+s_1 p_0)+s_0 p_0)+(c-s_0)(d-p_0)+\rho a
\big]\nonumber
\end{eqnarray}
{\footnotesize
\begin{eqnarray}
\lefteqn{A_3(a,b,c,d,e,\rho)=}\label{A3}
\\&&-(c-s_0)+C_n\big[ (a-\overline{p}_1)(c-s_0)+e+\rho a c+\rho(a+cd)\nonumber
\\&& +\rho^2c d+c^2(\rho a+\overline{p}_0)-s_0 \overline{p}_0 c-s_0((\rho a+\overline{p}_0)c-s_0\overline{p}_0)+(c-s_0)(d-p_0)+b\nonumber
\\&&-\rho(\rho a+\overline{p}_0)c s_1-\rho^2s_2(\rho a+\overline{p}_0)c+c^2d-s_0p_0 c-s_0(c d-s_0 p_0)-\rho cd (s_1+\rho s_2)\nonumber
\big]
\end{eqnarray}}
\begin{eqnarray}
A_4(a,b,c,d,e,\rho)= (d-p_0)+C_4(b+(c-s_0)).\label{A4}
\end{eqnarray}
\begin{eqnarray}
\lefteqn{A_5(a,b,c,d,e,\rho)=}\label{A5}
\\&&-e+E_n\big[ (a-\overline{p}_1)(c-s_0)+\rho ac+\rho(\rho a +\overline{p}_0)c +\rho a+\rho(\rho a+\overline{p}_0)\nonumber
\\&& +(c+1)d(\rho+\rho^2+\rho^3)+\rho c(c+1)(\rho a+\overline{p}_0+d)\nonumber
\big]
\end{eqnarray}
One can check $A_1,A_2,A_3,A_4$ and $A_5$ vanish for $a=\overline{p}_1=\Vert \psi_1\Vert_{Y\setminus R}$, $b=0$, $c=s_0=|\xi_0^n|$, $d=p_0=\Vert \psi_0\Vert_{H^1(R)}$, $e=0$, and $\rho=0$. Moreover, the determinant of the Jacobian matrix with respect to the first five variables at
this point, i.e., $(a,b,c,d,e,\rho)=(\overline{p}_1,0,s_0,p_0,0,0)$, is
\begin{eqnarray}
det\frac{\partial (A_1,A_2,A_3,A_4,A_5)}{\partial(a,b,c,d,e)}(\overline{p}_1,0,s_0,p_0,0,0)=1.
\end{eqnarray}
Thus the implicit function theorem for functions of several complex variables guarantees that the generating functions $a,b,c,d,e$ are analytic within a neighborhood of $\rho=0$ and we conclude that the series on the right hand side of the inequalities \eqref{majorant} are convergent.
Here $\overline{p}_0=\Vert \psi_0\Vert_{Y\setminus R}=|Y\setminus R|^{1/2}$ and the initial data $p_0$, $\overline{p}_1$ are bounded functions of $\xi_0^n$ on $I_n$. Choosing the maximum values of $p_0$, $\overline{p}_1$, $\xi_0^n$, as $\xi_0^n$ varies over $I_n$ delivers a majorizing sequence with radius of convergence $R$ depending only on $I_n$. It now follows that the series \eqref{xipowerseries} converges and that \eqref{convgncee} is summable for $0\leq\rho<R$.
\section{Power Series Solution of Cell Problem }
\label{powerseriessolutionofcellproblem}
In this section we show that the functions defined by the power series solve \eqref{variational2}. For $\xi_0^n$ in $I_n$ recall the series \eqref{upower}, \eqref{xipowerexpansion}
\begin{eqnarray*}
&&u=\sum_{m=0}^\infty \eta^m u_m \label{upowerrho}
\\&&\xi=\sum_{m=0}^\infty\eta^m\xi_m \label{xipowerexpansionrho}
\end{eqnarray*}
with $u_m=i^m\underline{u}_0\psi_m$ and $\eta=\rho\tau$ converge for $0<\rho<R$.
For $v\in H^1_{per}(Y) $ and $0\leq \rho< R$. We define
{\footnotesize
\begin{eqnarray}
&&a^\eta(v):=\int_H\tau^2(\xi-\epsilon_r\frac{\omega_p^2}{c^2})(\nabla_y+i\eta \hat{\kappa})u\cdot \overline{(\nabla_y+i\eta \hat{\kappa})v }\nonumber
\\&&+\int_P\tau^2\xi(\nabla_y+i\eta \hat{\kappa})u\cdot \overline{(\nabla_y+i\eta \hat{\kappa})v }\nonumber
\\&&+\int_R\eta^2(\xi-\epsilon_r\frac{\omega_p^2}{c^2})(\nabla_y+i\eta \hat{\kappa})u\cdot \overline{(\nabla_y+i\eta \hat{\kappa})v }-\int_Y\eta^2\xi(\xi-\epsilon_r\frac{\omega_p^2}{c^2})u\overline{v} .
\end{eqnarray}}
First observe that $a^\eta(v)$ has a convergent power series in $\eta$ that is obtained by inserting $u$ and $\xi$ into the above expression and expanding in powers of $\eta$ . The coefficients of this expansion are exactly the left-hand side of equation (\ref{summation2}). Since $\psi_m$ satisfies (\ref{summation2}), all coefficients of the power series expansion of $a^\eta(v)$ are zero and we conclude that $a^\eta(v)=0$ for all $v\in H^1_{per}(Y)$, $0\leq \rho< R$. Thus we conclude that the pair $(u,\xi)$ satisfies (\ref{variational2}).
Next we record the following property of the sequence $\{\xi_m\}$ that follows directly from the variational formulation of the problem and the power series solution.
\begin{theorem}
The functions $\xi_m$ are real for all $m$.
\label{realvalue}
\end{theorem}
\begin{proof}
Setting $v=u$ in (\ref{variational2}) delivers a quadratic equation for $\xi$. Since the discriminant greater than zero, we conclude that $\xi$ is real and it follows that $\xi_m$, $m=1,\ldots$, are real.
\end{proof}
\noindent With these results in hand, Theorems \ref{convergenceofxi} and \ref{summable} now follow from the convergence established in section \ref{convergence} together with Theorem \ref {Existencesequence}.
\section{Acknowledgments}
This research is supported by NSF grant DMS-0807265, DMS-1211066, AFOSR grant FA9550-05-0008, AFOSR MURI Grant FA95510-12-1-0489 administered through the University of New Mexico, and NSF EPSCOR Cooperative Agreement No. EPS-1003897 with additional support from the Louisiana Board of Regents.
|
\section{Introduction}
Measurements of the hadronic final state in deep-inelastic lepton-proton scattering (DIS) test
Quantum Chromodynamics (QCD), the theory of the strong force.
At moderate negative four-momentum
transfers squared $Q^2$ of a few ${\rm GeV}^2$,
the HERA $ep$ collider has extended the available kinematic range for deep-inelastic scattering to regions of small
Bjorken-$x\simeq 10^{-4}$.
This is the region of high parton densities in the proton, dominated by gluons and
sea quarks. At the large $\gamma^*p$ centre-of-mass energy available at small $x$, a transition is expected from
parton cascades ordered in transverse momentum, described by the
Dokshitzer-Gribov-Lipatov-Altarelli-Parisi (DGLAP) evolution equations \cite{DGLAP},
to cascades unordered in transverse momentum, described by the Balitsky-Fadin-Kuraev-Lipatov (BFKL) approach \cite{BFKL}.
\begin{figure}[hhh]
\center
\epsfig{file=d11-183f1.eps ,width=7cm}
\setlength{\unitlength}{1cm}
\caption{Generic diagram for deep-inelastic $ep$ scattering at small $x$. A
gluon cascade
evolves between the quark box, attached to the virtual photon, and the proton. The gluon longitudinal
momentum fractions and transverse momenta are labeled $x_i$ and $k_{Ti}$, respectively.}
\label{fig:diagram}
\end{figure}
A generic diagram for
parton evolution in a DIS process at low $x$, in which a gluon from the proton induces a QCD
cascade before an interaction with the virtual photon, is shown in figure~\ref{fig:diagram}.
In the DGLAP approximation the struck quark originates from a parton cascade ordered in
virtualities of the propagator partons. At low $x$ this implies a strong ordering in transverse
momentum, $k_T$, of the emitted partons, measured with respect to the proton direction.
In the BFKL approach there is no ordering in $k_T$ of the partons
along the ladder. Compared to the DGLAP scheme more gluons with sizable transverse
momentum are emitted near the proton direction.
For this reason energetic jets of high transverse momentum produced close to the proton direction in the laboratory frame, referred
to as the forward region, are considered to be especially sensitive to QCD
dynamics at low $x$~\cite{Mueller}.
Forward jet production was measured previously by the H1 and ZEUS collaborations.
In these measurements as well as in the present one, the
requirements on the forward jet and the phase space were chosen in such a way that the standard
DGLAP evolution is suppressed and the effects of BFKL dynamics are enhanced. Preference
for models which employ QCD evolution non-ordered in transverse momentum was
observed~\cite{h1fj00,h1fj01,h1fj02,zeusfj01,zeusfj02,zeusfj03,zeusfj04}.
One of the observables suggested to be
sensitive to BFKL dynamics~\cite{Bartels} is the azimuthal angle
difference, $\Delta \phi$, between the forward jet and
the scattered electron, defined in the laboratory frame.
In the Quark Parton Model (QPM) process $e+q \rightarrow e+q$ the simple two-body kinematics constrains the scattered electron
and the jet to be produced back-to-back, and thus predicts at the parton level $\Delta \phi = \pi$. Hadronisation effects induce some smearing to this parton level prediction.
Inclusion of higher order processes partially decorrelates the jet from the electron.
As a consequence,
for evolution schemes without ordering in transverse momentum, the decorrelation is expected
to increase with electron-jet rapidity distance, $Y$,
since the phase space for additional parton emissions increases.
The calculations employing the BFKL approach to the next-to-leading
order accuracy (NLO BFKL), indeed predict an increase of the azimuthal angle decorrelation with
the electron-jet rapidity distance \cite{NLOBFKL}.
This paper presents a study of low $x$ DIS interactions in which high transverse momentum jets
are produced in the forward region.
The forward jet cross sections and normalised distributions are measured
as a function of the azimuthal angle difference $\Delta \phi$ in three bins
of the rapidity separation $Y$ between the positron and the forward jet. The forward jet cross section
as a function of $Y$ is also measured.
Moreover, the measurements of the azimuthal correlations in $\Delta \phi$ are performed
using a subsample defined by a requirement of
an additional central jet. In comparison with the forward jet sample, this
subsample is expected to contain a higher fraction of forward jets
from additional gluon emissions.
The data set used for the analysis was collected with the H1 detector in the
year $2000$, when
positrons and protons collided with energies of $27.6$ ${\rm GeV}$
and $920$ ${\rm GeV}$, respectively, corresponding to a centre-of-mass energy of $\sqrt{s} = 319$ ${\rm GeV}$.
The integrated luminosity of
the data set is $38.2$ ${\rm pb}^{-1}$, which is
about fourteen times larger than in the previous
measurement of the azimuthal decorrelation of forward jets \cite{h1fj01}.
\section{QCD Calculations}
The measurements presented here are compared with predictions of
Monte Carlo (MC) programs and perturbative QCD calculations at
next-to-leading order (NLO).
The MC programs use first-order QCD matrix elements
and model higher order terms by parton showers in the leading logarithm approximation
or by quasi-classical gluon radiation from colour dipoles.
Three MC event generators, which adopt different QCD based approaches to model
the parton cascade, are used.
\begin{itemize}
\item
RAPGAP \cite{RAPGAP} matches first order QCD matrix elements to DGLAP based leading-log parton showers with $k_T$ ordering.
The factorisation and renormalisation scales are set to
$\mu_f = \mu_r = \sqrt{Q^2 + p^2_{\rm T}}$, where $p_{\rm T}$ is the transverse momentum of the two outgoing hard partons in
the centre-of-mass of the hard subsystem.
Predictions of RAPGAP are labeled DGLAP in the figures.
\item
DJANGOH \cite{DJANGOH} with ARIADNE includes an implementation of the Colour Dipole Model (CDM) \cite{CDM}, which has as its basic construct a colour dipole formed by the struck quark and the proton remnant. Subsequent parton emissions originate from a chain of independently radiating dipoles formed by the emitted gluons.
In this approach the transverse momenta of emitted gluons perform a random walk such that CDM provides a BFKL-like approach. The leading order partonic final state is corrected
to exactly reproduce the $O(\alpha_S)$ matrix elements.
The simulation
of DJANGOH/ARIADNE uses a set of colour dipole parameters tuned to describe measurements
of the hadronic final state in DIS at HERA \cite{thesis}.
The DJANGOH/ARIADNE predictions are referred to as CDM in the following.
\item
CASCADE \cite{CASCADE} implements the Ciafaloni-Catani-Fiorani-Marchesini (CCFM)
evolution \cite{CCFM} which aims to unify the DGLAP and BFKL approaches. It introduces
angular ordering of emissions to implement gluon coherence effects, and thus
in the high energy limit the CCFM evolution equation is almost equivalent to the BFKL
approach, while reproducing the DGLAP equations for large $x$ and high $Q^2$. CASCADE
uses off-shell leading order QCD matrix elements, supplemented with gluon emissions based
on the CCFM evolution equation, requiring an unintegrated gluon density function (uPDF) which takes the
transverse momenta of the propagators into account.
In this paper two different uPDF sets are used:
set A0 \cite{setA0} with only singular terms of the gluon splitting function
and J2003-set 2 \cite{set2} including also non-singular terms, labeled set 2 in the figures.
These parameterisations for the unintegrated gluon density were obtained using the CCFM
evolution equation to describe the structure function $F_2(x,Q^2)$ as measured
by H1 \cite{F2H1} and ZEUS \cite{F2ZEUS}.
Predictions of CASCADE are labeled CCFM in the figures.
\end{itemize}
To perform the hadronisation step, all of the above models use the Lund string fragmentation
scheme, as implemented in JETSET \cite{JETSET1} in the case of DJANGOH/ARIADNE and
in PYTHIA~\cite{PYTHIA} for RAPGAP and CASCADE, using a tuning based on LEP
$e^+e^-$ data \cite{ALEPH}.
The RAPGAP and DJANGOH/ARIADNE predictions are calculated using the HERAPDF1.0
\cite{HERAPDF} set of parton distribution functions (PDF).
The RAPGAP and DJANGOH/ARIADNE programs are interfaced with
HERACLES~\cite{HERACLES}, which allows the simulation of
QED-radiative effects.
These MC models are used to simulate detector effects in order to
determine the acceptance and efficiency for selected forward jet events in DIS.
Generated events are passed through a GEANT~\cite{GEANT} based simulation
of the H1 apparatus, which takes into account the running conditions of the
data taking.
Simulated events are reconstructed and analysed using the same program chain
as is used for the data.
The measurements of azimuthal
correlations are also compared to the fixed order NLO DGLAP
predictions of NLOJET++~\cite{NLOJET++}.
The NLOJET++ program is used here to calculate dijet production
at parton level in DIS at NLO($\alpha_S^2$) accuracy.
It should be noted that the jet search is performed on partons in the Breit frame (see section $3.2$), and therefore
the events contain at least one jet in addition to the forward jet.
The renormalisation and factorisation
scales are defined for each event and are set to
$\mu_r = \mu_f = \sqrt{(P_{\rm T,sc}^2 + Q^2) / 2}$,
where $P_{\rm T,sc}$ is the transverse momentum
of the forward jet or the average transverse momentum of the forward and central jet
in the forward jet sample and in the sample with an additional central jet, respectively.
The NLO calculations are performed using the
CTEQ6.6 \cite{CTEQ6.6} parameterisation of the parton distributions in the proton.
The NLOJET++ parton level cross sections are corrected for hadronisation effects
using the RAPGAP model.
The correction factors for hadronisation are estimated bin-by-bin by calculating the
ratio between the cross section for jets reconstructed
from stable hadrons (hadron level) and the parton level cross section.
The correction factors for hadronisation are in the range from $0.90$ to $1.08$, increasing with
rapidity distance $Y$.
The uncertainty of the NLOJET++ predictions due to
missing higher orders is estimated by applying a factor $2$ or $1/2$ to the
renormalisation and factorisation scales simultaneously.
\section{Experimental Method}
\subsection{H1 detector}
A detailed description of the H1 detector can be found elsewhere\cite{h1det01,h1det02,h1det03}.
The components of the detector which are most relevant for this analysis are briefly described below.
The origin of the H1 coordinate system is the nominal $ep$ interaction point. The direction
of the proton beam
defines the positive $z$-axis. Transverse momenta $p_{\rm T}$ and polar angles $\theta$ of all
particles are defined with respect to
this direction. The azimuthal angle $\phi$ defines the particle direction in the transverse plane.
The pseudorapidity is given by $\eta =-$ln (tan $\theta$/2).
The $ep$ interaction region is surrounded by the central tracking detector (CTD) consisting of two large concentric drift chambers,
operated inside a $1.16$ $\rm T$ solenoidal magnetic field.
Charged particles are measured in the angular range
$20^\circ < \theta < 160^\circ$ with a transverse momentum resolution of
$\sigma_{p_{\rm T}}/{p_{\rm T}} \approx 0.005\cdot p_{\rm T} [\rm GeV] \oplus 0.015$.
Information from the CTD is used to trigger events, to
locate the event vertex, and contributes to the reconstruction of the hadronic final state.
A highly segmented liquid argon (LAr) calorimeter is used to measure
the hadronic final state. It covers the range of
the polar angle $4^\circ < \theta < 154^\circ$
and offers full azimuthal coverage.
The LAr calorimeter consists of an electromagnetic section with lead absorbers and a hadronic section with steel absorbers.
The total depth of both sections varies between $4.5$ and $8$ interaction lengths in the
region $4^\circ < \theta < 128^\circ$, and between $20$ and $30$ radiation lengths in the region
$4^\circ < \theta < 154^\circ$ increasing towards the forward direction.
Test beam measurements of the LAr calorimeter modules
showed an energy resolution of
$\sigma_E/E \approx 0.50/\sqrt{E[\rm GeV]} \oplus 0.02$
for charged pions \cite{h1det04} and of $\sigma_E/E \approx 0.12/\sqrt{E[\rm GeV]} \oplus 0.01$ for
electrons \cite{h1det05}.
A lead/scintillating fiber calorimeter (SpaCal) \cite{h1det03} covers the region
$153^\circ < \theta < 177.5^\circ$. It has an electromagnetic and a hadronic section and is used to
measure the scattered positron and the backward hadronic energy flow.
The energy resolution, determined from test beam measurements \cite{h1det06},
is $\sigma_E/E \approx 0.07/\sqrt{E [\rm GeV] } \oplus 0.01$ for electrons.
The precision of the measurement of the polar angle of the positron,
improved using the backward drift chamber (BDC) situated in front of the SpaCal calorimeter, is
$1$ $\rm mrad$.
The luminosity determination is based on the measurement of the Bethe-Heitler process
$ep \rightarrow ep\gamma$ where the photon
is detected in a calorimeter located at $z = -103$ ${\rm m}$ downstream of the interaction region
in the positron beam direction.
\subsection{Event selection}
DIS events are selected using triggers based on electromagnetic energy deposits in the SpaCal calorimeter and
the presence of charged particle tracks in the central tracker. The trigger efficiency is determined using independently triggered data.
For DIS events with a forward jet, the trigger efficiency lies
between $60\%$ and $80\%$, and for the topology 'forward and central jet' it is
at the level of $80\%$.
The data set is restricted in inelasticity $y$, photon virtuality $Q^2$
and $x$: $0.1 < y < 0.7$, $5 < Q^2 < 85$ ${\rm GeV}^2$, $0.0001 < x < 0.004$.
In this analysis these variables are determined from measurements of the scattered positron energy
and its polar angle, and from the incident positron beam energy.
This phase space is chosen to ensure that the DIS kinematics are well determined and to
reduce the background from photoproduction.
The background from photoproduction and from events with large initial-state QED radiation is further reduced
by requiring
$35 < \Sigma_i(E_i - p_{z,i}) < 70$ $\rm GeV$. Here $E_i$ and $p_{z,i}$ are the energy and
longitudinal momentum of a particle $i$, respectively, and the sum extends over all
detected particles in the event.
Energy-momentum conservation requires that
$\Sigma_i(E_i - p_{z,i}) = 2 \cdot E^{0}_{e}$, where $E^{0}_{e}$ is the
positron beam energy.
Jets are identified from combined calorimeter and track objects \cite{object} using the $k_T$ cluster algorithm
in the longitudinally invariant inclusive mode~\cite{ktalgo} applied in the Breit frame.
The reconstructed jets are then boosted to the laboratory frame.
The measurements of forward jets are restricted to the phase space region where the transverse momentum of the jet is approximately equal to the photon virtuality, $P_{\rm T,fwdjet}^2 \approx Q^2$. This condition
suppresses the contribution of $k_T$-ordered DGLAP cascades with respect to
processes unordered in $k_T$~\cite{Mueller}. The selection
of forward jets with a large fraction of the proton energy, $x_{\rm fwdjet} \equiv E_{\rm fwdjet}/E_p$,
such that $x_{\rm fwdjet} \gg x$, enhances
the phase space for BFKL evolution with gluon cascades strongly ordered in fractional
longitudinal momentum. The above conditions are fulfilled by the requirement that the
analysed sample contains
at least one forward jet which satisfies the following criteria in the laboratory frame:
$P_{\rm T,fwdjet} > 6$~${\rm GeV}$, $ 1.73 < \eta_{\rm fwdjet} < 2.79$,
$x_{\rm fwdjet} > 0.035$ and $0.5 < P^2_{\rm T,fwdjet}/Q^2 < 6$.
Here $\eta_{\rm fwdjet}$ is the pseudorapidity of the forward jet.
If there is more than one jet fulfilling the above requirements, the jet with the largest pseudorapidity
is chosen.
The upper cut on $P^2_{\rm T,fwdjet}/Q^2$ is chosen so large in order to reduce
the contributions of
migrations from outside of the analysis phase space, which are due to the limited resolution of
the $P_{\rm T,fwdjet}$ measurement.
The subsample ``forward and central jet'' is selected by requiring an additional jet
in the central region of the laboratory frame. This jet is required to have a transverse
momentum $P_{\rm T, cenjet} > 4$ ${\rm GeV}$ and to lie in the pseudorapidity region $-1 < \eta_{\rm cenjet} < 1$. The central jet must have a large rapidity separation from the most forward jet
$\Delta \eta = (\eta_{\rm fwdjet} - \eta_{\rm cenjet}) > 2$. This condition enhances
the phase space for additional parton emissions between the two jets. If there is more than one
central jet, the one with the smallest $\eta_{\rm cenjet}$ is chosen.
A summary of the selection cuts, defining the DIS phase space
for the measurement, the forward jet sample and the subsample with an additional central jet,
is provided in table \ref{tab:disspace}.
With these requirements $13736$ and $8871$ events are selected for the forward jet and for the forward and central jet
analysis, respectively.
\renewcommand{\arraystretch}{1.15}
\begin{table}[tb]
\begin{center}
\begin{tabular}{|c|c|c|} \hline
{\bf DIS selection} & {\bf Forward jets} &{\bf Central jets}\\
\hline
$0.1 < y < 0.7 $&$1.73~ <~\eta_{\rm{fwdjet}}~< 2.79$&
$-1 < \eta_{\rm{cenjet}}< 1$ \\[0.2cm]
$5 < Q^2 < 85$ ${\rm GeV}^2$&$P_{\rm{T,fwdjet}}~>~6$ ${\rm GeV}$&$P_{\rm{T,cenjet}}> 4$ ${\rm GeV}$ \\[0.2cm]
$0.0001< x < 0.004 $&$x_{\rm{fwdjet}} >0.035$ &
$\Delta \eta = \eta_{\rm{fwdjet}} -\eta_{\rm{cenjet}} > 2$\\[0.2cm]
&$0.5 < P^2_{\rm{T,fwdjet}}/Q^2 < 6$&\\[0.2cm]
\hline
\end{tabular}
\end{center}
\caption{Summary of cuts defining the DIS phase space, the forward jet and the central jet selection.
If more than one forward jet is found, the jet with the largest $\eta_{\rm fwdjet}$ is chosen.
If there is more than one central jet, the one with the smallest $\eta_{\rm cenjet}$ is selected.}
\label{tab:disspace}
\end{table}
\subsection{Cross section determination}
In this measurement in addition to migrations between bins inside the measurement
phase space, there are considerable
migrations from outside of the analysis phase
space. This is taken into account in the calculation of the cross section corrected to the hadron level:
\begin{equation}
\sigma_i = \frac{N^{\rm data}_i - N^{\rm out}_i }{\epsilon_i \cdot {\cal L}}.
\label{eq:02}
\end{equation}
Here $N^{\rm data}_i$ is the number of observed events in bin $i$, $N^{\rm out}_i$
is the number of events from outside the measurement phase space reconstructed
in bin $i$, and $\epsilon_i$ is the efficiency in bin $i$. ${\cal L}$ is the total integrated luminosity.
$N^{\rm out}_i$ and $\epsilon_i$ are estimated using MC simulations.
The purities\footnote{The purity is defined as the ratio of the number of events generated and reconstructed in the bin to the number of events originating from the phase space of the analysis and reconstructed in that bin.}
in bins of the measured cross sections, as determined from the MC simulations, are at the level of $80\%$.
The efficiency factors $\epsilon_i$ are calculated according to the formula :
\begin{equation}
\epsilon_i=\frac{N^{\rm det}_{i} - N^{\rm out}_i}{N^{\rm had}_i} ,
\label{eq:04}
\end{equation}
where $N^{\rm det}_{i}$ and $N^{\rm had}_i$ are the numbers of events in bin $i$ at the
detector and at the hadron level, respectively.
For this approach to be valid, the shape of the distributions of all variables on which phase
space cuts are applied have to be well described by the MC simulations also in the
phase space extended beyond these cuts. This requirement is found to be satisfied by both
models considered here.
The efficiency factors are calculated as the ratio of the model prediction
at the detector level for a radiative MC
and at the hadron level for a non-radiative MC, i.e. the data are also corrected
for QED radiative effects. The efficiency factors are taken as the average of the factors estimated by the RAPGAP and DJANGOH/ARIADNE models.
The uncertainty of the efficiency factors is taken to be half of the difference
between the factors
calculated using the two MC models and is included in the systematic error.
\subsection{Systematic uncertainties}
The following sources of systematic uncertainties are considered :
\begin{itemize}
\item[-]
The model dependence of the bin-by-bin efficiency factors $\epsilon_i$
leads to systematic uncertainties between $2 \%$ and $6 \%$ for
the measured cross sections.
\item[-]
The LAr hadronic energy scale uncertainty of $4 \%$ for this analysis gives rise to the dominant
uncertainty of $7\%$ to $12 \%$ for the measured cross sections.
\item[-]
The uncertainty on the electromagnetic energy scale of the SpaCal of $1 \%$ results in an uncertainty of the measured cross sections below $3 \%$.
\item[-]
The uncertainty on the polar angle measurement of the scattered positron of
$1$ $\rm mrad$ has
a negligible effect on the cross section measurements.
\item[-]
The uncertainty on the determination of the trigger efficiency from the data, using independent
trigger samples,
leads to an uncertainty between $2\%$ and $4\%$ on the cross section measurements.
\item[-]
The measurement of the integrated luminosity is accurate to within $1.5\%$.
\end{itemize}
The total systematic uncertainty, adding all individual contributions
quadratically, amounts to $11-12 \%$ for the measured cross sections.
\section{Results}
The forward jet cross sections and their uncertainties
are given in table~\ref{tab:mytest1} and presented in
figures \ref{fig:azim1}-\ref{fig:azim3}.
Differential cross sections, $d\sigma/d\Delta \phi$, are presented as
a function of the azimuthal angle difference $\Delta \phi$
between the most forward jet and the scattered positron
in bins of the variable $Y = \ln(x_{\rm fwdjet}/x)$.
This variable approximates the rapidity distance
between the scattered positron and the forward jet.
For the selected data
sample the normalised shape distributions $1/ \sigma \cdot d\sigma/d\Delta \phi$ are also
determined, where $\sigma$ is
the integrated cross section in a given bin of $Y$. Furthermore, the forward jet cross
section is measured as a function of $Y$.
The cross section $d\sigma/d\Delta\phi$ as a function of $\Delta \phi$ is shown in
figure~\ref{fig:azim1} for three intervals of the variable
$Y$: $2.0 \leq Y < 3.4$, $3.4 \leq Y < 4.25$ and $4.25 \le Y \le 5.75$. These $Y$ bins
correspond to average $x$ values of $0.0024$, $0.0012$ and $0.00048$, respectively.
At higher values of $Y$
the forward jet is more decorrelated from the scattered positron.
The predictions of three QCD-based models with different underlying parton dynamics, discussed
in section 2, are compared with the data.
The cross sections are well described in shape and normalisation by CDM which has a
BFKL-like approach.
Predictions of RAPGAP, which implements DGLAP evolution, fall below the data,
particularly at large $Y$.
Calculations in the CCFM scheme as implemented in CASCADE using the uPDF set A0~\cite{setA0}
overestimate the measured cross section for large $\Delta \phi$ values in the two lowest
$Y$ intervals. However, this model provides as good a description as CDM
of the data in the highest $Y$ interval.
The shape of the $\Delta \phi$ distributions, $1 / \sigma \cdot d\sigma/d\Delta \phi$, is compared to the different MC predictions in the lower part
of figure~\ref{fig:azim1}, where the ratio $R$ is shown, defined as:
\begin{equation}
R = \left(\frac{1}{\sigma^{\rm MC}} \frac{d\sigma^{\rm MC}}{d\Delta \phi} \right) \, \Big/ \, \left(\frac{1}{\sigma^{\rm data}} \frac{d\sigma^{\rm data}}{d\Delta \phi}\right) \, .
\label{eq:1}
\end{equation}
The precision of the measurements is shown at $R=1$ where the statistical and
systematic uncertainties are indicated.
The systematic uncertainty is reduced in the ratio and contains only two
components added in quadrature: the model dependence of
the correction factors and the trigger efficiency uncertainty.
The ratio plots show that in the analysed phase space region the
shape of the $\Delta \phi$ distributions is well described by all MC models.
Since the shape predictions of the three models are very similar, this observable alone
cannot discriminate among the models.
It should be noted that the shape of the $\Delta \phi$ distributions
is rather insensitive to the PDF used for event generation. The shape distributions generated using
CTEQ6L, CTEQ6M \cite{CTEQ6L} and HERAPDF1.0 \cite{HERAPDF} differ on average by $1$-$2\%$.
However, the cross section normalisation is more sensitive to the choice of PDF with
differences up to $5\%$ for CDM and up to $20\%$ for RAPGAP at large $Y$.
Predictions of the CCFM model presented in figure~\ref{fig:azim2} indicate a significant
sensitivity to the choice of the uPDF. Set A0 and J2003-set 2 give quite different predictions
for the differential cross sections in all $Y$ intervals.
Set A0 provides a reasonable description of the measured cross sections, except for the
region of large $\Delta \phi$ in the two lowest $Y$ bins. Predictions using J2003-set 2 do
not describe the data, especially at higher $Y$, where the estimated cross sections are too low.
The shape of the $\Delta \phi$ distributions is reasonably well described by
the set A0. At low $Y$
it shows sensitivity to the unintegrated gluon density.
The cross section $d\sigma/dY$ as a function of the rapidity separation $Y$
is shown in figure~\ref{fig:azim3}.
The CDM model describes the data well
over the whole $Y$ range.
The DGLAP predictions fall below the data, but approach them at small $Y$.
The predictions of the CCFM model are above the data at small $Y$ but describe them
well at larger $Y$ corresponding to low values of $x$.
The forward and central jet cross sections and their uncertainties are given in table~\ref{tab:mytest2}.
The differential cross section $d\sigma/d\Delta\phi$ as a function of the azimuthal angle
difference $\Delta \phi$ is shown in figure~\ref{fig:azim4} in comparison with
the predictions of the three MC models.
The cross sections are measured in two intervals of $Y$,
$2.0 \le Y < 4.0$ and $4.0 \le Y \le 5.75$.
From figure~\ref{fig:azim4} it is observed that at lower $Y$ the predictions of all models
describe the cross sections reasonably well.
At high $Y$ all models undershoot the data: CCFM (set A0) is closest to the data,
the DGLAP and CDM predictions are below the measured cross section.
The ratio $R$ in the lower part of figure~\ref{fig:azim4} shows that the shape of the
$\Delta \phi$
distributions is well described by all MC models, as in the case of the forward jet measurements.
Comparisons of the measured $\Delta \phi$ distributions
with NLOJET++ predictions are shown in figures \ref{fig:nloinc} and \ref{fig:nlo}.
The calculations are performed at $O(\alpha_S^2)$ precision using
the CTEQ6.6 PDF \cite{CTEQ6.6} and
$\alpha_S(M_Z) = 0.118$.
Large theoretical uncertainties of up to $50\%$ from the variation of factorisation and
renormalisation scales are observed.
The size of the theoretical uncertainty indicates that in this phase space region
higher order contributions are expected to be important.
In the forward jet sample (figure \ref{fig:nloinc})
for all three ranges of $Y$ the data are
above the central NLO result but still within the theoretical uncertainty.
In the case of the forward and central jet sample
shown in figure \ref{fig:nlo}, the NLO calculation describes the
data at low $Y$. Only at high $Y$ in the regime of the BFKL evolution it is below the data, but again within the large
theoretical uncertainty.
In summary,
the correlation between the forward jet and the positron decreases with $Y$ and
the $\Delta \phi$ distributions are flat at high $Y$.
The measurements of the forward jet cross sections favour CDM and disfavour the RAPGAP model.
CASCADE provides a reasonable description of the data at large $Y$, but shows sizeable sensitivity
to the uPDF.
The shape of the measured $\Delta \phi$ distributions is well described by MC models based on different
QCD evolution schemes.
The similarity of the $\Delta \phi$ shapes of the MC predictions suggests that the forward jet predominantly
originates from the hard matrix elements which are similar in all three models.
However, MC studies with RAPGAP show
that $80\%$ of the forward jets are produced by parton showers. When the initial state parton
shower is switched off, the shape of the $\Delta \phi$ distribution is only
slightly changed, but the
normalisation is significantly reduced.
This indicates that the decorrelation in $\Delta \phi$
is mainly governed by the phase space requirements, in particular by the
rapidity separation $Y$,
and that the normalisation of the cross sections is mainly influenced by the
amount of soft radiation from
parton showers, which depends on the evolution scheme.
\section{Conclusions}
Measurements of DIS events at low $Q^2$ containing a high transverse momentum jet produced
in the forward direction, at small angles with respect to the proton beam, are presented.
Differential cross sections and normalised distributions are measured as a function of the azimuthal angle difference $\Delta \phi$ and the rapidity separation $Y$ between the forward jet and the scattered positron.
Investigations of the azimuthal correlation between the most forward jet and the outgoing positron
are performed in different regions of $Y$ for the forward jet sample
and for the subsample with an additional central jet. To test the sensitivity of the measured observables to QCD dynamics at low $x$, the data are compared to QCD models with different parton evolution approaches and to predictions
of next-to-leading order QCD calculations.
Measurements of the cross sections as a function of $\Delta \phi$ and $Y$ are best described by the BFKL-like CDM model, while
the DGLAP-based RAPGAP model is substantially below the data.
The CCFM-based CASCADE provides a reasonable description of the data but shows sizeable sensitivity to the unintegrated
gluon density. The shape of the $\Delta \phi$ distributions does not
discriminate further between different evolution schemes.
The fixed order NLO DGLAP predictions are in general below the data, but still
in agreement within the large theoretical uncertainties.
\section*{Acknowledgements}
We are grateful to the HERA machine group whose outstanding efforts have made this experiment possible. We thank the engineers and technicians for their work in constructing and maintaining the H1 detector, our funding agencies for financial support, the DESY technical staff for continual assistance and the DESY directorate for support and for the hospitality which they extend to the
non-DESY members of the collaboration.
|
\section{Introduction}
Neutrinos are unique messengers to study the high-energy universe as they are neutral and
stable, interact weakly and therefore travel directly from their point of creation to the Earth without absorption.
Neutrinos could play an important role in understanding the mechanisms of cosmic ray acceleration and their
detection from a cosmic source would be a direct evidence of the presence of hadronic acceleration.
The production of high-energy neutrinos has been proposed for several kinds
of astrophysical sources, such as active galactic nuclei (AGN), gamma-ray bursters (GRB), supernova
remnants and microquasars, in which the acceleration of hadrons may occur (see Ref.~\cite{bib:Becker} for a review).
Flat-Spectrum Radio Quasars (FSRQs) and BL Lacs, classified as AGN blazars, exhibit relativistic jets pointing almost directly towards
the Earth and are some of the most violent variable high energy phenomena in the Universe~\cite{bib:Blazars}. These sources are among the most likely sources of the observed ultra high energy
cosmic rays. Blazars typically display spectra with enhanced emission over two
energy ranges: the IR/X-ray and MeV/TeV peaks. The lower energy peak is generally agreed to be the product of synchrotron radiation from accelerated electrons. However,
the origin of the higher energy peak remains to be clarified. In leptonic models~\cite{bib:AGNleptonic}, inverse Compton scattering of synchrotron
photons (or other ambient photons) by accelerated electrons generates this high energy emission. In hadronic models~\cite{bib:AGNhadronic},
MeV-TeV gamma-rays and high energy neutrinos are produced through hadronic interactions of the high energy cosmic rays with radiation or gas clouds surrounding the source.
In the latter scenario, a strong correlation between the gamma-ray and the neutrino fluxes is expected. The gamma-ray light curves of bright blazars measured by the
LAT instrument on board the Fermi satellite reveal important time variability on timescales of hours to several weeks, with intensities
much larger than the typical flux of the source in its quiescent state~\cite{bib:FermiLATAGNvariability}.
This paper presents the results of the first time-dependent search for cosmic neutrino sources by the ANTARES telescope. The data sample used in
this analysis and the comparison to Monte Carlo simulations are described in Section 2,
together with a discussion on the systematic uncertainties. The point source search
algorithm used in this time-dependent analysis is explained in Section 3. The search results are presented in
Section 4 for ten selected candidate sources.
\section{ANTARES}
The ANTARES Collaboration completed the construction of a neutrino
telescope in the Mediterranean Sea with the connection of its twelfth detector line
in May 2008~\cite{bib:Antares}. The telescope is located 40 km off the Southern coast of France
(42$^{\circ}$48'N, 6$^{\circ}$10'E) at a depth of 2475 m. It comprises a three-dimensional array of
photomultipliers housed in glass spheres (optical modules~\cite{bib:OM}), distributed along twelve
slender lines anchored at the sea bottom and kept taut by a buoy at the top.
Each line is composed of 25 storeys of triplets of optical modules (OMs), each housing one 10-inch photomultiplier.
The lines are subject to the sea currents and can change shape and orientation. A positioning system
based on hydrophones, compasses and tiltmeters is used to monitor the detector geometry with an accuracy of $~10$~cm.
The main goal of the experiment is to search for high energy neutrinos with energies greater than 100~GeV by detecting muons produced
by the neutrino charged current interaction in the vicinity
of the detector. Due to the large background from downgoing atmospheric muons, the telescope is optimized
for the detection of upgoing muons as only they can originate from neutrinos.
Muons induce the emission of Cherenkov light in the sea water. The arrival time and intensity of the Cherenkov light on the OMs are digitized into hits and transmitted to shore.
Events containing muons are selected from the continuous deep sea optical backgrounds due to natural radioactivity and bioluminescence. A detailed description of the detector
and the data acquisition is given in~\cite{bib:Antares,bib:antaresdaq}.
The arrival times of the hits are calibrated as described in~\cite{bib:TimeCalib}. A L1 hit is defined either as a high-charge hit, or as hits separated by less
than 20~ns on OMs of the same storey. At least five L1 hits are required throughout the detector within a time window of 2.2~$\mu$s,
with the relative photon arrival times being compatible with the light coming from a relativistic particle. Independently,
events which have L1 hits on two sets of adjacent or next-to-adjacent floors are also selected.
The data used in this analysis were taken in the period from September 6 to
December 31, 2008 (54720 to 54831 modified Julian days, MJD) with the twelve line detector. This period overlaps with the availability of the first data from the LAT instrument
onboard the Fermi satellite. The corresponding effective live time is 60.8 days.
Atmospheric neutrinos are the main source of background in the search for astrophysical neutrinos. These upgoing neutrinos
are produced by the interaction of cosmic rays in the Earth's atmosphere. To account for this background, neutrino events were simulated according to the parametrization of
the atmospheric neutrino flux from Ref.~\cite{bib:horandel}. Only charged
current interactions of muon neutrinos and antineutrinos were considered. An additional source of background
is due to downgoing atmospheric muons mis-reconstructed as upgoing. Downgoing atmospheric muons were simulated with the MUPAGE package~\cite{bib:Mupage}.
In both cases, the Cherenkov light was propagated taking into account light absorption and scattering in sea water~\cite{bib:light}.
From the timing and position information of the hits, muon tracks are reconstructed
using a multi-stage fitting procedure, based on Ref.~\cite{bib:AAfit}. The initial fitting stages
provide the hit selection and starting point for the final fit. The final stage consists of a
maximum likelihood fit of the observed hit times and includes the contribution of optical
background hits.
Upgoing tracks are also required to have a good reconstruction quality. The latter is quantified by a parameter, $\Lambda$ which is based on the
value of the likelihood function obtained for the fitted muon (see Ref.~\cite{bib:AAfit} for details).
The cumulative distribution of $\Lambda$ for muons reconstructed as upgoing is shown
in Figure~\ref{fig:FitQuality} along with the simulated contributions from atmospheric muons and neutrinos.
The angular uncertainty obtained from the muon track fit is required to be smaller than 1 degree.
For this analysis, events are selected with $\Lambda>-5.4$. This value results in an optimal compromise between the atmospheric neutrino and muon background reduction and
the efficiency of the cosmic neutrino signal with an assumed spectrum proportional to $E_{\nu}^{-2}$,
where $E_{\nu}$ is the neutrino energy, which gives the best 5$\sigma$ discovery potential. The resulting sample consists of 628 events obtained in 60.8 days.
The simulations indicate that the selected sample contains 60~\% atmospheric neutrinos; the rest being mis-reconstructed atmospheric muons.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.9\textwidth]{lambda_flare_19102011.eps}
\caption{Track fit quality ($\Lambda$) distribution for upgoing events
in data (dots) and Monte Carlo samples (atmospheric muons: dashed line; atmospheric neutrinos: continuous line). Events are selected with an error estimate lower
than 1 degree. The green dashed vertical line corresponds to the optimized event selection ($\Lambda>-5.4$).}
\label{fig:FitQuality}
\end{figure}
The angular resolution of the reconstructed neutrino direction can not be determined directly from the data and has to be estimated from simulation. However,
comparison of data and Monte Carlo in which the time accuracy of the hits was degraded by up to 3~ns constrains the
uncertainty of the angular resolution to about 0.1$^{\circ}$~\cite{bib:AAfitps}. Figure~\ref{fig:Angres} shows the cumulative distribution of the
angular difference between the reconstructed muon direction and the neutrino direction for an assumed spectrum proportional to $E_{\nu}^{-2}$. For the
considered period, the median resolution is estimated to be 0.5 $\pm$ 0.1 degrees.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.8\textwidth]{papier_flare_fig_angres2_26102011.eps}
\caption{Cumulative distribution of the angle between the true Monte Carlo neutrino direction ($\alpha_{\nu_{MC}}$) and the reconstructed
muon direction ($\alpha_{rec}$) for an E$_{\nu}^{-2}$ flux of upgoing neutrino events selected for this analysis.}
\label{fig:Angres}
\end{figure}
The effective area for muon neutrinos is defined as the ratio between the rate of selected neutrino events and the cosmic neutrino flux. Figure~\ref{fig:Seff} shows the muon
neutrino and antineutrino effective area
of the ANTARES telescope as a function of the declination of the source, after
integrating over the energy with an assumed spectrum proportional to $E_{\nu}^{-2}$ between 10~GeV and 10~PeV. In the flux limits (see Section 4), a conservative uncertainty on the
detection efficiency of about 30~\% was taken into account. This number includes contributions on the uncertainty of the sea water optical parameters~\cite{bib:light} and the OM
properties such as efficiency and angular acceptance.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.8\textwidth]{fig_papier_seff_26102011.eps}
\caption{ANTARES muon neutrino and antineutrino effective area (continuous line) as a function of the declination of the source computed from the Monte Carlo simulation for an E$_{\nu}^{-2}$
flux of upgoing muons selected for this analysis. The product of the effective area by the visibility (i.e. fraction of the time the source is visible at the ANTARES location) is shown with the dashed line.}
\label{fig:Seff}
\end{figure}
\section{Time-Dependent Search Algorithm}
The time-dependent point source analysis is performed using an unbinned method based on a
likelihood ratio maximization. The data are parametrized as a
mixture of signal and background. The goal is to determine, at a given point in the sky and at a given
time, the relative contribution of each component and to calculate the probability to have a signal
above background in a given model. The likelihood ratio, $\lambda$, is the logarithm of the ratio of the probability density for the hypothesis
of signal and background ($H_{sig+bkg}$) over the probability density of only background ($H_{bkg}$):
\begin{equation}\label{eq:EQ_likelihood}
\lambda=\sum_{i=1}^{N} log\frac{P(x_{i}|H_{sig+bkg})}{P(x_{i}|H_{bkg})} = \sum_{i=1}^{N} log\frac{\frac{n_{sig}}{N}P_{sig}(\alpha_{i},t_{i}) + (1-\frac{n_{sig}}{N})P_{bkg}(\delta_{i},t_{i})}{P_{bkg}(\alpha_{i},t_{i})}
\end{equation}
where $n_{sig}$ is the unknown number of signal events determined by the fit and N is the total number of events in the considered data sample.
$P_{sig}(\alpha_{i},t_{i})$ and $P_{bkg}(\delta_{i},t_{i})$ are the probability density functions (PDF) for signal and background respectively.
For a given event \textit{i}, $t_{i}$, $\delta_{i}$ and $\alpha_{i}$ represent the time of the event, its declination and the angular separation
from the source under consideration.
The probability densities $P_{sig}$ and $P_{bkg}$ are factorized into a purely directional and a purely time-related component.
The shape of the time PDF for the signal event is extracted directly from the gamma-ray light curve assuming
proportionality between the gamma-ray and the neutrino fluxes. It is assumed that the muon neutrino velocity in vacuum is equal to that of light in vacuum.
For signal events, the directional PDF is described by the one
dimensional point spread function (PSF), which is the probability density of reconstructing an event at an angular distance $\alpha$ from the true source position.
The directional and time PDF for the background are derived from the data using the observed declination distribution of the selected events
and the observed one-day binned time distribution of all the reconstructed muons respectively.
Figure~\ref{fig:TimeDistri} shows the time distribution of all the reconstructed events and the selected upgoing events for this analysis. Once normalized to an integral
equal to 1, the distribution for all reconstructed events is used directly as the time PDF for the background.
Empty bins in the histograms correspond to periods with no data taking (i.e. detector in maintenance) or with very poor quality data (high bioluminescence or
bad calibration).
\begin{figure}[ht!]
\centering
\includegraphics[width=0.8\textwidth]{fig_papier_flare_timedistri_26102011.eps}
\caption{Time distribution of the reconstructed events. Upper histogram (black line): distribution of all reconstructed events. Bottom filled histogram (red): distribution of selected upgoing events.
}
\label{fig:TimeDistri}
\end{figure}
The statistical interpretation of the search result relies on simulated pseudo experiments (PE) in which the background events are randomly generated by
sampling the declination and the time from the parametrization $P_{bkg}(\delta_{i},t_{i})$ and the right ascension from a uniform distribution. Events from a
neutrino point source are simulated
by adding events around the desired coordinates according to the point spread function and the time distribution of the studied source. Systematic uncertainties
(cf Section 2) are incorporated directly into the pseudo experiment generation.
The null hypothesis corresponds to $n_{sig}=0$. The obtained value of $\lambda_{data}$ on the data is then
compared to the distribution of $\lambda(n_{sig}=0)$. Large values of $\lambda_{data}$ compared to the distribution of $\lambda(n_{sig}=0)$ reject the null hypothesis
with a confidence level (C.L.) equal to the fraction of the number of PE above $\lambda_{data}$. The fraction of
PE for which $\lambda(n_{sig}=0)$ is above $\lambda_{data}$ is referred to as the p-value. The discovery potential is then defined as the average number of signal events required to
achieve a p-value lower than 5$\sigma$ in 50~\% of the PEs. In the same way, the sensitivity is defined as the average signal required to obtain a p-value
less than that of the median of the $\lambda(n_{sig}=0)$ distribution in 90~\% of the PEs. In the absence of evidence of a signal, an upper limit on the
neutrino fluence is obtained and defined as the integral in energy and time of the flux upper limit with an assumed energy spectrum proportional to
$E_{\nu}^{-2}$ from 10 GeV to 10 PeV. The limits are calculated according to the classical (frequentist) method for upper limits~\cite{bib:Neyman}.
The performance of the time-dependent analysis was computed by applying this unbinned algorithm for a single source assuming a single square-shape flare with a width varying from 0.01 days to 84 days.
The solid line in Figure~\ref{fig:Nev5sigma} shows the average number of events required for a discovery from one source located at a declination of -40$^{\circ}$ as
a function of the width of the flare. The numbers in the black line are compared to that obtained without using the timing information (dashed line). The flare timing information yields
an improvement of the discovery potential by about a factor 2-3 with respect to a standard time-integrated point source search~\cite{bib:AAfitps}.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.8\textwidth]{papier_flare_figure_nevent_vs_deltaT2_26102011.eps}
\caption{Average number of events (solid line) required for a 5$\sigma$ discovery (50~\% probability) from a single source located at a
declination of -40$^{\circ}$ as a function of the width of the flare period ($\sigma_{t}$) for the 60.8 day analysis. These numbers are compared to that obtained without
using the timing information (dashed line).}
\label{fig:Nev5sigma}
\end{figure}
\section{Search for Neutrino Emission from Gamma-Ray Flares}
The time-dependent analysis was applied to bright and variable Fermi blazar sources reported in the first-year Fermi LAT catalogue~\cite{bib:Fermicatalogue} and
in the LBAS catalogue (LAT Bright AGN sample~\cite{bib:FermicatalogueAGN}). Sources were selected in the sky visible to ANTARES and that had at least one day binned gamma-ray flux
in the high state periods greater than 80x10$^{-8}$ photons cm$^{-2}$ s$^{-1}$ above 100~MeV and showed significant time variability on time scales of days to weeks in the
studied time period. A source is assumed variable in the LBAS catalogue when the observation has a probability of less than 1~\% of being a steady source.
This list includes six flat-spectrum radio quasars and four BL-Lacs. Only four bright and nearby sources in the considered sample, PKS2155-304~\cite{bib:TeVsources1},
PKS1510-089~\cite{bib:TeVsources2}, 3C279~\cite{bib:TeVsources3} and WComae~\cite{bib:TeVsources4}, have been detected by the ground Cherenkov telescopes
HESS, MAGIC or VERITAS. Table~\ref{tab:Sources} lists the characteristics of the ten selected sources.
\begin{table}[ht!]
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
Name & {OFGL name} & Class & {RA [$^{o}$]} & {Dec [$^{o}$]} & Redshift \\
\hline
\hline
{PKS0208-512} & {J0210.8-5100} & FSRQ & 32.70 & -51.2 & 1.003 \\
\hline
{AO0235+164} & {J0238.6+1636} & BLLac & 39.65 & 16.61 & 0.940 \\
\hline
{PKS0454-234} & {J0457.1-2325} & FSRQ & 74.28 & -23.43 & 1.003 \\
\hline
{OJ287} & {J0855.4+2009} & BLLac & 133.85 & 20.09 & 0.306 \\
\hline
{WComae} & {J1221.7+28.14} & BLLAc & 185.43 & 28.14 & 0.102 \\
\hline
{3C273} & {J1229.1+0202} & FSRQ & 187.28 & 2.05 & 0.158 \\
\hline
{3C279} & {J1256.1-0548} & FSRQ & 194.03 & -5.8 & 0.536 \\
\hline
{PKS1510-089} & {J1512.7-0905} & FSRQ & 228.18 & -9.09 & 0.36 \\
\hline
{3C454.3} & {J2254.0+1609} & FSRQ & 343.50 & 16.15 & 0.859 \\
\hline
{PKS2155-304} & {J2158.8-3014} & BLLac & 329.70 & -30.24 & 0.116 \\
\hline
\end{tabular}
\caption{List of bright variable Fermi blazars selected for this analysis~\cite{bib:FermicatalogueAGN}.}
\label{tab:Sources}
\end{center}
\end{table}
The light curves published on the Fermi web page for the monitored sources~\cite{bib:Fermimonitored} are used for this analysis. They correspond to
the one-day binned time evolution of the average gamma-ray flux above a threshold of 100~MeV since August 2008. The high state periods are defined
using a simple and robust method based on three main steps. Firstly, the baseline is determined with an iterative linear fit. After each fit, bins more than two sigma ($\sigma_{BL}$)
above the baseline (BL) are removed. Secondly, seeds for the high state periods are identified by searching for bins significantly above the baseline according to the criteria:
\begin{equation}\label{eq:EQ_selection}
(F - \sigma_{F}) > (BL + 2*\sigma_{BL})~~and~~ F > (BL + 3*\sigma_{BL})
\end{equation}
where F and $\sigma_{F}$ represent the flux and the uncertainty on this flux for each bin, respectively. For each seed, the adjacent bins for which the emission is
compatible with the flare are added if they satisfy: $(F - \sigma_{F}) > (BL + \sigma_{BL})$. Finally, an additional delay of 0.5 days is added before and after the flare in order to take into
account that the precise time of the flare is not known (1-day binned light curve). With this definition, a flare has a width of at least two days. Figure~\ref{fig:3C454} shows
the time distribution of the Fermi LAT gamma-ray light curve of 3C454.3 for almost two years of data and the corresponding selected high state periods. With the hypothesis that the neutrino
emission follows the gamma-ray emission, the signal time PDF is simply the normalized light curve of only the high state periods. The third column of Table~\ref{tab:Results}
lists the flaring periods for the ten sources found from September to December 2008.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.9\textwidth]{time_distribution_3C454.3_26102011.eps}
\caption{Gamma-ray light curve (black points) of the blazar 3C454.3 measured by the LAT instrument onboard the Fermi satellite above 100~MeV for almost two years
of data. The shaded histogram (blue) indicates the high state periods. The dashed line (red) represents the fitted baseline.
}
\label{fig:3C454}
\end{figure}
The results of the search for coincidences between flares and neutrinos are listed in Table~\ref{tab:Results}. For nine sources, no coincidences are found.
For 3C279, a single high-energy neutrino event is found in coincidence during a large flare in November 2008. Figure~\ref{fig:Result_3C279} shows the time
distribution of the Fermi gamma-ray light curve of 3C279 and the time of the coincident neutrino event. This event was reconstructed with 89 hits distributed on
ten lines with a track fit quality $\Lambda=-4.4$. The particle track direction is reconstructed at 0.56$^{\circ}$ from the source location.
The pre-trial p-value is 1.0~\%. However, the post-trial probability computed taking into account the ten searches is 10~\%; this occurrence is thus compatible with
a background fluctuation. In the absence of a discovery, upper limits on the neutrino fluence were computed and are shown in the last column of Table~\ref{tab:Results}.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.9\textwidth]{result_3C279_19102011.eps}
\caption{Gamma-ray light curve (dots) of the blazar 3C279 measured by the LAT instrument onboard the Fermi satellite above 100 MeV. The light shaded histogram (blue)
indicates the high state periods. The dashed line (green) corresponds to the fitted baseline. The red histogram displays the time of the associated ANTARES neutrino event.
}
\label{fig:Result_3C279}
\end{figure}
\begin{table}[ht!]
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
Source & Vis & {timePDF(MJD-54000)} & {LT} & {N($5\sigma$)} & $N_{obs}$ & {Fluence U.L.} \\
\hline
\hline
{PKS0208-512} & 1.0 & {712-5,722-4,745-7,}& 8.8 & 4.5 & 0 & 2.8 \\
& & {750-2,753-7,764-74,} & & & & \\
& & {820-2} & & & & \\
\hline
{AO0235+164} & 0.41 & {710-33,738-43,746-64,} & 24.5 & 4.3 & 0 & 18.7\\
& & {766-74,785-7,805-8,} & & & & \\
& & { 810-2} & & & & \\
\hline
{PKS1510-089} & 0.55 & {716-9,720-5,726-35,} & 4.9 & 3.8 & 0 & 2.8\\
& & {788-90,801-3} & & & & \\
\hline
{3C273} & 0.49 & {714-6,716-8,742-5} & 2.4 & 2.5 & 0 & 1.1 \\
\hline
{3C279} & 0.53 & {749-51,787-809,}& 13.8 & 5.0 & 1 & 8.2 \\
& & {812-5,817-21,824-6} & & & & \\
\hline
{3C454.3} & 0.41 & {713-51,761-5,767-9,} & 30.8 & 4.4 & 0 & 23.5 \\
& & {784-801} & & & & \\
\hline
{OJ287} & 0.39 & {733-5,752-4,760-2,}& 4.3 & 3.9 & 0 & 3.4 \\
& & {768-70,774-6,800-2,} & & & & \\
& & {814-6} & & & & \\
\hline
{PKS0454-234} & 0.63 & {743-5,792-6,811-3} & 6.0 & 3.3 & 0 & 2.9 \\
\hline
{WComae} & 0.33 & {726-9,771-3,790-2,}& 3.9 & 3.8 & 0 & 3.6 \\
& & {795-7,815-7} & & & & \\
\hline
{PKS2155-304} & 0.68 & {753-5,766-8,799-801,}& 3.1 & 3.7 & 0 & 1.6 \\
& & {828-30} & & & & \\
\hline
\end{tabular}
\caption{Results of the search for neutrino emission in the ten selected sources. The meaning of the columns is the following: Vis: fraction of the time the source is visible
at the ANTARES location;
timePDF: high state periods of the light curve; LT: corresponding ANTARES live time in days; N($5\sigma$): averaged number of events required for a 5$\sigma$
discovery (50~\% probability); $N_{obs}$: number of observed events in time/angle coincidence with the gamma-ray emission. Fluence U.L.: Upper limit (90~\% C.L.) on
the neutrino fluence in GeV~cm$^{-2}$.}
\label{tab:Results}
\end{center}
\end{table}
\section{Summary}
This paper presents the first time-dependent search for cosmic neutrinos using the data taken with the full twelve line ANTARES detector during the last four months of 2008.
For variable sources, time-dependent point searches are much more sensitive than time-integrated searches due to the large reduction of the background. This
search was applied to ten very bright and variable Fermi LAT blazars. One neutrino event was detected in time/direction coincidence with the gamma-ray emission in only one case, for a flare of 3C279
in November 2008, with a post-trial probability of 10~\%. Upper limits were obtained on the neutrino fluence for the ten selected sources.
\section{Acknowledgments}
The authors acknowledge the financial support of the funding agencies:
Centre National de la Recherche Scientifique (CNRS), Commissariat
\`a l'\'energie atomique et aux \'energies alternatives (CEA), Agence
National de la Recherche (ANR), Commission Europ\'eenne (FEDER fund
and Marie Curie Program), 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; Ministerio de Ciencia
e Innovaci\'on (MICINN), Prometeo of Generalitat Valenciana and MultiDark,
Spain. We also acknowledge the technical support of Ifremer, AIM and
Foselev Marine for the sea operation and the CC-IN2P3 for the computing facilities.
|
\section{Introduction}
Let $V$ be an $\ell$-dimensional Euclidean space.
Let $\Phi$ be an {\bf
irreducible
(crystallographic)
root system}
in the dual space $V^{*}$.
Fix
a set
$\Phi^{+} $
of {\bf positive roots}.
For any $\alpha\in\Phi^{+} $ and
$j\in\Z$,
the affine hyperplane
$$
H_{\alpha, j}
:=
\{x\in V \mid \alpha(x)=j\}
$$
is a parallel translation of $H_{\alpha} :=H_{\alpha, 0}$.
The arrangement
$\A(\Phi) := \{H_{\alpha} \mid \alpha\in \Phi^{+} \}
$
is called
the
{\bf
crystallographic arrangement}
of the type
$\Phi$.
\begin{define}
Let $k\in\Z_{>0} $.
An \textbf{extended Shi arrangement} $\mathit{Shi}^{k}$
of the type $\Phi$ is an affine arrangement defined by
$$
\mathit{Shi}^{k}
:=
\mathit{Shi}^{k}(\Phi^{+})
:=
\{H_{\alpha, j} ~\mid~ \alpha\in \Phi^{+},\,\,\,j\in\Z, \,-k+1 \leq j\leq k\}.
$$
\label{shi}
\end{define}
The extended Shi arrangements
for $k=1$ were introduced by
J.-Y. Shi \cite{Shi1, Shi2}
in his study of the Kazhdan-Lusztig
representation theory of the affine
Weyl groups.
For $k\geq 1$, they were
studied in \cite{Sta1, Athana1} among others.
Let $S:=S(V^{*} )$ be
the symmetric algebra of the dual space $V^{*} $ of $V$.
Recall that the {\bf
cone}
\cite[Definition 1.15]{OT0}
$$
{\mathcal S}^{k}
:=
{\mathcal S}^{k} (\Phi^{+})
:=
{\mathbf c}{\mathit{Shi}}^{k}
$$
over
$\mathit{Shi}^{k}$
is a central arrangement in
an
$(\ell+1)$-dimensional Euclidean space $E:=\R^{\ell+1}$.
Choose $\alpha_{H} \in E^{*} $ with
$H = \ker (\alpha_{H} )\in {\cal S}^{k} $.
Let ${\Der} (S)$ denote the $S$-module of derivations of $S$
to itself.
Since $V$ is embedded in $E$ as the affine hyperplane
defined by $z=1$,
we may consider the {\bf Ziegler restriction map} \cite{Z}
\begin{align*}
{\mathbf{res}} :
D_{0} ({\mathcal S}^{k})
\longrightarrow
D(\A(\Phi), 2k)
\end{align*}
by setting $z=0$.
Here
\begin{align*}
&D_0({\mathcal S}^{k})
:=
\{
\theta\in \Der(S(E^{*} ))~|~
\theta(\alpha_H) \in \alpha_H S(E^{*})
{\text{~for~each~}} H \in \mathcal{S}^{k},
\theta(z)=0 \},
\\
&D(\A(\Phi), 2k)
:=
\{\theta\in \Der(S(V^{*}))
~|~ \theta(\alpha)\in\alpha^{2k} S(V^{*} )
\text{~for each ~}\alpha\in\Phi^{+}
\}.
\end{align*}
Yoshinaga \cite{Y}
verified the
Edelman-Reiner conjecture
in \cite{EdRei}
by proving the surjectivity
of the Ziegler restriction map.
In particular,
the homogeneous part
$$
{\mathbf{res}}
:
D_0({\mathcal S}^{k})_{kh}
\longrightarrow
D(\A(\Phi), 2k)_{kh}
$$
of degree $kh$ of the Ziegler restriction map
is a linear isomorphism, where
$h$ denotes the Coxeter number.
On the other hand,
as we will
see in Proposition \ref{Xi},
$D(\A(\Phi), 2k)_{kh}$
is isomorphic to
$V$
as
a
$W$-module.
Now we are ready to introduce
the simple-root bases for
$D_{0} ({\mathcal S}^{k})_{kh}$.
\begin{definition}
\label{SRB}
Pick a $W$-isomorphism
$
\Xi :
V \tilde{\longrightarrow}
D(\A(\Phi), 2k)_{kh}.
$
Define
a linear isomorphism
$\Theta:= {\bf res}^{-1} \circ \Xi$:
\[
\xymatrix{
V
\ar[dr]^(0.55){{\!\!\!{\scalebox{1.05}{$\Xi$}}}}_(0.55){\rotatebox{-30}{\scalebox{1.1}{\!\!\!\!\!\!$\sim$}}}
\ar[r]^(0.4){\scalebox{1.05}{$\Theta$}}_(0.4){\scalebox{1.05}{$\sim$}}
&
D_{0}({\Shi}^{k})_{kh} \ar[d]^{\scalebox{1.05}{${\mathbf{res}}$}}_{\rotatebox{90}{$\sim$}}
\ar@{}[ld]|(0.3){\scalebox{1.05}{$~\circlearrowright$}}
\\
&
D({\mathcal{A}}(\Phi),2k)_{kh}. \\
}
\]
Let $\Delta:=\{\alpha_{1}, \dots , \alpha_{\ell} \}\subset V^{*} $
be the set of simple
roots.
Let $\Delta^{*} :=\{\alpha_{1}^{*} , \dots , \alpha_{\ell}^{*} \}\subset V$
denote the dual basis of $\Delta$.
Then
(1)
the derivations
$$
\varphi^{+}_{j} := \Theta(\alpha_{j}^{*} ) \,\,(j=1, \dots, \ell)
$$
are called a {\bf simple-root basis plus (SRB$+$)},
and
(2) the derivations
$$
\varphi^{-}_{j} := \sum_{p=1}^{\ell} I^{*}(\alpha_{j}, \alpha_{p})
\varphi^{+}_{p}\,\,(j=1, \dots, \ell)
$$
are called a {\bf simple-root basis minus (SRB$-$)}.
Here $I^{*} $ is the natural inner product on $V^{*} $
induced from the inner product $I$ on $V$.
\end{definition}
These two bases
are uniquely determined up to a nonzero constant multiple
because of Schur's lemma.
The bases have the following nice
properties:
\begin{theorem}
\label{characterization}
(1)
Let
$
\varphi^{+}_{1},
\dots,
\varphi^{+}_{\ell}
$ be an SRB$+$. Then
each $
\varphi^{+}_{i}(\alpha_{j}+kz)$ is divisible by $\alpha_{j}+kz$
whenever $i\neq j$.
Conversely, if derivations $
\phi^{+}_{1},
\dots,
\phi^{+}_{\ell}
\in
D_{0}({\Shi}^{k})_{kh}
$ are given and
each $\phi^{+}_{i}(\alpha_{j}+kz)$ is divisible by $\alpha_{j}+kz$
whenever $i\neq j$,
then
there exist nonzero constants
$
c^{+}_{1},
\dots,
c^{+}_{\ell}
$
satisfying
$\phi^{+}_{i} = c^{+}_{i} \varphi^{+}_{i}$ for any $i$.
\noindent
(2) Let $
\varphi^{-}_{1},
\dots,
\varphi^{-}_{\ell}
$ be an
SRB$-$.
Then
each derivation $\varphi^{-}_{i}$ is divisible by $\alpha_{i}-kz$.
Conversely, if derivations
$
\phi^{-}_{1},
\dots,
\phi^{-}_{\ell}
\in
D_{0}({\Shi}^{k})_{kh}
$ are given and
each derivation $\phi^{-}_{i}$ is divisible by $\alpha_{i}-kz$,
then
there exist nonzero constants
$
c^{-}_{1},
\dots,
c^{-}_{\ell}
$
satisfying
$\phi^{-}_{i} = c^{-}_{i} \varphi^{-}_{i}$ for any $i$.
\end{theorem}
The organization of this article is as follows.
In Section 2, we review a recent refinement \cite{AY2}
of Yoshinaga's freeness criterion \cite{Y}
before proving the two key results (Propositions \ref{BGamma+}
and \ref{BGamma-})
which we apply in Section 3 when
we prove
Theorem \ref{characterization}.
We also characterize the simple roots
in terms of the freeness of deleted/added
Shi arrangements in Theorem \ref{simplefree}.
The actions of a simple reflection on the SRB$+/-$
are studied as well.
In Section 4, we will describe a unique
$W$-invariant derivation
({\bf $k$-Euler derivation})
related to the Catalan arrangement in terms of the
SRB$+$.
\medskip
\noindent
\textbf{Acknowledgements}.
The first author is partially supported by JSPS Grants-in-Aid for Young Scientists
(B)
No. 24740012.
The second author is partially supported by JSPS Grants-in-Aid, Scientific Research
(A)
No. 24244001.
\section{Freeness criteria}
In the rest of the article we use \cite{OT0} as a general reference.
Let $\cal C$ be a central arrangement in an $(\ell+1)$-dimensional vector space
$E$.
Choose $\alpha_{H} \in E^{*} $
with $\ker \alpha_{H} = H\in\cal C$.
Let $\mathbf m$ be a {\bf
multiplicity} of $\cal C$:
\[
{\mathbf m} : {\cal C} \rightarrow {\Z_{>0}}.
\]
Let $S(E^{*})$ be the ring of polynomial functions on $E$.
Let $\Der(S(E^{*}))$ be the set of derivations of $S(E^{*})$ to itself:
\begin{align*}
\Der(S(E^{*})):= \{\theta : S(E^{*})\rightarrow
S(E^{*})~|~&\theta \text{~is $\R$-linear
and~} \theta(fg)=f\theta(g)+g\theta(f)\\
&\text{~for~any~} f, g\in S(E^{*}) \}.
\end{align*}
A derivation $\theta\in \Der(S(E^{*}))$
is said to be {\bf
homogeneous of degree} $d$ if
$\theta(\alpha)$ is a
homogeneous polynomial
of
degree $d$
for any $\alpha\in V^{*} $ unless $\theta(\alpha)=0.$
Define an $S(E^{*})$-module
\[
D({\cal C}, {\mathbf m})
:=
\{
\theta\in\Der(S(E^{*}))
~|~
\theta(\alpha_{H} )\in \alpha_{H}^{{\bf m}(H)} S(E^{*})
\text{~for each~}H\in {\cal C}
\}.
\]
When
$D({\cal C}, {\mathbf m})
$
is a free
$S(E^{*} )$-module, we say that a multiarrangement
$({\cal C}, {\bf m} )$ is {\bf free}.
We say that ${\cal C}$ is a {\bf free arrangement}
if $({\cal C}, {\bf 1} )$ is free.
Here ${\bf 1} $ indicates the constant multiplicity whose value
is equal to one.
When
$({\cal C}, {\bf m} )$ is {free},
$\mathbf{\exp}
({\cal C}, {\bf m} )
$
of {\bf exponents} denotes the set of degrees of homogeneous basis for
$D({\cal C}, {\mathbf m})
$.
We simply write
${\exp}({\cal C})$
instead of
$\exp
({\cal C}, {\bf 1} )
$
if $\mathcal C$ is a free arrangement.
For a fixed
hyperplane $H_{0}\in\calC $,
define a multiarrangement
$(\calC'', \bfz)$, which we call the {\bf Ziegler restriction}
\cite{Z}, by
\begin{align*}
\calC'' :=
\{
H_{0} \cap K ~|~ K\in \calC':=\calC\setminus\{H_{0} \}\},
\,\,
\bfz (X):= |\{
K\in\calC'~|~ X=K\cap H_{0}
\}|,
\end{align*}
where $\calC''$ is an arrangement in $H_{0} $
and
$X\in\calC''$.
For the intersection lattice $L(\calC)$
\cite[Definition 2.1]{OT0} of $\calC$
and any $Y\in L(\calC)$
define the {\bf
localization} $\calC_{Y} $ of $\calC$ at $Y$ by
$
\calC_{Y} :=\{H\in\calC~|~Y\subseteq H\}.
$
Let us present a recent refinement
of Yoshinaga's freeness criterion in
\cite{Y}:
\begin{theorem}
[\cite{AY2}]
\label{AbeYoshinagaCriterion}
Suppose $\ell+1>3$.
For a central arrangement $\calC$ and an
arbitrary hyperplane
$H_{0}\in\calC $, the following two conditions are
equivalent:
(1) $\calC$ is a free arrangement,
(2) (2-i) the Ziegler restriction $(\calC'', \bfz)$
is free
and
(2-ii)
$\calC_{Y}$ is free for any $Y\in L(\calC)$
such that $Y\subset H_{0} $ with $\codim_{H_{0}} Y =2$.
\end{theorem}
For a fixed hyperplane $H_{0} \in\mathcal C$, we may choose
a basis $x_{1}, x_{2},\dots , x_{\ell}, z $
for $E^{*} $ so that the hyperplane $H_{0} $ is defined by
the equation $z=0$.
Then the Ziegler restriction
$(\mathcal C'', \mathbf z)$ is a multiarrangement in $H_{0} $.
Let
\[
D_{0}(\mathcal C):=\{\theta\in D(\mathcal C)~|~
\theta(z)=0
\}.
\]
Then
\[
D(\mathcal C)
=
S\theta_{E} \oplus
D_{0} (\mathcal C),
\]
where
$\theta_{E}$ is the Euler derivation.
Note that $D_{0} (\mathcal C)$ is a free $S(E^{*} )$-module
if and only if
$\mathcal C$ is a free arrangement.
When
$\mathcal C$ is a free arrangement,
let $\exp_{0} (\mathcal C)$
denote the set of degrees of homogeneous basis for
$D_{0}(\mathcal C) $.
Note that the set $\exp_{0} (\mathcal C)$
does not depend upon the choice of $H_{0}. $
When we describe $\exp_{0} (\mathcal C) $,
we will
use the notation $a^{n} $ instead of listing $a, \dots, a$ ($n$ times).
\begin{theorem}
[Ziegler \cite{Z}]
\label{Ziegler}
The Ziegler restriction map
\[
{\mathbf{res}}:
D_{0} (\mathcal C)
\rightarrow
D(\calC'', \bfz)
\]
defined by setting $z=0$
is surjective if
$\mathcal C$ is
a free arrangement.
\end{theorem}
The following theorem was proved
in
\cite{AY} using
the shift isomorphism of
Coxeter multiarrangements:
\begin{theorem}[\cite{AY}, Corollary 12]
\label{AbeYoshinagaCorollary}
Let $\A$ be a Coxeter arrangement in an $\ell$-dimensional Euclidean space.
For a $\{0, 1\}$-valued multiplicity $\mathbf{m} :
\A \rightarrow \{0, 1\}$ and an integer $k>0$, the following conditions are
equivalent:
(1) a multiarrangement $(\A, \mathbf{m} )$ is free with
exponents $(e_{1} , \dots, e_{\ell} )$.
(2) a multiarrangement $(\A, 2k+\mathbf{m} )$ is free with
exponents $(kh+e_{1} , \dots, kh+e_{\ell} )$.
(3) a multiarrangement $(\A, 2k-\mathbf{m} )$ is free with
exponents $(kh-e_{1} , \dots, kh-e_{\ell} )$.
\end{theorem}
Now we go back to the situation in Section 1:
let $\Phi, \Phi^{+}$ and $ \Delta $ be an irreducible root system,
a set of postive roots, and the set of simple roots
respectively.
The following two Propositions \ref{BGamma+} and
\ref{BGamma-} are keys to our proof of
Theorem \ref{characterization}. They are dual to each other.
\begin{prop}
\label{BGamma+}
For any subset $\Gamma$ of the simple system
$\Delta$, the arrangement
$$
\B_{\Gamma}^{+} :=
\B_{\Gamma}^{+}(\Phi^{+}) :=
{\mathcal S}^{k}
\cup \bigcup_{\alpha\in \Gamma} \{{\mathbf c} H_{\alpha, -k }\}
$$
is a free arrangement
with
$${\rm exp}_{0} (\B_{\Gamma}^{+})
=
((kh+1)^{|\Gamma|}, (kh)^{\ell-|\Gamma|}).
$$
\end{prop}
\noindent
\textbf{proof}.
{\it Case 1.}
When $\ell=2$,
$\Phi$ is of the type either $A_{2} $, $B_{2} $
or $G_{2} $.
Then $\exp_{0} (\mathcal S^{k})=((kh)^{2})
=(kh,kh)$ and
$\Delta=\{\alpha_1,\alpha_2\}$.
For an affine $2$-arrangement $\A$ and an affine line $H_{0}$,
define
\[
\A\cap H_{0} :=
\{K\cap H_{0} \mid K\in \A, K\neq H_{0} \}.
\]
Then, by directly counting intersection points,
we get the following equalities:
\begin{align*}
|\mathit{Shi}^{k}
\cap H_{\alpha, -k}|&=kh~~(\alpha\in\Delta),\\
|(\mathit{Shi}^{k} \cup \{H_{\alpha_1,-k}\})
\cap H_{\alpha_{2} , -k}|&=kh+1.
\end{align*}
Thus
we
may verify the statement by
applying the addition theorem \cite{T} \cite[Theorem 4.49]{OT0}
to $\mathcal S^{k} $
for the types of
$A_{2}, B_{2}$ and $G_{2}$.
\medskip
{\it Case 2.}
Suppose that $\ell \ge 3$.
We will apply Theorem \ref{AbeYoshinagaCriterion}
by verifying the two conditions
(2-i) and (2-ii).
(2-i)
Note that
the Ziegler restriction of $\B_\Gamma^{+}$ to
the hyperplane $H_\infty$
at infinity
coincides with $(\A(\Phi), {\bfz}_\Gamma^{+} )$,
where
\[
\bfz_{\Gamma}^{+} (H_{\alpha}) =
|\{j ~|~ {\mathbf c}H_{\alpha, j} \in
{\mathcal B}_{\Gamma}^{+}\}|
=
\begin{cases}
2k+1 & {\text{~if~}} \alpha\in \Gamma\\
2k & {\text{~otherwise}}
\end{cases}
\,\,\,
\,\,\,
\,\,\,
(\alpha\in\Phi^{+}).
\]
If $\Gamma$ is empty, then
$\bfz_{\Gamma}^{+} = \bfz_{\emptyset}^{+} \equiv 2k$.
Note that $\Gamma$ is linearly independent
because it is a set consisting of simple roots.
Thus the arrangement
\[
\A(\Gamma) := \{H_{\alpha} ~|~ \alpha\in \Gamma\}
\]
is a free (Boolean)
subarrangement of $\A(\Phi)$.
Let $\chi_{\Gamma} $ be
the characteristic function of $\A(\Gamma)$ in $\A(\Phi)$:
\[
\chi_{\Gamma}
(H_{\alpha} ) =
\begin{cases}
1 & {\text{~if~}} \alpha \in \Gamma\\
0 & {\text{~otherwise}}
\end{cases}
\,\,\,
\,\,\,
\,\,\,
(\alpha\in\Phi^{+}).
\]
Since $\bfz_{\Gamma}^{+} =
\bfz_{\emptyset}^{+} + \chi_{\Gamma} $, we may apply
Theorem \ref{AbeYoshinagaCorollary}
to conclude
that
$(\A(\Phi), \bfz_\Gamma)$
is a free multiarrangement
with exponents $((kh+1)^{|\Gamma|},(kh)^{\ell-|\Gamma|})$.
(2-ii)
We will prove that
$(\B_\Gamma^{+})_{Y}$ is free for any
$Y \in L(\B_\Gamma^{+})$
such that $Y \subset H_\infty$
with $\codim_{H_{\infty}} Y=2$.
Define $X$ to be the unique subspace of $E$
such that
$X\in L(\A(\Phi))$ and
${\mathbf c}X \cap H_{\infty} = Y$.
Let
$X^{\perp} :=
\{\alpha\in V^{*} ~|~ \alpha|_{X} \equiv 0\}$.
Then
$
\Phi_{X}
:=
\Phi\cap X^{\perp}
$
is also a (not necessarily irreducible)
root system in $X^{\perp} $.
The set
of positive roots
of
$\Phi_{X}
$
is induced from $\Phi^{+} $:
$\Phi_{X}^{+}=\Phi^{+} \cap \Phi_{X}$.
It is not hard to see (e.g. \cite[Lemma 3.1]{AT})
that
$$
({\mathcal S}^{k})_{Y}
=
{\mathbf c}
\left(
\mathit{Shi}^{k}(\Phi^{+}_{X})
\times \emptyset_{X} \right),$$
where $\emptyset_{X} $ denotes the empty arrangement
in $X$.
Since $\dim X^{\perp} = 2$,
$\Phi_{X}$
is either
of the type
$A_{1} \times A_{1}$,
$A_{2} $,
$B_{2} $ or $G_{2} $.
{\it Case 2.1.}
When
$\Phi_{X}$
is
of the type
$A_{1} \times A_{1}$,
the arrangement
$
\mathit{Shi}^{k}(\Phi^{+}_{X})
$
is a product of two affine
$1$-arrangements.
Thus any subarrangement
of
$
({\mathcal S}^{k})_{Y}
$
is a free arrangement.
In particular,
$(\B_{\Gamma}^{+})_{Y} $
is a free arrangement.
{\it Case 2.2.}
Suppose
that
$\Phi_{X}$
is
of the type either
$A_{2}$,
$B_{2}$
or
$G_{2}$.
Suppose that $\alpha\in\Phi_{X} $ is a simple root of $\Phi$.
Then
$\alpha$
is also a simple root of
$\Phi_{X}$ because it cannot be expressed as a sum of two positive roots of
$\Phi_{X}$.
Thus $\Phi_{X}\cap\Gamma$ consists of simple roots of $\Phi_{X} $.
Therefore
\begin{align*}
(\B_{\Gamma}^{+})_{Y}
&
=
({\mathcal S}^{k})_{Y}
\cup
\bigcup_{\alpha\in \Phi_{X} \cap \Gamma} \{\mathbf{c}H_{\alpha, k} \}\\
&
=
{\mathbf c}
\left(
\left(
\mathit{Shi}^{k}(\Phi^{+}_{X})
\cup
\bigcup_{\alpha\in \Phi_{X} \cap \Gamma} \{H_{\alpha, k} \}
\right)
\times \emptyset_{X}
\right)
\end{align*}
is a free arrangement
because of {\it Case 1}.
Now we apply
Theorem \ref{AbeYoshinagaCriterion}
to complete the proof.
\owari
\begin{cor}
The vector space $D_{0}(\B_{\Gamma}^{+})_{kh} $
is $(\ell-|\Gamma|)$-dimensional.
\label{BGammadim+}
\end{cor}
\begin{prop}
\label{BGamma-}
For any subset $\Gamma$ of the simple system
$\Delta$, the arrangement
$$
\B_{\Gamma}^{-} :=
\B_{\Gamma}^{-}(\Phi^{+}) :=
{\mathcal S}^{k}
\setminus \bigcup_{\alpha\in \Gamma} \{{\mathbf c} H_{\alpha, k }\}
$$
is a free arrangement
with
$${\rm exp}_{0} (\B_{\Gamma}^{-})
=
((kh-1)^{|\Gamma|}, (kh)^{\ell-|\Gamma|}).
$$
\end{prop}
\noindent
\textbf{proof}.
{\it Case 1.}
When $\ell=2$,
$\Phi$ is of the type either $A_{2} $, $B_{2} $
or $G_{2} $.
Let $\exp_{0} (\mathcal S^{k})=((kh)^{2})$ and
$\Delta=\{\alpha_1,\alpha_2\}$.
Then, by directly counting intersection points,
we get the following equalities:
\begin{align*}
|\mathit{Shi}^{k}
\cap H_{\alpha,k}|&=kh~~(\alpha\in\Delta),\\
|(\mathit{Shi}^{k} \setminus \{H_{\alpha_1,k}\})
\cap H_{\alpha_{2} , k}|&=kh-1.
\end{align*}
Thus
we
may verify the statement by
applying the deletion theorem \cite{T} \cite[Theorem 4.49]{OT0}
to $\mathcal S^{k} $
for the types of
$A_{2}, B_{2}$ and $G_{2}$.
The rest is exactly the
same as the proof of
Proposition \ref{BGamma+} if one replaces
$\B_{\Gamma}^{+}$,
$kh+1$,
$H_{\bullet, -k}$,
$2k+1$,
$\cup$,
${\mathbf z}_{\Gamma}^{+} $,
${{\mathbf z}}_{\Gamma}^{+}+\chi_{\Gamma}$
with
$\B_{\Gamma}^{-}$,
$kh-1$,
$H_{\bullet, k}$,
$2k-1$,
$\setminus$,
${\mathbf z}_{\Gamma}^{-} $,
${\mathbf z}_{\Gamma}^{-}-\chi_{\Gamma}$
respectively.
\owari
\begin{cor}
The vector space $D_{0}(\B_{\Gamma}^{-})_{kh-1} $
is
$|\Gamma|$-dimensional.
\label{BGammadim-}
\end{cor}
\section{Proof of main results}
We will prove Theorem \ref{characterization} in this section.
Fix $k\in\Z_{>0} $ throughout in the rest of this article.
We first see that $V$ and $D(\A(\Phi), 2k)_{kh} $ are
$W$-isomorphic as mentioned in Section 1.
Let $F$ be the field of quotients of $S=S(V^{*})
=\R[x_{1} , \dots , x_{\ell}]$.
Recall a primitive derivation $D\in \Der(F)$ associated with $\A(\Phi)$:
$D$ satisfies
\[
D(P_{j} ) =
\begin{cases}
c \in \R^{\times} &\text{~if $j=\ell$ },\\
0 & \text{~if $1\leq j\leq \ell-1$ }.
\end{cases}
\]
Here $P_{1}, \dots , P_{\ell} $
are basic invariants of the invariant subring
$S^{W} $ with
\[
2=\deg P_{1} < \deg P_{2} \leq \dots\leq \deg P_{\ell-1} < \deg P_{\ell} =h.
\]
Then the choice of $D$ has the ambiguity of nonzero constant multiples.
Consider the Levi-Civita connection with respect to the inner product
$I$:
\begin{align*}
\nabla :
{\rm Der}(F) \times {\rm Der}(F) \longrightarrow
{\rm Der}(F),
\,\,\,\,
(\xi, \eta) \mapsto
\nabla_{\xi} \eta
\end{align*}
Note that
$$
\left(\nabla_{\xi}\eta\right)(\alpha)
=
\xi(\eta(\alpha))
\,\,\,(\xi, \eta\in \Der(F), \alpha\in V^{*} )
$$
because $I: V\times V \rightarrow \R$ is real
number-valued.
Consider $T:= \R[P_{1} ,\dots, P_{\ell-1}]$-linear
covariant derivative
$
\nabla_{D} : \Der(F) \rightarrow \Der(F).
$
By
\cite{AT10} it induces a $T$-linear bijection
\[
\nabla_{D}
:
D(\A(\Phi), 2k+1)^{W}
\tilde{\longrightarrow}
D(\A(\Phi), 2k-1)^{W}~~(k>0).
\]
The covariant derivative $\nabla_{D} $ was introduced by K. Saito (e.g. \cite{Sa93}) to study the flat structure (or the Frobenius manifold structure) of the
orbit space $V/W$.
Let $\theta_{E}:=\sum_{i=1}^{\ell} x_{i} (\partial/\partial x_{i})
$ denote the Euler derivation. Since $\theta_{E}\in D(\A(\Phi), 1)^{W}$,
one has
\[
\nabla_{D}^{-k} \theta_{E} \in D(\A(\Phi), 2k+1)^{W}
\]
which plays a principal role in this section.
For any $v\in V$, there exists a unique derivation
$\partial_{v} \in \Der(S)_{0} $ of degree zero
such that $$
\partial_{v} (\alpha) := \left<\alpha, v
\right>\,\,\,\,(\alpha\in V^{*} ).
$$
Thus we may identify
$V$ with $\Der(S)_{0} $ by the $W$-isomorphism
\[
V \longrightarrow \Der(S)_{0}
\]
defined by
$
v\mapsto \partial_{v}.
$
Let $\Omega(S)$ be the $S$-module of regular one-forms:
\[
\Omega(S)=S(dx_{1} )\oplus\dots\oplus S(dx_{\ell}).
\]
We may identify $V^{*} $ with $\Omega(S)_{0} $
as $W$-modules by the bijection $\alpha \mapsto d\alpha$.
Recall the $W$-invariant dual inner product
$
I^{*} : V^{*} \times V^{*} \rightarrow \R.
$
Define a $W$-isomorphism
\[
I^{*} : \Omega(S)_{0} \rightarrow \Der(S)_{0}
\]
by
$
\left(I^{*} (d\alpha)\right)
(\beta):= I^{*} (d\alpha, d\beta)
\,\,\,\,(\alpha, \beta\in V^{*}).
$
When a $W$-isomorphism $\Xi: \Der(S)_{0}
\rightarrow D(\A(\Phi), 2k)_{kh} $ is given,
we have the following commutative diagram in which the new maps
$\Xi^{*} $
and
$\Theta^{*} $ are defined:
$$
\xymatrix
{
&
{
D_{0}({\Shi}^{k})_{kh}}
\ar[ddd]^{\scalebox{1.1}{${\bf{res}}$}}_{\rotatebox{270}{\scalebox{3.0}[1.0]{$\widetilde{}$}}}
&
\\
&
{\Theta^{*}}~~~~~~~~~~~~
~~~~~~~~{\Theta}
&
\\
&
&
\\
&
{ D(\mathcal{A}(\Phi), 2k)_{kh}}
&
\\
&
\rotatebox{180}{\scalebox{1.3}{$\circlearrowright$}}
&
\\
{V^{*}=\Omega(S)_{0}
\ar[rr]_{\scalebox{1.1}{$I^{*}$}}}
\ar[ruu]^{\scalebox{1.1}{$\Xi^{*}$}}_{
\rotatebox{30}{\scalebox{3.0}[1.0]
{$\widetilde{}$}}}
\ar[ruuuuu]^{\scalebox{1.1} {}}_(0.53){{\rotatebox{180}{{\scalebox{1.3}{$\circlearrowright$}}}}}
\ar@{}[ru]|{\scalebox{1.1}{$W$-iso.}}
&
&
{
{\rm Der}(S)_{0}=V}
\ar[luu]^{\rotatebox{330}{\scalebox{3.0}[1.0]{$\widetilde{}$}}}_{
\scalebox{1.1}{$\Xi$}}
\ar[luuuuu]_{\scalebox{1.1} {}}^(0.53){\rotatebox{180}{\scalebox{1.3}{$\circlearrowright$}}}
\ar@{}[lu]|{\scalebox{1.1}{$W$-iso.}}
}
$$
\begin{prop}
\label{nablaD}
(1) For any primitive derivation $D$, define
\label{Xi}
\[
\Xi_{D} : \Der(S)_{0} {\longrightarrow} D(\A(\Phi), 2k)_{kh}
\]
by
\[
\Xi_{D} (\partial_{v})
:=
\nabla_{\partial_{v}} \nabla_{D}^{-k} \theta_{E}
\,\,\,\,(v\in V).
\]
Then $\Xi_{D} $ is a $W$-isomorphism.
\noindent
(2) Conversely, for any $W$-isomorphism
$
\Xi : \Der(S)_{0} \tilde{\longrightarrow} D(\A(\Phi), 2k)_{kh}
$,
there exists a unique primitive derivation
such that $\Xi = \Xi_{D} $.
\end{prop}
\noindent
{\bf proof}.
(1) was proved by Yoshinaga in \cite{Y02}.
(See \cite{T02} also.)
(2) follows from (1) and Schur's lemma.
\owari
\begin{remark}
\label{remark1.3}
(1)
The linear isomorphism $\Theta$
has the ambiguity of nonzero constant multiples
as in the case of the choice of
$\Xi $.
By Proposition \ref{nablaD}, the map $\Theta$ is uniquely determined
when a primitive derivation
$D$ is fixed.
\noindent
(2)
Note that
the arrangement
$
{\Shi}^{k}
$
is not $W$-stable.
Therefore
the $\ell$-dimensional vector space $D_{0}({\Shi}^{k})_{kh}$
is not naturally a $W$-module, while the $\ell$-dimensional vector spaces
$V$ and $D({\mathcal{A}}(\Phi),2k)_{kh}$ are both
$W$-modules.
\end{remark}
\bigskip
\noindent
{\bf{proof of Theorem \ref{characterization}.}}
Let $\Delta := \{\alpha_{1}, \dots, \alpha_{\ell} \}$
be the set of simple roots.
Let $\{\alpha_{1}^{*} , \dots, \alpha_{\ell}^{*} \}$
be the dual basis to $\Delta$: $\left<
\alpha_{i}, \alpha^{*}_{j} \right> = \delta_{ij} $
(Kronecker's delta).
Fix a primitive derivation $D$
and let $\Xi = \Xi_{D} $.
(1) Let $\varphi_{i}^{+} := \Theta (\partial_{\alpha_{i}^{*} } )
\,\,\,
(1\leq i\leq \ell)
$
be the SRB$+$.
We have
\begin{align*}
\left[{\bf{res}}(\varphi_{i}^{+})\right] (\alpha_{j})
&=
\left[{\bf{res}}(\Theta(\partial_{\alpha_{i}^{*}
}))\right]
(\alpha_{j})
=
\left[\Xi(\partial_{\alpha_{i}^{*} })\right]
(\alpha_{j})
=
(
\nabla_{\partial_{\alpha_{i}^{*} }}\nabla_{D}^{-k} \theta_{E}
)
(\alpha_{j} )\\
&=
\partial_{\alpha_{i}^{*} }\left(\left(\nabla_{D}^{-k} \theta_{E}
\right)
(\alpha_{j} )\right)
\in
\alpha_{j}^{2k+1} S
\end{align*}
because $\partial_{{\alpha}_{i}^{*}}(\alpha_{j})=0$ if $j\neq i$.
Define $\Gamma^{+}_{i}:=\Delta\setminus\{\alpha_{i} \}$.
Then we have the following commutative diagram
\[
\xymatrix{
D_{0} ({\mathcal S}^{k})_{kh}
\ar[r]^(0.43){\bf {res}}_(0.43){\scalebox{1.5}{$\widetilde{}$}}
&
D (\mathcal{A}(\Phi),2k)_{kh}
\\
D_{0} ({\mathcal B}_{\Gamma^{+}_{i} }^{+})_{kh}
\ar[r]^(0.43){\bf {res}}_(0.43){\scalebox{1.5}{$\widetilde{}$}}
\ar@{}[u]|{\scalebox{1.3}{$\bigcup$}}
&
D(\mathcal{A}(\Phi),{\bf z}^{+}_{\Gamma^{+}_{i}})_{kh}
\ar@{}[u]|{\scalebox{1.3}{$\bigcup$}}
}
\]
because of Proposition \ref{BGamma+}
and Theorem \ref{Ziegler}.
Since
$\varphi_{i}^{+}\in D_{0} (S^{k})_{kh} $
and
${\bf{res}}\left(\varphi_{i}^{+}\right)
\in
D(\mathcal{A}(\Phi),{\bf z}^{+}_{\Gamma^{+}_{i}})_{kh}
$,
we conclude that
$\varphi_{i}^{+}\in
D_{0} ({\mathcal B}_{\Gamma^{+}_{i} }^{+})_{kh}
$
by chasing the diagram above.
The ``uniqueness part''
follows from the equality
$$\dim D_{0} ({\mathcal B}_{\Gamma^{+}_{i} }^{+})_{kh} =1$$
which is a consequence of
Corollary \ref{BGammadim+} because
$|\Gamma^{+}_{i}|=\ell-1$.
(2) Recall
the SRB$-$:
\begin{align*}
\varphi_{i}^{-}
&:= \sum_{p=1}^{\ell}
I^{*}(\alpha_{i}, \alpha_{p}) \varphi_{p}^{+}
=
\Theta (
\sum_{p=1}^{\ell}
I^{*}(d\alpha_{i}, d\alpha_{p}) \partial_{\alpha_{p}^{*} }
)
=
\Theta (
I^{*}(d\alpha_{i})
)
\,\,\,\,
(1\leq i\leq \ell).
\end{align*}
We have
\begin{align*}
{\bf{res}}(\varphi_{i}^{-})
&=
{\bf{res}}\left(\Theta (I^{*}(d\alpha_{i}))\right)
=
\Xi(I^{*}(d\alpha_{i}))
=
\nabla_{I^{*}(d\alpha_{i})}\nabla_{D}^{-k} \theta_{E}.
\end{align*}
Let $s_{i} $ denote the orthogonal reflection
with respect to $\alpha_{i} $.
Since
$s_{i} (d\alpha_{i} )= -d\alpha_{i} $,
we have
\[
s_{i} (
{\bf{res}}(\varphi_{i}^{-})
)
=
-
{\bf{res}}(\varphi_{i}^{-}).
\]
Express
\[
{\bf{res}}(\varphi_{i}^{-}) = \sum_{p=1}^{\ell} f_{p}
\partial_{\alpha_{p}^{*} }
\,\,\,\,\,(f_{p} \in S).
\]
Recall $s_{i} (\partial_{\alpha^{*}_{p} }) = \partial_{\alpha^{*}_{p} }$
whenever $p\neq i$.
Thus we have
$s_{i} (f_{p})= -f_{p} $
whenever $p\neq i$.
Therefore
$f_{p}$ is divisible by $\alpha_{i}$
whenever $p\neq i$.
We also know that
\[
f_{i}
=
\left[{\bf{res}}(\varphi_{i}^{-})\right]
(\alpha_{i} )
=
(I^{*}(d\alpha_{i}))
\left(
(\nabla_{D}^{-k} \theta_{E})
(\alpha_{i} )\right)
\]
is divisible by $\alpha_{i}^{2k}$
because $\nabla_{D}^{-k} \theta_{E} (\alpha_{i} ) $ is divisible by
$\alpha_{i}^{2k+1}$.
Therefore
we conclude that $f_{i} $ is divisible by
$\alpha_{i} $ for any $i$.
Define $\Gamma^{-}_{i}:=\{\alpha_{i} \}$.
Then we have the following commutative diagram
\[
\xymatrix{
D_{0} ({\mathcal S}^{k})_{kh}
\ar[r]^{\bf {res}}_{\scalebox{1.5}{$\widetilde{}$}}
&
D (\mathcal{A}(\Phi),2k)_{kh}
\\
(\alpha_{i} -kz)\cdot
D_{0} ({\mathcal B}_{\Gamma^{-}_{i} }^{-})_{kh-1}
\ar[r]^(0.46){\bf {~~~~~~res}}_(0.46){\scalebox{1.5}
{~~~~$\widetilde{}$}}
\ar@{}[u]|{\scalebox{1.3}{$\bigcup$}}
&
\alpha_{i} \cdot D(\mathcal{A}(\Phi),{\bf z}^{-}_{\Gamma_{i}^{-} })_{kh-1}
\ar@{}[u]|{\scalebox{1.3}{$\bigcup$}}
}
\]
because of Proposition \ref{BGamma-}
and Theorem \ref{Ziegler}.
Since
$\varphi_{i}^{-}\in D_{0} (S^{k})_{kh} $
and
$$
{\bf{res}}\left(\varphi_{i}^{-}\right)
\in
\alpha_{i} \cdot D(\mathcal{A}(\Phi),{\bf z}^{-}_{\Gamma^{-}_{i}})_{kh-1}
,$$
we may conclude
$\varphi_{i}^{-}\in
(\alpha_{i} - kz)\cdot
D_{0} ({\mathcal B}_{\Gamma^{-}_{i}}^{-})_{kh-1}
$
by chasing the diagram above.
The ``uniqueness part'' follows from
the equality
$$\dim D_{0} ({\mathcal B}_{\Gamma^{-}_{i} }^{-})_{kh-1} =1$$
which is a consequence from
Corollary \ref{BGammadim-} because
$|\Gamma^{-}_{i}|=1$.
\owari
\medskip
The following proposition asserts that the simple roots
can be characterized by the freeness of an added/deleted
Shi arrangement:
\begin{theorem}
Let $\alpha\in \Phi^{+}$. Then
\noindent
(1)
the arrangement
$ {\mathcal S}^{k} \cup \{\mathbf{c}H_{\alpha, -k} \}$ is a
free
arrangement
if and only if $\alpha$ is a simple root,
and
\noindent
(2)
the
arrangement
$ {\mathcal S}^{k} \setminus \{\mathbf{c}H_{\alpha, k} \}$ is a
free arrangement
if and only if $\alpha$ is a simple root.
\label{simplefree}
\end{theorem}
\noindent
\textbf{proof}.
(1)
By Proposition \ref{BGamma+},
the ``if part'' is already proved.
Assume that $\alpha\in \Phi^{+}$ is a non-simple root.
We will prove that
$\calS^{k} \cup \{\mathbf{c}H_{\alpha, -k} \}$
is not free.
We may express
$\alpha=\beta_1+\beta_2$
with
$\beta_{i} \in \Phi^{+} \,\,(i=1, 2)$.
Let $H_i$ be the
hyperplane defined
by
$\beta_i = 0 \,\,(i=1, 2)$.
Then $X:=H_1 \cap H_2$
is of codimension two in $V$ because of
basic properties of the root systems.
As in the proof of Proposition \ref{BGamma+},
consider the two-dimensional
root system
$\Phi_X = \Phi \cap X^{\perp} $.
Note that
$\Phi_{X} $ is not of the type
$A_{1} \times A_{1} $ because
$\{\alpha, \beta_{1} , \beta_{2} \}\subseteq \Phi_{X} $.
Recall that the localization
$\left(\calS^{k} \cup \{\mathbf{c}H_{\alpha, -k} \}\right)_{Y}$
is free if
$\calS^{k} \cup \{\mathbf{c}H_{\alpha, -k} \}$
is free.
In the root system $\Phi_{X} $, $\alpha$ cannot be a simple
root since $\alpha$ is a sum of two positive roots
$\beta_{1} $ and $\beta_{2} $.
Thus we may assume that
$\Phi$ is a two-dimensional root system
without loss of generality.
In this case
the arrangement
$\calS^{k} \cup \{\mathbf{c}H_{\alpha, -k} \}$
is not free because of the addition theorem and
the equality
$$
|{Shi}^{k}
\cap H_{\alpha, -k}|=kh+1~~(\alpha\not\in\Delta),
$$
which can be verified by directly
counting intersection points.
(2)
Exactly the
same as the proof of
(1)
if one replaces
Proposition \ref{BGamma+},
$\cup$,
$kh+1$,
$H_{\bullet, -k}$,
addition theorem
with
Proposition
\ref{BGamma-},
$\setminus$,
$kh-1$,
$H_{\bullet, k}$,
deletion theorem
respectively.
\owari
\bigskip
Since the SRB bases uniquely exist,
it is
natural to pose the following problem:
\begin{problem}
Give an explicit construction of
the SRB$+$ and SRB$-$
for
$D_{0} ({\mathcal S}^{k})$
for every irreducible root system and every
$k\in \Z_{>0} $.
\end{problem}
This problem has been solved in
\cite{SuT, Su, GPT} for the SRB$-$
when $k=1$ and
the root system is either of the type $A_{\ell},
B_{\ell}, C_{\ell}, D_{\ell} $ or
$G_{2} $. The SRB$+$ can be obtained from the SRB$-$
by Definition \ref{SRB} (2).
\medskip
In the rest of this section we will describe the action of a
simple reflection on elements of the SRB$+/-$.
\begin{theorem}
\label{simplereflectionaction}
Let $\Delta = \{\alpha_{1}, \dots, \alpha_{\ell} \}$
be the set of simple roots.
Let $s_{i} $ denote the simple reflection
with respect to a simple root $\alpha_{i} $
with $s_{i}(\alpha_{i} )=-\alpha_{i}
\,\,(1\leq i\leq \ell)$.
(1) If $1\leq j\leq \ell$ with $i\neq j$, then
$s_{i} (\varphi_{j}^{+} ) = \varphi_{j}^{+} $.
(2) If we express
$
\varphi_{i}^{-}
=
(\alpha_{i} - kz)
\hat{\varphi}_{i}^{-},
$
then
$s_{i}(\hat{\varphi}_{i}^{-})
=
\hat{\varphi}_{i}^{-}.
$
\end{theorem}
\noindent
{\bf proof.}
Recall that
(e.g., see \cite[Ch. VI, \S 1, no. 6, Corollary 1]{bou})
\[
s_{i}
(
\Phi^{+} \setminus \{\alpha_{i} \}
)
=
\Phi^{+} \setminus \{\alpha_{i} \}.
\]
(1) When $i\neq j$,
$\varphi^{+}_{j}(\alpha_{i} + kz) $
is divisible by
$\alpha_{i}+kz $ because of Theorem \ref{characterization} (1).
Thus
$\left(
s_{i} (\varphi^{+}_{j})
\right)(\alpha_{i} - kz) $
is divisible by
$\alpha_{i}-kz $.
Using
$
s_{i}
(
\Phi^{+} \setminus \{\alpha_{i} \}
)
=
\Phi^{+} \setminus \{\alpha_{i} \},
$
we may conclude that
$
s_{i} (\varphi^{+}_{j})
\in
D_{0} (\calS^{k} ).
$
Consider the Ziegler restriction map
\[
\xymatrix{
D_{0} ({\mathcal S}^{k})_{kh}
\ar[r]^(0.43){\bf {res}}_(0.43){\scalebox{1.5}{$\widetilde{}$}}
&
D (\mathcal{A}(\Phi),2k)_{kh}
}
\]
Then we obtain
\begin{align*}
{\mathbf{res} } (s_{i}(\varphi_{j}^{+} ) )
=
s_{i}({\mathbf{res}} (\varphi_{j}^{+}) )
=
s_{i}(\Xi(\partial_{\alpha_{j}^{*} }) )
=
\Xi(s_{i}(\partial_{\alpha_{j}^{*} }) )
=
\Xi(\partial_{\alpha_{j}^{*} })
=
{\mathbf{res} } (\varphi_{j}^{+})
\end{align*}
because
$
s_{i}(\partial_{\alpha_{j}^{*} })
=
\partial_{\alpha_{j}^{*} }
$
whenever $i\neq j$.
Hence we have $s_{i}(\varphi_{j}^{+} )=
\varphi_{j}^{+}. $
(2) Define $\Gamma_{i}^{-} := \{\alpha_{i} \}$
as in the proof of Theorem \ref{characterization} (2).
Then $\hat{\varphi}_{i}^{-}\in D_{0}
(\B^{-}_{\Gamma_{i}^{-}} )_{kh-1}.$
Since
$$
\exp_{0} (\B^{-}_{\Gamma_{i}^{-}})
=
(
kh-1, \left(kh\right)^{\ell-1}
),$$
the Ziegler restriction map
\[
\xymatrix{
D_{0} (\B^{-}_{\Gamma^{-}_{i} })_{kh-1}
\ar[r]^(0.43){
\!\!\!\!\!\!\!\bf {res}}_(0.43){\scalebox{1.5}{
\!\!\!\!\!\!\!$\widetilde{}$}}
&
D (\mathcal{A}(\Phi),2k-\chi_{\Gamma_{i}^{-} } )_{kh-1}
}
\]
is bijective.
Note that the arrangement
$
\B^{-}_{\Gamma^{-}_{i} }
$
is $s_{i} $-stable because
\[
\B^{-}_{\Gamma^{-}_{i} }
=
\bigcup_{\beta\in\Phi^{+}\setminus \{\alpha_{i}\}}
\{
H_{\beta, p } ~|~
1-k\leq p\leq k
\}
\cup
\{
H_{\alpha_{i}, p } ~|~
1-k\leq p\leq k-1
\}.
\]
Then
$$s_{i} (
\hat{\varphi}^{-}_{i}
)
\in
D_{0}
(
\B^{-}_{\Gamma^{-}_{i} }
).
$$
We also verify
that
\[
{\mathbf{res}} (
\hat{\varphi}^{-}_{i}
)
=
\frac{1}{\alpha_{i} } \nabla_{I^{*}(d\alpha_{i} ) }
\nabla_{D}^{-k} \theta_{E}
\]
is $s_{i} $-invariant
because
$s_{i} (\alpha_{i} ) = -\alpha_{i}$
and
$s_{i} (d\alpha_{i} ) = -d\alpha_{i}$.
Therefore
\[
{\mathbf{res} }\left(
s_{i} (
\hat{\varphi}^{-}_{i}
)\right)
=
s_{i}\left(
{\mathbf{res} }
(
\hat{\varphi}^{-}_{i}
)\right)
=
{\mathbf{res} }
(
\hat{\varphi}^{-}_{i}
)
\]
and thus
$
s_{i} (
\hat{\varphi}^{-}_{i}
)
=
\hat{\varphi}^{-}_{i}.
$
\owari
\section{The $k$-Euler derivation }
Let $k\in \Z_{\geq 0} $.
An \textbf{extended Catalan arrangement} $\mathit{Cat}^{k}$
of the type $\Phi$ is an affine arrangement defined by
$$
\mathit{Cat}^{k}
:=
\mathit{Cat}^{k}(\Phi^{+})
:=
\{H_{\alpha, j} ~\mid~ \alpha\in \Phi^{+},\,\,\,j\in\Z, \,-k \leq j\leq k\}.
$$
Its cone
\[
{\cal C}^{k} := {\cal C}^{k}(\Phi^{+} ) := {\bf c}\mathit{Cat}^{k}
\]
was proved to be a free arrangement with
\[
\exp_{0} ({\cal C}^{k} ) =(kh+d_{1}, kh+d_{2}, \dots, kh+d_{\ell} )
\]
by Yoshinaga \cite{Y}.
Here the integers $d_{1} , d_{2} , \dots, d_{\ell} $ are the exponents
of the root system $\Phi$ with
$$d_{1} \leq d_{2} \leq \dots \leq d_{\ell}. $$
Since $1=d_{1} <d_{2} $, one has
$kh+1 < kh+d_{2} $.
Thus
we know that $D_{0} ({\cal C}^{k})_{kh+1} $ is a one-dimensional vector space.
\begin{definition}
We say that $\eta^{k}$ is the {\bf $k$-Euler derivation} if
$\eta^{k}$ is a basis for the vector space $D_{0} ({\cal C}^{k})_{kh+1} $.\end{definition}
Note that the ordinary Euler derivation
$\sum_{i=1}^{\ell} x_{i} (\partial/\partial x_{i} ) $ is a $0$-Euler derivation because
$\mathit{Cat}^{0} = \A(\Phi)$. The choice of a
$k$-Euler derivation has the ambiguity of nonzero constant multiples.
\begin{prop}
\label{etakinvariant}
A $k$-Euler derivation is $W$-invariant,
where the group $W$ acts trivially on the variable $z$.
\end{prop}
\noindent
{\bf proof.}
Note that
${\cal C}^{k} $
is stable under the action of $W$
unlike ${\cal S}^{k}$.
Since
the $W$-invariant derivation
$
\nabla_{D}^{-k} \theta_{E}
$
is a basis
for
$D (\mathcal{A}(\Phi),2k+1)_{kh+1},
$
we obatin
\[
D (\mathcal{A}(\Phi),2k+1)_{kh+1}
=
D (\mathcal{A}(\Phi),2k+1)_{kh+1}^{W}.
\]
Since the Ziegler restriction map
\[
\xymatrix{
D_{0} ({\mathcal C}^{k})_{kh+1}
\ar[r]^(0.43)
{\bf {\!\!\!\!\!\!\!res}}_(0.43){\scalebox{1.5}{
\!\!\!\!\!$\widetilde{}$}}
&
D (\mathcal{A}(\Phi),2k+1)_{kh+1}
}
\]
is a $W$-isomorphism,
we obtain
\[
D_{0} ({\mathcal C}^{k})_{kh+1}
=
D_{0} ({\mathcal C}^{k})_{kh+1}^{W}.
\]
\owari
\medskip
We may describe a $k$-Euler derivation in terms of the SRB$+$
$
\varphi^{+}_{1},
\dots,
\varphi^{+}_{\ell}
$
as follows:
\begin{theorem}
\label{etak}
The derivation
\[
\eta^{k} :=
\sum_{i=1}^{\ell}
(\alpha_{i} + kz) \varphi_{i}^{+}
\]
is a $k$-Euler derivation.
\end{theorem}
\noindent
{\bf{proof.}}
Note that ${\mathcal B}^{+}_{\Delta}$
is a subarrangement
of ${\mathcal C}^{k} $.
Consider a commutative diagram
\[
\xymatrix{
D_{0} ({\mathcal B}^{+}_{\Delta})_{kh+1}
\ar[r]^{\bf {\!\!\!\!\!\!\!\!\!res}}_{\scalebox{1.5}{$\!\!\!\!\! \widetilde{}$}}
&
D (\mathcal{A}(\Phi),{\bf z}_{\Delta}^{+})_{kh+1}
\\
D_{0} ({\mathcal C}^{k})_{kh+1}
\ar[r]^(0.43){\bf {res}}_(0.43){\scalebox{1.5}{$\widetilde{}$}}
\ar@{}[u]|{\scalebox{1.3}{$\bigcup$}}
&
D(\mathcal{A}(\Phi), 2k+1)_{kh+1}
\ar@{}[u]|{\scalebox{1.3}{$\bigcup$}}.
}
\]
Since $(\alpha_{i}+kz)\varphi^{+}_{i} \in
D_{0}({\mathcal B}^{+}_{\Delta})_{kh+1}
$
for any $i$ by Theorem \ref{characterization} (1),
we have $\eta^{k} \in
D_{0}({\mathcal B}^{+}_{\Delta})_{kh+1}
$.
We also have
\begin{align*}
{\mathbf{res}}\left(
\eta^{k}
\right)
&=
{\mathbf{res}}\left(
\sum_{i=1}^{\ell}
(\alpha_{i}+kz )\varphi_{i}^{+}
\right)\\
&=
\sum_{i=1}^{\ell}
\alpha_{i}
\Xi(\partial_{\alpha^{*} _{i}})
=
\sum_{i=1}^{\ell}
\alpha_{i}
\nabla_{\partial_{\alpha^{*}_{i}}}
\nabla_{D}^{-k} \theta_{E}
=
\nabla_{\theta_{E}}
\nabla_{D}^{-k} \theta_{E}
\\
&=
(kh+1)
\nabla_{D}^{-k} \theta_{E}
\in
D(\mathcal{A}(\Phi), 2k+1)_{kh+1}.
\end{align*}
Thus we may conclude that
$\eta^{k} \in D_{0} ({\mathcal C}^{k})_{kh+1}
$
by chasing the diagram above.
\owari
\begin{remark}
The logarithmic differential versions of Theorems \ref{characterization}
and
\ref{etak}
can be proved similarly in the dual setup.
For the definition of the logarithmic differential modules
$\Omega(\A(\Phi), 2k)~~(k \geq 1)$, see \cite{Z}.
For the Ziegler restriction map in the dual setup, see
\cite[Theorem 2.5]{Y05}.
\end{remark}
The following corollary of Theorem \ref{etak}
gives a formula for the action of a simple reflection on a
member of the SRB+.
This formula was not covered by Theorem \ref{simplereflectionaction}.
\begin{corollary}
\label{simplereflectionaction2}
Let $\Delta = \{\alpha_{1}, \dots, \alpha_{\ell} \}$
be the set of simple roots.
Let $s_{i} $ denote the simple reflection
with respect to a simple root $\alpha_{i} $
with $s_{i}(\alpha_{i} )=-\alpha_{i}
\,\,(1\leq i\leq \ell)$.
Then
\[
s_{i} (\varphi_{i}^{+})
=
\left(
\frac{\alpha_{i} + kz}{-\alpha_{i} + kz}
\right)
{\varphi}_{i}^{+}
+
\left(
\frac{\alpha_{i}}{-\alpha_{i} + kz}
\right)
\sum_{j\neq i}
\frac{2(\alpha_{i}, \alpha_{j}) }{(\alpha_{i}, \alpha_{i})}
{\varphi}_{j}^{+}.
\]
\end{corollary}
\noindent
{\bf proof.}
Since the $k$-Euler derivation
$
\eta^{k} =
\sum_{j=1}^{\ell}
(\alpha_{j} + kz) \varphi_{j}^{+}
$
is $W$-invariant by Proposition \ref{etakinvariant},
we obtain
\begin{align*}
&~~~\sum_{j=1}^{\ell}
(\alpha_{j} + kz) \varphi_{j}^{+}
=\eta^{k}
=s_{i} (\eta^{k})
=
\sum_{j=1}^{\ell}
(s_{i} (\alpha_{j}) + kz) s_{i} (\varphi_{j}^{+})\\
&=
(-\alpha_{i} + kz) s_{i} (\varphi_{i}^{+})
+
\sum_{j\neq i}
(\alpha_{j} - \frac{2(\alpha_{i}, \alpha_{j}) }{(\alpha_{i}, \alpha_{i})}
{\alpha}_{i}
+kz) \varphi_{j}^{+}.
\end{align*}
Here we used the formula in
Theorem \ref{simplereflectionaction} (1).
Solving this equation for $s_{i} (\varphi_{i}^{+})$, we may verify
the desired result.
\owari
|
\section{Introduction and qualitative discussion}\label{introduction}
The goal of this Lecture is to present a physical picture of the
process of continuous quantum measurement of a qubit in the circuit
quantum electrodynamics (cQED) setup
\cite{Wallraff-04,DiCarlo-10,Fragner-2008,Bertet,Siddiqi-2011}
(Fig.\ 1), extending or reformulating the previous theoretical
descriptions \cite{Blais-04,Clerk-RMP,Gambetta-06,Gambetta-08}.
Understanding of the qubit evolution in the process of measurement
is important for developing an intuition, which is useful in many
cases, in particular in designing various schemes of the quantum
feedback \cite{Wiseman-feedback,Hofmann-1998,Kor-feedback}.
When a quantum measurement is discussed \cite{Braginsky}, there
are usually two different types of questions to answer: we can
either focus on obtaining information on the initial state (before
measurement) or focus on the quantum state after the measurement
(i.e.\ evolution in the process of measurement). Let us emphasize
that we consider the latter problem here and essentially extend the
collapse postulate by describing continuous evolution ``inside'' the
collapse timescale.
In the cQED setup (Fig.\ 1) a qubit interacts with a GHz-range
microwave resonator, whose frequency slightly changes depending on
whether the qubit is in the state $|0\rangle$ or $|1\rangle$
\cite{Wallraff-04,DiCarlo-10,Fragner-2008,Bertet,Siddiqi-2011,Blais-04,Clerk-RMP,Gambetta-06,Gambetta-08}.
In turn, this frequency shift affects the phase (and in general
amplitude) of a probing microwave, which is transmitted through the
resonator (in another setup the microwave is reflected from the
resonator, but the difference is not important). The outgoing
microwave is amplified, and after that the rf signal is
downconverted by mixing it with the original microwave tone, so that
the low-frequency ($< 100$ MHz) output of the IQ mixer provides
information on the qubit state. The output noise is mainly
determined by the first amplifying stage, the pre-amplifier. With
recent development of nearly quantum-limited superconducting
parametric amplifiers \cite{Devoret-ampl,Lehnert-ampl,Siddiqi-2011},
it is natural to use them as pre-amplifiers \cite{Siddiqi-2011}
instead of cryogenic high-electron-mobility transistors (HEMTs)
\cite{Wallraff-04,DiCarlo-10,Fragner-2008,Bertet}, which usually
have noise temperature above 3 K.
\begin{figure}[tb]
\centering
\includegraphics[width=8.5cm]{LesHou-fg1.eps}
\caption{Schematic of the cQED setup. Microwave field of frequency
$\omega_m$ is transmitted through (or reflected from) the resonator
of frequency $\omega_r$, which slightly changes, $\omega_r\pm\chi$,
depending on the qubit state. After amplification the microwave is
sent to the IQ mixer, which produces two quadrature signals: $I(t)$
and $Q(t)$. For a phase-preserving amplifier we define $I(t)$ as the
quadrature carrying information on the qubit state, while for a
phase-sensitive amplifier we define $I(t)$ as corresponding to the
amplified quadrature.}
\label{fig1}
\end{figure}
Continuous quantum measurement in the cQED setup in some sense
falls in between a qubit measurement by a quantum point contact
(QPC) or a single-electron transistor (SET), the theory of which has
been developed over a decade ago \cite{Kor-99-01,Goan}, and
continuous quantum measurement in optics, for which the theory of
quantum trajectories has been developed even earlier
\cite{quant-traj}. Nevertheless, the cQED setup differs from both
these cases, and this is probably the reason why there is still a
confusion about the proper physical description of the measurement
process. The measurement by the QPC or SET is of the broadband type,
meaning that the monitored frequency band starts from zero. In
contrast, the cQED setup is of the narrowband type: we deal only
with a relatively narrow band around the probing microwave frequency
$\omega_m$. This necessarily involves two orthogonal quadratures
\cite{quadrature}: we work with rf signals of the type $A(t)\cos
(\omega_m t)+B(t)\sin(\omega_m t)$, and there are essentially two
signals $A(t)$ and $B(t)$ instead of only one in the broadband case.
In this sense the cQED setup is similar to the optical (especially
cavity QED) setup \cite{quant-traj}; however, there is an important
difference: in the cQED case the outgoing microwave is amplified
(Fig.\ 1) before being mixed with the original microwave, while
there is no amplification stage in the standard optical setup. The
operation will obviously depend on whether a phase-sensitive or a
phase-preserving amplifier is used, since a phase-preserving
amplifier necessarily adds the half quantum of noise into any
quadrature \cite{Haus-1962,Caves-1982,Dev-Likh,Clerk-RMP}. Notice
that the quantum trajectory theory for the cQED setup was developed
in \cite{Gambetta-08}; however, the amplifier stage was essentially
missing in the analyzed model.
In this Lecture we consider the simplest cQED case, assuming
dispersive regime \cite{Blais-04}, exactly resonant microwave
frequency, absence of the Rabi drive, and sufficiently wide
resonator and amplifier bandwidths for the Markov approximation.
Some generalizations are rather straightforward; however, our goal
is a simple picture in a simple case.
A description of continuous qubit measurement is essentially a
description of the quantum back-action. Following the same quantum
Bayesian framework as for the measurement by QPC/SET
\cite{Kor-99-01} (see \cite{Kor-rev} for review), we will discuss
two kinds of measurement back-action onto the qubit, which we name
here as ``spooky'' and ``realistic''. {\it The ``spooky'' (or
``quantum'', ``informational'', ``non-unitary'') back-action does
not have a physical mechanism} and therefore cannot be described by
the Schr\"odinger equation (in contrast to what people often think,
trying to find a mechanism for the quantum collapse); however, {\it
it is a common-sense consequence of acquiring information} on the
qubit state in the process of measurement. This is essentially the
same back-action which is discussed in the EPR paradox \cite{EPR}
and Bell inequality violation \cite{Aspect}; the only difference is
that in our case the information is incomplete and therefore we have
to use the quantum Bayes rule \cite{Caves-Bayes,Gardiner,Kor-rev}
instead of the projective collapse rule. In contrast, the
``realistic'' (or ``classical'', ``unitary'') back-action has a
physical mechanism: in the cQED case it is a fluctuation of the
number of photons in the resonator, which affects the phase of the
qubit state. The ``realistic'' back-action is usually discussed in
the standard theories of the cQED measurement
\cite{Blais-04,Clerk-RMP,Gambetta-06}. Actually, there is a certain
spookiness even in the ``realistic'' back-action (it may be affected
by a delayed choice, as discussed in Conclusion); however, we do not
want to emphasize it to keep the picture simple.
When we measure the $z$-coordinate of the qubit state on the Bloch
sphere (the basis states $|1\rangle$ and $|0\rangle$ correspond to
the North and South poles), then the ``spooky'' back-action changes
the $z$-coordinate and leads to the state evolution along the
meridian lines, while the ``realistic'' back-action leads to the
evolution around the $z$-axis, i.e.\ along the parallels.
It is important to notice that when the probing microwave leaves
the resonator after interaction with the qubit, one quadrature of
the microwave carries information about the qubit state, while the
orthogonal quadrature carries information on the fluctuating number
of photons in the resonator
\cite{Blais-04,Clerk-RMP,Gambetta-06,Gambetta-08}. Therefore, if a
phase-preserving amplifier is used, then the ``spooky'' and
``realistic'' back-actions are fully separated and correspond to two
orthogonal quadratures $I(t)$ and $Q(t)$ measured after the mixer
(it is trivial to choose the proper linear combinations of the I/Q
mixer outputs). The signals $I(t)$ and $Q(t)$ are necessarily noisy,
and the measurement back-actions are stochastic; however, there is a
correlation (full correlation in the ideal case) between the output
noise and the back-action noise in both channels. As a result
(derived later), for a quantum-limited phase-preserving amplifier
and absence of extra decoherence, the measured quadratures $I(t)$
and $Q(t)$ give us {\it full information about the back-action}, so
that a {\it random evolution of the qubit wavefunction can be
monitored precisely} (a useful analogy is with a Brownian particle
under a microscope: we cannot predict its motion, but we can monitor
it). This is what is needed, in particular, for arranging a perfect
quantum feedback control of the qubit state. It is interesting to
notice that for an ensemble-averaged evolution (in which the random
but monitorable qubit evolution is replaced by dephasing), exactly
one half of the ensemble dephasing $\Gamma$ comes from the
``spooky'' back-action, and the other half comes from the
``realistic'' back-action.
In the case of a phase-sensitive amplifier it is sufficient to
measure after the mixer only the quadrature which was amplified; let
us still denote it as $I(t)$, though now its phase is determined by
the amplifier instead of the microwave-qubit interaction. In this
case the ``spooky'' and ``realistic'' back-actions are in general
mixed (not separated), because there is only one output signal
$I(t)$. This situation exactly corresponds to the broadband
measurement by the QPC/SET with a correlation between the output and
``realistic'' back-action noises \cite{Kor-rev}. The situation
simplifies when the amplified quadrature is the one which carries
information about the qubit state ($z$-coordinate). Then in the
quantum-limited case the ``realistic'' back-action is fully absent:
we cannot measure the photon number fluctuation and correspondingly
it does not fluctuate (in the imperfect case the effect of the
remaining ``realistic'' back-action can be described by an extra
dephasing). So we are left with only the ``spooky'' back-action, and
the quantum measurement description coincides with the simpler
theory of measurement by a symmetric QPC \cite{Kor-rev}, which does
not produce the ``realistic'' back-action. In contrast, in the case
when the photon-number quadrature is amplified, we do not obtain any
information on the qubit $z$-coordinate, and therefore there is no
``spooky'' back-action, but only the ``realistic'' one. In a general
case, when the amplified quadrature makes an arbitrary angle
$\varphi$ with the qubit-information quadrature, both types of the
back-action are present, and their strength depends on $\varphi$. It
is important to mention that the ensemble dephasing rate $\Gamma$
does not depend on $\varphi$, as required by causality. In
particular, in the quantum-limited case the contribution $\Gamma
\cos^2\varphi$ comes from the ``spooky'' back-action, while
$\Gamma\sin^2\varphi$ comes from the ``realistic'' back-action.
Let us emphasize that both the phase-sensitive and
phase-preserving amplifiers permit exact monitoring of the qubit
state and therefore a perfect quantum feedback. The necessary
condition in both cases is that the detection system is
quantum-limited.
In the following sections a formal description of
the above discussed results is presented. We start with reviewing
the Bayesian approach for the broadband qubit measurement, then
briefly discuss the difference between phase-preserving and
phase-sensitive amplifiers, and then present the formalism of the
narrowband continuous measurement of a qubit in the cQED setup. In
Conclusion we briefly discuss generalizations of the formalism,
quantum feedback, and the causality principle. We note that our
approach can be converted into the formal language of the
positive-operator-valued-measure(POVM)-type generalized quantum
measurement \cite{N-C} (then separation of the ``spooky'' and
``realistic'' back-actions corresponds to the decomposition of the
measurement operator into diagonal and unitary parts -- see later),
and our results for the case of a phase-sensitive amplifier are very
similar to the results of Ref.\ \cite{Gambetta-08}.
\section{Broadband measurement}\label{broadband}
In this section we review the Bayesian formalism
\cite{Kor-99-01,Kor-rev} for the broadband measurement of a qubit,
considering only the simple case without additional evolution, and
thus emphasizing the main physical idea of the formalism. We start
with the broadband formalism because it is simpler than for the
narrowband (cQED) measurement and it can be used as a natural step
in understanding the cQED setup.
For definiteness let us assume that the qubit is a double quantum
dot populated with one electron (Fig.\ 2), and the states
$|0\rangle$ and $|1\rangle$ correspond to the electron localized in
one or the other dot. The qubit is measured by a small-transparency
tunnel junction (model of QPC), whose barrier height depends on the
electron location, so that the two qubit states correspond to
different average currents $I_0$ and $I_1$ through the QPC. The
voltage across the QPC is sufficiently large to make the detector
output classical (Markov approximation), and $|\Delta I|\ll |I_c|$,
where $\Delta I=I_1-I_0$ is the response and $I_c=(I_0+I_1)/2$ is
the mean value; this weak response assumption allows us to consider
the QPC current $I(t)$ as a quasicontinuous noisy signal (see
\cite{Kor-rev} for the detailed discussion of required assumptions;
the formalism needs only a minor change if $\Delta I\sim I_c$).
Then the output signal of the detector is
\be
I(t)=I_c + (\Delta I/2) \, z(t) +\xi (t), \,\,\,
S_\xi (\omega ) =S,
\ee
where $z=\rho_{11}-\rho_{00}$ is the $z$-component of the Bloch
sphere representation of the qubit density matrix $\rho(t)$, and
$\xi (t)$ is the white shot noise with spectral density $S=2eI_c$
(we use the single-sided definition for the spectral density, in
which the signal variance (``power'') corresponds to $\int_0^\infty
S(\omega)\, d\omega/2\pi$; the definition of $S$ is twice smaller in
Ref.\ \cite{Clerk-RMP} and $4\pi$ times smaller in Refs.\
\cite{Gambetta-06,Gardiner}). We emphasize that the detector signal
$I(t)$ is classical, and the qubit state $\rho(t)$ is practically
unentangled from the detector, but obviously depends on $I(t)$.
\begin{figure}[tb]
\centering
\includegraphics[width=5.0cm]{LesHou-fg2.eps}
\caption{Schematic of a broadband measurement setup:
double-quantum-dot qubit is measured by a QPC (tunnel junction). The
output signal $I(t)$ is the QPC current. }
\label{fig2}
\end{figure}
The detector Hamiltonian and the qubit-detector interaction
Hamiltonian are given in Refs.\ \cite{Kor-99-01,Kor-rev}, they are
not really important for our discussion here. For simplicity let us
assume that the qubit Hamiltonian is zero, $H_{qb}=0$, so that the
qubit evolution is due to the measurement only. In this case the
qubit evolution during time $t$ happens to be determined only by the
time-averaged value of the measured detector output
\be
\overline{I}_m (t) =\frac{1}{t} \int_0^t I(t')\, dt' ,
\ee
which would contain full information for a classical measurement.
Because of the correspondence principle, the evolution of the
diagonal elements of the qubit density matrix $\rho$ (${\rm Tr}
\rho=1$) should correspond to the classical evolution of
probabilities, which are given by the classical Bayes rule. The
Bayes rule says that an updated (a posteriori) probability of a
system state is proportional to the initial (a priori) probability
and the probability (likelihood) of the obtained measurement result
assuming this particular state. In our case $\overline{I}_m(t)$ is
the measurement result, and its probability for the qubit in the
basis state $|j\rangle$ has the Gaussian distribution
\be
P_{|j\rangle}(\overline{I}_m)=\frac{1}{\sqrt{2\pi D}} \, \exp
[-(\overline{I}_m-I_j)^2/2D], \,\, D=\frac{S}{2t},
\label{gaussian}\ee
where $D$ is the variance, which
decreases with the measurement time $t$. Therefore the
correspondence principle demands the Bayesian evolution
\be
\frac{\rho_{11}(t)}{\rho_{00}(t)} =
\frac{\rho_{11}(0)}{\rho_{00}(0)} \, \frac{\exp
[-(\overline{I}_m(t)-I_1)^2/2D]}{\exp
[-(\overline{I}_m(t)-I_0)^2/2D]},
\label{rho-diag}\ee
which in our terminology is due to the ``spooky'' back-action; it
cannot be described by the Schr\"odinger equation, but follows from
common sense.
If the phase of qubit state is not affected by the
measurement process (no ``realistic'' back-action), then an
arbitrary initial wavefunction $|\psi(0)\rangle=
\sqrt{\rho_{00}(0)}\, |0\rangle +e^{i\phi}\sqrt{\rho_{11}(0)}\,
|1\rangle$ becomes $|\psi(t)\rangle = \sqrt{\rho_{00}(t)}\,
|0\rangle +e^{i\phi}\sqrt{\rho_{11}(t)}\, |1\rangle$ with the same
phase $\phi$; therefore for an arbitrary mixed state we get
\be
\rho_{01}(t)= \rho_{01}(0)\, \frac{\sqrt{\rho_{00}(t)\,
\rho_{11}(t)}} {\sqrt{\rho_{00}(0)\, \rho_{11}(0)}}\, .
\label{rho-off}\ee
Equations (\ref{rho-diag}) and (\ref{rho-off}) describe the
``spooky'' back-action.
Now assume that due to the qubit-detector interaction (e.g.\
Coulomb interaction), each electron passing through the detector
rotates the qubit phase $\phi$ by a small amount $\Delta\phi$. From
the measured result $\overline{I}_m(t)$ we know exactly how many
electrons passed through [$n_e=\overline{I}_m(t)t/q$ with $q$ being
the electron charge], and can easily introduce the corresponding
phase factor into Eq.\ (\ref{rho-off}):
\be
\rho_{01}(t)= \rho_{01}(0)\, \frac{\sqrt{\rho_{00}(t)\,
\rho_{11}(t)}} {\sqrt{\rho_{00}(0)\, \rho_{11}(0)}}\, \exp[iK
\overline{I}_m(t)\, t] ,
\label{rho-off-2}\ee
where $K=\Delta\phi/q$. The non-stochastic factor $\exp(iKI_ct)$ can
be obviously ascribed to the qubit Hamiltonian; however, this is not
important here. The factor $\exp[iK \overline{I}_m(t)\, t]$ in Eq.\
(\ref{rho-off-2}) is the effect of the ``realistic'' back-action. It
may or may not be present in a particular physical situation; for
example, $K=0$ for measurement by a symmetric QPC, while $K\neq 0$
in an asymmetric QPC or SET case.
Finally, if there is an extra pure dephasing of a qubit with
rate $\gamma$, then Eq.\ (\ref{rho-off-2}) becomes
\be
\rho_{01}(t)= \rho_{01}(0)\, \frac{\sqrt{\rho_{00}(t)\,
\rho_{11}(t)}} {\sqrt{\rho_{00}(0)\, \rho_{11}(0)}}\, \exp[iK
\overline{I}_m(t)\, t] \, e^{-\gamma t} .
\label{rho-off-3}\ee
Equations (\ref{rho-diag}) and (\ref{rho-off-3}) is {\it the main
starting point of the Bayesian formalism} \cite{Kor-rev}. It is then
easy to include non-zero qubit Hamiltonian $H_{qb}$ by
differentiating Eqs.\ (\ref{rho-diag}) and (\ref{rho-off-3}) over
time (paying attention to whether the Stratonovich or It\^o
definition of the derivative is used) and adding terms due to
$H_{qb}$. Energy relaxation and other mechanisms of the qubit state
evolution can be included in the same way.
Actually, there are many ways to derive the Bayesian equations
(\ref{rho-diag}) and (\ref{rho-off-3})
\cite{Kor-99-01,Kor-rev,Goan,Jordan,Nazarov}, but we focus here only
on their meaning, not on their derivation.
Notice that averaging of Eqs.\ (\ref{rho-diag}) and
(\ref{rho-off-3}) over the measurement result $\overline{I}_m$
(i.e.\ ensemble averaging) with the probability distribution
\be
P(\overline{I}_m)=\rho_{00}(0)P_{|0\rangle}(\overline{I}_m)+
\rho_{11}(0)P_{|1\rangle}(\overline{I}_m)
\label{P(I)}\ee
gives the same evolution as for a
pure dephasing: the diagonal matrix elements of $\rho$ do not
evolve, while the off-diagonal element $\rho_{01}$ decays as
$\rho_{01}(0)\, e^{-\Gamma t}$ with the ensemble dephasing rate
\cite{Kor-rev}
\be
\Gamma = \frac{(\Delta I)^2}{4S}+ \frac{K^2S}{4}+\gamma ,
\label{Gamma}\ee
which has clear contributions from the ``spooky'' back-action,
``realistic'' back-action, and additional dephasing.
In the case $\gamma=0$ an initially pure qubit state remains
pure; in other words we can monitor evolution of a qubit
wavefunction. This property can be used as the definition of a
quantum-limited detector \cite{Kor-99-01,Kor-rev}. The quantum
efficiency $\eta$ can then be naturally defined as $\eta
=1-\gamma/\Gamma$. If by some reason the ``realistic'' back-action
is considered as dephasing (i.e.\ only in the averaged way), then
the quantum efficiency can be defined as $\tilde\eta=1-\gamma/\Gamma
- K^2S/4\Gamma$ (here the definitions of $\tilde\eta$ and $\eta$ are
exchanged compared with the definitions in \cite{Kor-rev}). In other
words, $\tilde\eta =(\Delta I)^2/4S\Gamma$ is the relative
contribution of only the ``spooky'' back-action in the ensemble
dephasing $\Gamma$. In particular, this definition is relevant to
the peak-to-pedestal ratio of the Rabi spectral peak \cite{Kor-osc},
which is equal to $4\tilde\eta$.
As an example, if
$\gamma=0$ and contributions in Eq.\ (\ref{Gamma}) from the
``spooky'' and ``realistic'' back-actions are equal to each other
(as in the cQED setup with a phase-preserving amplifier), then
$\eta=1$ but $\tilde{\eta}=1/2$.
A non-ideal detector ($\eta<1$) can be modeled in two equivalent
ways \cite{Kor-nonideal}: we either add an extra dephasing $\gamma$
to the qubit or we add an extra noise to the output of the ideal
detector. Only the total dephasing $\Gamma$, response $\Delta I$,
total output noise $S$, and correlation factor $K=
\delta\langle\phi\rangle/\delta(\overline{I}_m t)$ are the physical
(i.e.\ experimentally measurable) parameters, while distribution of
the non-ideality between the extra dephasing and additional output
noise is a matter of convenience [here $\phi={\rm arg(\rho_{01})}$,
and notation $\langle \phi\rangle$ reminds about averaging over
additional classical noise at the output]. We emphasize that the
Bayesian formalism deals only with the experimentally measurable
parameters $\Delta I$, $S$, $K$, $\Gamma$, and the output signal
$I(t)$.
In the ideal case ($\eta=1$) the evolution equations
(\ref{rho-diag}) and (\ref{rho-off-2}) can be translated into the
language of POVM-type generalized measurement. In this approach the
effect of measurement is described as \cite{N-C}
\be
|\psi (t)\rangle =\frac{M_R|\psi(0)\rangle}{||M_R|\psi(0)\rangle ||},
\,\,\,
\rho (t)=\frac{M_R\rho(0)M_R^\dagger}{\mbox{Tr}[M_R^\dagger
M_R\rho(0)]},
\ee
where $M_R$ is the so-called measurement (Kraus) operator,
corresponding to the result $R$. The probability of the result $R$
is $P_R=||M_R|\psi(0)\rangle||^2$ using wavefunctions or $P_R=\mbox
{Tr} [M_R^\dagger M_R\,\rho(0)]$ using density matrices; therefore
the POVM elements $M_R^\dagger M_R$ should satisfy the completeness
condition $\sum_R M_R^\dagger M_R = \openone$.
The relation between this
approach and the quantum Bayesian approach can be understood via the
operator decomposition
\be
M_R = U_R \sqrt{M_R^\dagger M_R},
\ee
where $U_R$ is unitary and the square root of the positive operator
$M_R^\dagger M_R$ is defined in the natural way in the diagonalizing
basis. It is easy to see that $\sqrt{M_R^\dagger M_R}$ is
essentially the quantum Bayes rule (in the diagonalizing basis); in
our terminology it corresponds to the ``spooky'' back-action, while
$U_R$ corresponds to the ``realistic'' back-action.
For the discussed setup the result $R$ is $\overline{I}_m(t)$, the
``spooky'' back-action $[M_R^\dagger M_R]^{1/2}$ should be
determined by the probabilities $P_{|j\rangle}(\overline{I}_m)$
given by Eq.\ (\ref{gaussian}), and the ``realistic'' back-action
$U_R$ is given by the phase factor in Eq.\ (\ref{rho-off-2}).
Therefore the corresponding measurement operator is
\begin{eqnarray}
&& M(\overline{I}_m) = \exp ( -i K \overline{I}_m t \,
\sigma_z/2)
\nonumber \\
&&\hspace{0.5cm} \times \left[\sqrt{P_{|0\rangle}(\overline{I}_m)}
\, |0\rangle \langle 0| + \sqrt{P_{|1\rangle}(\overline{I}_m)} \,
|1\rangle \langle 1|\right], \qquad
\end{eqnarray}
where $\sigma_z$ is the Pauli matrix.
\section{Phase-preserving vs.\ phase-sensitive
amplifiers}\label{amplifiers}
Before discussing microwave amplifiers, let us consider a
measurement of an oscillator, for example, a mechanical resonator
with frequency $\omega_r$ and mass $m$. This is a very well studied
problem \cite{Clerk-RMP,Braginsky}, so we will only discuss a way to
understand the results. A classical resonator position $x$
oscillates as $x_c(t)=A\cos(\omega_r t)+B\sin(\omega_r t)$ (in this
section $x$ stands for the usual spatial coordinate, not for the
Bloch sphere coordinate). The corresponding quantum state is called
the ``coherent state'' in the optical language; it is represented by
the wavefunction $\psi (x,t)=\psi_{gr}[x-x_c(t)]\exp(i p_c
x/\hbar)$, where $\psi_{gr}(x)$ is the ground state and
$p_c=m\dot{x}_c(t)$ is the classical momentum. So the coherent state
is essentially the ground state with oscillating center position.
Notice that continuous quantum measurement of a resonator position
can be described in the same Bayesian way \cite{Ruskov-resonator} as
in the previous section; for example the ``spooky'' back-action
gives the evolution
$\psi(x,t)=\psi(x,0)\exp[-(\overline{I}_m(t)-I(x))^2/4D]/{\rm
Norm}$, where $I(x)$ is the average detector signal for the
resonator position $x$, and ${\rm Norm}$ is normalization [see Eqs.\
(\ref{rho-diag}) and (\ref{rho-off})]. The time step $t$ in this
case should be chosen much shorter than $\omega_r^{-1}$ so that the
unitary evolution and evolution due to measurement may be simply
added.
Let us consider the following game: Charlie prepares an oscillator
in a coherent state with quadratures $A$ and $B$, gives it to David,
and David's goal is to find $A$ as accurately as possible. An
optimal strategy is rather obvious: David should make a projective
measurement of $x$ at time $t=2\pi n/\omega_r$ with any integer $n$
(to avoid contribution from the $B$-term), and the measurement
result is the best estimate of $A$ [if the measurement is done at
$t=(2\pi n+\pi)/\omega_r$, then the result should be multiplied by
$-1$]. Even though the strategy is optimal, the inaccuracy of
David's result is obviously the width (standard deviation)
$\sigma_{gr}=\sqrt{\hbar/2m\omega_r}$ of the ground state shape
$|\psi_{gr}(x)|^2$; in energy units this inaccuracy corresponds to
one half of the energy quantum.
Now assume that David cannot make projective measurements, but
only ``finite-strength'' (i.e.\ imprecise) measurements. The best
accuracy $\sigma_{gr}$ can still be achieved if the measurement is
done in the simple but very clever ``quantum non-demolition'' (QND)
way: many finite-strength measurements are made at times $t=2\pi
n/\omega_r$; this is called ``stroboscopic'' measurement
\cite{Braginsky}. Since the oscillator returns to the same state
after the period $2\pi/\omega_r$, the unitary evolution is not
important, and many finite-strength measurements (described by the
Bayesian equation above) are ``stacked'' to produce a strong,
essentially projective measurement. More generally, the necessary
condition to have the best accuracy $\sigma_{gr}$ for $A$ is that
the measurement is not sensitive to the quadrature $B$.
Now assume that David is only allowed to make a continuous
measurement with unmodulated weak strength (so that the inaccuracy
achieved after $\omega_r^{-1}$ is much larger than $\sigma_{gr}$).
Then the ``spooky'' back-action gets mixed with the unitary
evolution, essentially adding noise into the monitored evolution, so
that after a while the resonator state becomes mostly determined by
the back-action and almost not dependent on the initial state. As
the result, the best accuracy for measurement of $A$ becomes
$\sqrt{2}\, \sigma_{gr}$, which in energy units corresponds to two
half-quanta \cite{Braginsky}, that is twice worse than for the
projective or stroboscopic measurement. However, the continuous
monitoring gives us an information about $B$ in the same way as for
$A$, so the accuracy of $B$-measurement is also $\sqrt{2}\,
\sigma_{gr}$. Therefore, in some sense continuous phase-insensitive
measurement brings the same total information as the phase-sensitive
(e.g.\ stroboscopic) measurement; however, in our game only half of
this information is useful for the David's goal.
After discussing measurement of the resonator state it is easy
to understand the quantum limits for the high-gain microwave
amplifiers. Now suppose Charlie prepares a coherent state of a
microwave resonator with quadratures $A$ and $B$, gives it to David
to find $A$, and David uses an amplifier for amplification of the
microwave field, which slowly leaks from the resonator until it is
empty. There is only
classical signal processing after the amplifier, so amplification
is essentially the quantum measurement. The results are the same as
above \cite{Clerk-RMP,Braginsky,Caves-1982,Dev-Likh}: a
phase-sensitive amplifier, which amplifies only $A$-quadrature and
``de-amplifies'' (attenuates) $B$-quadrature can measure $A$ with
accuracy $\sigma_{gr}$, while a phase-preserving amplifier can
measure $A$ only with accuracy $\sqrt{2}\, \sigma_{gr}$ (and also
measures $B$ with the same accuracy $\sqrt{2}\, \sigma_{gr}$).
Technically, the accuracy is limited by the noise at the amplifier
output, so this noise should forbid measuring of $A$ with accuracy
better than $\sigma_{gr}$ by a phase-sensitive amplifier, and better
than $\sqrt{2}\, \sigma_{gr}$ by a phase-preserving amplifier.
Therefore in the quantum-limited case the output noise power of a
phase-preserving amplifier (per quadrature) is twice larger than for
a phase-sensitive amplifier with the same gain; this is often called
an ``additional noise'', corresponding to one half of the energy
quantum (one more half-quantum is present in both cases)
\cite{Clerk-RMP,Braginsky,Caves-1982,Dev-Likh}. It may be somewhat
confusing why this result does not depend on the rate with which the
microwave leaks from the resonator. So let us check the scaling: for
$k$ times slower leakage ($k$ times larger $Q$-factor) the microwave
amplitude is $\sqrt{k}$ times smaller, but accumulation time is $k$
times longer, therefore the measured signal for the quadrature $A$
is $\sqrt{k}$ times larger, which is the same factor as for the
noise accumulation. Therefore, the signal-to-noise ratio which
determines the $A$-accuracy does not depend on the leakage rate
(resonator bandwidth).
\section{Narrowband (cQED) measurement}\label{narrowband}
Using the above discussion of microwave amplifiers, it is easy
to extend the Bayesian approach for a broadband quantum measurement
to the narrowband cQED setup.
We consider the standard cQED setup
\cite{Wallraff-04,DiCarlo-10,Fragner-2008,Bertet,Siddiqi-2011,Blais-04,Clerk-RMP,Gambetta-06,Gambetta-08},
in which a qubit interacts with a microwave resonator, and assume
the dispersive regime with the Hamiltonian
\begin{equation}
H=(\hbar\omega_{qb}/2)\,\sigma_z +\hbar \omega_r a^\dagger a +\hbar
\chi a^\dagger a \sigma_z ,
\label{H-dispersive}\end{equation}
where $\omega_{qb}=\omega_{qb, bare}+\chi$ is the Lamb-shifted qubit
frequency with no photons in the resonator,
$\chi=g^2/(\omega_{qb,bare}-\omega_r)$ is the effective coupling
with $g$ being the Jaynes-Cummings coupling, $\omega_r$ is the bare
resonator frequency, Pauli operator $\sigma_z$ acts on the qubit
state in the energy basis $\{|0\rangle,|1\rangle\}$, and the
resonator creation/annihilation operators are $a^\dagger$ and $a$.
Notice that the resonator frequency increases by $2\chi$ when the
qubit state changes from $|0\rangle$ to $|1\rangle$; conversely, the
qubit frequency increases by $2\chi$ per each additional photon in
the resonator.
To measure the qubit state, a microwave field with frequency
$\omega_m$ is either transmitted through or reflected from the
resonator, then amplified and sent to the IQ mixer, which measures
both quadratures relative to the original microwave tone (Fig.\ 1).
The qubit state affects the resonator frequency and therefore
affects the phase (and in general amplitude) of the outgoing
microwave.
An elementary Fabry-P\'erot analysis gives the classical
(complex) microwave field $F_r$ inside the resonator:
\be
F_r =\frac{2 F_{in} t_{in}/\kappa\tau_{rt}}{1-2i(\omega_m-\omega_r)/
\kappa},
\label{Fabry-Perot}\ee
where $F_{in}$ is the applied incident field, $t_{in}$ is the
transmission amplitude of the barrier from the incident side,
$\kappa$ is the resonator bandwidth due to the microwave leakage
from the both sides (the $Q$-factor is $\omega_r/\kappa$), and the
round-trip time is $\tau_{rt}= 2\pi/\omega_r$ for a half-wavelength
resonator and $\tau_{rt}= \pi/\omega_r$ for a quarter-wavelength
resonator. A similar formula with the same denominator describes a
lumped resonator. In presence of the qubit, the resonator frequency
$\omega_r$ in this formula is substituted by $\omega_r\pm \chi$,
depending on the qubit state. Notice that for the quantum
measurement analysis {\it there is no difference between the cases
of transmission and reflection} for the same $F_r$ and $\kappa$,
because the field leaking from the resonator is determined only by
$F_r$ and $\kappa$. (The reflection case has a technical advantage
of dealing with a twice smaller outgoing microwave field for the
same measured signal.) However, an important parameter is the
collected fraction $\eta_{col}=\kappa_{col}/\kappa$ of the leaking
microwave power; we will often assume the ideal case $\eta_{col}=1$
(for the transmission setup this requires strongly asymmetric
coupling, $|t_{in}|\ll |t_{out}|$).
For simplicity we assume the resonant case, $\omega_m=\omega_r$,
then the ensemble qubit dephasing due to measurement is
\cite{Blais-04,Clerk-RMP}
\be
\Gamma = 8\chi^2\bar{n}/\kappa,
\label{Gamma-cQED}\ee
where $\bar{n}$ is the average number of photons in the resonator.
It is easy to include Rabi oscillations into the model; however, we
do not do it for simplicity and also for more transparent analogy
with Sec.\ \ref{broadband}, in which we considered a qubit with zero
Hamiltonian, evolving only due to measurement; this case exactly
corresponds to the cQED Hamiltonian (\ref{H-dispersive}) in the
rotating frame.
We will need several assumptions to describe the qubit state
evolution in the process of measurement. First, for the validity of
the dispersive approximation (\ref{H-dispersive}) we need
sufficiently large qubit-resonator detuning,
$|\omega_{qb}-\omega_r|\gg |g|$, and not too many photons in the
resonator, $\bar{n}\ll (\omega_{qb}-\omega_r)^2/g^2$ (we do not
consider the recently discovered nonlinear regime \cite{nonlinear}).
Second, to use the Markov approximation for the evolution we need
the so-called ``bad cavity'' assumption: $\Gamma \ll \kappa \ll
\omega_r$ (if the qubit evolves due to Rabi oscillations with
frequency $\Omega_R$, we also need $\Omega_R\ll \kappa$). This
assumption means that the photons leave the resonator much faster
than evolution of the qubit state, and therefore there is
practically no entanglement between the qubit and unmeasured
microwave field. This assumption also implies that the two resonator
states for the qubit states $|0\rangle$ and $|1\rangle$ are almost
indistinguishable, $\bar{n}(\chi/\kappa)^2\ll 1$. Third, we use the
``weak response'' assumption, which requires a small phase
difference between the two resonator states, $|\chi|/\kappa \ll 1$.
This means that each outgoing photon carries only a little
information about the qubit state. Notice that for $\bar{n}\agt 1$
the previous assumption $\kappa\gg\Gamma$ automatically implies the
weak response, and even for $\bar{n}\ll 1$ the weak response
assumption is not always needed. Fourth, we will neglect the qubit
energy relaxation due to measurement \cite{Blais-04,Clerk-RMP},
which can be added later.
A coherent state in the resonator with average $\bar{n}$ photons
and zero average phase corresponds to the oscillation of the field
expectation value $\langle F_r(t)\rangle=2\sqrt{\bar{n}}\,
\sigma_{gr} \cos (\omega_m t)$, where $\sigma_{gr}$ is the ground
state width (rms uncertainty) and we assume $\omega_m=\omega_r$.
(Notice that the amplitude $\sigma_{gr}$ corresponds to 1/4 photon.)
Interaction with the qubit slightly changes the phase,
$\cos(\omega_mt \mp 2\chi/\kappa)$, depending on the qubit state, so
that
\begin{eqnarray}
&& \langle F_r(t)\rangle = A\cos(\omega_m t)+B\sin (\omega_m
t),\,\,\,
A=2\sqrt{\bar{n}} \,
\sigma_{gr} , \quad
\nonumber \\
&&
B=\pm (4\chi /\kappa ) \sqrt{\bar{n}}\, \sigma_{gr} =
(4\chi /\kappa ) \sqrt{\bar{n}}\, \sigma_{gr} \, z,
\label{A-B-quadratures}\end{eqnarray}
where $z$ is the qubit Bloch coordinate.
Thus the small $B$-quadrature carries information about the qubit
state, while larger $A$-quadrature may give us information on the
fluctuations of the photon number in the resonator. In the optical
representation (Fig.\ 3) with axes $A/\sigma_{gr}$ and
$B/\sigma_{gr}$, the two resonator states for the qubit states
$|0\rangle$ and $|1\rangle$ are shown as two ``error circles''
\cite{Walls-Milburn} with rms uncertainty 1 along any direction and
distance $2\sqrt{\bar{n}}$ between the origin and circle centers (if
axes $A/2\sigma_{gr}$ and $B/2\sigma_{gr}$ are used, then the
distance is $\sqrt{\bar{n}}$, while the uncertainty is 1/2).
\begin{figure}[tb]
\centering
\includegraphics[width=5.0cm]{LesHou-fg3.eps}
\caption{Phase space representation: for each qubit state the coherent
state with $\langle F_r\rangle=A\cos(\omega_rt)+B\sin(\omega_rt)$ in
the resonator is shown \cite{Walls-Milburn} as an ``error circle''
with radius 1, shifted by $2\sqrt{\bar{n}}$ from the origin. Axes
are normalized by the standard deviation $\sigma_{gr}$ of the ground
state. The $B$-quadrature carries information about the qubit state,
the $A$-quadrature corresponds to the number of photons in the
resonator.}
\label{fig3}
\end{figure}
\subsection{Phase-sensitive amplifier}
Let us start with the case when a phase-sensitive amplifier is
used in the cQED setup. Also, we first assume the most ideal case:
the amplifier is quantum-limited, it amplifies the optimal
$B$-quadrature, there is no microwave collection loss
($\kappa_{col}/\kappa=1$), and there is no extra noise or dephasing.
Then, as discussed in the previous section, measuring once the
microwave contents of the resonator (by fully emptying it) we can
measure the $B$-quadrature with imprecision $\sigma_{gr}$.
Therefore, in the continuous measurement for time $t$ the
$B$-quadrature is measured with imprecision
$\sigma_{gr}/\sqrt{\kappa t}$, which converts into the imprecision
$\sqrt{\kappa/t}/(4\chi\sqrt{\bar{n}})$ of the qubit $z$-coordinate.
Following the language of Sec.\ \ref{broadband}, let us discuss the
signal and noise at the output of the setup. There are two outputs
of the IQ mixer; however, only the amplified quadrature carries an
information, so let us denote the corresponding output of the mixer
(or their linear combination) as $I(t)$. Then the response $\Delta
I=I_1-I_0$ corresponds to $\Delta z=2$ and $\Delta
B=8(\chi/\kappa)\sqrt{\bar{n}}\,\sigma_{gr}$. For measurement during
time $t$ the above variance $(\kappa/t)/(4\chi\sqrt{\bar{n}})^2$ of
the $z$-coordinate converts into the variance $(\kappa/t)(\Delta
I/8\chi \sqrt{\bar{n}})^2$ of the measured output
$\bar{I}_m=(1/t)\int_0^t I(t')\, dt'$. Equating it with $D=S/2t$, we
find the (single-sided) spectral density of the $I(t)$ noise:
\be
S_{min}=(\Delta I_{max})^2\kappa /(32 \chi^2 \bar{n}),
\label{S}\ee
where we replaced $S$ with $S_{min}$ and $\Delta I$ with $\Delta
I_{max}$ to remind that we consider the quantum-limited case, and
the response is maximized by amplifying the optimal quadrature.
Notice that since $\Delta I_{max}\propto \chi
\sqrt{\bar{n}/\kappa}$, the noise $S_{min}$ does not depend on the
qubit or resonator properties; it is essentially the amplified
vacuum noise and depends only on the amplifier gain.
\begin{figure}[tb]
\centering
\includegraphics[width=7cm]{LesHou-fg4.eps}
\caption{Relations between relevant quadratures. The quadratures
$A$ and $B$ are for the microwave field in the resonator. (a) For a
phase-sensitive amplifier the amplified quadrature makes an angle
$\varphi$ with the informational $B$-quadrature. The corresponding
quadrature at the mixer output is defined as $I(t)$ [the output
$Q(t)$ is then useless]. (b) For a phase-preserving amplifier we
define output quadratures $I(t)$ and $Q(t)$ as corresponding to the
resonator quadratures $B$ and $A$.}
\label{fig4}
\end{figure}
Obtaining information on the qubit $z$-coordinate via the signal $I(t)$
with response $\Delta I_{max}$ and noise $S_{min}$, we necessarily
cause the ``spooky'' back-action described by Eqs.\ (\ref{rho-diag})
and (\ref{rho-off}). As discussed in Sec.\ \ref{broadband}, this is
the consequence of the corresponding principle or just the common
sense. Now averaging the $\rho_{01}$ evolution in Eq.\
(\ref{rho-diag}) over the measurement result $\bar{I}_m$ with its
probability distribution (\ref{P(I)}) and (\ref{gaussian}), we see
that the ``spooky'' back-action dephases an ensemble of qubits with
the rate [see Eq.\ (\ref{Gamma})] $(\Delta
I_{max})^2/4S_{min}=8\chi^2\bar{n}/\kappa$. This rate coincides with
the total ensemble dephasing (\ref{Gamma-cQED}), and therefore the
qubit state cannot additionally fluctuate due to any other reason.
Thus we derived an important result: {\it in the ideal case with
phase-sensitive amplifier there is only the ``spooky'' back-action
and no ``realistic'' back-action}. This means that the {\it number
of photons in the resonator does not fluctuate} (otherwise there
would be an additional dephasing), which makes sense since we cannot
measure the $A$-quadrature, carrying information on the photon
number. Notice that it is also easy to prove this result when
$\omega_m\neq\omega_r$. Then from Eq.\ (\ref{Fabry-Perot}) we obtain
that the informational quadrature amplitude is multiplied by the
factor $[1+4(\omega_m-\omega_r)^2/\kappa^2]^{-1/2}$ compared with
Eq.\ (\ref{A-B-quadratures}). The response $\Delta I_{max}$ is
multiplied by same factor, while the noise $S_{min}$ does not
change. Therefore, the ``spooky'' back-action contribution into the
ensemble dephasing is multiplied by the factor
$[1+4(\omega_m-\omega_r)^2/\kappa^2]^{-1}$, which again coincides
with the result \cite{Blais-04,Clerk-RMP} for the total ensemble
dephasing $\Gamma$. This proves the absence of the ``realistic''
back-action for the non-resonant case $\omega_m\neq\omega_r$ as
well.
Now let us consider the case when an ideal phase-sensitive
amplifier amplifies the $A$-quadrature (we again assume
$\omega_m=\omega_r$ for simplicity). Then we do not get any
information on the qubit $z$-coordinate, and therefore there is no
``spooky'' back-action, but there is the ``realistic'' back-action
due to fluctuating number of photons. The description of evolution
in this case is essentially the standard description
\cite{Blais-04,Clerk-RMP}. Let us still denote with $I(t)$ the
output signal from the mixer, corresponding to the amplified
quadrature. For measurement during time $t$ we measure
$A$-quadrature with imprecision $\sigma_{gr}/\sqrt{\kappa t}$. This
is consistent with fluctuation of the number $N$ of emitted photons:
${\rm var}(N)=\bar{N}$, $\bar{N}=\bar{n} \kappa t$. The correlation
function of photon number in the resonator depends on time as $\exp
(-\kappa t/2)$ \cite{Blais-04,Clerk-RMP}, which means that each
extra photon inferred from $I(t)$ fluctuation spends (on average)
time $2/\kappa$ in the resonator and therefore changes the qubit
phase $\phi$ by $4\chi/\kappa$ (the correlation time $2/\kappa$ is
essentially the lifetime of the field, not power
\cite{Blais-04,Clerk-RMP}). Then $\phi$-variance is ${\rm var} (\phi
)=(4\chi/\kappa )^2 \bar{n}\kappa t$, and the corresponding ensemble
dephasing is ${\rm var} (\phi )/2t=8\chi^2\bar{n}/\kappa$. As
expected, this reproduces the standard result (\ref{Gamma-cQED}) for
the ensemble dephasing, while for individual qubit evolution we have
the above discussed correlation: each additional photon inferred
from $I(t)$ fluctuation changes $\phi$ by $4\chi/\kappa$. For the
same amplifier gain and noise as for measuring $B$-quadrature, we
get $\delta \sqrt{\bar{n}}=(4\chi/\kappa) \sqrt{\bar{n}}\,(\delta
I_m)/\Delta I_{max}$, and therefore the correlation is $K=\delta
\langle \phi \rangle/\delta (\overline{I}_m t) = 32
(\chi^2/\kappa)\, \bar{n}/\Delta I_{max}=\Delta I_{max}/S_{min}$. It
is easy to check that $K^2S_{min}/4$ [see Eq.\ (\ref{Gamma})]
coincides with ensemble dephasing $\Gamma$ from Eq.\
(\ref{Gamma-cQED}), as expected for the presence of only
``realistic'' back-action.
Finally, assume that the phase-sensitive amplifier amplifies the
quadrature, which makes angle $\varphi$ with the optimal
$B$-quadrature and angle $\pi/2-\varphi$ with the $A$-quadrature.
The measured signal $I(t)$ still denotes the output of the IQ mixer,
corresponding to the amplified quadrature; now it gives information
about both $B$ and $A$ quadratures, with the factors $\cos\varphi$
and $\sin\varphi$, respectively. Combining the ``spooky'' and
``realistic'' back-actions, we get the same formulas as for the
broadband detection of Sec.\ \ref{broadband}:
\begin{eqnarray}
&& \frac{\rho_{11}(t)}{\rho_{00}(t)} =
\frac{\rho_{11}(0)}{\rho_{00}(0)} \, \exp [
\tilde{I}_m(t) \Delta I/D], \,\,\, D=\frac{S}{2t}, \qquad
\label{ph-sens-diag}\\
&& \frac{\rho_{01}(t)}{\rho_{01}(0)}= \frac{\sqrt{\rho_{00}(t)\,
\rho_{11}(t)}} {\sqrt{\rho_{00}(0)\, \rho_{11}(0)}}\, \exp [i K
\tilde{I}_m(t)\, t] ,
\label{ph-sens-off}\\
&& \Delta I= I_1-I_0= \Delta I_{max}\cos\varphi, \,\,\,
K=K_{max} \sin\varphi , \qquad
\label{ph-sens-DeltaI}\\
&& K_{max} = \Delta I_{max}/S,
\label{ph-sens-Kmax}\\
&& \tilde I_m(t) = \frac{1}{t} \int_0^\infty I(t')\, dt'
- \frac{I_0+I_1}{2},
\label{ph-sens-Im}\\
&& I(t)=\frac{I_0+I_1}{2} + \frac{\Delta I}{2} \, z (t) +\xi(t),
\,\,\, S_\xi(\omega )=S.
\label{ph-sens-I(t)}\end{eqnarray}
Here we introduced $\tilde I_m$ by subtracting the constant
$(I_0+I_1)/2$ from $\overline{I}_m$ and did a simple algebra to
convert Eq.\ (\ref{rho-diag}) into Eq.\ (\ref{ph-sens-diag}); the
qubit rotating frame corresponds to $\bar{n}$ photons, $K_{max}$ is
the above discussed correlation for $A$-quadrature amplification,
$\Delta I_{max}$ is the response for $B$-quadrature amplification,
and $S=S_{min}$. Notice that the total ensemble dephasing
(\ref{Gamma}) does not depend on $\varphi$:
\be
(\Delta I_{max}\cos\varphi )^2/4S_{min}+
(K_{max}\sin\varphi )^2 S_{min}/4= \Gamma.
\ee
So far we discussed only the ideal case. There are several
mechanisms for non-ideality. First, the qubit may have additional
environmental dephasing $\gamma_{env}$. This will lead to the extra
factor $e^{-\gamma_{env} t}$ in Eq.\ (\ref{ph-sens-off}) and
increase the ensemble dephasing $\Gamma$ by $\gamma_{env}$.
Following the definitions in Sec.\ \ref{broadband}, the
corresponding quantum efficiency is $\eta_{env}=(1+\gamma_{env}
\kappa/8\chi^2\bar{n})^{-1}$. Second, not all microwave power
leaking from the resonator may be collected and amplified. This can
be characterized by the collection efficiency
$\eta_{col}=\kappa_{col}/\kappa$ and multiplies the response $\Delta
I$ and correlation $K$ by the factor $\sqrt{\eta_{col}}$, while not
affecting the output noise $S$. Third, if the phase-preserving
amplifier is not quantum-limited, it introduces additional noise
$S_{add}$ compared with the quantum limit $S_{min}$ [given by Eq.\
(\ref{S}) when $\eta_{col}=1$]. The corresponding amplifier
efficiency is $\eta_{amp}=S_{min}/(S_{min}+S_{add})$. This does not
affect $\Delta I$ but multiplies $K$ by $\eta_{amp}$ (because for
uncorrelated Gaussian-distributed random numbers $x_1$ and $x_2$,
the averaging of $x_1$ for a fixed sum $x_1+x_2$ gives the
correlation $\langle x_1\rangle/(x_1+x_2)={\rm var}(x_1)/[{\rm
var}(x_1)+{\rm var}(x_2)]$).
If all three mechanisms of the non-ideality are present, then the
evolution can still be described \cite{Kor-nonideal} by Eqs.\
(\ref{ph-sens-diag})--(\ref{ph-sens-I(t)}), but $S$ is now the total
(experimental) output noise, $\Delta I_{max}$ is the experimental
response for $\varphi=0$, so that $S=(\Delta I_{max})^2
\kappa/(32\chi^2\bar{n}\,\eta_{col}\,\eta_{amp})$, the correlation
$K=\delta\langle\phi\rangle /\delta (\tilde{I}_mt)$ is still given
by Eqs.\ (\ref{ph-sens-DeltaI})--(\ref{ph-sens-Kmax}), and the only
change is the extra factor in Eq.\ (\ref{ph-sens-off}):
\begin{eqnarray}
&& \frac{\rho_{01}(t)}{\rho_{01}(0)}= \frac{\sqrt{\rho_{00}(t)\,
\rho_{11}(t)}} {\sqrt{\rho_{00}(0)\, \rho_{11}(0)}}\, \exp [i K
\tilde{I}_m(t)\, t] \, e^{-\gamma t} , \qquad
\label{ph-sens-off-2}\\
&& \gamma =\Gamma -(\Delta I_{max})^2/4S,
\label{ph-sens-gamma}\end{eqnarray}
where the ensemble dephasing is now $\Gamma
=8\chi^2\bar{\eta}/\kappa +\gamma_{env}$. We emphasize that the
qubit evolution depends only on the experimentally measurable
parameters $\Delta I_{max}$, $S$, $\Gamma$, $\varphi$, and the
output signal $I(t)$.
The quantum efficiencies can be expressed as
\be
\eta=\frac{(\Delta
I_{max})^2}{4S\Gamma}=\eta_{amp}\,\eta_{col}\,\eta_{env} , \,\,\,
\tilde{\eta}=\eta \cos^2\varphi,
\ee
where as in Sec.\ \ref{broadband}, $\eta$ is the relative
contribution to $\Gamma$ from both the ``spooky'' and ``realistic''
back-actions, while $\tilde\eta$ is the relative contribution from
only the ``spooky'' back-action. The definition of $\tilde\eta$
corresponds to replacing the ``realistic'' back-action factor
$\exp(iK\tilde{I}_m t)$ in Eq.\ (\ref{ph-sens-off-2}) with the
corresponding ensemble dephasing $\exp (-K^2St/4)$. As mentioned in
Sec.\ \ref{broadband}, the peak-to-pedestal ratio of the spectral
peak of contiuous Rabi oscillations is
$4\tilde{\eta}=4\eta\cos^2\varphi$.
\subsection{Phase-preserving amplifier}
Now assume that a phase-preserving amplifier is used (this
includes parametric amplifier, HEMT, etc.). Now both the
$A$-quadrature and $B$-quadrature of Eq.\ (\ref{A-B-quadratures})
are amplified independently with the same gain. Correspondingly,
both quadratures at the IQ mixer output carry physical information
instead of only one quadrature in the case of a phase-sensitive
amplifier. Let us denote with $I(t)$ the output of the IQ mixer,
corresponding to the $B$-quadrature; thus $I(t)$ provides
information on the qubit $z$-coordinate. The output signal for the
orthogonal quadrature is denoted $Q(t)$; it corresponds to the
$A$-quadrature in the resonator and provides information on the
fluctuating number of photons. The main difference from the case of
a phase-sensitive amplifier is that now the ``spooky'' and
``realistic'' back-actions are related to two different output
signals: $I(t)$ and $Q(t)$.
Let us start with the quantum-limited case and assume
an amplifier with the same gain as in the phase-sensitive case, so
that the $I(t)$-channel response is the same as the optimal
phase-sensitive response, $\Delta I=\Delta I_{max}$. The ``spooky''
back-action is always described by the quantum Bayes formulas
(\ref{rho-diag})--(\ref{rho-off}), but now the noise $S$ of the
output $I(t)$ is twice larger than the value (\ref{S}) for the
phase-sensitive amplifier, $S=2S_{min}$ (see discussion in Sec.\
\ref{amplifiers}); therefore the ``spooky'' evolution is twice
slower than in the phase-sensitive case with $\varphi=0$.
The signal $Q(t)$ has the same noise $S=2S_{min}$; it is again
twice larger than for the phase-sensitive case with
$\varphi=\pi/2$; therefore the correlation factor
$K=\delta\langle\phi\rangle/\delta [\int_0^t Q(t')\, dt']$ for the
``realistic'' back-action is twice smaller: $K=K_{max}/2$ (this
reduction is similar to the effect of a non-ideal amplifier
discussed above). We see that $K=\Delta I/S$, and the ensemble
dephasing is at least $(\Delta I)^2/4S+K^2S/4=(\Delta
I)^2/2S=(\Delta I_{max})/4S_{min}$. This again coincides with
$\Gamma=8\chi^2\bar{n}/\kappa$, and therefore there can be no
additional evolution of the qubit besides these ``spooky'' and
``realistic'' back-actions.
Thus in the ideal case the qubit evolution is
\begin{eqnarray}
&& \frac{\rho_{11}(t)}{\rho_{00}(t)} =
\frac{\rho_{11}(0)}{\rho_{00}(0)} \, \exp [
\tilde{I}_m(t) \Delta I/D], \,\,\, D=\frac{S}{2t}, \qquad
\label{ph-pres-diag}\\
&& \frac{\rho_{01}(t)}{\rho_{01}(0)}= \frac{\sqrt{\rho_{00}(t)\,
\rho_{11}(t)}} {\sqrt{\rho_{00}(0)\, \rho_{11}(0)}}\, \exp [i K
\tilde{Q}_m(t)\, t] ,
\label{ph-pres-off}\\
&& \Delta I= I_1-I_0, \,\,\, K = \Delta I/S,
\label{ph-pres-K}\\
&& \tilde Q_m(t) = \frac{1}{t} \int_0^\infty Q(t')\, dt'
- \langle Q\rangle, \,\,\, S_Q=S_I=S,
\label{ph-pres-Q}\end{eqnarray}
where $\langle Q\rangle$ is the average value of $Q(t)$ (which
depends on $\bar{n}$), $\tilde{I}_m(t)$ is defined by Eq.\
(\ref{ph-sens-Im}), and the channels $I(t)$ and $Q(t)$ both have the
same (uncorrelated) noise $S=(\Delta I)^2\kappa/(16\chi^2\bar{n})$.
Notice that $(\Delta I)^2/4S=K^2S/4=4\chi^2\bar{n}/\kappa$, and
therefore {\it in the phase-preserving case the ensemble dephasing
$\Gamma$ contains equal contributions $\Gamma/2$ from the ``spooky''
and ``realistic'' back-actions}. We emphasize that Eqs.\
(\ref{ph-pres-diag})--(\ref{ph-pres-off}) still {\it allow us to
monitor a qubit wavefunction} if the initial qubit state is pure.
A non-ideal case can be analyzed in the same way as for the
phase-sensitive amplifier. An extra dephasing $\gamma_{env}$ of the
qubit is described by
$\eta_{env}=(1+\gamma_{env}\kappa/8\chi^2\bar{n})^{-1}$, imperfect
collection efficiency is described by
$\eta_{col}=\kappa_{col}/\kappa$, and the amplifier efficiency is
$\eta_{amp}$. We define $\eta_{amp}=S_{ql}/S$ for a phase-preserving
amplifier by comparing its output noise $S$ (per quadrature) with
the quantum limit for a phase-preserving amplifier:
$S_{ql}=2S_{min}$, so that $\eta_{amp}=1$ in the quantum-limited
case. We also define $\tilde\eta=S_{min}/S$ by comparison with a
phase-sensitive amplifier having the same gain, so that
$\tilde{\eta}_{amp}=\eta_{amp}/2$ and obviously
$\tilde{\eta}_{amp}\leq 1/2$. Similarly to the phase-preserving
case, incomplete microwave collection multiplies the response
$\Delta I$ and correlation $K$ by the factor $\sqrt{\eta_{col}}$ but
does not change the noise $S$; the extra noise in the amplifier
multiplies $K$ by $\eta_{amp}$ but does not change $\Delta I$.
The qubit evolution can still be described by Eqs.\
(\ref{ph-pres-diag})--(\ref{ph-pres-Q}) with the only change in Eq.\
(\ref{ph-pres-off}):
\begin{eqnarray}
&& \frac{\rho_{01}(t)}{\rho_{01}(0)}= \frac{\sqrt{\rho_{00}(t)\,
\rho_{11}(t)}} {\sqrt{\rho_{00}(0)\, \rho_{11}(0)}}\, \exp [i K
\tilde{Q}_m(t)\, t] \, e^{-\gamma t} , \qquad
\label{ph-pres-off-2}\\
&& \gamma=\Gamma-2(\Delta I)^2/4S,
\label{ph-pres-gamma}\end{eqnarray}
where now $S$ is the total (experimental) noise per quadrature,
$\Delta I$ is the experimental response, and $\Gamma
=8\chi^2\bar{\eta}/\kappa +\gamma_{env}$ is the total ensemble
dephasing. The qubit evolution is determined by the parameters
$\Delta I$, $S$, $\Gamma$, and output signals $I(t)$ and $Q(t)$.
The quantum efficiencies are
\be
\eta=\eta_{amp}\eta_{col}\eta_{env}=(\Delta I)^2/2S\Gamma, \,\,\,
\tilde\eta =\eta/2.
\ee
Here $\eta$ is the fraction of $\Gamma$ due to the contribution from
both the ``spooky'' and ``realistic'' back-actions. The efficiency
$\tilde\eta =\tilde{\eta}_{amp}\eta_{col}\eta_{env}$ is the fraction
from only the ``spooky'' contribution; it corresponds to replacing
the term $\exp (i K \tilde{Q}_m\, t)$ in Eq.\ (\ref{ph-pres-off-2})
with the dephasing term $\exp(-K^2t/4S)$. In particular, the
peak-to-pedestal ratio of the Rabi spectral peak for the signal
$I(t)$ is $4\tilde\eta=2\eta$.
Let us mention that Eqs.\
(\ref{ph-pres-diag})--(\ref{ph-pres-K}) for the ideal
phase-preserving case can also be obtained from Eqs.\
(\ref{ph-sens-diag})--(\ref{ph-sens-Kmax}) for the phase-sensitive
case in the following way.
Let us think about a phase-preserving
amplifier as a phase-sensitive amplifier, in which the angle
$\varphi$ rapidly changes with time, and we have to average over
$\varphi$. When coefficients $\cos\varphi$ and $\sin\varphi$ in Eq.\
(\ref{ph-sens-DeltaI}) are substituted into Eqs.\
(\ref{ph-sens-diag}) and (\ref{ph-sens-off}), we see a natural
formation of the quadratures $\tilde{I}_m$ and $\tilde{Q}_m$ of the
phase-preserving setup. Then the exponential factor in Eq.\
(\ref{ph-sens-diag}) becomes $\exp (\tilde{I}_m \Delta I_{max}/D)$,
and the exponential factor in Eq.\ (\ref{ph-sens-off}) becomes
$\exp (i K_{max}\tilde{Q}_m t)$. Now let us take into account that
the average response is $\Delta I=\overline{\cos^2\varphi}\, \Delta
I_{max}=\Delta I_{max}/2$, and the phase-sensitive amplifier noise
$S$ splits equally between the $I(t)$ and $Q(t)$ quadratures (the
orthogonal, de-amplified quadrature is noiseless). The mutual
cancellation of these two factors of 2 leads to the same form of
Eq.\ (\ref{ph-pres-diag}) as in Eq.\ (\ref{ph-sens-diag}) and the
relation $K=\Delta I/S$ in Eq.\ (\ref{ph-pres-K}).
One more way to understand the relation between the ideal
phase-sensitive and phase-preserving cases is the following. Instead
of using a phase-preserving amplifier, let us split the outgoing
microwave into two equal parts and use phase-sensitive amplifiers
with $\varphi=0$ and $\varphi=\pi/2$ in the two channels. To keep
the same noise $S$ per channel, we increase the gain by the factor
$\sqrt{2}$, that also compensates the signal loss at the splitter.
Then the channel $\varphi=0$ produces the ``spooky'' back-action
(\ref{ph-pres-diag}), while the channel $\varphi=\pi/2$ informs us
of the ``realistic'' back-action (\ref{ph-pres-off}), and the
relation (\ref{ph-pres-K}) between $K$ and $\Delta I$ is the same as
between $K_{max}$ and $\Delta I_{max}$ in (\ref{ph-sens-Kmax}).
In the ideal case ($\eta=1$) the qubit evolution description
can be translated into the language of the POVM-type measurement. In
the same way as in Sec.\ \ref{broadband}, Eqs.\ (\ref{ph-pres-diag})
and (\ref{ph-pres-off}) can be converted into the measurement
operator
\begin{eqnarray}
&& M(\overline{I}_m,\tilde Q_m) = \exp ( -i K \tilde{Q}_m t \,
\sigma_z/2)
\nonumber \\
&&\hspace{0.5cm} \times \left[\sqrt{P_{|0\rangle}(\overline{I}_m)}
\, |0\rangle \langle 0| + \sqrt{P_{|1\rangle}(\overline{I}_m)} \,
|1\rangle \langle 1|\right], \qquad
\label{meas-op-QED}\end{eqnarray}
where the probabilities $P_{|j\rangle}$ are given by Eq.\
(\ref{gaussian}). Similarly, Eqs.\ (\ref{ph-sens-diag}) and
(\ref{ph-sens-off}) for the case of a phase-sensitive amplifier can
be converted into the same measurement operator (\ref{meas-op-QED}),
in which $\tilde Q_m$ is replaced with $\tilde I_m$.
\section{Conclusion}
We have presented a simple physical picture of the qubit
evolution due to its measurement in the circuit QED setup. The
``spooky'' back-action is universal, it is caused by gradual
extraction of information about the qubit state. The ``realistic''
back-action is due to a specific mechanism: fluctuation of the
photon number in the resonator. For a phase-sensitive amplifier the
qubit evolution is described by Eqs.\ (\ref{ph-sens-diag}) and
(\ref{ph-sens-off-2}); it is determined by the output signal $I(t)$,
which corresponds to the amplified quadrature. For a
phase-preserving amplifier the evolution is described by Eqs.\
(\ref{ph-pres-diag}) and (\ref{ph-pres-off-2}); it is determined by
two output signals: $I(t)$ and $Q(t)$, where $I(t)$ now corresponds
to the quadrature, which provides information about the qubit state
($B$-quadrature in the resonator) and $Q(t)$ corresponds to the
orthogonal $A$-quadrature, which gives us a record of the photon
number fluctuations in the resonator.
While the cQED setup significantly differs from both the
broadband quantum measurement setup \cite{Kor-rev} and the standard
optical setup \cite{quant-traj}, we see that the description of the
qubit evolution is exactly the same as in both these cases if a
phase-sensitive amplifier is used. The description is only slightly
different when a phase-preserving amplifier is used: we should
assign the ``spooky'' and ``realistic'' back-actions to the separate
output signals $I(t)$ and $Q(t)$ instead of only one signal. It is
also useful to think about the phase-preserving case via the model
in which we split the outgoing microwave (the quantum signal) into
two equal parts and then use 90-degree-shifted phase-sensitive
amplifiers for these two channels.
We intentionally considered only the simplest case, because most
of the further steps and generalizations are quite straightforward
\cite{Kor-rev}. In particular, it it very simple to include Rabi
oscillations and energy relaxation of the qubit state. For that we
have to take time-derivative of the evolution equations and add the
terms due to Rabi oscillations and energy relaxation. If the
Stratonovich definition of the derivative is used, we get the
equations of the Bayesian formalism \cite{Kor-rev}; if the It\^o
derivative is used, we get the equations of the quantum trajectory
formalism \cite{quant-traj,Gambetta-08}. Generalization to
measurement of several entangled qubits is also straightforward
\cite{Kor-rev}. We have considered only the resonant case,
$\omega_m=\omega_r$; however, generalization to the case
$\omega_m\neq \omega_r$ is quite simple (see \cite{Gambetta-08}): we
just need a different definition of the informational $B$-quadrature
and photon-fluctuation $A$-quadrature. In our formalism we
implicitly assumed sufficiently wide bandwidth of the amplifier
(much larger than the ensemble dephasing $\Gamma$ and Rabi frequency
$\Omega_R$). If this is not the case, the formalism should change
significantly. However, we believe that in most of the practical
cases we can take this effect into account by adding a classical
narrowband filter to the classical signal at the amplifier output;
this will correspond to passing the signals $I(t)$ and $Q(t)$
through the low-pass filters. A much more serious change of the
theory is required when the resonator bandwidth $\kappa$ is
comparable to $\Gamma$ or $\Omega_R$; this still has to be done.
Understanding the difference between the ``spooky'' and
``realistic'' back-actions is important for designing the quantum
feedback control of the Rabi oscillations \cite{Kor-feedback}. The
simplest case is when a phase-sensitive amplifier amplifies the
informational $B$-quadrature. Then there is no ``realistic''
back-action, and the feedback loop should only modulate the
amplitude of the Rabi drive (i.e.\ the Rabi frequency $\Omega_R$);
this case was well studied for the broadband setup
\cite{Kor-feedback}. The situation is different for a
phase-preserving amplifier. Then we need two feedback channels: the
first (usual) channel should modulate the Rabi frequency $\Omega_R$
to compensate the ``spooky'' back-action, while the second channel
should compensate the ``realistic'' back-action by modulating the
qubit frequency $\omega_{qb}$ or the the frequency of the Rabi drive
$\omega_R$. The controller for the second feedback channel is quite
simple: it should compensate the contribution $iK\tilde{Q}(t)$ to
the qubit phase derivative $\dot{\phi}(t)$ due to the $K$-term in
Eq.\ (\ref{ph-pres-off-2}). Therefore the controller is
\be \Delta(\omega_{qb}-\omega_{R})=
-K[Q(t)-\langle Q\rangle],
\ee
i.e. we should directly apply the signal $Q(t)$ to modulate
$\omega_{qb}$ or $\omega_R$. The second feedback channel essentially
eliminates the $K$-term in Eq.\ (\ref{ph-pres-off-2}) and decreases
the ensemble dephasing $\Gamma$ by $K^2S/4=(\Delta I)^2/4S$.
Correspondingly, in the absence of the first (main) feedback channel
the peak-to-pedestal ratio of the Rabi peak increases from 2 to 4 in
the quantum-limited case. The first feedback channel should be the
same as for the broadband setup; it depends on the signal $I(t)$ and
can be realized using various ideas for the controller (``direct'',
Bayesian, ``simple'', etc. \cite{Kor-feedback}). Notice that {\it
without the second channel the feedback performance is determined by
the quantum efficiency $\tilde\eta$, while with the second channel
it is determined by
$\tilde{\tilde\eta}=\tilde\eta/(1-\eta+\tilde\eta)$ } (this is one
more combination of the terms in Eq.\ (\ref{Gamma}), which can be
used for the definition \cite{Kor-rev} of quantum efficiency). The
case of a phase-sensitive amplifier, which amplifies a non-optimal
quadrature ($\varphi\neq 0$, $\varphi\neq \pi/2$) is similar to the
case of a phase-preserving amplifier, but both feedback channels
should start with the same signal $I(t)$. {\it In both the
phase-sensitive and phase-preserving setups a perfect feedback
control is possible in the quantum-limited case $\eta=1$.}
Discussion of the ``spooky'' and ``realistic'' back-actions in
the cQED setup necessarily raises the question of causality. When
the microwave leaves the resonator, it does not yet ``know'' in
which way it will be measured (phase-preserving or phase-sensitive,
which angle $\varphi$, etc.). Moreover, when a circulator is used
for the outgoing microwave, the field in the resonator and the qubit
can never ``know'' in a realistic way which method of measurement is
used. Nevertheless, the qubit evolution strongly depends on the
measurement method. As we discussed, the ``spooky'' evolution moves
the qubit state along the meridians of the Bloch sphere, the
``realistic'' back-action moves the state along the parallels, and
the measurement method determines whether the qubit experiences the
``spooky'' or ``realistic'' back-action (or their combination). In
this sense the ``realistic'' back-action is not fully realistic: it
has the physical mechanism, but whether this mechanism works or not
is determined in a spooky way. The causality requires that we cannot
pass a ``useful'' information to the qubit by choosing the
measurement method. This means that {\it the ensemble-averaged
evolution of the qubit cannot depend on the measurement method (this
is the general requirement of causality in quantum mechanics)}. It
is surely satisfied in our cQED setup.
The author thanks Michel Devoret, Konstantin Likharev, Patrice
Bertet, and Farid Khalili for useful discussions. The work was
supported by ARO MURI grant W911NF-11-1-0268 and by NSA/IARPA/ARO
grant W911NF-10-1-0334.
|
\section{Introduction}
Both axions and weakly interacting particles (WIMPs) are considered forms of
cold dark matter (CDM). Furthermore, until recently, axions and WIMPs were
thought to be indistinguishable on observational grounds, i.e. indistinguishable
on the basis of purely astronomical data. The discovery \cite{CABEC} that dark
matter axions form a Bose-Einstein condensate (BEC) has changed this view since
axion BEC is claimed to have observable consequences \cite{CABEC,case,Li7}.
This raises the question: When do axions behave as ordinary CDM and when
do they not?
To start off it is worth emphasizing that, at the fundamental level,
axions and WIMPs are very different. The surprise is really that they
have similar properties as far as large scale structure is concerned.
Both axions and WIMPs are described by quantum fields. Furthermore,
both are excellently described by classical limits of quantum fields.
But the classical limits are different in the two cases: WIMPs are in
the classical particle limit whereas (decoupled) axions are in the classical
field limit. In the classical particle limit one takes $\hbar \rightarrow 0$
while keeping $E = \hbar \omega$ and $\vec{p} = \hbar \vec{k}$ fixed. Since
$\omega, \vec{k} \rightarrow \infty$, the wave nature of the quanta disappears.
WIMPs are to excellent approximation classical point particles. In the
classical field limit, on the other hand, one takes $\hbar \rightarrow 0$
for constant $\omega$ and $\vec{k}$. $E = {\cal N} \hbar \omega$ and
$\vec{p} = {\cal N} \hbar \vec{k}$ are held fixed by letting the quantum
state occupation number ${\cal N} \rightarrow \infty$. This is the limit
in which quantum electrodynamics becomes classical electrodynamics. It is
the appropriate limit for (decoupled) cold dark matter axions because they
are a highly degenerate Bose gas. The axion states that are occupied have
huge occupation numbers, ${\cal N} \sim 10^{61}$ \cite{CABEC}. The need
to restrict to {\it decoupled} axions will be explained shortly.
So axions and WIMPs are fundamentally different even if both can
legitimately be called CDM. The distinction is not just academic,
and is certainly important if axions thermalize, i.e. if axions
find a state of larger entropy through self-interactions. Recall that,
whereas statistical mechanics makes sense of the behaviour of large
aggregates of classical particles (it was invented by Boltzmann to derive
the properties of atoms in the gaseous state) it fails to make sense of
classical fields. In thermal equilibrium every mode of a classical
field would have average energy $k_{\rm B} T$. As Rayleigh pointed out,
the energy density is infinite then at finite temperature due to the
contributions from short wavelength modes. Thus the application of
statistical mechanics to classical field theory (classical electrodynamics
in particular) is in direct disagreement with observation. As is well-known,
the disagreement is removed because of, and only because of, quantum mechanics.
In summary, if the axions are decoupled (i.e. do not interact and hence do not
thermalize), they behave to excellent approximation like a classical field.
However, if the axions thermalize, they are not described by a classical field.
Instead they form a Bose-Einstein condensate, an essentially quantum-mechanical
phenomenon.
\section{Axion Bose-Einstein Condensation}
Axions were originally introduced \cite{PQWW} as a solution to the
strong CP problem. It was later found that they are a cold dark
matter candidate \cite{axdm}. Cold axions are produced when the
axion mass turns on during the QCD phase transition. The critical
time, defined by $m(t_1) t_1 = 1$, is
$t_1 \simeq 2 \cdot 10^{-7}~{\rm sec}~(f / 10^{12}~{\rm GeV})^{1 \over 3}$,
where $f$ is the axion decay constant. The zero temperature axion mass is
given in terms of $f$ by
\begin{equation}
m \simeq 6 \cdot 10^{-6}~{\rm eV}~{10^{12}~{\rm GeV} \over f}~~\ .
\label{mass}
\end{equation}
At temperatures well above 1 GeV, the axion mass is practically zero.
It increases from zero to $m$ during the QCD phase transition. The
cold axions are the quanta of oscillation of the axion field that result
from the turn on of the axion mass. They have number density \cite{axdm}
\begin{equation}
n(t) \sim {4 \cdot 10^{47} \over {\rm cm}^3}~
\left({f \over 10^{12}~{\rm GeV}}\right)^{5 \over 3}
\left({a(t_1) \over a(t)}\right)^3
\label{numden}
\end{equation}
where $a(t)$ is the cosmological scale factor. Because the axion
momenta are of order ${1 \over t_1}$ at time $t_1$ and vary with
time as $a(t)^{-1}$, the velocity dispersion of cold axions is
\begin{equation}
\delta v (t) \sim {1 \over m t_1}~{a(t_1) \over a(t)}
\label{veldis}
\end{equation}
{\it if} each axion remains in whatever state it is in, i.e. if axion
interactions are negligible. We refer to this case as the limit of
decoupled cold axions. If decoupled, the average state occupation
number of cold axions is
\begin{equation}
{\cal N} \sim~ n~{(2 \pi)^3 \over {4 \pi \over 3} (m \delta v)^3}
\sim 10^{61}~\left({f \over 10^{12}~{\rm GeV}}\right)^{8 \over 3}~~\ .
\label{occnum}
\end{equation}
Clearly, the effective temperature of cold axions is much smaller than
the critical temperature
\begin{equation}
T_{\rm c} = \left({\pi^2 n \over \zeta(3)}\right)^{1 \over 3}
\simeq 300~{\rm GeV}~\left({f \over 10^{12}~{\rm GeV}}\right)^{5 \over 9}~
{a(t_1) \over a(t)}
\label{Tc}
\end{equation}
for BEC. Axion number violating processes, such as their decay to two
photons, occur only on time scales vastly longer than the age of the
universe. The only condition for axion BEC that is not manifestly
satisfied is thermal equilibrium.
Axions are in thermal equilibrium if their relaxation rate $\Gamma$
is large compared to the Hubble expansion rate $H(t) = {1 \over 2t}$.
However, the usual techniques of non-equilibrium statistical mechanics
are not applicable to dark matter axions. On the one hand, cold axions
are highly condensed in phase space (${\cal N} >> 1$), which greatly
exaggerates the quantum effect of Bose-enhancement in their scattering
processes. On the other, because their energy dispersion is very small,
they are outside the `particle kinetic regime'. The picture of instantaneous
collisions breaks down, and the usual Boltzmann equation no longer applies.
The particle kinetic regime is defined by $\delta \omega >> \Gamma$ where
$\delta \omega$ is the energy dispersion of the particles. Axions are in
the opposite regime, $\delta \omega << \Gamma$, which we call the `condensed
regime'. Thermalization in the condensed regime is discussed in detail in
ref. \cite{therm}, which gives an estimate for the relaxation rate of cold
dark matter axions due to their gravitational self-interactions:
\begin{equation}
\Gamma \sim 4 \pi G n m^2 \ell^2
\label{rate}
\end{equation}
where $\ell \sim (m \delta v)^{-1}$ is the axion correlation length.
$\Gamma(t)/H(t)$ is of order
$5 \cdot 10^{-7}(f/10^{12}~{\rm GeV})^{2 \over 3}$
at time $t_1$ but grows as $t a^{-1}(t) \propto a(t)$. Thus
gravitational interactions cause the axions to thermalize and
form a BEC when the photon temperature is of order
500 eV~$(f/10^{12}~{\rm GeV})^{1 \over 2}$.
The question is then whether axion BEC has observable consequences.
It is shown in refs. \cite{CABEC,therm} that cold dark matter axions
behave as ordinary cold dark matter on all scales of observational
interest when they are non-interacting. Observable differences between
cold axions and ordinary CDM occur only when the axions self-interact
or interact with other species. Before Bose-Einstein condensation, cold
axions are described by a free classical field and are indistinguishable
from ordinary cold dark matter on all scales of observational interest.
After Bose-Einstein condensation, almost all axions are in the same state.
In the linear regime of evolution of density perturbations and within the
horizon, the lowest energy state is time independent and no rethermalization
is necessary for the axions to remain in the lowest energy state. In that
case, axion BEC and ordinary CDM are again indistinguishable on all scales
of observational interest \cite{CABEC}. However, beyond first order
perturbation theory and/or upon entering the horizon, the axions
rethermalize to try and remain in the lowest energy available
state. Axion BEC behaves differently from CDM then and the
resulting differences are observable.
An example of an observable distinction between axion BEC and ordinary
CDM is given in the next section. Another possible observable distinction
is the cooling of cosmic photons by thermal contact with the axion BEC. If
this happens, the baryon-to-photon ratio at nucleosynthesis and the effective
number of neutrinos (a measure of the radiation density at recombination) are
modified compared to their values in the standard cosmological model \cite{Li7}.
\section{Tidal torquing with axion BEC}
Let us consider axion BEC dark matter as it is about to fall into the
gravitational potential well of a galaxy. The gravitational field of
neighbouring galaxies applies a tidal torque \cite{TTT} to the axion BEC.
Under what conditions is thermalization by gravitational self-interactions
sufficiently fast that the condensed axions remain in the lowest energy
available state as the space-time background evolves? Following the
arguments or ref. \cite{therm}, we expect that the axion BEC rethermalizes
provided the gravitational forces produced by the BEC are larger than the
typical rate $\dot{p}$ of change of axion momenta required for the axions
to remain in the lowest energy state. The gravitational forces are of order
$4 \pi G n m^2 \ell$. In this case, the correlation length $\ell$ must be
taken to be of order the size $L$ of the region of interest since the
gravitational fields due to axion BEC outside the region do not help the
thermalization of the axions within the region. Hence the condition is
\begin{equation}
4 \pi G n m^2 L \gtrsim \dot{p}~~~\ .
\label{thercon}
\end{equation}
The self-similar infall model \cite{selfsim} was used \cite{therm}
to estimate $L$ and $\dot{p}$ as functions of time. Furthermore,
assuming that most of the dark matter is axions, the Friedmann
equation implies
\begin{equation}
4 \pi G n m \simeq {3 \over 2} H(t)^2 \simeq {2 \over 3 t^2}
\label{Fried2}
\end{equation}
after equality between matter and radiation. It is found \cite{therm}
that Eq.~(\ref{thercon}) is satisfied at all times from equality till
today by a margin of order 30.
We conclude that the axion BEC does rethermalize before falling
into the gravitational potential well of a galaxy. Most axions
go to the lowest energy state consistent with the total angular
momentum acquired from neighboring inhomogeneities through tidal
torquing \cite{TTT}. That state is a state of rigid rotation on
the turnaround sphere, implying $\vec{\nabla} \times \vec{v} \neq 0$
where $\vec{v}$ is the velocity field of the infalling axions. In
contrast, the velocity field of WIMP dark matter is irrotational.
The inner caustics of galactic halos are different in the two cases.
Axions produce caustic rings \cite{crdm,sing} whereas WIMPs produce
the `tent-like' caustics described in ref.~\cite{inner}. There is
evidence for the existence of caustic rings in various galaxies at
the radii predicted by the self-similar infall model. For a review
of this evidence see ref. \cite{MWhalo}. It is shown in ref. \cite{case}
that the phase space structure of galactic halos implied by the evidence
for caustic rings is precisely and in all respects that predicted by the
assumption that the dark matter is a rethermalizing BEC.
\section{Acknowledgments}
This work was supported in part by the U.S. Department of Energy
under grant DE-FG02-97ER41209.
\section{Bibliography}
\begin{footnotesize}
|
\section{Introduction}
It is well known from a large set of astrophysical observables that after primordial recombination (which occurred at a redshift of $z\sim1100$) the universe ``reionized'' at a redshift $z > 6$. It is common practice in Cosmic Microwave Background (CMB) studies to parametrize the reionization as an instantaneous process occurring at some redshift $z_r$, with $4<z_{r}<32$, and to marginalize over $z_r$ when deriving constraints on the other cosmological parameters. In the absence of any precise astrophysical model of the reionization process, the electron ionization fraction $x_e(z)$ is parametrized by $z_r$ in the following way: $x_e(z)=1$ for $z \ll z_r$ (possibly $x_e(z)=1.08$ or $x_e(z)=1.16$ for $z<3$ in order to take into account the first and second Helium ionization) and $x_e(z)<2\times 10^{-4}$ for $z>z_r$ in order to join the ionization fraction value after the recombination. In the following we will refer to this parametrization as ``sudden'' or ``instantaneous'' reionization. With this choice of parametrization there exist a one-to-one relation between the redshift of sudden reionization $z_r$ and the electron scattering optical depth $\tau_{e}$. The most recent constraints on the optical depth that come from the analysis of the Wilkinson Microwave Anisotropy Probe team on their seven-year data (WMAP7), in which it is assumed a sudden reionization scenario, is $\tau_{e}=0.088\pm0.015$.
However, as already noticed, e.g. in \cite{mortonson}, and further emphasized by our previous works (\cite{Pandolfi1} and \cite{Pandolfi2}), the assumption of a general reionization scenario could affect the extraction of the constraints of cosmological parameters. In particular, we studied the effects of non-instantaneous reionization on the two principal inflationary parameters (the scalar spectral index of primordial perturbations $n_s$ and the tensor-to-scalar ratio parameter $r$), and on the optical depth $\tau_{e}$.
The method used in the above cited works to describe a general reionization scenario, developed in Ref. \cite{mortonson}, is based on a principal components (PC) analysis of the reionization history, $x_e(z)$\ . PCs provide a complete basis for describing the effects of reionization on large-scale $E$-mode polarization spectrum.
Following Ref.\ \cite{mortonson}, one can treat $x_e(z)$\ as a free function of redshift by decomposing it into its principal
components:
\begin{equation}
x_e(z)=x_e^f(z)+\sum_{\mu}m_{\mu}S_{\mu}(z),
\label{eq:xez}\end{equation}
where the principal components $S_{\mu}(z)$ are the eigenfunctions of the Fisher matrix describing the dependence of the polarization spectra on $x_e(z)$; the $m_{\mu}$ are the PC amplitudes for a particular reionization history, and $x_e^f(z)$ is the WMAP {\it fiducial } model for which the Fisher matrix is computed and from which the PCs are obtained.
Therefore the amplitude of eigenmode $\mu$ for a perturbation around the fiducial reionization history
$\delta x_e(z)\equiv x_{e}(z)-x_{e}^{f}(z)$ is
\begin{equation}
m_{\mu}=\frac{1}{z_{max}-z_{min}}
\int _{z_{min}}^{z_{max}} dz~S_{\mu}(z)\delta x_{e}(z).
\label{eq:xetommu}
\end{equation}
In Refs. \cite{Pandolfi1} and \cite{Pandolfi2} we made use of the publicly available $S_{\mu}(z)$ functions and varied the amplitudes $m_{\mu}$ for the first five eigenfunctions (i.e. for $\mu=1,...,5$). The principal components were computed only in the range of redshifts $z\in[6-30]$.
In what follows we refer to this parametrization of reionization as the ``Principal Components''(PC) reionization.
Since the ionization fraction is bounded in $0 < x_e(z) < 1$ (neglecting helium reionization and the small residual ionized fraction after recombination) in the range of redshifts in which PCs are defined, it is necessary to impose some limits on the amplitudes of the eigenmodes of equation (\ref{eq:xetommu}) to let the reionization fraction be within these limits, if only for the definition of reionization fraction. In Ref. \cite{mortonson} the authors find the ranges of values for the amplitudes $m_{\mu}$ compatible with $x_e(z) \in[0, 1]$ for all the redshifts in range of interest. In \cite{Pandolfi1} and \cite{Pandolfi2} we performed a Monte Carlo Markov Chains analysis assuming a flat prior on (only) the ranges of values of the amplitudes $m_{\mu}$ whose linear combination with the function $S_{\mu}$ give a $x_e(z)$\ in the allowed range. These values are reported in left part of Table \ref{tab:TableIV} and are labeled ``PC Bounds''.
However, these limits for the values of the PC amplitudes are a necessary but not sufficient condition for the reionization fraction to lie in $0<x_e(z)<1$. In fact, as noticed also by \cite{mortonson}, if any $m_{\mu}$ violates those bounds $x_e(z)$\ is guaranteed to be unphysical in some redshift range, but the opposite is not true, because the full reionization history depends on the linear combinations of the product of the amplitudes times their corresponding PC principal component. Indeed, even if all the amplitudes $m_{\mu}$ satisfy the bounds reported in Table \ref{tab:TableIV}, $x_e(z)$ could assume an unphysical value for some redshifts.
To overcome this potential problem, we have added in the version of the \texttt{cosmomc} package used in \cite{Pandolfi1} and \cite{Pandolfi2} the condition that the value of $x_e(z)$ computed at each step of a Markov Chain must be in the range $0 < x_e(z) < 1$ for every $z$. In these studies, this was the only ``physicality'' condition imposed on the possible reionization history. However, experimental data gathered in the last few years can be used to discard at least some of the possible $x_e(z)$\ histories on well understood (astro)physical grounds. It is now possible to use reionization histories that are physically motivated and tested with known probes of the reionization epoch, such as the Gunn-Peterson optical depth, or the distribution in redshift of the Ly$\alpha$ emitters.
In this work we adopt the results of a well-tested semi-analytical reionization model proposed in Refs. \cite{CF2005} and \cite{CF2006} (in what follows we will refer to this model as the CF model). This model takes into account a large number of parameters and physical processes that are involved in modeling reionization, including (e.g.) the radiative and chemical feedbacks of the first sources of ionizing light on the evolution of the intergalactic medium (IGM), and constrain the model by comparing it with a variety of observational data, such as the redshift evolution of Lyman Limit Systems (LLS), the IGM temperature and the cosmic star formation density. Thus we will be able to build up an ensemble of reionization histories that is more robust from both the theoretical and the observational point of view, rather then rely on purely phenomenological, albeit model-independent, parameterization schemes as the PCs.
We will combine the CF model with a standard $\Lambda$CDM cosmological model and we perform a Monte Carlo Markov Chains analysis of the joint CMB and reionization data. We will thus be able to test the impact of considering a detailed physical model for reionization on the constraints of the cosmological parameters, and conversely to test the dependence of the CF model on the underlying cosmological model.
At the end of such analysis we will moreover derive the subsequent constraints on the amplitudes of the reionization principal components $m_{\mu}$ (applying directly the equation \ref{eq:xetommu}). By construction then, these limits on the values of amplitudes of the principal components will be compatible and constrained both by the CMB and by the astrophysical probes of the reionization process.
The main objectives of the present work are then:
\begin{itemize}
\item Verify the impact of considering a data-constrained and realistic reionization model on the determination cosmological parameters.
\item Verify the impact on the constraints of the reionization parameters produced by variations of the cosmological parameters, i.e. refraining from fixing them a priori from the most updated best fit values of the WMAP experiment.
\item Obtain the PC amplitudes $m_{\mu}$ from the allowed reionization histories.
\end{itemize}
As such an analysis with combined cosmological parameters characterizing the background evolution of the universe and astrophysical parameters modeling the reionization history has not yet been made, it is worthwhile to explore their mutual implications on the extraction of the constraints of the two ensemble of parameters.
\begin{table*}[htbp]
\begin{tabular}{l|c|c}
Parameter &\ \ Mean\ \ &\ $95\%$\ c.l.\ limits\ \\
\hline
\hline
$\Omega_m$ & 0.2733 &[0.2260, 0.3305]\\
$\Omega_bh^2$ & 0.2184 &[0.0208, 0.0229]\\
$h$ & 0.6984 &[0.6553, 0.7422]\\
$n_s$ & 0.9579 &[0.9330, 0.9838]\\
$\sigma_8$ & 0.7941 &[0.7434, 0.8491]\\
\hline
$\epsilon_{II}$ & 0.0037 &[0.0016, 0.0067]\\
$\epsilon_{III}$ & 0.0165 &[0.0000, 0.0398]\\
$\lambda_{0}$ & 3.0152 &[1.0000, 5.1739]\\
\hline
$\tau_{e}$ & 0.0803 & [0.0625, 0.1042]\\
$z_r$ &6.7469 & [5.8563, 8.2000]\\
\end{tabular}
\caption{Mean and 95\% c.l. constraints on the cosmological, astrophysical and derived parameters obtained with the reionization parametrized with the CF model of reionization.}
\label{tab:TableI}
\end{table*}
\section{Analysis}
\label{sec:anal}
The details of the CF model are summarized in Ref. \cite{MCF2010}; in the present work we assume the following settings:
\begin{itemize}
\item We consider here a flat $\Lambda$CDM cosmology described by a set of cosmological parameters:
\begin{equation}
\label{parameter}
\{\Omega_m,\Omega_bh^2, h, \sigma_8, n_s\},
\end{equation}
where $\Omega_m$ is the total matter density relative to the critical density, $\Omega_bh^{2}$ is the baryonic matter density, $h$ is the reduced Hubble parameter $H_0=100 h$, $\sigma_8$ is the r.m.s. density fluctuation in spheres of radius $8h^{-1}$ Mpc and $n_s$ is the scalar spectral index of primordial perturbations. We want to stress that these cosmological parameters are considered here as free parameters, so that they are not assumed a priori, as in \cite{MCF2010}.
\item The CF reionization model contains additional three free parameters. These are
$\epsilon_{\rm II, III} = [\epsilon_* f_{\rm esc}]_{\rm II,III}$, the product of the star-forming efficiency (fraction of baryons within collapsed haloes going into stars) $\epsilon_*$ and the fraction of photons escaping into the IGM $f_{\rm esc}$ for PopII and PopIII stars;
the normalization $\lambda_0$ of the ionizing photons mean free path (see Ref. \cite{MCF2010} for details). In what follows we refer to these three parameters as the ``astrophysical'' parameters, to distinguish them from the five ``cosmological'' ones described described in the previous point.
\item The ranges of variation adopted for the three free astrophysical parameters are $\epsilon_{\rm II} \in [0; 0.02]$, $\epsilon_{\rm III} \in [0; 0.1]$, $\lambda_0 \in [1;10]$.
\item The observational data used to compute the likelihood analysis are (i) the photo-ionization rates $\Gamma_{\rm PI}$ obtained using Ly$\alpha$ forest Gunn-Peterson optical depth observations and a large set of hydrodynamical simulations \cite{Bolton} and (ii) the redshift distribution of LLS ${\rm d} N_{\rm LL}/{\rm d} z$ in the redshift range of $0.36 < z < 6$ \cite{Songaila}.
The data points are obtained using a large sample of QSO spectra. For details, see Ref.\ \cite{MCF2011}.
\item In order to make the analysis self-consistent, the WMAP7 constraint on the total electron scattering optical depth $\tau_{e}$ is not considered in this analysis. This prevents a possible loophole in our analysis: WMAP7 constraints on $\tau_{e}$ have been obtained using the assumption of instantaneous reionization at $z=z_r$. Once this idealized evolution of $x_e(z)$\ is dropped (this paper), the value of $\tau_{e}$ must be a byproduct of the new analysis rather than being inserted artificially as an external constraint into it. Moreover, as already pointed out in Ref. \cite{MCF2010}, the CMB polarization spectra are sensitive to the shape of the reionization history and considering a more general reionization scenario could lead to a tighter optical depth constraint than derived by WMAP7 \cite{Pandolfi1}.
\item Finally, we impose the prior that reionization should be completed by $z=5.8$ to match the flux data of Ly$\alpha$ and Ly$\beta$ forest.
\end{itemize}
With these hypotheses we have then modified the Boltzmann CAMB code \cite{camb} to incorporate the CF model and performed a MCMC analysis based on an adapted version of the public available MCMC package \textsc{cosmomc} \cite{Lewis:2002ah}. Our basic data set is the seven--yr WMAP data \cite{wmap7} (temperature and polarization), on top of which we add two ``astrophysical'' datasets, i.e. the LLS redshift evolution, $dN_{LL}/dz$ Ref. \cite{Songaila}, and the Gunn-Peterson optical depth measurements presented in Ref. \cite{Bolton}. To extract the constraints on free parameters from such combined data set we consider a total likelihood function $L \propto \exp (-\mathcal{L})$ made up by two parts:
\begin{equation}
{\mathcal L} = \frac{1}{2}\sum_{\alpha=1}^{N_{\rm obs}} \left[
\frac{{\cal J}_{\alpha}^{\rm obs} - {\cal J}_{\alpha}^{\rm th}}{\sigma_{\alpha}}
\right]^2 + {\cal L'}
\end{equation}
where $\mathcal{L}'$ refers to the WMAP7 likelihood function and is computed using the routine supplied by the WMAP team; ${\cal J}_{\alpha}$ represents the set of $N_{\rm obs}$ observational points referring to Gunn-Peterson optical depth LLS distribution data; finally, $\sigma_{\alpha}$ are the corresponding observational error-bars. We constrain the free parameters by maximizing ${\mathcal L}$ with flat priors on the allowed parameter ranges and the aforementioned prior on the end of reionization at $z=5.8$.
The Monte Carlo-Markov Chain convergence diagnostics are done on 4 chains applying the Gelman and Rubin ``variance of chain mean''/``mean of chain variances'' $R$
statistic for each parameter. We considered the chains to be converged at $R-1<0.03$.
\begin{table*}[htbp]
\begin{tabular}{l|c|c|c}
Parameter &\ \ WMAP7\ \ &\ \ WMAP7 + PC &\ WMAP7\ +\ ASTRO \\
\hline
\hline
$\Omega_m$ & $0.266\pm0.029$ & $0.243\pm0.032$ &$0.273\pm0.027$ \\
$\Omega_bh^2$ & $0.02258^{+0.00057}_{-0.00056}$ & $0.02321\pm0.00076$ & $0.02183\pm 0.00054$ \\
$h$ & $0.710\pm0.025$ & $0.735\pm 0.033$ & $0.698\pm0.023$ \\
$n_s$ & $0.963\pm0.014$ & $0.994\pm0.023$ &$0.958\pm 0.013$ \\
$\sigma_8$ & $0.801\pm0.030$ & ----- &$0.794\pm 0.027$ \\
\hline
$\tau_{e}$ & $0.088\pm0.015$ & $0.093\pm0.010$ & $0.080 \pm 0.012$ \\
$z_r$* & $10.5\pm 1.2$ & ----- & $6.7\pm 0.6$ \\
\end{tabular}
\caption{Comparison of the 68\% c.l. posterior probabilty contraints obtained for different parametrizations of reionization.\\
* The $z_r$ parameter has a different definition in the different reionization scenarios (see text for details).}
\label{tab:TableII}
\end{table*}
\begin{table*}[htbp]
\begin{tabular}{l|c|c|c|c}
Parameter &WMAP7 + ASTRO Mean & WMAP7 + ASTRO $95\%$ c.l. limits & CF Mean & CF $95\%$\ c.l.\ limits\ \\
\hline
\hline
$\epsilon_{II}$ & 0.0037 &[0.0016, 0.0067] & 0.003 &[0.001, 0.005]\\
$\epsilon_{III}$ & 0.0165 &[0.0000, 0.0398] & 0.020 &[0.0000, 0.043] \\
$\lambda_{0}$ & 3.0152 &[1.0000, 5.1739] & 5.310 &[2.317, 9.474]\\
\hline
$\tau_{el}$ & 0.0803 & [0.0625, 0.1042]& $\equiv0.088\pm0.015$ & $\equiv0.088\pm0.015$ \\
$z_r$ & 6.7469 &[5.8563, 8.2000]& 6.762 & [5.800, 7.819]\\
\end{tabular}
\caption{Comparison between the mean value and the 95\% c.l posterior constraints between the present work (WMAP7 + ASTRO) and the CF model, Ref. \cite{MCF2010} (MCF). }
\label{tab:TableIII}
\end{table*}
\section{Results}
\label{Res}
The results of the MCMC analysis described above are summarized in Table \ref{tab:TableI}, where we list the marginalized posterior probabilities at 95\% confidence level (c.l.) errors on the free cosmological and astrophysical parameters. We also report the constraints for two derived parameters: the electron scattering optical depth $\tau_{e}$ and reionization redshift $z_r$, to be intended as the redshift at which the reionization is 99\% complete.
In Table \ref{tab:TableII} we show the 68\% c.l. constraints obtained by the WMAP team for the standard 6-parameter $\Lambda$CDM model (``WMAP7'') and the constraints obtained on the cosmological parameters from the present analysis (``WMAP7 + ASTRO'').
As we can see from the Table \ref{tab:TableII} the results of our work mildly differ from the WMAP7 results for the parameters of the standard $\Lambda$CDM model. The most sensitive parameter for the presence of the ``astrophysical'' datasets (LLS and Gunn-Peterson data) is $\Omega_bh^2$ whose mean values in the two cases differ by more than a standard deviation from each other. It is important to note that even when considering a complex reionization history implying three new parameters the errors remain practically the same as in the standard case.
Table \ref{tab:TableII} reports the results obtained in \cite{Pandolfi1} for the WMAP7 dataset with the PC reionization (``WMAP7 + PC'').
This method produces two main differences with respect to the WMAP7 + ASTRO case: the first is related to the constraints obtained for $n_s$. In \cite{Pandolfi1} the constraints for the scalar spectral index were compatible with $n_s=1$, i.e. the Harrison-Zel'dovich (HZ) primordial power spectrum, when instead WMAP7+ASTRO excludes the value $n_s=1$ at $>3\sigma$. The second difference concerns $\tau_{e}$ in the two cases: for WMAP7 + PC this quantity is in the range $\tau_{e}=0.093\pm0.010$, while the WMAP+ASTRO case gives a mean value lower by $>1-\sigma$, i.e. $\tau_{e}=0.080\pm0.012$.
Note that in the WMAP7 + PC case we did not consider constraints on the $\sigma_8$ parameter, so in Table \ref{tab:TableII} the corresponding value is missing.
There is a caveat in comparing the constraints obtained on $z_r$. Indeed, in the WMAP7 case $z_r$ is the redshift at which the universe undergoes an instantaneous and complete reionization process. In the more realistic, extended reionization scenarios considered here instead, $z_r$ is defined as the redshift at which the IGM is 99\% re-ionized by volume. With this clarification in mind, WMAP7+ASTRO results predict $5.8 <z_r < 8.2$ at 95\% c.l. (see Table \ref{tab:TableI}).
In Table \ref{tab:TableIII} we report the 95\% c.l. posterior probability constraints for the reionization parameters $\epsilon_{II}$, $\epsilon_{III}$ and $\lambda_0$ obtained in the present work (WMAP7 + ASTRO case, cosmological parameters free to vary) compared to those obtained in Ref. \cite{MCF2010} in which the cosmological parameters were fixed to the WMAP7 best fit values (CF case). Figure \ref{bf} shows the comparison between the best-fit model for the $x_e(z)$\ evolution for the two cases of WMAP7 + ASTRO and CF. For the WMAP7 + ASTRO case, full hydrogen reionization is not only achieved earlier than in the CF model, but the evolution is faster, resulting in an initially
lower $x_e(z)$\ above $z=8$. These differences are entirely induced by the fact that we have now allowed the cosmological parameters to vary together with the astrophysical ones, but they are relatively small. The fact that the astrophysical parameters do not show much dependence on cosmology is understandable because the cosmological parameters affect the reionization process mostly through structure formation. The next obvious step is to include large scale structure information in the analysis. In conclusion, including astrophysical datasets in the analysis seems to lead to relatively important effects on the extraction of the cosmological parameters.
\begin{figure}[htbp]
\centering
\includegraphics[width=6cm, angle=-90]{wmap7_astro_cf.eps}
\caption{Ionization histories for the best-fit model for the two cases WMAP7+ASTRO (red dotted solid curve) and CF (green solid curve) \cite{MCF2010}.}
\label{bf}
\end{figure}
\subsection{PC amplitude reconstruction}
\begin{table*}[htbp]
\begin{tabular}{lcc}
Parameter & PC Bounds & Astrophysical Bounds\\
\hline
\hline
$m_{1}$ &$[-0.1236, 0.7003]$& $[-0.1229, -0.0866]$ \\
$m_{2}$ &$[-0.6165, 0.2689]$& $[-0.2594, 0.0002]$\\
$m_{3}$ &$[-0.3713, 0.5179]$& $[0.0763,\ 0.2941]$\\
$m_{4}$ &$[-0.4729, 0.3817]$& $[-0.2107, -0.1080]$\\
$m_{5}$ &$[-0.3854, 0.4257]$& $[0.0418,\ 0.1319]$\\
\hline
\end{tabular}
\caption{Ranges of variation for the amplitudes of the principal component, in the case of the Principal Components and in the case of the 99\% c.l. reconstructed amplitudes of the present analysis (see text for details).}
\label{tab:TableIV}
\end{table*}
For each reionization history allowed by the MCMC likelihood analysis, we use eq. (\ref{eq:xetommu}) to reconstruct the amplitudes of the first five PC amplitudes, $m_{\mu}$, with $\mu=1...5$. By construction now, the amplitudes $m_{\mu}$ not only fulfill the necessary physicality conditions (see Sec. 1) but also they are compatible with the additional astrophysical data sets considered in this analysis, i.e. the Ly$\alpha$ Gunn-Peterson test and the LLS redshift distribution.
\begin{figure}[htbp]
\centering
\includegraphics[width=5cm]{m1_m2_thin_99.eps}
\includegraphics[width=5cm]{m2_m3_thin_99.eps}
\includegraphics[width=5cm]{m3_m4_thin_99.eps}
\includegraphics[width=5cm]{m4_m5_thin_99.eps}
\caption{68\% and 99\% reconstructed c..l. constraints for the values of the PC amplitudes computed from CF model and eq. \ref{eq:xetommu} (top layer, pink). Background contours refer to 68\% and 95\% c.l. constraints obtained in \cite{Pandolfi1} with the PC reionization for WMAP7 (bottom layer, blue), WMAP7+QUAD+ACBAR+BICEP (CMBAll, next layer up, red), CMB All + LRG-7 (next layer, green) and simulated Planck data (next layer, yellow), respectively.}
\label{m_i}
\end{figure}
In Fig. \ref{m_i} we show the two dimensional 68\% and 99\% c.l constraints for the amplitudes $m_{\mu}$ obtained here compared with those obtained in \cite{Pandolfi1} for which we show the two dimensional 68\% and 95\% c.l. distributions for each of the cases considered.
We choose to report the 99\% c.l. instead of the usual 95\% c.l. limits to be as conservative as possible in showing the reionization histories allowed by the MCMC likelihood analysis. The color (layer) code is the following: in pink (top layer) there is the case WMAP7 + ASTRO considered in the present work. In the background there are the cases considered in \cite{Pandolfi1}: in blue is the WMAP7 case (bottom layer) , in red (next layer up) is the case called ``CMB All'' ( i.e. WMAP7 + ACBAR + BICEP+ QUAD + BOOMERanG), green (next layer) is CMB All + LRG-7 and yellow (next layer) is simulated Planck data. \cite{Pandolfi1} considered an ensable of CMB dataset along with WMAP7, and also we forecasted future constraints from the Planck experiment, simulating a set of mock data with a fiducial model given by the best fit WMAP5 model with the following experimental noise:
\begin{equation}
N_{\ell} = \left(\frac{w^{-1/2}}{\mu{\rm K\mbox{-}rad}}\right)^2
\exp\left[\frac{\ell(\ell+1)(\theta_{\rm FWHM}/{\rm rad})^2}{8\ln 2}\right],
\end{equation}
\noindent where $w^{-1/2}$ is the temperature noise level (a factor $\sqrt{2}$ larger for polarization noise) and $\theta$ is the beam size.
For the Planck mission we use $w^{1/2}=58 \mu K$ and $\theta_{\rm FWHM}=7.1'$ equivalent to expected sensitivity of the $143$ GHz channel.
The region spanned by PC amplitude values is much smaller than that allowed by when the PC bounds only are imposed. The 99\% c.l. constraints values are reported in the right part of the Table (\ref{tab:TableIV}) (``Astrophysical Bounds''). As seen from Table (\ref{tab:TableIV}) the amplitudes of all the principal components (except for $m_2$) obtained with the above procedure are constrained at 99\% c.l. to take a definite sign, negative for $m_1$ and $m_4$ and positive for $m_3$ and $m_5$. Moreover, even if the 99\% c.l. upper bound of $m_2$ is positive, this second amplitude is mostly constrained to be always negative. These results are in qualitative agreement with \cite{Pandolfi1}, who also found that the same amplitude signature, albeit with errors large enough that the 95\% c.l. bounds encompass values of both possibile signs.
\section{CONCLUSIONS}
\label{sec:concl}
With the aim of constraining the evolution of cosmic reionization, we have extended previous work based on the use
of Principal Components analysis. The main novelty of the present work is represented on one hand by complementing
available CMB data with additional astrophysical results from quasar absorption line experiments, as the Gunn-Peterson
test and the redshift evolution of Lyman Limit Systems. In addition, we have for the first time explored the effects of a
joint variation of both the cosmological ($\Omega_m,\Omega_bh^2, h, \sigma_8, n_s$) and astrophysical ($\epsilon_{\rm II},
\epsilon_{\rm III}, \lambda_0$, see Sec \ref{sec:anal} for their physical meaning) parameters. Note that, differently from the vastly
used approach in the literature, we do not impose a priori any bound on the electron scattering optical depth $\tau_e$,
which instead we calculate a posteriori. This is to prevent a possible loophole in the calculation, as the WMAP determination
of such quantity is based on the assumption of an instantaneous reionization which we do not make here.
Including a realistic (i.e physically motivated) reionization history in the analysis induces mild changes in the
cosmological parameter values deduced through a standard WMAP7 analysis. Particularly noteworthy are the
variations in $\Omega_bh^2 = 0.02258^{+0.00057}_{-0.00056}$ (WMAP7) vs. $\Omega_bh^2 = 0.02183\pm 0.00054$ (WMAP7 + ASTRO),
and the new constraints for the scalar spectral index, for which WMAP7+ASTRO excludes the Harrison-Zel'dovich value $n_s=1$ at $>3\sigma$.
Finally, the e.s. optical depth values is considerably decreased with respect to the standard WMAP7, i.e. $\tau_{e}=0.080\pm0.012$.
We conclude that inclusion of astrophysical datasets, allowing to robustly constrain the reionization history, in the extraction procedure
of cosmological parameters leads to relatively important differences in the final determination of their values.
\section*{AKNOWLEDGMENTS}
SP is grateful for the hospitality and support during his research at Scuola Normale Superiore in Pisa in
March-July 2011. Support was given by the Italian Space Agency through the ASI contracts Euclid-IC (I/031/10/0).
|
\section*{Supporting Online Material}
{\Large \bf A Reservoir of Ionized Gas in the Galactic Halo to Sustain Star Formation in the Milky Way}
\\
\vspace{0.3 cm}
{\Large Nicolas Lehner, J. Christopher Howk} \\
\vspace{0.3 cm}
\normalsize{Department of Physics, University of Notre Dame,}\\
\normalsize{225 Nieuwland Science Hall, Notre Dame, IN 46556, USA}\\
\end{center}
\paragraph*{Observations and Analysis.}
Our COS and STIS program consists of $23$ early-type and post-asymptotic giant branch (PAGB) field stars as well as globular cluster (GC) stars obtained from our Cycle $17$ {\em HST}\ program (\#$11592$). The criteria for assembling this stellar sample was that the stars are at $|z|\ga 3 $ kpc and are bright enough to be observed at a rate of $1$ {\em HST}\ orbit per star. We also searched the literature and the MAST archive for any additional stars at $|z|\ga 3 $ kpc observed with STIS or COS, which added $5$ stars to the sample. (We note that although BD+38$^\circ$2182 is at $z=+3.2$ kpc, this sight line was initially targeted for the HVC Complex M, and hence is not included here.) In Table~S\ref{t-star}, we list the distance and positions of the $28$ stars. Stars associated with GCs based on their location and velocity were placed at distances of the hosting GC. For all the field stars, distances are spectroscopic parallaxes based on models of stellar atmospheres, which produce stellar distances accurate to $\sim$\,$20\%$. The distances were derived by \cite{ramspeck01} and \cite{smoker03} and references therein.
Sixteen stars were observed with COS with gratings G130M and G160M (resolution $R\simeq 17,000$), and the 5 brightest stars of our sample were obtained with STIS with the echelle E140M mode ($R\simeq 45,800$) using the $0.2\arcsec$ square aperture to maximize throughput. Although the COS resolution is $\sim 2.7$ lower than the STIS resolution, it was adequate for searching for and detecting HVC absorption since the high-velocity absorption is sufficiently separated from the lower velocity Galactic gas. The wavelength ranges covered by COS and STIS observations are $1134$ to $1796$ \AA\ and $1170$ to $1730$ \AA, respectively. However, the three faintest stars (PG1002+506, PG0914+001, and NGC104-UIT14) were only observed with the COS G160M giving wavelength coverage $1382\le \lambda \le 1747$ \AA. In these wavelength intervals, some or all of the following atomic and ionic species are available: O$\;${\small\rm I}\relax\ $\lambda$1302, C$\;${\small\rm II}\relax\ $\lambda$$1334$, C$\;${\small\rm IV}\relax\ $\lambda\lambda$$1548, 1550$, S$\;${\small\rm II}\relax\ $\lambda\lambda$$1250, 1253, 1259$, Si$\;${\small\rm II}\relax\ $\lambda$$1190, 1193, 1260, 1304, 1526$, Si$\;${\small\rm III}\relax\ $\lambda$$1206$, Si$\;${\small\rm IV}\relax\ $\lambda\lambda$$1393, 1402$, Al$\;${\small\rm II}\relax\ $\lambda$$1670$, Fe$\;${\small\rm II}\relax\ $\lambda$1608. All these atomic and ionic species were used to search high-velocity absorption components in the COS and STIS data.
Information about the COS instrument and COS spectra can be found in \cite{froning09,dixon10}. Information about STIS can be found in the STIS {\it HST}\ Instrument Handbook \cite{dressel07}. Standard reduction and calibration procedures were performed using the reference files available as of October 2010 or later. The proper alignment of the individual spectra was achieved through a cross-correlation technique. Data from individual exposures, including all grating configurations for COS and echelle orders for STIS, were combined to produce a single spectrum.
\begin{table}[!t]
\begin{minipage}{15 truecm}
\caption{Stellar Sample \label{t-star}}
\footnotesize
\begin{tabular}{lccccccccc}
\hline
\hline
{Name} & {$l$}& {$b$} &
$d$ & $z$ & $v \sin i$& $v^\star_{\rm LSR}$& $v^{\rm HVC}_{\rm LSR}$ &$Q$ & Note\\
& ($^\circ$)& ($^\circ$) & (kpc)& (kpc) & (${\rm km\,s}^{-1}$) & (${\rm km\,s}^{-1}$) & (${\rm km\,s}^{-1}$) & & \\
(1) & (2)& (3) & (4)& (5) & (6) & (7) & (8) & (9) & \\
\hline
NGC6723-III60 & 0 & $ -17 $ & 8.7 & $ -2.6 $ & $\cdots$ & $ -101 $ & $ -90 $ & 0 & \\
NGC5904-ZNG1 & 4 & $ +47 $ & 7.2 & $ +5.3 $ & $\cdots$ & $ +65 $ & $ -140,-120 $ & 1 & $a$ \\
PG1708+142 & 35 & $ +29 $ & 21.0 & $ +10.0 $ & $\cdots$ & $ -28 $ & No & 1 & \\
PG1704+222 & 43 & $ +32 $ & 6.9 & $ +3.6 $ & $\cdots$ & $ -22 $ & No & 1 & \\
PG1610+239 & 41 & $ +45 $ & 8.4 & $ +5.9 $ & 75 & $ +109 $ & No\,$:$ & 0 & \\
PHL346 & 41 & $ -58 $ & 8.7 & $ -7.4 $ & 45 & $ +66 $ & No & 0 & \\
vZ\,1128 & 43 & $ +79 $ & 10.2 & $ +10.0 $ & $\cdots$ & $ -137 $ & No & 0 & $a$ \\
NGC6205-Barnard29& 59 & $ +41 $ & 7.7 & $ +5.04 $ & $\cdots$ & $ -227 $ & $ -107 $ & 1 & $a$ \\
PG1511+367 & 59 & $ +59 $ & 3.8 & $ +3.2 $ & 75 & $ +118 $ & No & 1 & \\
NGC6341-326 & 68 & $ +35 $ & 8.2 & $ +4.7 $ & $\cdots$ & $ -101 $ & $ -94: $ & 0 & \\
PG2219+094 & 73 & $ -40 $ & 6.6 & $ -4.2 $ & 225 & $ -17 $ & No & 1 & \\
HS1914+7139 & 103 & $ +24 $ & 14.9 & $ +6.0 $ & 250 & $ -25 $ & $ -165,-112$ & 1 & $b,c$ \\
PG0009+036 & 105 & $ -58 $ & 10.8 & $ -9.1 $ & 350 & $ +150 $ & No & 1 & \\
PG0122+214 & 133 & $ -41 $ & 9.6 & $ -6.2 $ & 117 & $ +24 $ & $ -160,-91 $ & 1 & \\
PG0832+675 & 148 & $ +35 $ & 7.5 & $ +4.3 $ & $\cdots$ & $ -70 $ & $ -123 $ & 1 & $b,d$ \\
PG1002+506 & 165 & $ +51 $ & 13.9 & $ +10.8 $ & 350 & $ 0 $ & $ -102,+101 $ & 1 & \\
HD233622 & 168 & $ +44 $ & 4.7 & $ +3.3 $ & 240 & $ +9 $ & No & 1 & $a$ \\
PG0855+294 & 196 & $ +39 $ & 6.5 & $ +4.1 $ & 100 & $ +73 $ & $ +93,+107 $ & 1 & \\
PG0955+291 & 200 & $ +52 $ & 5.5 & $ +4.3 $ & 190 & $ +72 $ & No & 1 & \\
PG1243+275 & 207 & $ +89 $ & 6.2 & $ +6.2 $ & $\cdots$ & $ +107 $ & $ (+105) $ & 0 & $e$ \\
PG0934+145 & 219 & $ +43 $ & 8.5 & $ +5.8 $ & 50: & $ +105 $ & $ (+100) $ & 0 & $f$ \\
PG0914+001 & 232 & $ +32 $ & 16.0 & $ +8.4 $ & 350 & $ +80 $ & $ +100,+170 $ & 1 & \\
EC10500-1358 & 264 & $ +40 $ & 5.2 & $ +3.3 $ & 100 & $ +84 $ & $ +97 $ & 0 & \\
SB357 & 301 & $ -81 $ & 7.9 & $ -7.8 $ & 180 & $ +52 $ & No & 1 & \\
NGC104-UIT14 & 306 & $ -45 $ & 4.5 & $ -3.2 $ & $\cdots$ & $ +55 $ & No & 1 & $b,g$ \\
PG1323-086 & 317 & $ +53 $ & 15.8 & $ +12.6 $ & $\cdots$ & $ -41 $ & $ -91 $ & 1 & \\
NGC5824-ZNG1 & 333 & $ +22 $ & 32.0 & $ +12.0 $ & $\cdots$ & $ -20 $ & $ -160: $ & 0 & \\
HD121968 & 334 & $ +56 $ & 3.8 & $ +3.1 $ & 160 & $ +137 $ & No & 1 & $a$ \\
\hline
\end{tabular}
(1): Name of the background star. \\
(2) and (3): Galactic longitude and latitude of the star. \\
(4) and (5): Distance and vertical $z$-height of the star. The stellar distances are accurate to $\sim$\,20\% \cite{ramspeck01,smoker03}. \\
(6): Projected rotational velocity of the star. If no value is listed, the star is a PAGB or GC star. \\
(7) and (8): Radial velocity of the star and the HVC if present.
The velocities are averaged over multiple transitions.
The uncertainties in absolute velocity scales are $\sim$10 ${\rm km\,s}^{-1}$\ for COS and $\le 3$ ${\rm km\,s}^{-1}$\ for STIS, except where there is a colon. A colon means the result is tentative in view of the low signal-to-noise of the data. \\
(9): Quality flag; $Q=0$ means either low S/N ratio ($< 15$) or a possible stellar contamination. \\
{\sc Notes ---} $a$: Not initially in our {\it HST}\ Cycle 17 program.
$b$: HVC H$\;${\scriptsize \rm I}\relax\ emission is observed.
$c$: The H$\;${\scriptsize \rm I}\relax\ emission seen toward HS1914+7139 at $-118$ ${\rm km\,s}^{-1}$\ is associated with the Outer Spiral Arm, but there is also a second HVC along this line of sight with no H$\;${\scriptsize \rm I}\relax\ emission ({\it 18}).
$d$: For PG0832+675, the HVC velocity in absorption does not coincide with the H$\;${\scriptsize \rm I}\relax\ emission velocity, and therefore we assume that they are not directly related.
$e$: high-velocity absorption is observed but it is very likely stellar.
$f$: there are possibly both stellar and interstellar high-velocity components at the same velocity.
$g$: Toward NGC104-UIT14, we did not find any evidence of HVC in absorption while it is seen in emission at about $+150$ ${\rm km\,s}^{-1}$. This suggests that this HVC seen in H$\;${\scriptsize \rm I}\relax\ emission is at $d>4$ kpc ($z<-3.5$ kpc) or that the pencil-like beam through the COS aperture did not cover the emission seen in the much larger H$\;${\scriptsize \rm I}\relax\ beam.
\end{minipage}
\end{table}
\normalsize
\clearpage
One concern with the use of stars as background continuum sources is the possible stellar contamination. In particular, the PAGB and evolved GC stars may have narrow atmospheric lines that can mimic interstellar absorption. We minimize this issue by requiring interstellar HVC absorption to be present in multiple transitions and by using synthetic spectra derived from model stellar atmospheres to identify possible contaminating stellar lines. The most metal-rich evolved stars in our samples are NGC6723-III60, PG0943+145, and PG1243+275. For these stars, model atmospheres were calculated by Dr. P. Chayer in a manner similar to that described in \cite{sonneborn02} using the publicly available TLUSTY and SYNSPEC codes of Drs. I. Hubeny and T. Lanz (http://nova.astro.umd.edu/). The models used the stellar atmosphere parameters previously derived by \cite{moehler98,rolleston99}. The stellar contamination is, however, less of a concern for the rapidly rotating B-type stars in our sample (see the large projected rotational velocity $v \sin i$ in Table~S\ref{t-star}), except for PHL346, a relatively slowly rotating and extremely metal rich star. Yet, even in that case it is possible to rule out the presence of any HVC in this sightline following our methods. We are therefore confident that the detected HVCs are not stellar, except possibly for two stars (see Table~S\ref{t-star}). We also emphasize that the HVC absorption is found blueshifted and redshifted relative to the star velocity. Thus the HVCs are not circumstellar material (see, also, the ionization model and other arguments in {\it 17}).
From the whole sample of stars, we also built a sample ($Q=1$, see Table~S\ref{t-star}) in which the S/N ratio is more uniform with S/N\,$\ge 15$, i.e., a sensitivity $W_\lambda \ge 15$ m\AA\ near Si$\;${\small\rm II}\relax\ $\lambda$1526. We also require no stellar contamination for the $Q=1$ sample, i.e., any star with some detected narrow photospheric features at $90 \le |v^\star_{\rm LSR}|\la 170$ ${\rm km\,s}^{-1}$\ that correspond to the same species used to search HVC absorption is rejected from the $Q=1$ sample (e.g., vZ\,1128).
Finally, the signal-to-noise in the spectra of $3$ stars (PG1610+239, NGC6341-326, and NGC5824-ZNG1) is much below than in the spectra of the other stars. We therefore consider the identification or non-detection of the HVCs along these sightlines tentative and mark those with colons in Table~S\ref{t-star} to emphasize this.
\paragraph*{Extragalactic Sample.}
We built our extragalactic sample based on earlier works where the background AGNs were observed with {\it HST}/STIS E140M ({\it 14,15}). By using these works and references therein and analyzing some spectra ourselves, we retrieved or estimated the velocities of the detected HVCs toward AGNs. We summarize the velocities in Table~S\ref{t-extra}, and as in the stellar sample, there are sometimes several HVCs at different velocities along a given sightline. We note that the spectra in the stellar and AGN samples have similar sensitivity, and both samples have a similar size. We also emphasize that as for the stellar sample, we require that a given HVC is detected in absorption in at least two different ions. For this reason we do not confirm all the Si$\;${\small\rm III}\relax\ HVCs listed by ({\it 14}). As for the stellar sample we also searched each sightline for possible HVCs seen in H$\;${\small\rm I}\relax\ $21$-cm emission using the LAB data ({\it 21}); the sightlines with H$\;${\small\rm I}\relax\ emission are listed in Table~S\ref{t-extra}. In Table~S\ref{t-extra}, we also mark the sightlines that pass through the Magellanic Stream.
\begin{table}[!t]
\begin{minipage}{9.7 truecm}
\caption{Extragalactic Sample \label{t-extra}}
\footnotesize
\begin{tabular}{lcccc}
\hline
\hline
Name& $l$& $b$ & $v^{\rm HVC}_{\rm LSR}$ & H$\;${\scriptsize \rm I}\relax\ 21 cm \\
& ($^\circ$) & ($^\circ$) & (${\rm km\,s}^{-1}$) & Emission \\
(1) & (2)& (3) & (4)& (5) \\
\hline
PKS2155-304 & 17 & $ -52 $& $ -130 $ & No \\
NGC5548 & 31 & $ +70 $& No & No \\
Mrk509 & 35 & $ -29 $& $ -284,+139 $ & Yes \\
PG1444+407 & 69 & $ +62 $& No & No \\
PHL1811 & 47 & $ -44 $& $ -208,-159 $ & No \\
NGC7469 & 83 & $ -45 $& $ -355 $ & No$^a$ \\
3C351 & 90 & $ +36 $& $ -185,-118 $ & No \\
UGC12163 & 92 & $ -25 $& $ -245 $ & No$^a$ \\
H1821+643 & 94 & $ +27 $& $ -130 $ & No \\
Mrk876 & 98 & $ +40 $& $ -185,-137 $ & Yes \\
Mrk335 & 108 & $ -41 $& $ -416,-342,-118 $ & No$^a$ \\
Mrk279 & 115 & $ +46 $& $ -145 $ & Yes \\
PG1259+593 & 120 & $ +58 $& $ -125 $ & Yes \\
HS0624+690 & 145 & $ +23 $& $ -108 $ & Yes \\
NGC4051 & 148 & $ +70 $& No & No \\
NGC4151 & 155 & $ +75 $& $ +145 $ & No \\
PG0953+414 & 179 & $ +51 $& $ -145,+125 $ & No \\
TON28 & 200 & $ +53 $& No & No \\
PKS0405-123 & 204 & $ -41 $& No & No \\
PG1116+215 & 223 & $ +68 $& $ +100,+180 $ & No \\
TONS210 & 224 & $ -83 $& $ -243,-167 $ & No \\
HE0226-4410 & 253 & $ -65 $& $ +99,+148,+175,+193 $ & No \\
PG1211+143 & 267 & $ +74 $& $ +174,+191 $ & No \\
PG1216+069 & 281 & $ +68 $& $ +199,+216 $ & No \\
3C273 & 289 & $ +64 $& No & No \\
PKS0312-77 & 293 & $ -37 $& $+97,+160,+210,+295$ & Yes$^{a,b}$ \\
Q1230+0115 & 291 & $ +63 $& $ +103,+284 $ & No \\
NGC4593 & 297 & $ +57 $& $ +98,+258 $ & No \\
PKS1302-102 & 308 & $ +52 $& No & No \\
Mrk1383 & 349 & $ +55 $& No & No \\
\hline
\end{tabular}
(1): Name of the background QSO/AGN. \\
(2) and (3): Galactic longitude and latitude of the QSO/AGN. \\
(4): Radial velocity of the HVC if present. \\
(5): LAB H$\;${\scriptsize \rm I}\relax\ emission present at similar high velocities seen in absorption. \\
{\sc Notes ---} $a$: Sightline passing through the Magellanic Stream. \\
$b$: Sightline passing through the Magellanic Bridge.
\end{minipage}
\end{table}
\normalsize
\paragraph*{Ionization Fraction for the HVCs in the Stellar Sample.}
By comparing the column densities of Si$\;${\small\rm II}\relax\ and O$\;${\small\rm I}\relax, we can estimate the ionization fraction in the HVCs. Indeed O$\;${\small\rm I}\relax\ is an excellent tracer of H$\;${\small\rm I}\relax\ because O$\;${\small\rm I}\relax\ and H$\;${\small\rm I}\relax\ have nearly identical ionization potentials and are strongly coupled through charge exchange reactions. On the other hand, the ionization potential of Si$\;${\small\rm II}\relax\ is larger than that of H$\;${\small\rm I}\relax\ ($13.6$ eV) allowing this ion to be present in both neutral and ionized gas. In Table~S\ref{t-col}, we summarize the column densities of O$\;${\small\rm I}\relax\ and Si$\;${\small\rm II}\relax\ and their ratios corrected for the relative solar abundance for the HVCs where the blending with lower velocity gas or stellar lines is not an issue. The column densities were estimated through $N = \int 3.768\times 10^{14} \ln[F_c(v)/F_{\rm obs}(v)]/(f\lambda) dv$ cm$^{-2}$, where $F_c(v)$ and $F_{\rm obs}(v)$ are the modeled continuum flux and observed flux as a function of velocity, respectively, $f$ is the oscillator strength of the considered atomic transition and $\lambda$ (in \AA) its wavelength \cite{savage91}. When O$\;${\small\rm I}\relax\ is not detected, we estimated a $3\sigma$ upper limit of the column density following \cite{lehner08}. The absorption features are weak enough that saturation is not an issue. For $5$ HVCs, $[{\mbox O$\;${\small\rm I}\relax}/{\mbox Si$\;${\small\rm II}\relax}]$ ratios imply ionization fraction H$\;${\small\rm II}\relax/(H$\;${\small\rm I}\relax$+$H$\;${\small\rm II}\relax)\,$\ge 57\%$--$95\%$. We emphasize that we only consider Si$\;${\small\rm II}\relax, but other higher ionization stages are known to be present as revealed by the detection of Si$\;${\small\rm III}\relax\ and Si$\;${\small\rm IV}\relax. This implies that $[{\mbox O$\;${\small\rm I}\relax}/{\mbox Si$\;${\small\rm II}\relax}]$ gives a strict lower limit on the ionization fraction. For the other $5$ HVCs, we could only estimate $3\sigma$ upper limits based on the non-detection of O$\;${\small\rm I}\relax. They are all consistent with a significant ionization fraction. Therefore the HVCs seen in the foreground stars in our sample can be labeled iHVCs.
\begin{table}[!t]
\begin{minipage}{11.5 truecm}
\caption{Ionization Fraction for the HVCs in the Stellar Sample \label{t-col}}
\footnotesize
\begin{tabular}{lcccc}
\hline
\hline
Name & $v^{\rm HVC}_{\rm LSR}$ & $\log N({\mbox O$\;${\scriptsize \rm I}\relax})$& $\log N({\mbox Si$\;${\scriptsize \rm II}\relax})$ & $[{\mbox O$\;${\scriptsize \rm I}\relax}/{\mbox Si$\;${\scriptsize \rm II}\relax}]$$^a$ \\
& (${\rm km\,s}^{-1}$) & $[{\rm cm}^{-2}]$& $[{\rm cm}^{-2}]$ & \\
\hline
NGC5904-ZNG1 & $ -140 $ & $ 13.12 \pm 0.06 $ & $ 13.32 \pm 0.03 $ & $ -1.38 \pm 0.07 $ \\
NGC5904-ZNG1 & $ -120 $ & $ 13.09 \pm 0.06 $ & $ 13.03 \pm 0.03 $ & $ -1.12 \pm 0.07 $ \\
HS1914+7139 & $ -165 $ & $ 13.83 \pm 0.08 $ & $ 13.10 \pm 0.10 $ & $ -0.45 \pm 0.13 $ \\
PG0122+214 & $ -160 $ & $ <13.44 $ & $ 12.36 \pm 0.05 $ & $ < -0.10 $ \\
PG0122+214 & $ -91 $ & $ 14.17 \pm 0.04 $ & $ 13.36 \pm 0.04 $ & $ -0.37 \pm 0.06 $ \\
PG1002+506 & $ -102 $ & $ <13.34 $ & $ 12.47 \pm 0.12 $ & $ < -0.31 $ \\
PG1002+506 & $ +101 $ & $ <13.41 $ & $ 12.42 \pm 0.15 $ & $ < -0.19 $ \\
PG0855+294 & $ +86 $ & $ <12.98 $ & $ 12.50 \pm 0.04 $ & $ < -0.70 $ \\
PG0855+294 & $ +107 $ & $ <12.91 $ & $ 12.64 \pm 0.03 $ & $ < -0.91 $ \\
PG0914+001 & $ +100 $ & $ 13.96 \pm 0.04 $ & $ 13.15 \pm 0.03 $ & $ -0.36 \pm 0.05 $ \\
\hline
\end{tabular}
{\sc Notes ---} $a$: $[{\mbox O$\;${\scriptsize \rm I}\relax}/{\mbox Si$\;${\scriptsize \rm II}\relax}] \equiv \log (N({\mbox O$\;${\scriptsize \rm I}\relax})/N({\mbox Si$\;${\scriptsize \rm II}\relax})) - \log (A_{\rm O}/A_{\rm Si})_\odot) $ where the solar abundances for O and Si are from \cite{asplund09}. Values preceded by ``$<$" are 3 sigma upper limits (O$\;${\scriptsize \rm I}\relax\ absorption is not detected). Errors are at the 1$\sigma$ level.
\end{minipage}
\end{table}
\normalsize
|
\section*{Appendix}
\begin{center}
{\bf \small {APPENDIX}}
\end{center}
This appendix contains eight sections. The first shows that the relative entropy distance to the Gibbs state is an asymptotically continuous function. The next four sections discuss in detail the state transformation protocols for the case of two-level systems. Section II presents a distillation protocol for quasi classical states, while section III describes a formation protocol also for quasi classical states. The following two sections extend these protocols to the case of arbitrary nonstationary two-level resources. Then in Section VI we outline how the results can be easily generalized to higher dimensions by considering as an example the distillation protocol for quasi classical states. Section VII discusses some characteristics of the exhaust states produced in these protocols. Finally, Section VIII discusses the equivalence of our formulation to other models of thermodynamics and the degree of control one needs to implement our thermal operations.
\section{Extensivity and asymptotic continuity}
To show that $\rent{\rho}{\gamma}$ is asymptotically-continuous, we make use of the following, from~\cite{horodecki_locking_2005+}:
\begin{proposition}
Suppose a function $f$ satisfies (1) ``approximate affinity''
\begin{align}
|pf(\rho)+(1-p)f(\sigma)-f(p\rho+(1-p)\sigma))|\leq c,
\end{align}
for some constant $c>0$ and any $p$ such that $0\leq p\leq 1$, and (2) ``subextensivity'' $f(\rho)\leq M\log d$, where $M>0$ is constant and $d={\rm dim}(\mathcal{H})$ for $\mathcal{H}$ the state space on which $\rho$ has support. Then $f$ is asymptotically continuous:
\begin{align}
|f(\rho_1)-f(\rho_2)|\leq M\| \rho_1-\rho_2\|_1\log d+4c.
\end{align}
\end{proposition}
The entropy relative to the Gibbs state, $f(\rho):=\rent{\rho}{\sigma}$, satisfies both conditions. To see the first, let $\tau=\sum_k p_k\tau_k$ for some arbitrary set of density operators $\{ \tau_k \}$ and probability distribution $p_k$ and let $\omega$ be another arbitrary density operator. Then
\begin{align}
\rent{\tau}{\omega}&={\rm Tr}\left[\sum_k p_k\left(-\tau_k\log\omega+\tau_k\log\tau\right)\right]\\
&={\rm Tr}\left[\sum_k p_k\left(\tau_k\log\tau_k-\tau_k\log\omega+\tau_k\log\tau-\tau_k\log\tau_k\right)\right]\\
&=\sum_k p_k \rent{\tau_k}{\omega}+\sum_k p_k S(\tau_k)-S(\tau),
\end{align}
where in the second line we have added and subtracted ${\rm Tr}(\tau \log \tau_k)$. Since $S(\tau)\leq \sum_k p_kS(\tau_k)+H(p_k)$ where $H(p_k)$ denotes the Shannon entropy for the distribution $p_k$, and since the relative entropy is convex, this implies
\begin{align}
0\leq \sum_k p_k \rent{\tau_k}{\omega}-\rent{\tau}{\omega}\leq H(p_k).
\end{align}
Finally, letting $\omega=\gamma$ and $\{ \tau_k\} = \{\rho,\sigma\}$ with distribution $(p,1-p)$, we find
\begin{align}
p\rent{\rho}{\gamma}+(1-p)\rent{\sigma}{\gamma}-\rent{p\rho+(1-p)\sigma}{\gamma}\leq h_2(p)\leq 1,
\end{align}
where $h_2$ is the binary entropy function.
The fact that the second condition is satisfied, i.e., that the entropy relative to the Gibbs state is subextensive, follows from the fact that
the maximum energy of the system is extensive. First, note that $\rent{\rho}{\gamma}=\beta F_\beta(\rho)-\beta F_\beta(\gamma)$, where $F_\beta(\rho)=\langle H\rangle_\rho-\frac{1}{\beta}S(\rho)$ is the free energy. Thus, the maximum of $\rent{\rho}{\gamma}$ occurs for $\rho=\ketbra{E_{\rm max}}$ where $\ket{E_{\rm max}}$ is the eigenstate of maximum energy. Direct calculation shows
\begin{align}
\rent{\ketbra{E_{\rm max}}}{\gamma}&={\rm Tr}\left[\ketbra{E_{\rm max}}\left(\log \ketbra{E_{\rm max}}-\log \gamma\right)\right]=-\bra{E_{\rm max}}\log\gamma\ket{E_{\rm max}}\\
&=\beta E_{\rm max}+\log Z_\beta=\beta E_{\rm max}+\log {\sum_k e^{-\beta E_k}}\leq \beta E_{\rm max}+\log d.
\end{align}
Here we have assumed that the energy values $E_j>0$. When the maximum energy is extensive, i.e.\ $E_{\rm max}\leq K\log d$ for some constant $K$, we obtain $\rent{\ketbra{E_{\rm max}}}{\gamma}\leq M\log d$ for $M=\beta K+1$.
\section{Distillation of quasiclassical states}
For simplicity of presentation, we consider qubit systems with the Hamiltonian given by $H=\sum_i \ket{1}_i \bra{1}$, where the sum runs over all involved qubits. We start with $l$ copies of the Gibbs state $\gamma$ and $n$ copies of the resource state $\rho$, where
\begin{eqnarray} \begin{aligned}
\rho=(1-p) \ket{0}\bra{0}+p \ket{1}\bra{1};\quad
\gamma=(1-q) \ket{0}\bra{0}+q\ket{1}\bra{1},
\end{aligned} \end{eqnarray}
with $q=e^{-\beta}/(1+ e^{-\beta})$ and $\beta$ the inverse temperature, which we take as a constant parameter. The aim is to obtain the maximal number of copies possible of qubits in the pure excited state $\ket{1}$ by implementing a unitary that commutes with $H$ and taking the partial trace over some subsystem,
\begin{eqnarray} \begin{aligned}
\gamma^{\otimes l} \otimes \rho^{\otimes n} \to \sigma^{(k)}\otimes \ket{1}\bra{1}^{\otimes m}.
\end{aligned} \end{eqnarray}
We denote $R=\frac{m}{n}$ (the rate of distillation) and
$\epsilon=\frac{n}{l}$ (the ratio between the number of used Gibbs states
and the number of resource states). The Gibbs states are free, so
we accept that $\epsilon$ asymptotically vanishes.
In the protocol we shall use the fact that up to a small error (vanishing for a large number of qubits)
\begin{eqnarray} \begin{aligned}
\rho^{\otimes n} \approx \sum_t p_t P_t
\end{aligned} \end{eqnarray}
where $t$ run over strongly typical types,
i.e.\ the types containing strings with the number of $1$'s
within the interval $(n p - O(\sqrt{n}), np +O(\sqrt{n}))$,
and $P_t$ denotes the projector onto type $t$. Similarly
\begin{eqnarray} \begin{aligned}
\gamma^{\otimes l}\approx \sum_t q_t Q_t
\end{aligned} \end{eqnarray}
again with $q_t\approx qn$. The errors in both approximations are smaller than $2^{-\sqrt{n}}$ when quantified by the trace norm.
For simplicity we shall first pretend that both $\gamma^{\otimes l} $ and $\rho^{\otimes n}$ consist of a single type. Then further we will show how to extend the argument to a mixture of types.
So we start with a tensor product of two types (one from Gibbs,
the other from $\rho$), i.e.\ an equal mixture of strings of length $l+n$. The string consist of two substrings:
the first has $ql$ $1$'s and the second has $pn$ $1$'s:
\begin{eqnarray} \begin{aligned}
\overbrace{000 \ldots 0\,\underbrace{11 \ldots 1}_{lq}}^l\,
\overbrace{00 \ldots 0\,\underbrace{111 \ldots 1}_{np}}^n.
\end{aligned} \end{eqnarray}
There are roughly
\begin{eqnarray} \begin{aligned}
2^{l h(q)}\times 2^{n h(p)}
\label{eq:entpq}
\end{aligned} \end{eqnarray}
such strings (with the error being a multiplicative $\text{poly}(n)$ factor).
We now apply a unitary transformation to these strings to map them into strings of the same total length which have $m$ $1$'s
to the right:
\begin{eqnarray}
&& \overbrace{000 \ldots 0\,\underbrace{11 \ldots 1}_{lq}}^l\,
\overbrace{00 \ldots 0\,\underbrace{111 \ldots 1}_{np}}^n \quad\to\qua
\underbrace{00000 \ldots 0000\,\underbrace{111 \ldots 11}_{rk}}_{k}\,
\underbrace{111 \ldots 1}_{m}
\end{eqnarray}
where
\begin{eqnarray} \begin{aligned}
k=l+n-m
\label{eq:con-dim}
\end{aligned} \end{eqnarray} (conservation of dimension), and $r$ and $m$ are about to be determined.
First, $r$
is fixed by {\it conservation of energy}, which requires that the number of $1$'s is conserved:
\begin{eqnarray} \begin{aligned}
lq+ np=rk+m.
\label{eq:con-en}.
\end{aligned} \end{eqnarray}
Then {\it unitarity} requires that
there are at least as many strings of length $k$
with $rk$ $1$'s as the number of initial strings \eqref{eq:entpq}:
\begin{eqnarray} \begin{aligned}
2^{k h(r)}\geq 2^{l h(q)+n h(p)}.
\label{eq:con-ent}
\end{aligned} \end{eqnarray}
Roughly speaking this is {\it conservation of entropy}.
Using \eqref{eq:con-en} and \eqref{eq:con-ent} we obtain
that our transformation is possible if
\begin{eqnarray} \begin{aligned}
h(q) + \epsilon h(p)\leq (1+ \epsilon -R \epsilon)\,h\left(\frac{q+ \epsilon p -\epsilon R}{1 +\epsilon -\epsilon R}\right),
\end{aligned} \end{eqnarray}
where recall that $\epsilon$ is the ratio of Gibbs states used,
and $R$ the ratio of pure excited states obtained.
We now expand this with respect to $\epsilon$ to first order and let $\epsilon\to 0$. This means that we take many more Gibbs states than the resource states. In other words, we presume a heat reservoir that is arbitrarily larger than the size of our system. As a result, recalling that $q=e^{-\beta}/(1+ e^{-\beta})$, we obtain that the following rate can be achieved
\begin{eqnarray} \begin{aligned}
R=\frac{h(q)-h(p)+\beta (p-q)}{h(q)+\beta(1-q)}=\frac{S(\rho || \gamma)}
{S( \ket{1}\bra{1} || \gamma)},
\end{aligned} \end{eqnarray}
i.e., the protocol achieves the upper bound obtained by the monotonicity argument in the main text.
So far we have worked with a single type. But our initial state is actually a mixture of products of types $Q_t \otimes P_{t'}$.
We thus apply the protocol separately to each type,
with the same number $m$ of required output excited states,
for all types, with this number being fitted to
the product of less numerous types (i.e.\ the one with the smallest number of $1$'s, namely $(np-O(\sqrt{n}))(lq-O(\sqrt{l}))$ $1$'s).
Note, however, that the variations will disappear asymptotically, as we divide the equations by $l$.
Also the approximation to the number of strings in each type by the exponential of the entropy is correct up to a multiplicative polynomial factor, which is also irrelevant asymptotically.
\section{Formation of quasiclassical states}
We are going to construct the formation protocol in three stages.
The first is to show that for a particular type $T_q$ of the $l$ copies of the Gibbs state, we can
create a particular type $T_p$ of the state we want to form. We then show
that we can do this for all types of the Gibbs state. Finally we show
how to correctly get the distribution over types of the target state.
We shall need the following useful lemma:
\begin{lemma}
\label{lemma:birkoff}(Birkhoff primitive) The following operation can
be done by means of thermal operations with arbitrary accuracy:
\begin{eqnarray} \begin{aligned}
\rho \to \sum_k p_k U_k \rho U_k^\dagger
\label{eq:bir}
\end{aligned} \end{eqnarray}
where $p_k$ is an arbitrary probability distribution,
and the $U_k$ commute with the Hamiltonian. In particular,
a random permutation of the systems is a valid operation.
\end{lemma}
{\bf Remark.} The accuracy depends on
the number of Gibbs states that are used, but in our paradigm they are for free.
{\bf Proof.} First, let us note that the following unitary transformation preserves energy:
\begin{eqnarray} \begin{aligned}
\sum_i \ketbra{i}\otimes \tilde U_i
\end{aligned} \end{eqnarray}
provided that $\ket{i}$ are eigenvectors of the reservoir Hamiltonian and the $\tilde U_i$ commute with the system Hamiltonian. To obtain the required transformation, we take the initial state of the reservoir to be $l$ copies of the Gibbs state. Let $q_i$ denote the probability distribution of single strings.
We now divide the set of eigenvectors of the reservoir Hamiltonian into sets, denoted $S_k$, such that the sum of the probability distribution over $i$ within each set yields approximately the probability $p_k$ from \eqref{eq:bir}, that is,
$\sum_{i\in S_k} q_i \approx p_k$.
This can be done with arbitrary accuracy by taking $l$ large enough, since $q_i\leq \max\{q,1-q\}^l$.
Then, for every $i\in S_k$, we set $\tilde U_i=U_k$.
\hfill$\square$
We want to form $n$ copies of the state
$\rho=(1-p) \ketbra{0}+p\ketbra{1}$
from a pure excited state. Let us first show how we can form any of
the typical types of the state.
\vspace{0.2 cm}
{\bf Formation of a maximally mixed state over a fixed type.}
\vspace{0.2 cm}
Consider a fixed type $T_p$ of the state of $n$ systems that we want to create, and let
it be one with $np$ $1$'s. For other types, the reasoning is the same and the asymptotic
rate will be the same. We start with $l$ copies of the Gibbs state $\gamma$, and $m$
copies of the excited state $\ket{1}$. We consider a final exhaust system
consisting of $k$ two-level systems.
Consider a typical type $T_q$ of the Gibbs state; it has $lq$ $1$'s.
In that type, there are $\approx 2^{l h(q)}$ strings up to some $2^{\sqrt{l} h(q)}$ factor. We want to map these initial strings onto the $N$ final strings in the type $T_p$
as follows. Take $\{u_i\}_i$ to be the set of strings in $T_p$ and for each
string $u_i$ consider some set $\{v^i_j\}_j$ of strings on the exhaust system. We
now map each of the initial strings to some string $u_iv^i_j$. This is illustrated in Figure~1
\begin{figure}
\includegraphics[width=8cm]{formation-eps-converted-to}
\label{fig:formation}
\caption{Mapping of strings in the formation protocol}
\end{figure}
For sets $\{v^i_j\}_j$ corresponding to different values of $i$, we can take the number of strings in each set to be the same
or off by $1$, simply by assigning the strings in an order determined
by fixing $j=1$, then incrementing the $i$ register until $i=N$, then incrementing
the $j$ register by $1$, reseting $i$ to $1$ and again incrementing the $i$ register until $i=N$ and repeating. This is all done to ensure that when we
trace out the exhaust system, we get an even mixture over permutations
within the type class.
We can now use the analogues of equations
(\ref{eq:con-dim}), (\ref{eq:con-en}) and (\ref{eq:con-ent})
to ensure that we can perform the unitary which implements this mapping:
\begin{eqnarray} \begin{aligned}
m+l=n+k
\label{eq:revcon-dim}
\end{aligned} \end{eqnarray}
\begin{eqnarray} \begin{aligned}
lq+ m=rk +np
\label{eq:revcon-en}
\end{aligned} \end{eqnarray}
\begin{eqnarray} \begin{aligned}
2^{l h(q)} \leq 2^{k h(r)+n h(p)}
\label{eq:revcon-ent}
\end{aligned} \end{eqnarray}
We now take $l\propto m^\alpha$ with $1<\alpha<2$ and do as before. We take $\alpha>1$ so that we can
take $\epsilon=n/l$ small, and $\alpha<2$ so that $\sqrt{l}$ is sublinear
in $m$ and we can ignore such terms.
This maps type $T_q$ of the $l$ Gibbs states onto the type $T_p$ of $\rho^{\otimes n}$.
We can map each of the initial Gibbs types onto $T_p$
in this manner using a unitary $U_{pq}$. For each such mapping, the above three equations will change,
but only by some $\sqrt{l}$ factors which we took to be sublinear in $m$.
We can thus choose $m$ to ensure conservation of energy in the worst case of Eq. (\ref{eq:revcon-en}),
and $k h(r)$ is chosen to ensure the inequality Eq. (\ref{eq:revcon-ent}) in
the worst case.
We now need to implement $U_{pq}$ conditioned on the initial type $T_q$.
Since the typical initial types are on orthogonal and diagonal subspaces,
we first do a conditional copying of the type class $q$ onto an initialised
register. We then act $U_{pq}$ conditioned on this register.
Since the number of types is polynomial in $l$, the
register only needs to be of size $\log{l}$, and thus this
resource does not matter as it is sublinear in $m$. It is an interesting question whether the formation
protocol can be made to work without this sublinear supply of pure states.
We denote by $U_p$ the unitary that creates a particular type $T_p$. To get the distribution
over types, we simply use the Birkoff primitive of Lemma
\ref{lemma:birkoff} to implement $\sum_p \ket{p}\bra{p} \otimes U_{p}$. This is irreversible, but since the number
of typical types is polynomial in $n$, the rate of entropy that is created
by this procedure is negligible, i.e.\ logarithmic in $n$. It is not hard to see that Eqs. (\ref{eq:revcon-dim}-\ref{eq:revcon-ent}) give the required rate (i.e.\ the inverse of the distillation rate).
\section{Distillation of arbitrary states}
We now extend distillation to the case where our states are not
diagonal in the energy eigenbasis.
Consider a state $\rho=p \ketbra{\phi_1}+(1-p)\ketbra{\phi_2}$.
The average energy of the state is $\mean{E}= p|\bracket{\phi_1}{1}|^2+ (1-p) |\bracket{\phi_2}{1}|^2$.
As before, we consider $n$ copies of $\rho$ and $l$ copies of systems in a Gibbs state $\gamma$.
Regarding $\rho^{\otimes n}$, only the blocks with energy $E\in [n \mean{E} - \sqrt{n}, n\mean{E} + \sqrt{n}]$
will be relevant, i.e.\ we have
\begin{eqnarray} \begin{aligned}
{\rm Tr} \left( \sum_E P_E \rho^{\otimes n} \right) \geq 1- 2^{-O(n)}
\end{aligned} \end{eqnarray}
where the sum runs over $E \in [n \mean{E} - O(\sqrt{n}), n\mean{E} + O(\sqrt{n})]$,
with $P_E$ the projector onto the energy $E$ eigenspace
(this follows from Eq. \eqref{eq:tails-av-E}, proven in Section \ref{sec:form_arb}).
Our protocol has two stages:
\begin{itemize}
\item[(i)] unitary rotation within energy blocks of a resource system (consisting of $n$ qubits) solely.
\item[(ii)] drawing work by string permutations on the total system resource (with $n$ qubits) plus heat bath
(with $l$ qubits)
\end{itemize}
We write down the resource state in the energy eigenbasis. As said above, only blocks
with energy $\approx nE$ will appear. We will use the fact that
the state is, up to exponentially vanishing error, equal to its projection onto the typical subspace, having
dimension $2^{nS(\rho)+O(\sqrt{n})}$ where $S(\rho)$ is the von Neumann entropy of $\rho$. Therefore, within every block the rank of the state is not larger than $\approx 2^{n S}$ (as a projection cannot increase the rank).
Now stage $(i)$ is the following: within each energy block, we apply unitary rotation, which diagonalizes the state restricted to the block
in the energy basis. Then there is stage $(ii)$, in which
we apply the protocol of distillation of quasiclassical states as in Section \ref{sec:distq},
i.e.\ we permute strings in such a way that the output strings have $m$ $1$'s to the right.
Such a protocol produces $m$ systems in a pure excited state (note that all coherences
initially present in the state are now left in the garbage).
Using the same notation as in Section \ref{sec:distq}, for a product of two single types, e.g.
with $lq$ and $n\mean{E}$ $1$'s, respectively, the constraints now become
\begin{eqnarray} \begin{aligned}
&&k=m+m-l \nonumber \\
&&lq + n\mean{E}= rk +m\nonumber\\
&& 2^{k h(r)}\geq 2^{l h(q)+n S(\rho)}
\end{aligned} \end{eqnarray}
As before, taking the limit $\frac{n}{l}\to 0$ we obtain
that any rate $R$ is achievable, provided it satisfies
\begin{eqnarray} \begin{aligned}
R\leq \frac{h(q)-S(\rho)+\beta (\mean{E}-q)}{h(q)+\beta(1-q)}=\frac{S(\rho ||\gamma)}
{S(\ketbra{1} || \gamma)}.
\end{aligned} \end{eqnarray}
Thus also for states that are not quasiclassical, we can reach the upper bound given by the relative entropy distance from the Gibbs state.
\section{Formation of arbitrary states}
\label{sec:form_arb}
We now show that we can achieve reversibility even in the case of states that are not quasiclassical. To do so, however, we must allow the use of a sublinear amount of states that are a superposition over energy eigenstates. This is a reasonable assumption since the rate at which such states are consumed vanishes in the asymptotic limit. This is very similar
to the fact that in entanglement theory, distillation requires no communication
but formation requires a sublinear amount of it. Or, instead, formation requires a state which is a superposition over different amounts
of entanglement. The superposition over different amount of entanglement (known as entanglement spread \cite{HP97, HW03, HL04, Har09}),
is analogous to the superposition over energy eigenstates in the present context. In the athermality context, distillation of work requires no superposition over energy eigenstates, but formation does.
Suppose we want to implement some unitary
\begin{eqnarray} \begin{aligned}
U=\sum_{ij}u_{ij}\ket{E_i}\bra{E_j}
\end{aligned} \end{eqnarray}
that does not conserve energy. We introduce a state that acts as a reference frame for time,
\begin{eqnarray} \begin{aligned}
\ket{H}=\sum f(h)\ket{h},
\label{eq:qrefsys}
\end{aligned} \end{eqnarray}
where $\ket{h}$ denotes an energy eigenstate, and we implement
\begin{eqnarray} \begin{aligned}
U^{\textrm{inv}}=\sum_{ij}u_{ij}\ket{E_i}\bra{E_j}\otimes \ket{h-E_i+E_j}\bra{h}
\,\,
\label{eq:energyswap}
\end{aligned} \end{eqnarray}
on the system and reference frame. If we are interested in implementing $U$ on the state
\begin{eqnarray} \begin{aligned}
\ket{\psi}=\sum_{i\in \cal{S}} c_i\ket{E_i},
\end{aligned} \end{eqnarray}
then we do so by implementing $U^{\textrm{inv}}$ on $\ket{\psi}\otimes \ket{H}$. Note that we must ensure that the reference frame system has energy levels with gaps of size $|E_i-E_j|$ for every transition appearing in $U$.
If the energy spread of $f(h)$ is large compared to the largest value of $|E_i-E_j|$ in an energy transition induced by $U$, then the state of the reference frame is not disturbed
very much in the process.
For the problem in which we are interested, this is indeed the case because on the typical subspace, the variation in energy is sublinear in the number of copies of the state we want to create. To see how this works by way of example, note that if in Eq.~\eqref{eq:qrefsys}, we take $f(h)$ to be $1/\sqrt{N}$ for energies $h \in \{1,...,N\}$ and $f(h)=0$ otherwise,
then removing a unit of energy and adding it to another system, does not change the state of Eq.~\eqref{eq:qrefsys} much. i.e.\
the inner product between $\ket{H}$ and $\sum_h \ket{h-1}\bra{h} \ket{H}$ approaches $1$, because
\begin{eqnarray} \begin{aligned}
\sum_{h=1}^N \frac{1}{\sqrt{N}} \bra{h}\sum_{h=0}^{N-1}\frac{1}{\sqrt{N}} \ket{h}=1-\frac{1}{N},
\end{aligned} \end{eqnarray}
which approaches $1$ for large $N$. States like that of Eq.~\eqref{eq:qrefsys} therefore allow us to lift the superselection rule for
energy, without being consumed much in the process.
This gives some insight into embezzling states~\cite{ent-embezzling}. These are resource states that are often used in entanglement theory in similar situations. For instance, one can use a state similar in form to Eq.~\eqref{eq:qrefsys}
\begin{eqnarray} \begin{aligned}
\ket{E}=\sum f(k)\ket{\phi}_{AB}^{\otimes k}\ket{00}_{AB}^{\otimes (n-k)}
\label{eq:embezzling}
\end{aligned} \end{eqnarray}
which is a superposition of a different number of entangled EPR pairs $\ket{\phi}_{AB}$. These
states can be used to implement operations which
need to create superpositions over amounts of entanglement (entanglement spread). Just as removing one unit of energy, doesn't change the state of Eq.~\eqref{eq:qrefsys} much, likewise,
removing one EPR pair from the state of Eq.~\eqref{eq:embezzling} and adding it to another system doesn't change the embezzling state by much. We can embezzle energy, just as one can embezzle entanglement. We therefore see that a superposition over some resource can create an embezzling state for that resource, and will allow us to lift some superselection rule or restriction.
With this small superposition over energy states, let us now show that we can create an arbitrary state at a rate given by the relative
entropy distance to the Gibbs state. Let $\rho := p \ket{\phi_1}\bra{\phi_1} + (1 - p) \ket{\phi_2}\bra{\phi_2}$ and
\begin{equation}
\rho^{\otimes n} = \sum_{k, g} p_k \ket{\Psi_{k, g}}\bra{\Psi_{k, g}},
\end{equation}
with
\begin{equation} \label{formPsi}
\ket{\Psi_{k, g}} := \pi_{g} \ket{\phi_1}^{\otimes k} \otimes \ket{\phi_2}^{\otimes n - k},
\end{equation}
for $\pi_g$ a permutation.
The idea of the protocol is as follows: we will first create a diagonal
state
\begin{eqnarray} \begin{aligned}
\varrho_n=\sum p_k \ket{{t_k, s_g}}\bra{{t_k, s_g}}
\end{aligned} \end{eqnarray} which has the same spectrum as $\rho^{\otimes n}$ and
where each eigenstate has the same average energy as an eigenstate in the
typical subspace of $\rho^{\otimes n}$. From the result
of the previous section it is not hard to see that
this can be done at a rate given
by the relative entropy distance of $\rho$ to the Gibbs state, since in
the limit of many copies, the regularised relative entropy distance is the same. We would then like to rotate
the diagonal basis to the $\Psi_{k, s}$-basis. This cannot be done by
unitaries which commute with the Hamiltonian unless we allow for a reference
frame $\ket{H}$ which is a superposition over energy states. We then want to show
that the reference frame which allows us to break the energy superselection
rule is consumed at a vanishingly small rate. We do so by showing that
the reference frame superposition is over a size sublinear in $n$. This
can be understood as coming from the fact that in the typical subspace,
the superposition over different types is sublinear.
We consider only typical $\ket{\Psi_{k, g}}$ with $k \in \text{Typ}_{\rho} := [np - \sqrt{n}, np + \sqrt{n}]$. Then
\begin{equation} \label{typpicalPsits}
\left \Vert \rho^{\otimes n} - \sum_{k \in \text{Typ}_{\rho}, g} p_k \ket{\Psi_{k, g}}\bra{\Psi_{k, g}} \right \Vert_1 \leq 2^{- \Omega(\sqrt{n})}
\end{equation}
Let also
\begin{equation}
\label{lineardecompPsi}
\ket{\Psi_{k, g}} := \sum_{t', s'} \c \ket{t', s'}
\end{equation}
where $\ket{t', s'}$ is an eigenstate of the Hamiltonian with energy $t'$ ($s'$ labels the degeneracy). From Eq. (\ref{formPsi}) it follows that the sum in Eq. (\ref{lineardecompPsi}) will be peaked around only a few energy values $t'$. Indeed, with
\begin{equation}
\ket{\phi_1} := a\ket{0} + b\ket{1},
\end{equation}
and
\begin{equation}
\ket{\phi_2} := b \ket{0} - a\ket{1},
\end{equation}
set $\text{Typ} := [ nE_t - \sqrt{n}, nE_t + \sqrt{n} ]$, where $E_t := \left( (n - t)|b|^{2} + t |a|^{2} \right) /n$. Then
\begin{equation}
\left \Vert \sum_{t' \notin \text{Typ}_t, s'}\c \ket{t', s'} \right \Vert = 2^{- \Omega(\sqrt{n})}
\label{eq:tails-av-E}
\end{equation}
Note that since $E_t := \left( (n - t)|b|^{2} + t |a|^{2} \right) /n$ the
degeneracy of each energy state $\ket{t_k, g_s}$ is at least as large
as the degeneracy of $\ket{\Psi_{ks}}$.
Now we construct the reference frame.
Let $\ket{w}$ be an energy eigenstate with energy $n(p |b|^{2} + (1 - p)|a|^{2}) - n^{2/3}$. It is needed to pad the dimension of the reference frame, since
although the probability that it happens is vanishingly small, the unitary does
connect states with large energy difference. After the protocol, we will see that $\ket{w}$ will hardly be changed, and thus
is only used as a catalyst. We define the reference system as follows
\begin{equation}
\ket{H} := \frac{1}{\sqrt{|H|}} \sum_{h \in H} \ket{h}
\end{equation}
with $\ket{h} := \ket{h'}\otimes \ket{w}$, where$\ket{h'}$ is an energy eigenstate of energy $h'$ and $H := \{ 0, ..., 2 n^{2/3} \}$.
Consider the energy preserving unitary
\begin{equation}
U := \sum_{h, t, s, t', s'}\c \ket{t', s'}\bra{t_k, s_g} \otimes \ket{h + t - t'}\bra{h}.
\end{equation}
Then in the sequel we prove that
\begin{eqnarray} \label{main}
&& \left \Vert U \left(\sum_{k \in \text{Typ}_{\rho} ,s} p_k \ket{t_k, s_g}\bra{t_k, s_g} \otimes \ket{H} \bra{H}\right)U^{\cal y} - \rho^{\otimes n} \otimes \ket{H}\bra{H} \right \Vert_1 \nonumber \\ &\leq& O(n^{-1/6}).
\end{eqnarray}
We first analyze the action of $U$ in $\ket{s, t} \otimes \ket{H}$:
\begin{eqnarray}
U \left( \ket{t_k,s_g} \otimes \ket{H} \right) &=& \sum_{t', s'} \c \ket{t', s'}\otimes \left( \frac{1}{\sqrt{|H|}} \sum_{h \in H} \ket{h + t_k - t'} \right) \nonumber \\ &=& \ket{\nu_1} + \ket{\nu_2} + \ket{\nu_3}
\end{eqnarray}
where the non-normalized pure states $\ket{\nu_k}$ are given by
\begin{equation}
\ket{\nu_1} := \sum_{t' \in \text{Typ}_t, s'}\c \ket{t', s'} \otimes \ket{H},
\end{equation}
\begin{equation}
\ket{\nu_2} := \sum_{t' \in \text{Typ}_t, s'} \c \ket{t', s'} \otimes \ket{\text{err}_{t'}}
\end{equation}
with
\begin{equation}
\ket{\text{err}_{t'}} := \frac{1}{\sqrt{|H|}} \sum_{h \in H} \ket{h + t - t'} - \ket{H},
\end{equation}
and
\begin{equation}
\ket{\nu_3} := \sum_{t' \notin \text{Typ}_t, s'} \c \ket{t', s'} \otimes \left( \frac{1}{\sqrt{|H|}} \sum_{h \in H} \ket{h + t - t'} \right).
\end{equation}
Set $t_k=E_k$ and let us take $s_g=g$. We can do the latter since as we
mentioned, the degeneracy
of $\ket{t_k,s_g}$ is larger than the degeneracy of $\ket{\Psi_{k,g}}$.
Then,
\begin{eqnarray} \label{closetoright}
\Vert U \left( \ket{t_k, s_g} \otimes \ket{H} \right) - \ket{\Psi_{k, g}} \otimes \ket{H} \Vert &\leq& \Vert \ket{\nu_1} - \ket{\Psi_{k, g}} \otimes \ket{H} \Vert \nonumber \\ &+& \Vert \ket{\nu_2} \Vert + \Vert \ket{\nu_1} \Vert.
\end{eqnarray}
We now show that the three terms in the R.H.S. are small. For $\ket{\nu_2}$ we first note that for $t' \in \text{Typ}_t$
\begin{equation}
\Vert \ket{\text{err}_{t'}} \Vert \leq n^{-1/6}.
\end{equation}
by taking the worst case.
Then
\begin{eqnarray}
\Vert \ket{\nu_2} \Vert ^{2} &=& \sum_{t' \in \text{Typ}_t, s'} |\c|^{2} \Vert \ket{\text{err}_{t'}} \Vert^{2} \nonumber \\ &\leq& \max_{t' \in \text{Typ}_t} \Vert \ket{\text{err}_{t'}} \Vert^{2} \leq n^{-1/3}.
\end{eqnarray}
For $\ket{\nu_3}$, in turn, we have
\begin{eqnarray}
\Vert \ket{\nu_3} \Vert^{2} &\leq& \sum_{t' \notin \text{Typ}, s'} |\c|^{2} \Vert \ket{\text{err}_{t'}} \Vert^{2} \nonumber \\ &\leq& \sum_{t' \notin \text{Typ}, s'} |\c|^{2} \leq 2^{- \Omega(\sqrt{n})}.
\end{eqnarray}
Finally, for $\ket{\nu_1}$,
\begin{eqnarray}
\Vert \ket{\nu_1} - \ket{\Psi_{k, g}} \otimes \ket{H} \Vert &=& \left \Vert \sum_{t' \in \text{Typ}_t, s'} \c \ket{t', s'} - \ket{\Psi_{k, g}} \right \Vert \nonumber \\ &\leq& 2^{- \Omega(\sqrt{n})}.
\end{eqnarray}
From Eq. (\ref{closetoright}) it thus follows that
\begin{equation}
\Vert U \left( \ket{t_k, s_g} \otimes \ket{H} \right) - \ket{\Psi_{k, g}} \otimes \ket{H} \Vert \leq O(n^{-1/6}).
\end{equation}
Since $\Vert \ket{\psi}\bra{\psi} - \ket{\phi}\bra{\phi} \Vert_1 \leq \sqrt{2}\Vert \ket{\psi} - \ket{\phi} \Vert$ for every two states $\ket{\psi}, \ket{\phi}$, we find
\begin{eqnarray}
&& \left \Vert U \left( \ket{t_k, s_g}\bra{t_k, s_g} \otimes \ket{H}\bra{H} \right) U^{\cal y} - \ket{\Psi_{k, g}}\bra{\Psi_{k, g}} \otimes \ket{H}\bra{H} \right \Vert_1 \nonumber \\ &\leq& O(n^{-1/6}).
\end{eqnarray}
Eq. (\ref{main}) then follows from the triangle inequality for trace-norm and Eq. (\ref{typpicalPsits}).
\section{Distillation of quasiclassical states in arbitrary dimensions}
\label{sec:distq}
In this section we present the details of the distillation protocol for quasiclassical states for the general case of $d$-dimensional systems. This is presented as an example of how the results can be extended to arbitrary dimensions using arguments very similar to those used in the two-dimensional case.
The input to the protocol consists of $n$ copies of the initial resource $\rho$ and $\ell$ copies of the Gibbs state $\gamma$ of the same Hamiltonian $H$. Since the states are quasiclassical, the overall state of the input is fully described by a collection of strings $\mathbf{s}\in\{0,\dots,d{-}1\}^{n+\ell}$ listing the energy level occupied by each system, each string weighted by its probability of occurrence. Here $d$ is the total number of energy levels of $H$, and the first $n$ entries of $\mathbf{s}$ correspond to the state of $\rho$ and the remainder correspond to $\gamma$. If some of the energy levels are degenerate, we simply work in the eigenbasis of $\rho$ to remove the degeneracy in labeling.
Because permutations within each of the two substrings do not change the overall probability, we can therefore instead work with the collection of occupation frequencies $\mathbf{f}$ of the state, which describe the number of systems in the ground state, first excited state, second excited state, and so on (divided by the total number of systems), again each weighted by an appropriate probability.
We would now like to define a protocol which creates as many standard resources in the form of work as possible. In the qubit case the standard resource had a very simple form, namely the excited state of the Hamiltonian. Here, however, the setup is more cumbersome, as there are $d-1$ excited states and no guarantees that their energy differences are in any way commensurate (as, e.g. in the case of a harmonic oscillator, where the energy of the state $\ket{2}$ can be transferred to two instances of the state $\ket{1}$). To handle this issue most simply, we imagine that, in addition to the resource and thermal states, we also have a work system at our convenience. The work system is capable of accepting arbitrary amounts of energy, i.e.\ it has energy transitions which precisely correspond to those of $H$, but it cannot accept any entropy. Now the goal of the protocol is to change the occupation of the energy levels so as to transfer as much energy to the work system as possible.
Let us now restrict attention to a fixed occupation frequency $\mathbf{f}_\rho$ of the resource and a fixed $\mathbf{f}_\gamma$ for the Gibbs state.
We will later design the protocol so that it works for every such frequency pair which has appreciable probability. Suppose that we now change the occupation numbers by an amount described by the vector $-n\mathbf{x}$ (whose prefactor is chosen for later convenience). This results in a new occupation vector $\boldsymbol{\nu}$ defined by
\begin{align}
\label{eq:nudef}
(n+\ell)\boldsymbol{\nu}\equiv n\mathbf{f}_\rho+\ell\mathbf{f}_\gamma-n\mathbf{x}.
\end{align}
For this to be an allowable transformation in our framework, this mapping must satisfy two constraints: energy conservation and unitarity. In contrast to the qubit case, here the input and output dimensions are equal by design.
Energy conservation is simply enforced by requiring the work system to take up the change in energy of the input systems. Using the vector $\mathbf{H}$ to describe the energy of each energy level, the initial energy is given by $E_{\rm in}= n\mathbf{H}\cdot\mathbf{f}_\rho+\ell\mathbf{H}\cdot\mathbf{f}_\gamma$ while the final energy is $E_{\rm out}=n\mathbf{H}\cdot\mathbf{f}_\rho+\ell\mathbf{H}\cdot\mathbf{f}_\gamma-n\mathbf{H}\cdot\mathbf{x}$. Energy conservation is then the statement that the work extracted is given by $W=E_{\rm in}-E_{\rm out}=n\mathbf{H}\cdot\mathbf{x}$.
Unitarity is enforced by making sure that the total number of configurations (strings) consistent with each occupation vector is conserved by the process. The total number of possible input strings in this case, $N_{\rm in}$,
is just the product of the multinomial cofficients using the frequency vectors:
\begin{align}
N_{\rm in}=M(n\mathbf{f}_\rho)M(\ell\mathbf{f}_\gamma)=\frac{n!}{(n(\mathbf{f}_\rho)_0)!\cdots(n(\mathbf{f}_\rho)_{d{-}1})!}\frac{\ell!}{(\ell(\mathbf{f}_\gamma)_0)!\cdots(\ell(\mathbf{f}_\gamma)_{d{-}1})!}
\end{align}
The maximum number of strings $N_{\rm out}=M((n+\ell)\boldsymbol{\nu})$ which can be created in the $n+\ell$ systems given the new occupation frequency $\boldsymbol{\nu}$ is the multinomial coefficient of the new occupation frequency vector,
\begin{align}
N_{\rm out}=M((n+\ell)\boldsymbol{\nu})=\frac{(n+\ell)!}{((n+\ell){\nu}_0)!\cdots((n+\ell){\nu}_{d{-}1})!}.
\end{align}
Therefore, a sufficient condition for unitarity is $N_{\rm in}\leq N_{\rm out}$, or $M(n\mathbf{f}_\rho)M(\ell\mathbf{f}_\gamma)\leq M((n+\ell)\boldsymbol{\nu})$.
\begin{comment}
The task of constructing the forward protocol can be formulated a little more compactly in terms of $\mathbf{x}$, as follows. First note that $\mathbf{1}\cdot\boldsymbol{\nu}=1$, where $\mathbf{1}$ is the vector of ones, since $\boldsymbol{\nu}$ is a frequency. This implies $\mathbf{1}\cdot\mathbf{x}=0$, and simply reflects the fact that increasing (decreasing) the occupation of one level means decreasing (increasing) the occupation of a different level. Then we have
\begin{align}
\max_{\mathbf{x}}\,\, n\mathbf{H}\cdot\mathbf{x}\quad\text{subject to}\quad N(n\mathbf{f}_r+m\mathbf{f}_g-n\mathbf{x})\geq N(n\mathbf{f}_r,m\mathbf{f}_g),\, \mathbf{1}\cdot\mathbf{x}=0.
\end{align}
\end{comment}
It can be shown that the multinomial coefficients obey the bounds
\begin{align}
\frac{e}{(ne)^d}\frac{1}{f_1\cdots f_d}2^{nH(\mathbf{f})}\leq \frac{n!}{(nf_1)!\cdots (nf_d)!}\leq \frac{n}{e^d-1}2^{nH(\mathbf{f})},
\end{align}
and therefore $M(n\mathbf{f})\approx 2^{nH(\mathbf{f})\pm O(\log n)}$.
The {unitarity} condition then becomes
\begin{align}
\label{eq:unitdef}
nH(\mathbf{f}_\rho)-O(\log n)+\ell H(\mathbf{f}_\gamma)-O(\log \ell)\leq (n+\ell)H(\boldsymbol{\nu})+O(\log (n+\ell))
\end{align}
Defining $\ep=\frac{n}{\ell}$ we may express this as
\begin{align}
\ep H(\mathbf{f}_\rho) + H(\mathbf{f}_\gamma) \leq (1+\ep)H(\boldsymbol{\nu}) + O(\tfrac{\log \ell}{\ell}).
\end{align}
Using the expression for $\boldsymbol{\nu}$ from \eqref{eq:nudef} and assuming that $\epsilon\ll 1$ gives
\begin{align}
H(\boldsymbol{\nu}) &= H\left(\frac{n\mathbf{f}_\rho+\ell\mathbf{f}_\gamma-n\mathbf{x}}{n+\ell}\right)\\
&=H(\mathbf{f}_\gamma)+\ep\left[\left(\mathbf{f}_\gamma+\mathbf{x}-\mathbf{f}_\rho\right)\cdot\mathbf{1}-H(\mathbf{f}_\gamma)+\left(\mathbf{x}-\mathbf{f}_\rho\right)\cdot\log\mathbf{f}_\gamma\right]+O(\ep^2)\\
&=H(\mathbf{f}_\gamma)-\ep\left(\mathbf{f}_\rho-\mathbf{x}\right)\cdot\log\mathbf{f}_\gamma-\ep H(\mathbf{f}_\gamma)+O(\ep^2).
\end{align}
Here $\mathbf{1}$ is the vector of all ones, and we have made use of the fact that $\mathbf{f}\cdot\mathbf{1}=1$ for any frequency vector $\mathbf{f}$, which also implies $\mathbf{x}\cdot\mathbf{1}=0$. Combining this with \eqref{eq:unitdef} we obtain the relation
\begin{align}
\label{eq:ucondition}
-\mathbf{x}\cdot\log\mathbf{f}_\gamma\leq D(\mathbf{f}_\rho||\mathbf{f}_\gamma)+O(\tfrac{\log \ell}{\ell}).
\end{align}
The next step is to fix the protocol to the worst case among the {likely} frequency vectors $\mathbf{f}_\rho$ and $\mathbf{f}_\gamma$. Their probabilities sharply peaked around the individual distributions $\boldsymbol{\rho}$ and $\boldsymbol{\gamma}$, respectively. Specifically, fixing an error parameter $\delta$, the probability that $||\mathbf{f}_\rho-\boldsymbol{\rho}||_1\geq \delta$ is less than a quantity of order $e^{-n\delta^2}$.
The variations of likely $\mathbf f_\rho$ from $\boldsymbol \rho$ itself are again $O(\tfrac1{\sqrt{n}})$ as in the argument presented in the main text (there the statement was phrased in terms of the number of $1$'s and not the type class or frequency distribution itself). Thus we may choose $\ell=(Rn^{3/2})$ to ensure that \eqref{eq:nudef} and \eqref{eq:unitdef} hold with $\mathbf f_\rho$ replaced with $\boldsymbol \rho$ and similarly for $\gamma$, at least to terms sublinear in $n$. We conclude that even in the worst, but still probable case, we have
\begin{align}
-\mathbf{x}\cdot\log\boldsymbol{\gamma}\leq D(\boldsymbol{\rho}||\boldsymbol{\gamma})-O(\tfrac{1}{\sqrt{n}}).
\end{align}
Now $\log\gamma =-\beta\mathbf{H}-\mathbf{1}\log Z$, so this condition becomes
\begin{align}
\beta\mathbf{H}\cdot\mathbf{x}\leq D(\boldsymbol{\rho}||\boldsymbol{\gamma})-O(\tfrac{1}{\sqrt{n}}).
\end{align}
This equation gives the minimum amount of extractable work among all the likely frequencies, which is taken to be the target amount for the process. As the extraction is unitary for each frequency, and these correspond to disjoint quantum states, we can thus find a unitary for the entire input capable of generating
$\tfrac{1}{\beta}\left[D(\boldsymbol{\rho}||\boldsymbol{\gamma})-O(\tfrac{1}{\sqrt{n}})\right]$ units of useful work per input resource state, with probability greater than $1-O(\tfrac{1}{\sqrt{n}})$.
\section{Structure of the exhaust state}
When doing a transformation of $\rho^{\otimes n}$ into $\sigma^{\otimes m}$, at the end of the protocol we actually obtain $\sigma^{\otimes m} \otimes \pi_{k}$ and we trace out $\pi_{k}$, which lives in $k = \Omega(n)$ copies of the system. Although $\pi_n$ is usually far away, in fidelity, from many copies of a Gibbs state, we show that its reductions are very close to a Gibbs state. The main observation is that because $\pi_{n}$ should be useless for extracting more copies of $\sigma$ at a non-zero rate, we must have
\begin{equation}
S(\pi_k || \rho_{\beta}^{\otimes k}) \leq k^{1 - \delta},
\end{equation}
for $\delta > 0$. But by subadditivity of the entropy we have
\begin{eqnarray}
S( \pi_k || \rho_{\beta}^{\otimes k}) &=& - S(\pi_k) - \sum_{l=1}^{k} {\rm Tr}(\rho_{k, l}\log \rho_{\beta}) \nonumber \\ &\geq& - \sum_{l=1}^{k} S(\pi_{k, l}) - \sum_{l=1}^{k} {\rm Tr}(\rho_{k, l}\log \rho_{\beta}) \nonumber \\ &=& \sum_{l=1}^{k} S(\pi_{k, l} || \rho_{\beta}),
\end{eqnarray}
where $\pi_{k, l} := {\rm Tr}_{\backslash l} (\pi_{k})$ is the reduced state of $\pi_k$ that is obtained by partial tracing all the systems except the $l$-th one. Let us assume for simplicity that all the $\pi_{k, l}$ are identical. Then
\begin{equation}
S(\pi_{k, 1} || \rho_{\beta}) \leq k^{-\delta},
\end{equation}
which by Pinsker's inequality implies
\begin{equation}
\Vert \pi_{k, 1} - \rho_{\beta} \Vert_{1} \leq \Omega(k^{-2 \delta}).
\end{equation}
More generaly, repeating the same argument for larger blocks we get that
\begin{equation}
\Vert {\rm Tr}_{L, L+1, ..., k}\left( \pi_{k} \right) - \rho_{\beta}^{\otimes L} \Vert_{1} \leq \Omega(Lk^{-2 \delta}).
\end{equation}
\section{Equivalence and degree of control for thermal operations }
Here, we address two questions. The first is how our paradigm, where we use unitaries $V$ which commute with the total Hamiltonian $H$, relates to other
approaches. The second is how much control an experimenter needs over the choice of unitaries $V$. To answer the first question, consider a common
approach to thermodynamics, which is to manipulate thermodynamical systems using an external apparatus. In this model, the systems are manipulated using
a time-dependent Hamiltonian, $H(t)$. Another approach is to add an interaction term $H_{\rm int}$ between various systems we are
trying to manipulate (e.g. the resource, and the heat bath), and then bring these systems into contact with one another. Let us now see that these are equivalent
to considering unitaries $V$ which commute with the original Hamiltonian $H$.
First, observe that in the case of a time-dependent Hamiltonian $H(t)$, we can simply include the clock as one of our systems. Letting $\tau$ be the coordinate operator of the clock system and $\Pi_\tau$ such that $[\tau,\Pi_\tau]=-i$, define $H_{\rm indep}=H(\tau)+\Pi_\tau$. The $\tau$ observable faithfully records the time $t$, as can be seen by solving the Heisenberg
equations of motion to get $\tau(t)=\tau(0)+ t$. Now consider a joint density matrix for the system plus clock of the form $\xi(t)=\rho(t)\otimes \proj t$, where $\ket{t}$ are the eigenstates of $\tau$. The time-independent Hamiltonian $H_{\rm indep}$ acting on the state $\xi(t)$ will generate the equation of motions of $H(t)$ acting on $\rho(t)$, but will also conserve energy.
To see this, recall that the product rule of derivatives gives
\begin{align}
\frac{d\xi(t)}{dt}=\frac{d\rho(t)}{dt}\otimes\proj{t}+\rho(t)\otimes \frac{d}{dt}\proj{t},
\end{align}
while the Heisenberg equation of motion gives
\begin{align}
\frac{d\xi(t)}{dt}&=i[H_{\rm indep},\xi(t)]\nonumber\\
&=i[H(t),\rho(t)]\otimes\proj{t}+\rho(t)\otimes[\Pi_\tau,\proj{t}]
\end{align}
Comparing the above two equations we have $\dot\rho(t)=i[H(t),\rho(t)]$ as claimed. That a system in a pure state stays in a pure state, can be achieved by having the clock have a large coherent superposition over energy levels, thus the change in it's state can be made arbitrarily small, as explained in Section \ref{sec:form_arb}.
We thus can go from
a picture with a changing Hamiltonian, to one with a fixed one. The model with time-dependent Hamiltonian is therefore equivalent to the one considered here, with fixed Hamiltonian.
Likewise, in the case where an interaction term is added, we can take the
total Hamiltonian to be $H_{\rm tot}=H+H_{\rm int}$ and assume that initially, $(H+H_{\rm int})\ket\psi\approx H\ket\psi$ i.e.\ the systems are initially far apart.
They can then evolve unitarily, such that the systems interact, and then move far enough apart that the interaction terms are negligible again.
In such a picture, an eigenstate of the initial Hamiltonian $H$ will evolve into an eigenstate of $H$ with the same energy
(by conservation of energy, and the fact that the interaction is negligible at initial and final times). Thus, all that happens here is that
eigenstates of fixed energy evolve to other eigenstates of the same energy, and this can be accomplished by means of a fixed Hamiltonian $H$
and a unitary $V$ which commutes with it. We thus see that also the picture of adding interaction terms
is equivalent to having a fixed Hamiltonian $H$, and operations $V$ which commute with it.
Similarly, the application of a unitary during some time period can be made via application of a fixed Hamiltonian.
One can include an internal clock $\tau$ which merely acts as a catalyst and thus
have some fixed Hamiltonian
\begin{align}
H_{\rm tot}=H+H_{\rm int}g(\tau)+\Pi_\tau
\label{eq:unitarytoham}
\end{align}
which effectively implements $e^{-iH_{\rm int}t}$ over some time interval determined by the function $g(\tau)$.
Here $\Pi_\tau$ is conjugate to $\tau$ and one can verify via the Heisenberg equations of motion that $\tau$ depends linearly on $t$.
Since $[H_{\rm int},H]=0$ one can also verify via the Heisenberg equations of motion for $\Pi_\tau$ that there is no
backreaction or energy exchange to the clock at late times provided $g(\tau)$ is chosen such that $\int_{t_i}^{t_f} g'(\tau)=0$ and
$g(\tau)$ and $\tau(t)$ chosen such that $g(\tau)=0$ before $t=t_i$ and after $t=t_f$.
One might be concerned that if we perturb the Hamiltonian slightly, our work extraction will not be robust. To see this,
let us consider the case where we don't succeed in implementing our unitary exactly, but rather some $H_{\rm int}$ which does not completely
commute with the original Hamiltonian $[H,H_{\rm int}]=-i\delta$. Viewed internally, will see that this is equivalent to allowing a violation of conservation of energy by amount $\delta$ -- something which is interesting to study in its own right. Now the equations of motion for $\tau$ are unchanged, and thus
the unitary $e^{-iH_{\rm int}t}$ is still implemented. However there is some backreaction on the clock. Solving the Heisenberg equations of motion for
$\Pi$, we find that there is a small momentum kick to the clock
\begin{align}
\Pi(t_f)-\Pi(t_i) &=\int^{t_f}_{t_i} g'(t) H_{\rm int}(t)\nonumber \\
&= H_{\rm int}(t_f)- H_{\rm int}(t_i)
\end{align}
where we have taken $g(\tau)$ to be $1$ between $t_i$ and $t_f$ and $0$ everywhere else. At each cycle some amount of energy gets stored in the clock, depending
on the initial and final states of the system. If we allow these to fluctuate, then the transfer is some $\delta$ and the clock undergoes a random walk.
We thus find that if we run the extraction process as a cycle, where we repeat the process over several cycles, then at each cycle we still exactly
implement $e^{-iH_{\rm int}t}$, it's just that it now has some tiny non-commuting part with the original Hamiltonian, resulting in some energetic backreaction to the clock, and thus some entropy being stored in
the clock. However, the extracted work grows linearly, and $\delta$ can be made arbitrarily small. After $n$ cycles,
the moment of the clock has undergone a random walk, of order $\sqrt{n}$, an amount which is negligible compared to the extractable work in the
case of many cycles.
Let us now turn to the second question. It might appear that an experimenter who wished to implement our protocols would need
to very carefully manipulate all the many degrees of freedom of the $n$ systems and the heat bath. However, this is not the case, as we will now demonstrate explicitly using the example of work distillation. There, we were mapping eigenstates which had a type $lq$ on the heatbath $\gamma$, and $pn$ on the resource $\rho$
to microstates which had type $rk$ on the garbage $\sigma$, and $m$ $1$'s on the work system. However, although for any implementation
of the protocol, we need a particular mapping of strings (i.e.\ microstates) of these initial types, to strings of the final type,
{\bf any mapping} will do. The only important thing which is required is just that the unitary operation
map the initial types to the final types. Thus an experimenter who wishes to
implement the protocol, does not need fine-grained control over the mapping of microstates within one type to microstates within another type.
She only needs to know that the unitary maps one type into another. In other words, there are an exponentially large number of
possible implementations of our protocols each of which
map particular strings within the initial types to particular strings in the final types. However, it doesn't matter which implementation is chosen,
and the experimenter thus does not need the fine degree of control that is required to achieve a particular implementation.
We can think of the type as being like a macroscopic variable such as the total magnetisation of a composite system, or its total energy (indeed it is the latter).
In the distillation protocol, we map the macroscopic variables of energy on two large systems ($\gamma^{\otimes l} $ and $\rho^{\otimes n}$) to the macroscopic
variable of energy on the final system. Any unitary which accomplishes this will successfully implement our protocol. Thus, the experimenter only
needs control over the macroscopic variables, not the microscopic ones.
Equivalence of these paradigms is discussed in more detail, and in the case of finite systems, in \cite{HOcrude}.
|
\section{Decision Tree}
This section is devoted to study the efficiency of the decision tree algorithm described in the main text.
Some results in this section are analytical and some are numerical.
In all our numerical investigations (unless explicitly stated otherwise) we used the decision tree of the main text (for two qubits)
and in cases when going through the whole tree did not reveal entanglement
we augmented it with additional measurements of those correlations which were not performed until that moment.
The order of the additional measurements also results from the correlation complementarity (anti-commutation relations)~\cite{CORRELATION_COMPLEMENTARITY}.
With every remaining measurement we associate the ``priority'' parameter
\begin{equation}
P_{ij} = \sum_{k \ne i} P_{ij}(T_{kj}) + \sum_{l \ne j} P_{ij}(T_{il}),
\end{equation}
that depends on the measured correlation tensor elements of the decision tree in the following way
\begin{equation}
P_{ij}(T_{mn}) = \Big\{
\begin{array}{rl}
T_{mn}^2 & \textrm{ if } T_{mn} \textrm{ was performed before,} \\
0 & \textrm{ else.}
\end{array}
\end{equation}
According to the correlation complementarity there is a bigger chance that this correlation is significant if the value of the corresponding parameter is small.
Therefore, the correlations $T_{ij}$ with lower values of $P_{ij}$ are measured first.
Let us consider the following example.
The measured correlations of the decision tree are as follows: $T_{zz}=0.7$, $T_{xx}=0.1$, and $T_{yy}=0.4$.
Therefore, $P_{xy} = P_{yx} = T_{xx}^2 + T_{yy}^2 = 0.17$ , $P_{xz} = P_{zx} = T_{xx}^2 + T_{zz}^2 = 0.5$, and $P_{zy} = P_{zy} = T_{zz}^2 + T_{yy}^2 = 0.65$.
Accordingly the order of the remaining measurements is as follows:
first measure $xy$, then $yx$, next $xz,zx,yz$ and $zy$.
\subsection{Two qubits}
\subsubsection{Werner states}
As an illustration of how the decision tree works for a well-known class of mixed states we first consider Werner states.
It turns out that not all entangled states of the family can be detected.
Consider a family of states
\begin{equation}
\rho = p \proj{\psi^-} + (1-p) \frac{1}{4} \openone,
\label{WERNER_STATE}
\end{equation}
where $\ket{\psi^-} = \frac{1}{\sqrt{4}}(\ket{01} - \ket{10})$ is the Bell singlet state,
$\frac{1}{4} \openone$ describes the completely mixed state of white noise, and $p$ is a probability.
Its correlation tensor, written in the same coordinate system for Alice and Bob, is diagonal with entries $T_{xx} = T_{yy} = T_{zz} = -p$, arising from the contribution of the entangled state.
The states (\ref{WERNER_STATE}) are entangled if and only if $p > \frac{1}{3}$,
whereas the decision tree reveals that these states are entangled for $p > \frac{1}{\sqrt{3}} \approx 0.577$.
This is because only three elements contribute to the criterion.
Note that the value of the decision parameter is irrelevant here.
Furthermore, a random choice of local coordinate directions does not help.
Although more steps would be involved in the decision tree the sum of squared correlations is invariant under local unitary operations
and therefore only for $p > \frac{1}{\sqrt{3}}$ entanglement is detected for the Werner state.
\subsubsection{Entangled state mixed with colored noise}
An exemplary class of density operators for which the decision tree detects all entangled states is provided by:
\begin{equation}
\gamma = p \proj{\psi^-} + (1-p) \proj{01},
\end{equation}
where entanglement is mixed with colored noise $\ket{01}$ bringing anti-correlations along local $z$ axes.
This state has the following non-vanishing elements of its correlation tensor $T_{xx} = T_{yy} = -p$ and $T_{zz} = -1$.
Therefore, the decision tree allows detection of entanglement for this class of states in two steps.
Note that the state is entangled already for an infinitesimal admixture of the Bell singlet state as can be shown for a random choice of local coordinate systems.
\subsubsection{Random mixed states}
Fig. \ref{SI_FIG_RANDOM_MIXED} shows how the efficiency of the algorithm grows with the purity of tested states.
The efficiency is measured by the fraction of detected entangled states obtained from extensive Monte Carlo sampling in the two qubit state space.
For nine steps the algorithm detects all the pure states. This is expected because Eq. (1) of the main text is a necessary and sufficient condition
for the detection of entanglement of pure states.
\begin{figure}[h]
\includegraphics[width=0.40\textwidth]{SD-2-mixed.pdf}
\caption{Efficiency of the decision tree for two qubit random mixed states.}
\label{SI_FIG_RANDOM_MIXED}
\end{figure}
\subsection{Many qubits}
\subsubsection{Algorithm for generation of the tree}
The principle behind the decision tree is the correlation complementarity~\cite{CC}.
Correlation complementarity states that for a set of dichotomic anti-commuting operators $\{\alpha_1,\dots, \alpha_k \}$
the following trade-off relation is satisfied by all physical states:
\begin{equation}
T_{\alpha_1}^2 + \dots T_{\alpha_k}^2 \le 1,
\end{equation}
where $T_{\alpha_1}$ is the average value of observable $\alpha_1$ and so on.
Therefore, if one of the average values is maximal, $\pm 1$, the other anti-commuting observables have vanishing averages.
This motivates taking only sets of commuting operators as different branches of the decision tree.
\subsubsection{Decision tree for three qubits}
This tree is an example of the application of the algorithm presented in the previous section.
Fig. \ref{DT_3QUBITS} shows only one branch of the whole tree.
A numerical simulation reveals that the correlation measurement along local Bloch vectors
gives correlations close to the maximal correlations of a pure multi-qubit state in more than $80\%$ of the cases.
Therefore, these local directions give an excellent starting point of the decision tree.
Exemplarily, the branch begins with $T_{xxx}$ assuming that this correlation is big.
If the local Bloch vectors indicate correlation along a different directions
it is advantageous to correspondingly change the elements of the decision tree (see main text).
\begin{figure}[h]
\includegraphics[width=0.42\textwidth]{dt-3.pdf}
\caption{One branch of the decision tree for three qubits.}
\label{DT_3QUBITS}
\end{figure}
|
\section{Introduction}
\noindent
With increased precision of lattice calculations it becomes necessary to investigate and
to remove the approximations that are often made. One such approximation is the use of unphysically heavy sea quark masses, another one is the omission of electromagnetic effects. From the mass differences of charged and uncharged pions or between the nuclei one would expect these to be of the order of one to a few MeV.
Recent lattice calculations \cite{precDs1,precDs2} of pseudoscalar $D_s$ decay constants have reached an accuracy of one percent, a regime where electromagnetic effects may become significant. The systematic errors stated
in these calculations include an estimate of QED effects that is based on electromagnetic shifts of
the $D_s$ meson mass \cite{precDs2}.
The methods initially introduced in \cite{Duncan96} make it possible to include QED effects on the lattice explicitly and thereby enable us to differentiate between isospin breaking through charge and different up and down quark masses. These techniques have been successfully applied in calculations of the light hadron spectrum \cite{Duncan96,Blum2010,Portelli2010} and used to estimate the $u$,$d$ mass difference. We
deviate from these references, using a compact QED action, to
investigate electromagnetic effects on pseudoscalar decay constants.
\section{Methods}
\noindent
Electromagnetic effects on the lattice are included by multiplying the QCD SU(3)-links with QED U(1)-links, which are given by $U_{QED,\mu}(x)=\exp\left( \mathrm{i} e B_{\mu}(x) \right)$, where $e$ is the charge of the corresponding quark. The calculation of fully unquenched SU(3)$\times$U(1)$\rightarrow$U(3) configurations is practically unfeasible and unnecessary since sea quark charge effects are suppressed by an
additional factor of $\alpha_{QED}\approx 1/137$. Instead, we use quenched QED configurations
and multiply these with $N_{\text{f}}=2$ QCD configurations. This is correct up to $\mathcal{O}(\alpha_{QED})$.
\begin{figure}
\centering
\includegraphics[width=0.88\textwidth]{./vr.pdf}
\caption{Coulomb potential: measured data and prediction of lattice pertubation theory.}
\label{fig:vr}
\end{figure}
Instead of using a noncompact action as in previous lattice studies of electromagnetic effects \cite{Duncan96,Blum2010,Portelli2010}, we employ the compact Wilson gauge plaquette action, so we
can deal with these degrees of freedom on the same footing as with the SU(3) gauge links.
Starting from a cold (unit link) lattice, we employ the heatbath algorithm to obtain
the $B_{\mu}(x)$ phases by rescaling random variables $\theta_x$ that are distributed
according to $P(\theta_x) \sim \exp( w_x \cos \theta_x)$ where the weights $w_x$ depend on the coupling and on the local staple. We follow the procedure described in \cite{Hattori92}. The lattice is divided into 4-cubes with side length 2 and our algorithm loops over single sites of every 4-cube simultaneously. We find integrated autocorrelation times of the plaquette of $\mathcal{O}(1)$. Nonetheless, to be on
the safe side, only every 200$^{\text{th}}$ QED configuration is used. The first 3000 configurations are
discarded such that the ensemble is sufficiently thermalized. The improvement coefficient of the fermionic
QCD Sheikholeslami-Wohlert action $c_{sw,QCD}$ has been determined nonperturbatively.
Since the QED interactions are quenched this coefficient will not be affected by them (and in the
unquenched case it would receive an $\mathcal{O}(\alpha_{QED})$ shift only).
To be consistent to leading order in the QED coupling we set $c_{sw,QED}=1$.
We treat the U(1)-links as a background and set their coupling to the physical value.
Because quark charges are multiples of one third of the positron charge $e_{p^+}$, we rescale the coupling such that the smallest (nonzero) charge is $e = \frac{e_{p^+}}{3}$. This means that $\beta^R\approx 3^2/(4\pi\alpha_{QED})\approx 98$.
This can then easily be rescaled to $e_q\,e$ with $e_q=\{-1,\pm2\}$ to calculate propagators of other charges.
We use the bare value $\beta=99$ (that corresponds to the above renormalized coupling)
for the generation of
the U(1)-links. The previously mentioned heatbath algorithm deals well with the resulting,
very narrow distribution $P(\theta_x)$. Since we are using a compact formulation of the
action there are unphysical 4-point and higher order vertices, which renormalize the coupling.
This is no problem since this effect is purely multiplicative and does not introduce any
running. This means that $\beta^R=Z\beta$, with a renormalization factor
$Z=1+\mathcal{O}(1/\beta)$. We determine this factor nonperturbatively by
comparison between the static lattice QED potential and the analytic perturbative expectation
that, without Fermions, should be exact in noncompact QED. This is visualized in Fig.~\ref{fig:vr}
for $\beta=6$.
From the multiplicative constant that is fitted, we determine $Z\approx 0.99$ at our
$\beta$-value.
To suppress large finite-size effects for correlation functions of charged particles, that
are not gauge invariant with respect to U(1), zero modes of the electromagnetic field need to be
subtracted \cite{QEDFiniteSize08}. This is achieved by a global gauge transformation \cite{Bogolubsky99}.
In addition, we fix the U(1)-links to Landau gauge, to improve the signal.
Including electromagnetic effects additively renormalizes the mass of charged quarks in the
Wilson formulation. We choose to set the quark masses in such a way that their renormalized
values are (approximately) the same. The strange $\kappa$ is fixed by tuning the mass of the
hypothetical $s\bar{s}$ pseudoscalar to the experimental value,
\begin{align}
m^2_{K^0} + m^2_{K^+} - m^2_{\pi^\pm} \simeq (690\unit{MeV})^2 = m^2_{s\bar{s}}.
\end{align}
This is done for the charges $e_\text{strange}=\{0,\pm1,\pm2\}$. The charm mass parameter ($\kappa_{\text{charm}}$) is only determined for $e_\text{charm}=\{0,\pm2\}$.
It is fixed by tuning $D_s=c\bar{s}$ to its physical mass for uncharged charm and strange quarks as well as for $e_\text{charm}=2$ and $e_\text{strange}=-1$. To check the tuning of the charm $\kappa$s we also calculated the mass combination $M_\text{charm} = \frac{1}{4}\left(3 m_{J/\Psi}+m_{\eta_c}\right)$.
\begin{center}
\begin{tabular}{|c|c|}
\hline
$e_\text{charm}$ & $M_\text{charm}$ \\
\hline
exp. & $3.0678(3) \unit{GeV}$ \\
0 & $3.002(2) \unit{GeV}$ \\
$\pm2$ & $3.024(2) \unit{GeV}$ \\
\hline
\end{tabular}
\end{center}
The difference between the 3.024~GeV and the experimental value may be attributed to the unphysical
sea quark content and finite size effects. However, the difference between our two
calculations shows that there is some inconsistency in the tuning of the charm quark mass between
the two procedures of about 10~MeV. This is no surprise since the total charges of the two mesons differ. We regard tuning of the $\bar{c}c$ combination
the cleaner procedure since this does not depend on the strange quark mass and we will implement this in the future.
We also simulate 2 lighter quarks of all five charges to enable extrapolations to physical light quark masses
for pions and for the $D^{0,\pm}$. We tune the symmetric and therefore uncharged combinations $l\bar{l}$ for all charges $e_l=\{0,\pm1,\pm2\}$ to the same values $m_{l\bar{l}}^2$. To reduce the noise we average over $\pm B_{\mu}$ \cite{Blum2010} which is equivalent to averaging over $\pi^\pm$. Our $\kappa$-values are listed in the table below.
\begin{center}
\begin{tabular}{|c|llll|}
\hline
$e_q$ & $\kappa_{l_1}$ & $\kappa_{l_2}$ & $\kappa_\text{strange}$ & $\kappa_\text{charm}$ \\
\hline
0 & 0.13629 & 0.136013 & 0.135676 & 0.123019 \\
$\pm1$ & 0.136337 & 0.13606 & 0.135722 & - \\
$\pm2$ & 0.136477 & 0.136199 & 0.135861 & 0.123086 \\
\hline
\end{tabular}
\end{center}
We use spin-explicit, complex $Z_2$ random wall sources \cite{OneEndSEM2008}. This enables us to average over the spatial volume. We use 3 noise sources per configuration and analyze a total of 200
$N_\text{f}=2$ nonperturbatively improved Sheikholeslami-Wohlert configurations \cite{QCDSFgen} at $\beta=5.29$, $\kappa=0.1355$ on a $24^348$ volume. The sea pion mass reads $m_{\pi,sea}\simeq 750\unit{MeV}$ and
we use $a=0.086\unit{fm}$ as our lattice spacing. All 2-point functions are calculated for a local and a Wuppertal smeared sink. Wuppertal smearing was done using the product of separately APE smeared QCD- and QED-links.
The decay constants are extracted from $\pi\pi$ and $\pi A_4$ correlators and have been improved and renormalized according to \cite{Gockeler:1997fn}, neglecting $\alpha_{QED}$ corrections since we consistently work at
$\mathcal{O}(\alpha_{QED})$. We are using the Chroma software system \cite{CiteChroma} which has been modified to support QCD+QED calculations.
\section{Analysis}
\noindent
The available data can also be used to extract the up/down quark masses and the pion mass splitting, which is done to verify the correctness of the approach. To do this all pseudoscalar mass data must be fitted to a modified chiral expression which takes QED into account. Instead of using the simple formula \cite{Duncan96}
\begin{align}
m_{PS}^2 = A_0 Q^2 + \left(B_0 + B_1 Q^2\right) \left( m_q + m_{\bar{q}} \right),
\label{eq:totalChiPT}
\end{align}
where $Q=e_q+e_{\bar{q}}$, to fit charged and uncharged data, we alternatively subtract the uncharged mesons (with the parameters obtained by the $\kappa_c$ fit) and fit this difference to
\begin{align}
\Delta m_{PS}^2 = A_0 Q^2 + B_1 Q^2 \left( m_q + m_{\bar{q}} \right).
\end{align}
Once the fit parameters $A_0,B_0$ and $B_1$ are determined the experimental values of $m_{K^{0,\pm}}$ and $m_{\pi^\pm}$ can be used to determine the masses of up, down and strange quarks. The physical $\pi$ mass difference is retrieved by reinserting all this information into eq. (\ref{eq:totalChiPT}).
The result is in good agreement with experimental value (Tab. \ref{tab:qmassdiff}) despite a relatively large $\chi^2/$dof$=4.5$. We plan to address this in future studies, for example by including higher order terms in the chiral fit.
\begin{table}
\begin{center}
\begin{tabular}{|c|cc|}
\hline
& lattice result & experimental \\
\hline
$m_d/m_u$ & $1.80(4)$ & - \\
$\Delta m_{\pi}$ & $4.4(8)\unit{MeV}$ & $4.5936(5)\unit{MeV}$ \\
\hline
\end{tabular}
\end{center}
\caption{Quark masses and the pion mass differences. The errors are only statistical.}
\label{tab:qmassdiff}
\end{table}
The calculated decay constants differ by about 10~\% from the
experimental results, which is not surprising since this is only a partially quenched study on an ensemble with a relatively large sea quark mass, at one lattice spacing. By computing differences between differently charged pseudoscalars at the same mass scale we are able to extract
the QED contributions. However, the decay constant will depend on the mass and this will depend on the
charge of the particle. To disentangle these two effects we perform the comparison, matching the
uncharged squared mesons masses $m^2_{PS,light}(Q=0)$ of each of the two constituent quarks.
As an example consider the positively charged Kaon, which consists of a light quark $l$ with charge $2$ and a strange antiquark with charge $1$: The corresponding scale would be
$m^2_{PS,light}(Q=0)=\frac{1}{2}( m^2_{l\bar{l}} + m^2_{s\bar{s}} )$. This is equivalent to using
the average light quark mass, but more correlations between the data remain, because the
$\kappa_{crit}(e)$-values do not have to be determined. Preserving these correlations is necessary
if differences significantly smaller than the noise of the absolute signals are extracted.
In the case of the charm quark the Gell-Mann-Oakes-Renner relation does not apply anymore
so that we cannot perform chiral fits and moreover, it would not be sensible to match
squared pseudoscalar masses. The charm quark matching has not been done as yet so that in this case we cannot
compare the decay of the physical $D_s$ meson to that of a $D_s$ with electrically neutral
quarks. However, we can compare the decay constants obtained for combinations of the
$e_{\text{charm}}=2$ charm quark with differently charged light quarks.
In the case of light mesons we restrict ourselves to (approximately) equal valence
quark masses and compare the other decay constants with $f_{\pi^0} = \frac{1}{2}(f_{u\bar{u}} + f_{d\bar{d}})$, where $u$ and $d$ quark masses are approximately equal and have the appropriate (physical) charges. Neglecting disconnected loops for $\pi^0$ is correct to $\mathcal{O}(\alpha_{QED}^2)$. The effects of isospin breaking by different $u$ and $d$ quark masses have not been investigated. Under these conditions we are able to extract the difference $f_{\pi^0}-f_{\pi^\pm} = 0.09(3)\unit{MeV}$ from a fit to correlated data. This estimate is already more precise than the experimental value.
\begin{figure}
\centering
\subfigure[Pions]{
\label{fig:fll}
\includegraphics[width=0.47\textwidth]{./diff_ll.pdf}
}
\subfigure[charmed pseudo scalars]{
\label{fig:fhl}
\includegraphics[width=0.47\textwidth]{./d0_dplus.pdf}
}
\caption{Differences of the decay constants between uncharged and charged mesons. All errors are only statistical.}
\end{figure}
In the case of the $D$ mesons we are limited to the difference between the $D^0$ and charged $D^\pm$: $f_{D^0}-f_{D^\pm}=0.79(11)\unit{MeV}$. If the light mass is set to the strange mass, we can define the QED contribution as the difference between the charged $D_s$ and a hypothetical, uncharged $D_s^0$ meson, where the antistrange has charge $e_{\bar{s}}=-2$: $f_{D_s^0}-f_{D_s}=0.95(4)\unit{MeV}$. The order of magnitude and the sign of our result agree with naive expectations.
\section{Conclusion}
\noindent
We are able to resolve differences between decay constants of differently charged mesons
of about 0.1~MeV. We determine the electromagnetic effect on $D$, $D_s$ and $\pi$ decay constants.
The difference between charge and uncharged $\pi$ decay constants is more precise than experimental
results. We find $f_{D_s}$ to decrease by about 1~MeV, due to electromagnetic effects.
\section*{Acknowledgments}
\noindent
We thank
J. Najjar, S. Collins and L. Castagnini
for their help.
This work was supported by the GSI Hochschulprogramm (RSCHAE), the European Union
under Grant Agreement number 238353 (ITN STRONGnet)
and by the Deutsche
Forschungsgemeinschaft SFB/Transregio 55.
Computations were performed on Regensburg's Athene HPC Cluster.
|
\section{Introduction}
\begin{figure}[br]
\centering
\includegraphics[scale=0.5]{q-scans_6meV.png}
\caption{(color online) Scans through [H, 0, \nicefrac{1}{2}] with E = 6 meV, taken above and below \sampTN. The black horizontal bar shows the estimated resolution width. The shift of the spin fluctuations from the commensurate to incommensurate position (when warming from the ordered to paramagnetic state) has not been previously reported.}
\label{fig:H-scan}
\end{figure}
Superconductivity in the recently-discovered iron-based superconductors (FeSCs) \cite{Kamihara2008} often appears at high transition temperatures. These compounds contain a simple square lattice of iron atoms, coordinated with pnictogen or chalcogen atoms forming planes of tetrahedra. Interplanar layers differ between families, consisting of either metal oxides, alkaline earth atoms, alkali atoms, or nothing at all as in the ``11'' compounds on which we focus here. The parent compounds of most families become orthorhombic and antiferromagnetic (AFM) at low temperatures, and superconductivity can be induced by doping with electrons, holes, isoelectronically, or by the application of pressure (see, e.g., \cite{Paglione2010}). Their high-\Tc\ superconductivity may be related to the magnetic order \cite{Mazin08}, which seems to be the result of Fermi surface nesting \cite{Singh08}. There is also intrinsic interest in magnetism in these materials, but it has not been as extensively explored.
While nesting explains many experiments, the magnetic order in the $x = 0$ endpoint of the Fe$_{1+y}$Te$_{1-x}$Se$_{x}$\ series is a highly unusual commensurate ``bicollinear'' magnetic structure \cite{Bao09} with a wavevector along $\mathbf{Q}_{AFM} = (0.5, 0, 0.5)$. Neither density functional theory (DFT) \cite{Subedi08} nor photoemission measurements \cite{Xia2009} find any evidence for nesting at this wavevector, which indicates the presence of local moments. These moments must interact with each other via competing exchange interactions \cite{Turner09}, which should lead to incommensurate order. The mechanism behind the commensurate bicollinear order is one of the main unresolved issues in the effort to understand the FeSCs.
Here we report results of a detailed inelastic neutron scattering (INS) investigation of the region near $\mathbf{q}_{AFM}$\ in a high-quality single-crystal sample of Fe$_{1.08}$Te. We find strong evidence for competition between commensurate and incommensurate ordering in the form of spin excitations which abruptly shift from the incommensurate \Qinc\ $\approx$ [0.45, 0, 0.5] to the commensurate wavevector $\mathbf{Q}_{AFM} = (0.5, 0, 0.5)$\ when passing below the N\'eel temperature \TN\ (see Fig. \ref{fig:H-scan}). We interpret this unusual behavior in terms of a lock-in transition driven by crystalline anisotropy, which causes bicollinear order.
The INS measurements were performed on a single crystal of Fe$_{1.08}$Te\ ($\approx$ 0.1 cc, $\approx$ 2 g, mosaic spread 2$^{\circ}$). It was grown by the Bridgman technique, and the excess iron content was determined by Patterson refinement of single-crystal x-ray diffraction data to be $y$ = 0.08. The N\'eel temperature was defined as the steepest slope of the magnetic Bragg peak intensity with temperature, was found to be \sampTN, consistent with other reports \cite{Liu10, Lipscombe11, StockRodriguez} for this value of $y$. The crystal structure at room temperature is tetragonal (space group 129, P4/nmm) with lattice constants $a$ = $b$ = 3.823(3) \AA\ and $c$ = 6.282(6) \AA. Although this compound becomes monoclinic (space group 11, P2$_1$/m) at low temperature \cite{Li09-1}, the primary effect is a shift of the Te atoms within the unit cell \cite{Bao09}; the rotation of the $c$-axis away from $90^\circ$ is quite small, and amounts to a broadening of certain Bragg peaks (see Fig. \ref{fig:mag_vs_temp}b), so we describe the measurements in the tetragonal [HHL] notation.
\begin{figure}[h]
\centering
\includegraphics[scale=0.45]{magnetic_vs_temp.png}
\caption{(color online) Temperature dependence of the magnetic and structural behavior.
{\bf (a)} Peak intensities of the incommensurate magnetic excitation and the [1.5, 0, 0.5] magnetic Bragg peak as a function of temperature. Inset: map of reciprocal space, showing the locations of the scans.
{\bf (b)} Width of the [0, 0, 4] nuclear Bragg peak (an indicator of monoclinic splitting) as a function of temperature, showing the structural transition at 67.5 K. This is plotted together with the intensity of the [2.5, 0, 0.5] magnetic Bragg peak, demonstrating the close coincidence between the structural and magnetic transitions.}
\label{fig:mag_vs_temp}
\end{figure}
The neutron measurements were performed on the 1T1 triple-axis spectrometer at the Laboratoire L\'eon Brillouin, Saclay, France. The sample was mounted in a standard displex cryostat (base temperature T = 11 K) in the [H0L] plane. The measurements were done with vertically and horizontally focusing PG crystals as monochromator and analyzer, respectively. A graphite filter was used to reduce $\lambda$/2 contamination. The scans were performed using a fixed final wavevector of \kf\ = 2.662 \AA\ (\Ef\ = 14.7 meV), corresponding to an energy resolution of 0.8 meV at the elastic line. Fitting was done using the Fityk program \cite{Wojdyr10}. All constant-$\mathbf{q}$\ scans were fit with a background function (either constant or linear) and Gaussians for the peaks. Elastic scans were fit with a constant background and Voigt profile.
We measured inelastic and elastic magnetic scattering from our sample near $\mathbf{q}_{AFM}$, obtaining a detailed picture of the evolution of the magnetic order, the crystal structure, and the magnetic excitations as a function of temperature in the vicinity of the first-order phase transition from the high-temperature paramagnetic to the low-temperature AFM phase.
Figure \ref{fig:H-scan} shows constant-energy scans, at E = 6 meV, along [H, 0, \nicefrac{1}{2}] above and below \TN. At 50 K the peak is commensurate and comparable in width to the resolution function, whereas at 70 K it is broad and incommensurate.
Figure \ref{fig:mag_vs_temp}a details the peak scattering intensity at 2 meV, which is well below the low-T spin gap, taken along the [H, 0, \nicefrac{1}{2}]-direction. This is plotted together with the temperature dependence of the magnetic Bragg peak. The phase transition at
67.5 K
is clearly evident: the intensity at 2 meV builds up on cooling towards the phase transition, then drops abruptly as the spin gap opens and the magnetic Bragg peak appears. The rapid increase of magnetic intensity is consistent with a first-order phase transition. We also measured the temperature dependence of the mosaic of the [004] lattice reflection, which is proportional to the amount of the monoclinic distortion. Figure \ref{fig:mag_vs_temp}b demonstrates that the magnetic transition is concurrent with the structural phase transition to the monoclinic phase.
Figure \ref{fig:energy_dep} shows $S(q,\omega)$, and summarizes the energy- and wavevector-dependence of the magnetic inelastic scattering above and below \TN. Above \TN\ it is incommensurate, broad, and ungapped. Below \TN\ it is commensurate, narrow, and gapped. Both have nearly vertical dispersion within the experimental uncertainty.
\begin{figure}[t]
\centering
\includegraphics[scale=0.45]{energy_dep.png}
\caption{ (color online) Energy dependence of the spin excitations.
{\bf (a,b)} $S(q,\omega)$ at T = 70 K (a) and 11 K (b). The data have been smoothed and background-subtracted. Intensity is plotted on a logarithmic scale (arb. units).
{\bf (c)} Linewidths of the incommensurate (solid green circles) and commensurate (open blue diamonds) peaks, as a function of energy.
{\bf (d)} Constant-Q scans taken at the center of the incommensurate and commensurate spin excitations.}
\label{fig:energy_dep}
\end{figure}
The temperature dependence of the signal with L is shown in Fig. \ref{fig:L-scan}. The magnetic scattering at 2 meV is nearly absent at 65 K, because of the spin gap opening at the phase transition. Upon heating in the paramagnetic phase, the signal broadens along the L-direction, although the integrated intensity stays roughly the same. The signal broadens only slightly in the H-direction (not shown). Based on this observation, we conclude that the buildup of intensity in H-integrated scans at 2 meV towards \TN\ shown in Fig. \ref{fig:mag_vs_temp} results from an increase in the magnetic correlation length along the $c$-axis, as opposed to an increase of total scattering.
\begin{figure}[b]
\centering
\includegraphics[scale=0.65]{L-scans.png}
\caption{ (color online) Scans along the $c$-axis direction. The black horizontal bar shows the estimated FWHM of the resolution function. The integrated intensity is approximately constant for temperatures above \TN, indicating that the number of spins is constant, but the correlation length decreases with T.}
\label{fig:L-scan}
\end{figure}
The most important feature of the observed behavior is that incommensurate fluctuations give rise to commensurate order, or more generally, fluctuations appear at one wavevector while ordering occurs at another.
Our measurements demonstrate that, somewhat surprisingly, the shift in the inelastic peak is energy independent all the way up to the highest observed energy of 10 meV.
Existing theories do not provide a cromulent explanation for the observed behavior. According to the orbital scheme as discussed, for example, in Ref. \cite{Turner09}, in the tetragonal phase the highest energy electron can occupy either d$_{xz}$ or d$_{yz}$, orbitals which are otherwise empty. This orbital degeneracy implies that the entire spin system is free to rotate in the $x$-$y$ plane with no in-plane anisotropy. In such a model, the bicollinear spin order results from a combination of anisotropic exchange and the biquadratic spin interaction. However, the predicted gapless spin-wave dispersion is inconsistent with the large spin gap that we and others have observed \cite{Lipscombe11, StockRodriguez}. Using DFT, Ma \emph{et al}. \cite{Ma09} have found the lowest-energy \emph{commensurate} phase to be bicollinear, even when the lattice is tetragonal. They did not calculate whether incommensurate ordering is more favorable than commensurate (DFT requires calculations using large supercells to understand the competition between commensurate and incommensurate order).
Our results (as well as those of previous experiments) may be captured by a simple model \cite{Choi_Chen_Radzihovsky}, in which the spin rotation symmetry is explicitly broken by the single-ion anisotropy. The low-temperature monoclinic phase features both an orthorhombic distortion with shortening along the $b$-axis, as well as a monoclinic shearing of the Te-plans along the $a$-axis. Together these distortions break the symmetry along the $a$- and $b$-axes. Thus the crystal field environment becomes anisotropic, and the degeneracy between the d$_{xz}$ and d$_{yz}$ orbitals is lifted.
This single-ion anisotropy is at the heart of magnetoelastic coupling,
causing an Ising-type behavior in which the spins are locked along the $b$-axis, which in turn opens a gap in the spin-wave spectrum (in agreement with experiments). In the tetragonal or orthorhombic phase with incommensurate magnetic order, the exchange interaction becomes dominant, and the magnetoelastic coupling does not lift the degeneracy between the d$_{xz}$ and d$_{yz}$ orbitals. In such a Heisenberg spin system, the various Js set the periodicity of the spin system, but the spins are free to rotate with respect to the lattice, and thus the spin excitations are gapless, and the fluctuating moments should be isotropic. However, spin fluctuations in the related system BaFe$_{2-x}$Ni$_{x}$As$_{2}$\ were found to have a preferred axis \cite{Lipscombe10}, lending support to our interpretation of a Ising-type behavior in which the spin axis is coupled to the lattice.
\begin{figure}[b]
\centering
\includegraphics[scale=0.35]{anisotropy.png}
\caption{ (color online)
Schematic of the magnetic energy in Fe$_{1+y}$Te\ at zero temperature as a function of the ordering wavevector, for two values of the excess iron concentration $y$. The exchange energy, $E_{exchange}$, is represented as a parabola (dashed line); the structural anisotropy at the commensurate wavevector lowers the total energy, $E_{total}$, further by ${\Delta}E_{anisotropy}$. The system always settles at the global energy minimum, marked with an ``X''.
{\bf (a)} For values of $y <$ 0.12, the minimum of $E_{exchange}$ is close to the commensurate wavevector, and so the global minimum is at the \emph{commensurate} position. Above \TN\ the anisotropy well becomes filled, thus spin fluctuations appear at \Qinc\ = [0.45, 0, 0.5], the exchange energy minimum.
{\bf (b)} For $y >$ 0.12, the energy minimum for $E_{exchange}$ is far from the commensurate position, which puts the global minimum at an \emph{incommensurate} wavevector near \Qinc\ = [0.38, 0, $\nicefrac{1}{2}$] \cite{Bao09,StockRodriguez}.
}
\label{fig:anisotropy}
\end{figure}
Fig. \ref{fig:anisotropy} shows a schematic of the energy as a function of the ordering wavevector, \Qorder, and illustrates the competition between the high-temperature incommensurate paramagnetic and low-temperature commensurate magnetic states. In the tetragonal phase the interactions between the electrons favor an incommensurate phase whose wavevector is not too far from [$\nicefrac{1}{2}$, 0, $\nicefrac{1}{2}$]. This is represented by a dotted parabola on the energy vs. \Qorder\ plot, which has a minimum at an incommensurate wavevector. Here we are not concerned with the microscopic mechanism that causes the system to choose such an incommensurate wavevector (although such a wavevector is not difficult to achieve, given the interplay between different exchange, superexchange, and double-exchange pathways \cite{Turner09}).
However, the commensurate wavevector is special, because it allows the system to lower energy by an amount ${\Delta}E_{anisotropy}$ through the monoclinic lattice distortion, which in turn allows the magnetic order to align the moments according to the spin anisotropy. Thus $E_{total}$ has a dip at the commensurate $\mathbf{q}$. If that dip is close to the minimum of the exchange interactions, then it will be the global minimum, and the ground state will become bicollinear through the monoclininc distortion (this distortion is probably cooperative and related to the strong magnetoelastic coupling \cite{Paul2011-2}, since J$_1$ and J$_2$ will split under the monoclinic and orthorhombic distortions, respectively). At temperatures greater than the anisotropy gap ($\approx$ 6 meV, corresponding to 69 K), the spins are thermally excited above the gap, and the advantage it provides to the commensurate order disappears. Thus the incommensurate fluctuations we observe in the paramagnetic phase are the critical fluctuations leading up to a standard second-order transition, and in the absence of the lock-in transition the system would order at this incommensurate wavevector; but this transition is interrupted by the first-order transtition to the commensurate state.
This lock-in scenario is expected to take place in the low-Fe doping regime. When the excess iron doping is increased beyond the critical concentration $y \approx$ 0.12, the system enters a mixed-phase regime \cite{RodriguezStock, Zaliznyak11-2}, consistent with two minima (at the commensurate and incommensurate wavevectors). With further Fe-doping, the system becomes single-phase, and the incommensurability increases rapidly to \Qinc\ = [0.38, 0, $\nicefrac{1}{2}$] \cite{RodriguezStock}. The orthorhombic distortion is also a function of doping, and decreases beyond the critical doping \cite{Bao09}, which suggests the reduction of anisotropy.
This can be understood as a doping-driven (as opposed to temperature-driven) classical commensurate-incommensurate transition \cite{Choi_Chen_Radzihovsky}, in which the cost of the exchange energy required to form the bicollinear phase increases with doping, eventually becoming greater than the energy gained by the anisotropy gap (see Fig. \ref{fig:anisotropy}). Doping should change the incommensurate wavevector preferred by the exchange energy; for lower values of iron doping with $y <$ 0.12, this dependence can be found by examining the fluctuations just above \TN\ (below \TN, the system is of course gapped and commensurate).
This doping-driven commensurate-incommensurate transition is expected to be continuous and can be interpreted as the condensation of the domain walls (or solitons) of the commensurate state. One consequence of this transition is that at finite temperature a soliton liquid is expected to occur between the commensurate state (soliton vacuum) and the incommensurate state (soliton line crystal)
\cite{Coppersmith1982, Pokrovsky1986}.
The recent observation of anomalous hysteresis in the thermal expansion of Fe$_{1.13}$Te has been attributed to strong soliton pinning \cite{Roessler2011}.
To conclude, we have observed clear signatures of a competition between commensurate and incommensurate magnetism in a well-characterized sample of Fe$_{1.08}$Te. We find incommensurate fluctuations above \TN, which give way to commensurate order below \TN\ with a gap in the excitation spectrum. This behavior, as well as other previously unexplained observations, can be understood in terms of a lock-in transition induced by the spin anisotropy gap which is present in the monoclinic bicollinear phase and absent in the tetragonal phase.
\begin{acknowledgments}
D.P and D.R. were supported by the DOE, Office of Basic Energy Sciences under Contract No. DE-SC0006939. The extended visit of D.P. to KIT, during which part of this work was performed, was supported by the International Institute for Complex Adaptive Matter, and by NSF grants DMR-0847782, DMR-0820579, and DMR-0844115. L.R. and G.C. were supported through NSF grant No. DMR-1001240. The authors are grateful to I. Mazin and S. Maekawa for valuable discussions, and to J. L. Niedziela for critical reading of the manuscript.
\end{acknowledgments}
|
\section{Introduction.}
\label{intro}
Mixing and transport processes are fundamental to determine the
physical, chemical and biological properties of the oceans.
From plankton dynamics to the evolution of pollutant spills,
there is a wide range of practical issues that benefit from a
correct understanding and modeling of these processes. Although
mixing and transport in the oceans occur in a wide range of
scales, mesoscale and sub-mesoscale variability are known to
play a very important role \citep{Mahadevan2008,Klein2009}.
Mesoscale eddies are especially important in this aspect
because of their long life in oceanic flows, and their
stirring and mixing properties. In the southern Benguela, for
instance, cyclonic eddies shed from the Agulhas current can
transport and exchange warm waters from the Indian Ocean to the
South Atlantic \citep{Byrne1995,Lehahn2011}.
Moreover, mesoscale
eddies have been shown to drive important biogeochemical
processes in the ocean such as the vertical flux of nutrients
into the euphotic zone \citep{McGillicuddy1998,Oschlies1998}.
Another effect of
eddy activity seems to be the intensification
of mesoscale and sub-mesoscale variability due to the
filamentation process where strong tracer gradients are created
by the stretching of tracers in the shear- and strain-dominated
regions in between eddy cores \citep{Elhmaidi1993}.
Studies of the vertical structure of such eddies in the
Benguela region (e. g.
\citet{Doglioli2007} and \citet{Rubio2009288}) have shown that
they can extended to one thousand meters deep waters.
In the last decades new developments in the description and
modelling of oceanic mixing and transport from a Lagrangian
viewpoint have emerged \citep{Mariano2002,Lacasce2008}.
These Lagrangian approaches have become more and more frequent
due to
the increased availability of detailed knowledge of the
velocity field from Lagrangian drifters, satellite measurements
and computer models. In particular, the very relevant concept
of Lagrangian Coherent Structure (LCS)
\citep{Haller2000,Haller2000b} is becoming crucial for the
analysis of transport in flows. LCSs are structures that
separate regions of the flow with different dynamical behavior.
They give a general geometric view of the dynamics, acting as a
(time-dependent) roadmap for the flow. They are templates
serving as proxies to, for instance, barriers and avenues to
transport or eddy boundaries
\citep{Boffetta2001,Haller2000b,Haller2002,dOvidio2004,dOvidio2009,
Mancho2006b}.
The relevance of the three-dimensional structure of LCSs begins
to be unveiled in atmospheric contexts
\citep{DuToit2010,Tang2011,Tallapragada2011}. In the case of
oceanic flows, however, the identification of the LCSs and the
study of their role on biogeochemical tracers transport has
been mostly restricted to the marine surface
\citep{dOvidio2004,Waugh2006, dOvidio2009,BeronVera2008}. This
is mainly due to two reasons: a) tracer vertical displacement
is usually very small with respect to the horizontal one; and
b) satellite data of any quantity (temperature, chlorophyll,
altimetry for velocity, etc..) are only available from the
observation of the ocean surface.
Oceanic flows can be considered mainly two-dimensional, because
there is a great disparity between the horizontal and vertical
length scales, and they are strongly stratified due to the
Earth's rotation. There are, however, areas in the ocean where
vertical motions are fundamental. Firstly there are the
so-called upwelling regions, which are the most biologically
active marine zones in the world \citep{Rossi2008,Pauly1995}.
The reason is that due to an Ekmann pumping mechanism close to
the coast, there is a surface uprising of deep cold waters rich
in nutrients, inducing a high proliferation of plankton
concentration. Typically, vertical velocities in upwelling
regions are much larger than in open ocean, but still one order
of magnitude smaller than horizontal velocities. Another
example where there are significant vertical processes are
mesoscale eddies producing submesoscale structures
(frontogenesis), which are responsible for strong ageostrophic
vertical process, in addition to the vertical exchange thought
to occur at the eddy interior \citep{Klein2009}. Thus, the
identification of the three-dimensional (3d) LCSs in these
areas is crucial, as well as understanding their correlations
with biological activity. Another reason to include the third
dimension in LCS studies is to investigate the vertical
variation in their properties.
The main objective of this
paper is the characterization of 3d LCSs, extracted in an
upwelling region, the Benguela area in the Southern Atlantic
Ocean. For this goal we use Finite-Size Lyapunov Exponents
(FSLEs). FSLEs \citep{Aurell1997,Artale1997} measure the
separation rate of fluid particles between two given distance
thresholds. LCSs are computed as the ridges of the FSLE field
\citep{dOvidio2004,Molcard2006,Haza2008,dOvidio2009,Poje2010,Haza2010}.
The rigorous definition of LCS as ridges of a Lagrangian
stretching measure was given for the Finite-Time Lyapunov
Exponents (FTLE) in \citet{Shadden2005} and \citet{Lekien2007},
which are closely related to FSLEs. More recently, hyperbolic LCS
have been defined independently of such stretching measures by
\citet{Haller2011}.
Following many previous studies \citep{dOvidio2004,Molcard2006,
dOvidio2009,Branicki2009} we adopt the mathematical results
for Finite-Time Lyapunov Exponents (FTLE) to FSLE, assuming
them to be valid. In particular, we assume that LCS are identified
with ridges \citep{Haller2001}, i.e., the local extrema of
the FTLE field, and also we expect, in accordance to the
results in \citet{Shadden2005} and \citet{Lekien2007} for FTLEs, that the
material flux through these LCS is small and that they are
transported by the flow as quasi-material surfaces.
To confirm that our identification of LCSs
with ridges of the FSLE field, we perform (in Sect. III) direct
particle trajectory integrations that show that the computed
LCS really organize the tracer flow. In our work, we will
emphasize the numerical methodology since up to now FSLEs have
only been computed for the marine surface (an exception is
\citet{Ozgokmen2011}). We then focus on a particular eddy very
prominent in the area at the chosen temporal window and study
the stirring and mixing on it's vicinity. Some previous results
for Lagrangian eddies were obtained by \citet{Branicki2010} and
\citet{Branicki2011}, applying the methodology of lobe dynamics
and the turnstile mechanism to eddies pinched off from the Loop
Current. In this paper we focus on FSLE fields and the
associated particle trajectories to study transport in and out
of the chosen mesoscale eddy. Since this is a first attempt to
study 3d oceanic LCS, more general results (on Benguela and
other upwelling regions) are left for future work.
To circumvent the lack of appropriate observational data in the
vertical direction, we use velocity fields from a numerical
simulation.
They are high resolution simulations from the ROMS model (see section
\ref{sec:data} below) thus
appropriate to study regional-medium scale basins.
The paper is organized as follows:
In section II we describe the data and methods. In section III
we present our results. Section IV contains a discussion of the
results and Section V summarizes our conclusions.
\section{Data and Methods.}
\label{sec:data}
\subsection{Velocity data set.}
\label{subsec:velocity}
The Benguela ocean region is situated off the west coast of
southern Africa. It is characterized by a vigorous coastal
upwelling regime forced by equatorward winds, a substantial
mesoscale activity of the upwelling front in the form of eddies
and filaments, and also by the northward drift of Agulhas
eddies.
The velocity data set comes from a regional ocean model
simulation of the Benguela Region \citep{LeVu2011}. ROMS
\citep{Shchepetkin2003,Shchepetkin2005} is a split-explicit
free-surface, topography following model. It solves the incompressible
primitive equations
using the Boussinesq and hydrostatic approximations. Potential
temperature and salinity transport are included by coupling
advection/diffusion schemes for these variables. The model was
forced with climatological data. The data set area extends from
12\textdegree S to 35\textdegree S and from 4\textdegree E to
19\textdegree E (see Fig. 1). The velocity field $\mathbf{u}=(u,v,w)$
consists of two years of daily averaged zonal ($u$),
meridional ($v$), and vertical velocity ($w$) components,
stored in a three-dimensional grid with an horizontal
resolution of $1/12$ degrees $\sim 8$ km, and $32$ vertical
terrain-following levels
using a stretched vertical coordinate where the
layer thickness varies, increasing from the surface
to the ocean interior.
Since the ROMS model considers the hydrostatic approximation
it is important to note that
\citet{Amala2006222}, when
comparing results from non-hydrostatic and hydrostatic
versions of the same model of vertical motions
at submesoscale fronts, found that while instantaneous
vertical velocities structures differ, the averaged
vertical flux is similar in both hydrostatic and non-hydrostatic
simulations.
\begin{figure}
\includegraphics[width=\columnwidth]{Fig1.pdf}
\caption{
Benguela ocean region. The velocity field domain is limited by the continuous black line.
The FSLE calculation area is limited by the dash-dot black line. Bathymetric contour lines are
from ETOPO1 global relief model \citep{AmanteETOPO1} starting a 0 m depth up to 4000 m at 500 m
interval.}
\label{Fig1}
\end{figure}
\subsection{Finite-Size Lyapunov Exponents.}
\label{subsec:fsle}
In order to study non-asymptotic dispersion processes such as
stretching at finite scales and time intervals, the Finite Size
Lyapunov Exponent \citep{Aurell1997,Artale1997} is
particularly well suited. It is defined as:
\begin{linenomath*}
\begin{equation}\label{FSLE_1}
\lambda=\frac{1}{\tau}\log\frac{\delta_{f}}{\delta_{0}},
\end{equation}
\end{linenomath*}
where $\tau$ is the time it takes for the separation between
two particles, initially $\delta_{0}$, to reach $\delta_{f}$.
In addition to the dependence on the values of $\delta_0$ and
$\delta_f$, the FSLE depends also on the initial position of
the particles and on the time of deployment. Locations (i.e.
initial positions) leading to high values of this Lyapunov
field identify regions of strong separation between particles,
i.e., regions that will exhibit strong stretching during
evolution, that can be identified with the LCS
\citep{Boffetta2001,dOvidio2004,Joseph2002}.
In principle, for computing FSLEs in three dimensions one just
needs to extend the method of \citet{dOvidio2004}, that is, one
needs to compute the time that fluid particles initially
separated by $\delta_0=[(\delta x_0)^2+ (\delta y_0)^2+(\delta
z_0)^2]^{1/2}$ need to reach a final distance of
$\delta_f=[(\delta x_f)^2+ (\delta y_f)^2+(\delta
z_f)^2]^{1/2}$. The main difficulty in doing this is that in
the ocean vertical displacements (even in upwelling regions)
are much smaller than the horizontal ones, and so do not
contribute significantly to total particle dispersion
\citep{Ozgokmen2011}. By the time the horizontal particle
dispersion has scales of tenths or hundreds of kilometers
(typical mesoscale structures are studied using $\delta_f
\approx 100 km$ \citep{dOvidio2004}), particle dispersion in
the vertical can have at most scales of hundreds of meters and
usually less. This means that the vertical separation will not
contribute significantly to the accumulated distance between
particles. In addition, since length scales in the horizontal
and vertical differ by several orders of magnitude, one faces
the impossibility of assigning equal $\delta_0$ to the
horizontal and vertical particle pairs. It should be noted
however that these shortcomings arise from the different
scales of length and time that characterize horizontal and
vertical dispersion processes in the ocean, and so should not
be seem as intrinsic limitations of the method. For non-oceanic
flows a direct generalization of FSLEs is straightforward.
Thus, in this paper we implemented a quasi three-dimensional
computation of FSLEs. That is, we make the computation for
every (2d) ocean layer, but where the particle trajectories
calculation use the full 3d velocity field. I.e., at each
level (depth) we set $\delta z_O=0$, and the final distance is
computed without taking the vertical distance between
particles. It is important to note that, since we allow the
particles to evolve in the full 3d velocity field, we take into
account vertical quantities such as vertical velocity shear
that may influence the horizontal separation between particle
pairs.
There are other possible approaches to the issue of different
scales in the vertical and horizontal. One way is to assign anisotropic
initial and final displacements in the FSLE calculation
(i. e., including a $\delta z_0$ and $\delta z_f$ much smaller than
the horizontal initial and final separations). A second approach is to use
different weights for the horizontal and vertical separations in the calculations
of the distance, perhaps in combination with the first. We have cheked both alternatives and
found that, with reasonable choices of initial and final distances and
distance metrics, the results were equivalent to the quasi-3d computation.
The reason is that actual dispersion is primarily
horizontal as commented above.
More in detail, a grid of initial locations $\textbf{x}_{0}$ in
the longitude/latitude/depth geographical space
$(\phi,\theta,z)$, fixing the spatial resolution of the FSLE
field, is set up at time $t$. The horizontal distance among the
grid points, $\delta_0$, was set to $1/36$ degrees ($ \approx
3$ km), i.e. three times finer resolution than the velocity
field \citep{Ismael2011}, and the vertical resolution (distance
between layers) was set to $20$ m in order to have a good
representation of the vertical variations in the FSLE field.
Particles are released from
each grid point and their three dimensional trajectories
calculated. The distances of each particle with respect to the
ones that were initially neighbors at an horizontal distance
$\delta_0$ are monitored until one of the horizontal
separations reaches a value $\delta_{f}$. By integrating the
three dimensional particle trajectories backward and forward in
time, we obtain the two different types of FSLE maps: the
attracting LCS (for the backward), and the repelling LCS
(forward) \citep{dOvidio2004,Joseph2002}. We obtain in this way
FSLE fields with a horizontal spatial resolution given by
$\delta_0$. The final distance $\delta_f$ was set to $100$ km,
which is, as already mentioned, a typical length scale for
mesoscale studies. The trajectories were integrated for a
maximum of $T=178$ days (approximately six months) using an
integration time step of $6$ hours. When a particle reached the
coast or left the velocity field domain, the
FSLE value at its initial position and initial time was set to
zero. If the interparticle horizontal separation remains
smaller than $\delta_{f}$ during all the integration time, then
the FSLE for that location is also set to zero.
The equations of motion that describe the evolution of particle
trajectories are
\begin{linenomath*}
\begin{eqnarray}
\frac{d\phi}{dt} &=& \frac{1}{R_z}\frac{u(\phi,\theta,z,t)}{cos(\theta)},\label{INTX}\\
\frac{d\theta}{dt} &=& \frac{1}{R_z}v(\phi,\theta,z,t),\label{INTY}\\
\frac{dz}{dt} &=& w(\phi,\theta,z,t),\label{INTZ}
\end{eqnarray}
\end{linenomath*}
where $\phi$ is longitude, $\theta$ is latitude and $z$ is the
depth. $R_z$ is the radial coordinate of the moving particle
$R_z=R-z$, with $R=6371$ km the mean Earth radius. For all
practical purposes, $R_z\approx R$. Particle trajectories are
integrated using a $4^{th}$ order Runge-Kutta method. For the
calculations, one needs the (3d) velocity values at the current
location of the particle. Since the six grid nodes surrounding
the particle do not form a regular cube, direct trilinear
interpolation can not be used. Thus, an isoparametric element
formulation is used to map the nodes of the velocity grid
surrounding the particles position to a regular cube, and an
inverse isoparametric mapping scheme \citep{Yuan1994} is used
to find the coordinates of the interpolation point in the
regular cube coordinate system.
\subsection{Lagrangian Coherent Structures.}
\label{LCS}
In 2d, LCS practically coincide with (finite-time) stable and
unstable manifolds of relevant hyperbolic structures in the
flow \citep{Haller2000,Haller2000b,Joseph2002}. The structure
of these last objects in 3d is generally much more complex than
in 2d \citep{Haller2001,Pouransari2010}, and they can be
locally either lines or surfaces. As commented before, however,
vertical motions in the ocean are slow. Thus, at each fluid
parcel the strongest attracting and repelling directions should
be nearly horizontal. This, combined with the incompressibility
property, implies that the most attracting and repelling
regions (i.e. the LCSs) should appear as almost vertical
surfaces, since the attraction or repulsion should occur
normally to the LCS. As a consequence, the LCSs will have a
``curtain-like" geometry, with deviations from the vertical due
to either the orientation of the most attracting or repelling
direction deviating from the horizontal, or when strong vertical
shear produces variations along the vertical in the most
repelling or attracting regions in the flow. We expect the LCS
sheet-like objects to coincide with the strongest hyperbolic
manifolds when these are two dimensional, and to contain the
strongest hyperbolic lines.
The curtain-like geometry of the LCS was already commented in
\citet{Branicki2010b}, \citet{Branicki2010}, or
\citet{Branicki2011}. In the latter paper it was shown that, in a
3d flow, these structures would appear mostly vertical when the
ratio of vertical shear of the horizontal velocity components
to the average horizontal velocities is small. This ratio also
determines the vertical extension of the structures. In
\citet{Branicki2010}, the argument was used to construct a 3d
picture of hyperbolic structures from the computation in a 2d
slice. In the present paper we confirm the curtain-like
geometry of the LCSs, and show that they are relevant to
organize the fluid flow in this realistic 3d oceanic setting.
This is done in the next section by comparing actual particle
trajectories with the computed LCSs.
Differently than 2d, where LCS can be visually identified as
the maxima of the FSLE field, in 3d the ridges are hidden
within the volume data. Thus, one needs to explicitly compute
and extract them, using the definition of LCSs as the ridges of
the FSLEs. A ridge $L$ is a co-dimension 1 orientable,
differentiable manifold (which means that for a
three-dimensional domain $D$, ridges are surfaces) satisfying
the following conditions \citep{Lekien2007}:
\begin{enumerate}
\item The field $\lambda$ attains a local extremum at $L$.
\item The direction perpendicular to the ridge is the
direction of fastest descent of $\lambda$ at $L$.
\end{enumerate}
Mathematically, the two previous requirements can be expressed
as
\begin{linenomath*}
\begin{eqnarray}\label{RREQ}
\mathbf{n}^\mathrm{T}\nabla\lambda &=&0\label{first},\\
\mathbf{n}^\mathrm{T}\mathbf{H}\mathbf{n}=
\min_{\lVert \mathbf{u} \rVert =1}\mathbf{u}^\mathrm{T}\mathbf{H}\mathbf{u} &<& 0,\label{second}
\end{eqnarray}
\end{linenomath*}
where $\nabla\lambda$ is the gradient of the FSLE field $\lambda$,
$\mathbf{n}$ is the unit normal vector to $L$ and
$\mathbf{H}$ is the Hessian matrix of $\lambda$.
The method used to extract the ridges from the scalar field
$\lambda(\mathbf{x}_0,t)$ is from \citet{Schultz2010}. It uses
an earlier \citep{Eberly1994} definition of ridge in the
context of image analysis, as a generalized local maxima of
scalar fields. For a scalar field $f:\mathbb{R}^n \rightarrow
\mathbb{R}$ with gradient $\mathbf{g}=\nabla f$ and Hessian
$\mathbf{H}$, a \textit{d}-dimensional height ridge is given by
the conditions
\begin{linenomath*}
\begin{equation}\label{RIDG1}
\forall_{d<i\leq n} \quad \mathbf{g}^\mathrm{T}\mathbf{e}_i=0 \ \textrm{and}\ \alpha_i<0,
\end{equation}
\end{linenomath*}
where $\alpha_i, i \in \lbrace 1, 2, \dotsc, n \rbrace$, are
the eigenvalues of $\mathbf{H}$, ordered such that $\alpha_1
\geq \dotsc \geq \alpha_n$, and $\mathbf{e}_i$ is the
eigenvector of $\mathbf{H}$ associated with $\alpha_i$. For
$n=3$, (\ref{RIDG1}) becomes
\begin{linenomath*}
\begin{equation}\label{RIDG2}
\mathbf{g}^\mathrm{T}\mathbf{e}_3=0 \ \textrm{and}\ \alpha_3<0.
\end{equation}
\end{linenomath*}
This ridge definition is equivalent to the one given by
(\ref{RREQ}) since the unit normal $\mathbf{n}$ is the
eigenvector (when normalized) associated with the minimum eigenvalue of
$\mathbf{H}$. In other words, in $\mathbb{R}^3$ the
$\mathbf{e}_1, \mathbf{e}_2$ eigenvectors point locally along the ridge
and the $\mathbf{e}_3$ eigenvector is orthogonal to it.
The ridges extracted from the backward FSLE map approximate the
attracting LCS, and the ridges extracted from the forward FSLE
map approximate the repelling LCS. The attracting ones are the
more interesting from a physical point of view
\citep{dOvidio2004,dOvidio2009}, since particles (or any
passive scalar driven by the flow) typically approach them and
spread along them, giving rise to filament formation. In the
extraction process it is necessary to specify a threshold $s$
for the ridge strength $|\alpha_3|$, so that ridge points whose
value of $\alpha_3$ is lower (in absolute value) than $s$ are
discarded from the extraction process. Since the ridges are
constructed by triangulations of the set of extracted ridge
points, the $s$ threshold greatly determines the size and shape
of the extracted ridge, by filtering out regions of the ridge
that have low strength. The reader is referred to
\citet{Schultz2010} for details about the ridge extraction
method. The height ridge definition has been used to extract
LCS from FTLE fields in several works (see, among others,
\citet{Sadlo2007}).
\section{Results}
\label{sec:results}
\subsection{Three dimensional FSLE field}
\label{subsec:3d}
The three dimensional FSLE field was calculated for a $30$ day
period starting September 17, with snapshots taken
every $2$ days. The fields were calculated for an area of the
Benguela ocean region between latitudes 20\textdegree S and
30\textdegree S and longitudes 8\textdegree E to 16\textdegree
E (see figure \ref{Fig1}). The area is bounded at NW by the
Walvis Ridge and the continental slope approximately bisects
the region from NW to SE. The western half of the domain has
abyssal depths of about 4000 m. The calculation domain extended
vertically from $20$ up to $580$ m of depth. Both backward and
forward calculations were made in order to extract the
attracting and repelling LCS.
\begin{figure}
\includegraphics[width=\columnwidth]{Fig2.pdf}
\caption{
Vertical profile of 30 day average backward
and forward FSLE. The 30 day average field was spatially averaged at each layer over the FSLE calculation area
to produce the vertical profiles. The backward FSLE average is shown in continuous and the forward FSLE is shown in
dashed.}
\label{Fig2}
\end{figure}
Figure \ref{Fig2} displays the vertical profile of the average
FSLE for the 30 day period. There are small differences between
the backward and the forward values due to the different
intervals of time involved in their calculation. But both
profiles have a similar shape and show a general decrease with
depth. There is a notable peak in the profiles at about 100 m
depth that indicates increased mesoscale variability (and
transport, as shown in Sect. \ref{subsec:eddy} at that depth).
A snapshot of the attracting LCSs for day 1 of the calculation
period is shown in figure \ref{Fig3}. As expected, the
structures appear as thin vertical curtains, most of them
extending throughout the depth of the calculation domain. The
area is populated with LCS, denoting the intense mesoscale
activity in the Benguela region. As already mentioned, in three
dimensions the ridges are not easily seen, since they are
hidden in the volume data. However the horizontal slices of the
field in figure \ref{Fig3} show that the attracting LCS fall on
the maximum backward FSLE field lines of the 2d slices. The
repelling LCS (not shown) also fall on the maximum forward FSLE
field lines of the 2d slices.
\begin{figure}
\includegraphics[width=\columnwidth]{Fig3_Rev.pdf}
\caption{
Attracting LCS (blue) for day 1 of the calculation period, together with
horizontal slices of the backward FSLE field at 120 m and 300 m
depth. Colorbar refers to colormap of horizontal slices.
The units of the colorbar are $day^{-1}$.
}
\label{Fig3}
\end{figure}
Since the $\lambda$ value of a point on the ridge and the
ridges strength $\alpha_3$ are only related through the
expressions (\ref{RIDG1}) and (\ref{RIDG2}), the relationship
between the two quantities is not direct. This creates a
difficulty in choosing the appropriate strength threshold for
the extraction process. A too small value of $s$ will result in
very small LCS that appear to have little influence on the
dynamics, while a greater value will result in only a partial
rendering of the LCS, limiting the possibility of observing
their real impact on the flow. Computations with several values
of $s$ lead us to the optimum choice $s=20\: day^{-1}m^{-2}$,
meaning that grid nodes with $\alpha_3 < -20\: day^{-1}m^{-2}$ were
filtered out from the LCS triangulation.
We have seen in this section
an example of
how the ridges of the $3d$ FSLE
field, the LCS, distribute in the Benguela ocean region. Their
ubiquity shows their impact on the transport and mixing
properties. In the next section we concentrate on the
properties of a single 3d mesoscale eddy.
\subsection{Study of the dynamics of a relevant mesoscale eddy}
\label{subsec:eddy}
\begin{figure}
\includegraphics[width=\columnwidth]{Fig4.pdf}
\caption{
Trajectory (advancing from NE to SW) of the eddy center inside the calculation domain. Circles
indicate the center location during the 30 day FSLE calculation period, and squares previous and
posterior positions. Bathymetric lines same
as in figure \ref{Fig1}.}
\label{Fig4}
\end{figure}
Let us study a prominent cyclonic eddy observed in the data
set. The trajectory of the center of the eddy was tracked and
it is shown in figure \ref{Fig4}. The eddy was apparently
pinched off at the upwelling front. At day $1$ of the FSLE
calculation period its center was located at latitude
24.8\textdegree S and longitude 10.6\textdegree E, leaving the
continental slope, and having a diameter of approximately $100$
km. One may ask: what is its vertical size? is it really a
barrier, at any depth, for particle transport?
To properly answer these questions the eddy, in particular its
frontiers, should be located. From the Eulerian point of view
it is commonly accepted that eddies are delimited by closed
contours of vorticity and that the existence of strong
vorticity gradients prevent the transport in and out of the
eddy. Such transport may occur when the eddy is destroyed or
undergoes strong interactions with other eddies
\citep{Provenzale1999}. In a Lagrangian view point, however, an
eddy can be defined as a region delimited by intersections and
tangencies of LCS, whether in 2d or 3d space. The eddy itself
is an elliptic structure
\citep{Haller2000b,Branicki2010,Branicki2011}. In this
Lagrangian view of an eddy, the transport inhibition to and
from the eddy is now related to the existence of these
transport barriers delimiting the eddy region, which are known
to be quasi impermeable.
Using the first approach, i.e., the Eulerian view, the vertical
distribution of the $Q$-criteria \citep{Hunt1988,Jeong1995} was
used to determine the vertical extension of the mesoscale eddy.
The $Q$ criterium is a 3d version of the Okubo-Weiss criterium
\citep{Okubo1970,Weiss1991} and measures the relative strength
of vorticity and straining. In this context, eddies are
defined as regions with positive $Q$, with $Q$ the second
invariant of the velocity gradient tensor
\begin{linenomath*}
\begin{equation}\label{Q1}
Q=\frac{1}{2}(\|\mathbf{\Omega}\|^2-\|\mathbf{S}\|^2),
\end{equation}
\end{linenomath*}
where
$\|\mathbf{\Omega}\|^2=tr(\mathbf{\Omega}\mathbf{\Omega}^\mathrm{T})$,
$\|\mathbf{S}\|^2=tr(\mathbf{S}\mathbf{S}^\mathrm{T})$ and
$\mathbf{\Omega}$, $\mathbf{S}$ are the antisymmetric and
symmetric components of $\nabla\mathbf{u}$.
Using $Q=0$ as the Eulerian eddy boundary, it can be seen from
Fig. \ref{Fig5} that the eddy extends vertically down to, at
least, $600$ m.
\begin{figure*}
\begin{center}
\includegraphics[width=0.8\textwidth]{Fig5_Rev.pdf}
\end{center}
\caption{Colormap of $Q$-criterium. White contours have $Q=0$.
Day $1$ of the $30$ day FSLE calculation period. Left panel:
Latitude $24.5^{\circ}S$; Rigth panel: Longitude $10.5^{\circ}E$. Colorbar
values are $Q\times10^{10}\:s^{-2}$.}
\label{Fig5}
\end{figure*}
\begin{figure}
\includegraphics[width=\columnwidth]{Fig6_Rev.pdf}
\caption{$Q$-criterium map at 200 m depth together with patches of backward (blue)
and forward (green) FSLE values. Black dashed lines have $Q=0$. FSLE patches contain the highest
60\% of FSLE values.
Colorbar values are $Q\times10^{10}\:s^{-2}$.The eddy we study
is the clear region in between points H1 and H2.}
\label{Fig6}
\end{figure}
Let us move to the Lagrangian description of eddies, which is
much in the spirit of our study, and will allow us to study
particle transport: eddies can be defined as the {\it region
bounded by intersecting or tangent repelling and attracting
LCS} \citep{Branicki2010,Branicki2011}. Using this criterion,
and first looking at the surface located at 200 m depth, we see
in Fig. \ref{Fig6} that certainly the Eulerian eddy seems to be
located inside the area defined by several intersections and
tangencies of the LCS. This eddy has an approximate diameter of
$100$ km. In the south-north direction there are two
intersections that appear to be hyperbolic points (H1 and H2 in
figure \ref{Fig6}). In the West-East direction, the eddy is
closed by a tangency at the western boundary, and a
intersection of lines at the eastern boundary. The eddy core is
devoid of high FSLE lines, indicating that weak stirring occurs
inside \citep{dOvidio2004}. As additional Eulerian properties,
we note that near or at the intersections H1 and H2 the
$Q$-criterium indicates straining motions. In the case of H2,
figure \ref{Fig5} (right panel) indicates high shear up to 200 m
depth.
The fact that the hyperbolic regions H1 and H2 lie in strain
dominated regions of the flow ($Q<0$) highlights the connection
between hyperbolic particle behavior and instantaneous hyperbolic
regions of the flow. The ridges of the FSLE field, however, do not
remain in the negative $Q$ regions but cross into rotation dominated regions
with $Q>0$. This indicates that there are some differences between the
Eulerian view (Q) and the Lagrangian view (FSLE). It is the latter that
can be understood in terms of particle behaviour as limiting regions
of initial conditions (particles) that stay away from hyperbolic
regions for long enough time \citep{Haller2000b}.
In 3d, the eddy is also surrounded by a set of attracting and
repelling LCS (figure \ref{Fig7}), calculated as explained in
Subsection \ref{LCS}. The lines identified in figure \ref{Fig6}
are now seen to belong to the vertical of these surfaces.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{Fig7.pdf}
\end{center}
\caption{3d LCSs around the mesoscale eddy at day $1$ of the $30$ day FSLE calculation period.
Green: repelling LCS; Blue: attracting LCS.}
\label{Fig7}
\end{figure}
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.7\textwidth]{Fig8.pdf}
\end{center}
\caption{Three dimensional view of the evolution of elliptic
patches released at different depths inside of the eddy at day
1 of the 30 day FSLE calculaton period. Top left: day 3; Top
right: day 13; Bottom left: day 19: Bottom right: day 29. Red:
40 m; Yellow: 100 m; Cyan: 200 m; Magenta: 300 m; Grey: 400 m;
Black: 500 m. Attracting LCS are shaded in blue while repelling LCS
are shaded in green.} \label{Fig8}
\end{figure*}
Note that the vertical extent of these surfaces is in part
determined by the strength parameter used in the LCS extraction
process, so their true vertical extension is not clear from the
results presented here. On the south, the closure of the
Lagrangian eddy boundary extends down to the maximum depth of
the calculation domain, but moving northward it is seen that
the LCS shorten their depth. Probably this does not mean that
the eddy is shallower in the North, but rather that the LCS are
losing strength (lower $|\alpha_3|$) and portions of it are
filtered out by the extraction process. In any case, it is seen
that as in two-dimensional calculations, the LCS delimiting the
eddy do not perfectly coincide with its Eulerian boundary
\citep{Joseph2002}, and we expect the Lagrangian view to be
more relevant to address transport questions.
In the next paragraphs we analyze the fluid transport across the eddy
boundary. Some previous results for Lagrangian eddies were
obtained by \citet{Branicki2010} and \citet{Branicki2011}.
Applying the methodology of lobe dynamics and the turnstile
mechanism to eddies pinched off from the Loop Current,
\citet{Branicki2010} observed a net fluid entrainment near the
base of the eddy, and net detrainment near the surface, being
fluid transport in and out of the eddy essentially confined to
the boundary region. Let us see what happens in our setting.
We consider six sets of $1000$ particles each, that were
released at day $1$ of the FSLE calculation period, and their
trajectories integrated by a fourth-order Runge-Kutta method
with a integration time step of $6$ hours. The sets of
particles were released at depths of $50$, $100$, $200$, $300$,
$400$ and $500$ m. In figure \ref{Fig8} we plot the particle
sets together with the Lagrangian boundaries of the mesoscale
eddy viewed in 3d. A top view is shown in figure \ref{Fig9}. As
expected, vertical displacements are small.
At day $3$ (top left panel of figures \ref{Fig8} and
\ref{Fig9}) it can be seen that there is a differential
rotation (generally cyclonic, i.e. clockwise) between the sets
of particles at different depths. The shallower sets rotate
faster than the deeper ones. This differential rotation of the
fluid particles could be viewed, in a Lagrangian perspective,
as the fact that the attracting and repelling strength of the
LCS that limit the eddy varies with depth. Note that the six
sets of particles are released at the same time and at the same
horizontal positions, and thereby their different behavior is
due to the variations of the LCS properties along depth.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.75\textwidth]{Fig9.pdf}
\end{center}
\caption{Top view of the evolution of particle patches and LCSs
shown in Fig. 8. Top left: day 3; Top right: day 13; Bottom
left: day 19: Bottom right: day 29. Colors as in figure 8.}
\label{Fig9}
\end{figure*}
\begin{figure*}[h!]
\begin{center}
\includegraphics[width=0.75\textwidth]{Fig10.pdf}
\end{center}
\caption{Top view of the initial stages of evolution of the
particle patches and LCSs of Figs. 8 and 9. Top left: day 7;
Top right: day 9; Bottom left: day 11: Bottom right: day 13.
Colors as in figure 8.} \label{Fig10}
\end{figure*}
At day $13$ the vortex starts to expel material trough
filamentation (Figs.\ref{Fig8} and \ref{Fig9}, top right
panels). A fraction of the particles approach the southern
boundaries of the eddy from the northeast. Those to the west of
the repelling LCS (green) turn west and recirculate inside the
eddy along the southern attracting LCS (blue). Particles to the
east of the repelling LCS turn east and leave the eddy forming
a filament aligned with an attracting (blue) LCS. At longer
times trajectories in the south of the eddy are influenced by
additional structures associated to a different southern eddy.
At day 29 (bottom right panels) the same process is seen to
have occurred in the northern boundary, with a filament of
particles leaving the eddy along the northern attracting (blue)
LCS. The filamentation seems to begin earlier at shallower
waters than at deeper ones since the length of the expelled
filament diminishes with depth. However all of the expelled
filaments follow the same attracting LCS. Figure \ref{Fig10}
shows the stages previous to filamentation in which the LCS
structure, their tangencies and crossings, and the paths of the
particle patches are more clearly seen. Note that the LCS do
not form fully closed structures and the particles escape the
eddy through their openings. The images suggest lobe-dynamics
processes, but much higher precision in the LCS extraction
would be needed to really see such details.
This filamentation event seems to be the only responsible for
transport of material outside of the eddy, since the rest of
the particles remained inside the eddy boundaries. To get a
rough estimate of the amount of matter expelled in the
filamentation process we tracked the percentage of particles
leaving a circle of diameter 200 km centered on the eddy center.
In Fig. \ref{Fig11} the time evolution of this percentage is
shown for the particle sets released at different depths. The
onset of filamentation is clearly visible around days 9-12 as a
sudden increase in the percentage of particles leaving the
eddy. The percentage is maximum for the particles located at
100 m depth and decreases as the depth increases. At 400 and
500 m depth there are no particles leaving the circle. There is
a clear lag between the onset of filamentation between the
different depths: the onset is simultaneous for the 40 m and
100 m depths but occurs later for larger depths.
\section{Discussion.}
The spatial average of FSLEs defines a measure of stirring and
thus of horizontal mixing between the scales used for its
computation. The larger the average, the larger the mixing
activity \citep{dOvidio2004}. The general trend in the vertical
profiles of the average FSLE (Fig. \ref{Fig3}) shows a
reduction of mesoscale mixing with depth. There is however a
rather interesting peak in this average profile occurring at
100 m, i.e. close to the thermocline. It could be related to
submesoscale processes that occur alongside the mesoscale ones.
Submesoscale is associated to filamentation (the thickness of
filaments is of the order of $10$ km or less), and we have seen
that the filamentation and the associated transport intensity
(Fig. \ref{Fig11}) is higher at $100$ m depth. It is not
clear at the moment what is the precise mechanism responsible
for this increased activity at around 100 m depth (perhaps
associated to instabilities in the mixed layer), but we note
that the intensity of shearing motions (see the $Q$ plots in
\ref{Fig5}) is higher in the top 200 meters. Less intense
filamentation could be caused by reduction of shear in depths
larger than these values.
\begin{figure}[ht]
\includegraphics[width=84mm]{Fig11.pdf}
\caption{Percentage of particles outside a $200$ km diameter
circle centered at the eddy center, as a function of time.
}
\label{Fig11}
\end{figure}
From an Eulerian perspective, it is thought that vortex
filamentation occurs when the potential vorticity (PV) gradient
aligns itself with the compressional axis of the velocity
field, in strain coordinates
(\citet{Louazel2004};\citet{Lapeyre1999}). This alignment is
accompanied by exponential growth of the PV gradient magnitude.
The fact that the filamentation occurs along the attracting LCS
seems to indicate that this exponential growth of the PV
gradient magnitude occurs across the attracting LCS.
In the specific spatiotemporal area we have studied, and in
particular, for the eddy on which we focussed our analysis, we
have confirmed that the structure of the LCSs is
``curtain-like", so that the strongest attracting and repelling
structures are quasivertical surfaces. Their vertical extension
would depend of the physical transport properties, but it is
also altered by the particular threshold parameter selected to
extract the LCSs. These observations imply that transport and
stirring occurs mainly on the horizontal, which is a reasonable
result considering the disparity between horizontal and
vertical velocities in the ocean, and its stratification.
However, we should mention that our results are not fully
generalizable to all ocean situations, and that any ocean area
or oceanic event should be studied in particular to reveal the
shape of the associated 3d LCS.
Some comments follow about the nature of vertical transport structures.
FSLEs are suited to the identification of hyperbolic structures (structures that exhibit
high rates of transversal stretching or compression in their vicinity).
The question is if one can expect that structures responsible for vertical transport
will also exhibit substantial (vertical) stretching.
This is not so clear in the ocean for the reasons already indicated.
If one considers the case (relevant to our work) of purely isopycnal flow,
then strong vertical stretching would be associated with a rapid divergence of isopycnic surfaces.
In the case of coastal upwelling, for instance, the lifted isopycnic surfaces move vertically
in a coherent fashion, so one should not expect strong vertical divergence of particles flowing
along neighbouring isopycnic surfaces. This is just an example of the fact
that it is possible that coherent vertical motions do not
imply the presence of hyperbolic coherent structures such as those the FSLE may indicate.
Another possible limitation worth mentioning is the velocity
field resolution and its
relation to the intensity of the vertical velocity. It is accepted that in fronts
or in the eddy periphery, vertical velocities are significantly greater than,
for instance, in the eddy interior. These zones of enhanced vertical transport
correspond to submesoscale features that were not adequately captured in
the velocity field used in this work due to its coarse resolution, since
submesoscale studies usually have resolutions $<10$ km
(the literature on this subject is quite
large, so we refer the reader to \cite{Klein2009} and \cite{Levy2008}
).
In any case, a most important point for the LCS we have computed
is that in 3d, as in 2d,
they act as pathways and barriers to transport,
so that they provide a skeleton organizing the transport
processes.
\section{Conclusions}
Three dimensional Lagrangian Coherent Structures were used to
study stirring processes leading to dispersion and mixing at
the mesoscale in the Benguela ocean region. We have computed 3d
Finite Size Lyapunov Exponent fields, and LCSs were identified
with the ridges these fields. LCSs appear as quasivertical
surfaces, so that horizontal cuts of the FSLE fields gives
already a quite accurate vision of the 3d FSLE distribution.
These quasivertical surfaces appear to be coincident
with the maximal lines of the FSLE field (see fig. \ref{Fig3}) so
that surface FSLE maps could be indicative of the position of 3d LCS,
as long as the vertical shear of the velocity
does not result in a significant deviation
of the LCS with respect to the vertical.
Average FSLE values generally decrease with depth, but we find
a local maximum, and thus enhanced stretching and dispersion,
at about 100 m depth.
We have also analyzed a prominent cyclonic eddy, pinched off
the upwelling front and study the filamentation dynamics in
3d. Lagrangian boundaries of the eddy were made of
intersections and tangencies of attracting and repelling LCS
that apparently emanating from two hyperbolic locations North
and South of the eddy. The LCS are seen to provide pathways and
barriers organizing the transport processes and geometry. This
pattern extends down up to the maximum depth were we calculated
the FSLE fields ($\sim 600$ m), but the exact shape of the
boundary is difficult to determine due to the decrease in ridge
strength with depth. This caused some parts of the LCS not to
be extracted. The inclusion of a variable strength parameter in
the extraction process is an important step to be included in
the future.
The filamentation dynamics, and thus the transport out of the
eddy, showed time lags with increasing depth. This arises from
the vertical variation of the flow field. However the
filamentation occurred along all depths, indicating that in
reality vertical sheets of material are expelled from these
eddies.
Many more additional studies are needed to further clarify the
details of the geometry of the LCSs, their relationships with
finite-time hyperbolic manifolds and three dimensional lobe
dynamics, and specially their interplay with mesoscale and
submesoscale transport and mixing processes.
\section*{Acknowledgements} Financial support from Spanish MICINN
and FEDER through project FISICOS (FIS2007-60327) and from CSIC Intramural
project TurBiD is acknowledged. JHB acknowledges financial support of the
Portuguese FCT (Foundation for Science and Technology) through
the predoctoral grant SFRH/BD/63840/2009. We thank the LEGOS
group for providing us with 3D outputs of the velocity fields
from their coupled BIOBUS/ROMS climatological simulation. The
ridge extraction algorithm of \citet{Schultz2010} is available
in the \texttt{seek} module of the data visualization library
\texttt{Teem} (http://teem.sf.net).
\bibliographystyle{model2-names}
|
\section{Introduction}
Simulating Wilson type fermions at small quark masses is challenging,
in particular because of the potential instabilities caused by the
fluctuations of the low modes of the Dirac operator. A direct study of
the low-lying spectrum of Wilson Dirac operators reveals that a
spectral gap forms for large volume
lattices~\cite{DelDebbio:2005qa}. However, occasional near-zero modes
may appear since chiral symmetry is explicitly broken by the
discretization. In simulations based on the hybrid Monte-Carlo
algorithm (HMC,~\cite{Duane:1987de}), this effect can lead to large
`spikes' in the history of the molecular dynamics Hamiltonian
violation. Some time ago, Palombi and L\"uscher
proposed~\cite{Luscher:2008tw} to reweight the fermion determinant so
that the zero modes are suppressed by construction. The low-mode
contribution to any observable is faithfully restored by including the
reweighting factor in the ensemble average. For a review of other
applications of reweighting with various fermion discretizations, see
the contribution of A.\ Hasenfratz at this conference.
In the simplest version of the idea, a twisted mass term is added to
the Dirac operator,
\begin{equation}
D(\mu) = D_W + i\mu\gamma_5 \,,
\label{eq:Diracmu}
\end{equation}
where $D_W$ is the $\mathcal{O}(a)$ improved Wilson Dirac operator
(see e.g.\ \cite{Luscher:1996sc}). For $N_{\rm f}=2$, the weight
\begin{equation}
W = \det \left( \frac{D_W^\dagger D_W}{D_W^\dagger D_W + \mu^2} \right)
\end{equation}
should be included in the expectation value of the observable $\mathcal O$,
\begin{equation}\label{eq:reweight}
\left< \mathcal{O} \right> =\frac{ \left< \mathcal{O} W \right>_\mu}{\left<W\right>_\mu} \,,
\end{equation}
where $\left<\cdots\right>_\mu$ stands for the ensemble average
for the modified Dirac operator $D(\mu)$.
Evaluating the reweighting factor $W$ exactly is normally not
possible, nor is it in fact required; instead it can be calculated
stochastically. One may add a set of $N$ pseudo-fermion fields
($\eta_k,\ k= 1,\dots,N$) to the theory with action
\begin{equation}
S_\eta = \sum_{k=1}^N (\eta_k,\eta_k) \,,
\end{equation}
and the reweighting factor is replaced by
\begin{equation}
W_N = \frac{1}{N}\sum_{k=1}^N\exp\left\{ \left(\eta_k,
\left[1-\frac{D_W^\dagger D_W +\mu^2}{D_W^\dagger D_W}\right]
\eta_k \right)\right\}.\label{eq:W}
\end{equation}
The simulation with respect to the modified Dirac operator proceeds as before
and the reweighting factor $W$ is estimated, for each gauge configuration,
according to Eq.\ (\ref{eq:W}) with $N$ randomly chosen pseudofermion fields.
Obviously, the larger the number of pseudofermion fields, the more accurate the
estimate. Our tests show that 40 pseudofermion fields are sufficient to yield
satisfactory reweighted measurements of meson correlators on $32^3\times64$
lattices. As is visible from Eq.\ (\ref{eq:W}), fluctuations in the reweighting
factor $W$ that strongly suppress the contribution of certain configurations to
the statistical average can occur if the original Dirac operator $D_W$ admits
very small eigenvalues.
Convincing arguments were given by Palombi and L\"uscher that the
reweighting factor would not end up being exponentially small in the
volume, as one would at first expect. This follows from the fact that
the fluctuations of the lowest eigenvalues become smaller when the
volume is increased~\cite{DelDebbio:2005qa} (possibly with the
exception of a few). Nevertheless an explicit test is needed to
ascertain that the twisted-mass reweighting idea works in practice,
and this is the subject of this work.
\section{Implementation}
We use the domain decomposition preconditioning \cite{Luscher:2005rx} of the
hybrid Monte-Carlo algorithm (DDHMC). Our simulation code is based on
L\"uscher\rq{s} open-source DDHMC code~\cite{CLScode}. In this algorithm, the
whole lattice is divided into check-board coloured blocks. Using $\Omega$ or
$\Omega^*$ to denote the union of white or black blocks respectively, the Dirac
operator assumes the form
\begin{equation}
D_W = \left( \begin{array}{ll}
D_\Omega & D_{\partial \Omega} \\
D_{\partial\Omega^*} & D_{\Omega^*}\\
\end{array}
\right) \,,
\end{equation}
where $D_\Omega,\;D_{\Omega^*}$ denotes the Dirac operator respectively on block
$\Omega,\;\Omega^*$ with Dirichlet boundary condition, and
$D_{\partial \Omega}$ is the sum of all hopping terms from the
exterior boundary $\partial\Omega$ of $\Omega$ to the boundary
$\partial\Omega^*$ of $\Omega^*$. The fermion determinant is
factorized into local block parts and a global part. For $N_{\rm f}=2$ it reads
\begin{equation} \label{eq:fact1}
\det D_W^\dagger D_W = \det R^\dagger R\;\cdot\; \prod_\Lambda\; \det D^\dagger_\Lambda D_\Lambda\,,
\end{equation}
where $\Lambda$ runs through every block and
\begin{equation}
R=1-P_{\partial\Omega^{*}}D_{\Omega}^{-1}D_{\partial\Omega}D_{\Omega^{*}}^{-1}D_{\partial\Omega^{*}}.
\end{equation}
With even-odd preconditioning on every block determinant, Eq.\ (\ref{eq:fact1}) can be further written as
\begin{equation} \label{eq:fact2}
\det R^{\dagger}R \cdot
\prod_{{\Lambda}}
\left[\det\left(Q_{ee}^{\dagger}Q_{ee}\right)\det\left(Q_{oo}^{\dagger}Q_{oo}\right)
\det\left(Q_{ee}^{-1}\widehat{Q}\right)^{\dagger}
\left(Q_{ee}^{-1}\widehat{Q}\right)\right] \,,
\end{equation}
where
\begin{equation}
\widehat{Q}=Q_{ee}-Q_{eo}Q_{oo}^{-1}Q_{oe}
\end{equation}
and the site ordering has been chosen so that $D_\Lambda$ has the form
\begin{equation}
\gamma_5 D_\Lambda = \left(\begin{array}{cc}
Q_{ee} & Q_{eo} \\
Q_{oe} & Q_{oo}
\end{array}\right)\,.
\end{equation}
The forces of the MD evolution steps are
\begin{align}
(\omega,F_\Lambda) & =
2\mbox{Re}\left(\widehat{Q}^{-1}Q_{ee}\phi_\Lambda,\delta_{\omega}
\left(\widehat{Q}^{-1}Q_{ee}\right)\phi_\Lambda\right)-2\mbox{Re}\mbox{Tr}
\left(\frac{\delta Q_{ee}}{Q_{ee}}+\frac{\delta Q_{oo}}{Q_{oo}}\right)\,, \\
(\omega,F_R) & =
2\mbox{Re}\left(R^{-1}\phi_R,\delta R^{-1}\phi_R\right)\,,
\end{align}
where $\phi_\Lambda$ is the pseudofermion fields supported on the
even sites of block $\Lambda$ and $\phi_R$ is the pseudofermion field
residing on the block boundaries.
The modifications required to introduce the twisted mass term are relatively benign.
For the modified Dirac operator Eq.\ (\ref{eq:Diracmu}), the fermion determinant becomes
\begin{equation}
\det D(\mu)^\dagger D(\mu) = \det (D_W-i\mu\gamma_5)(D_W+i\mu\gamma_5)
\end{equation}
where the reweighting parameter $\mu$ appears with opposite sign for up and down quarks.
The forces take the form
\begin{align}
\left(\omega,F_\Lambda \right) & =
2\mbox{Re}\left(\widehat{Q}(\mu)^{-1}(Q_{ee}+i\mu)\phi_\Lambda,\delta_{\omega}
\left(\widehat{Q}(\mu)^{-1}(Q_{ee}+i\mu)\right)\phi_\Lambda\right)
\\ & \quad
-2\mbox{Re}\mbox{Tr}\left(\frac{Q_{ee}\delta Q_{ee}}{Q_{ee}^{2}+\mu^{2}}
+\frac{Q_{oo}\delta Q_{oo}}{Q_{oo}^{2}+\mu^{2}}\right)\,,\nonumber \\
(\omega,F_R) & =
2\mbox{Re}\left(R(\mu)^{-1}\phi_R,\delta_\omega R(\mu)^{-1}\phi_R\right)\,,
\end{align}
where
\begin{align}
\widehat{Q}(\mu) &=Q_{ee}+i\mu-Q_{eo}(Q_{oo}+i\mu)^{-1}Q_{oe} \,, \\
R^{-1}(\mu) &= 1-P_{\partial\Omega^{*}}D(\mu)^{-1}D_{\partial\Omega^{*}} \,.
\end{align}
Note that to calculate the global force $F_R$, one needs to solve the full
Dirac equation in every integration step. It is therefore important to solve
the equation efficiently. The deflation
accelerated~\cite{Luscher:2007se,Luscher:2007es},
Schwartz-alternating-procedure (SAP,~\cite{Luscher:2003vf}) preconditioned GCR
algorithm is applied. Since the deflation subspace need not be exact, it is
constructed by a relaxation process, starting from a set of random quark fields
$\psi_l$, which are updated iteratively according to
\begin{equation}
\psi_l \rightarrow `` D_W^{-1}{``} \;\psi_l \,,
\end{equation} until the condition
$
||D_W\psi_l|| \leq M\psi_l
$
is satisfied, where M is in the range of low eigenvalues of
$(D_W^\dagger D_W)^{1/2}$, and $`` D_W^{-1}{``}$ is an iterative procedure that
approximates the inverse of $D_W$. In the current implementation~\cite{CLScode}, it
is done with the SAP. The operator $D_W$ is used in the relaxation
procedure, the deflation efficiency for the operator $D(\mu)$ is equally good, because
the deflation subspace needs not to be exact and $\mu$ normally is small in practice.
The little Dirac operator, i.e. restriction of the Dirac operator $D(\mu)$
to the deflation subspace, is then specified as matrix
\begin{equation}
A_{kl}(\mu) = \left(\psi_k, D(\mu) \psi_l \right) .
\end{equation}
The Dirac equation $D(\mu) \psi(x) =\eta(x) $ is separated into two equations by acting with
projectors $P_L(\mu)$ and $1-P_L(\mu)$ from the left, where
\begin{equation}
P_L(\mu) \psi(x) = \psi(x) - \sum_{k,l=1}^N D(\mu)\psi_k(x) A(\mu)^{-1}_{kl} (\psi_l, \psi).
\end{equation}
\section{Testing}
We have tested the reweighting method for two flavors of $\mathcal
O(a)$ improved Wilson fermions on large lattices ($24^3\times48$ and
$32^3\times64$) with fine lattice spacings (respectively 0.07 and 0.08
fm). The masses of the lightest pseudoscalar mesons in these ensembles
are approximately 360 and 300 MeV respectively. Some details of the
simulation, for instance the molecular dynamics trajectory length
$\tau$, step size $\delta\tau$, the number of configurations
(ncfg) of the generated ensemble and acceptance rate of the HMC
simulations are listed in table \ref{tab:simu}.
The reweighting parameter $\mu$ must be chosen with some care. Larger
$\mu$ accelerates the algorithm but leads to larger fluctuations in
the reweighting factor $W$. We have tested two values of $a\mu$,
0.00569 and 0.003, on $24^3\times48$ lattices. The reweighting
factors $W$ are calculated according to Eq.\ (\ref{eq:W}). Their
Monte-Carlo history is displayed in Fig.\ \ref{fig:W}, after being
renormalized so as to have an average value of one.
For $a\mu=0.00569$ (left plot),
the reweighting factor $W$ fluctuates strongly: about 20\%
configurations receive a weight that is smaller than 0.01. By contrast, for
$a\mu=0.003$ (right plot), the reweighting factor fluctuates moderately
about the mean value. We may use the kurtosis
\begin{equation}
K = \frac{\mu_4}{\sigma^4}-3 \,,
\end{equation}
where $\mu_4$ is the fourth central moment and $\sigma$ is the
standard deviation, to quantify the distribution of the reweighting
factor $W$. For $a\mu=0.003$, the kurtosis is $K\simeq -0.11$,
showing that the fluctuations of $W$ is close to normal distribution;
while for $a\mu=0.00569$, kurtosis rises to $K \simeq 12.8$. Our
choice of $\mu$ on the $32^3\times64$ lattice and the corresponding
kurtosis is also listed in table \ref{tab:simu}.
\begin{table}[b]
\begin{center}
\begin{tabular}{|c|c|c|c||c|c|c|c||c|c|}
\hline
Volume & $\beta$ & $a$/fm & $m_\pi$/MeV & ncfg & $\tau$ & $\delta \tau$ & acc. rate & $a\mu$ & $K$ \\
\hline
$24^3\times48$ & 5.3 & 0.07 & 360 & 140 & 0.5 & 0.028 & 0.83 & 0.003 & -0.11 \\
\hline
$32^3\times64$ & 5.2 & 0.08 & 300 & 120 & 2 & 0.013 & 0.78 &0.001 & -0.37 \\
\hline
\end{tabular}
\end{center}
\caption{Parameters of the simulation, as explained in the text. The step size
$\delta \tau$ is for updating the global force $F_R$.}
\label{tab:simu}
\end{table}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.45\textwidth]{W-D5mu000569.pdf}
\includegraphics[width=0.45\textwidth]{W-D5mu0003.pdf}
\end{center}
\caption{Reweighting factor $W$ on $24^3\times48$
lattices for $a\mu=0.00569$ (left) and 0.003 (right).}
\label{fig:W}
\end{figure}
In our simulations, we have tuned the integration step size of molecular
trajectories to achieve an acceptance rate of $\approx 80\%$. For comparison,
we have simulated the same set of physical parameters with the standard method
($\mu=0$), using the same molecular trajectory length. We have tuned the
acceptance rate to be similar, $\sim$ 80\%. In order to do so, the integration
step size had to be roughly 10\% smaller. Although both simulations thus ran at
similar acceptance rates, we observe that the molecular dynamics trajectories
are more stable when reweighting is used compared with the standard method. We
plot the changes in Hamiltonian $\Delta H$ of molecular dynamics for each
trajectory in Fig.\ \ref{fig:dH} for simulations on the $32^3\times64$ lattice
(reweighting method on the left and standard method on the right). For the
reweighting method, the history of $\Delta H$ is fairly stable, while for the
standard method spikes show up more frequently; the highest spikes are up to
three orders of magnitude higher compared to the reweighted algorithm.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.45\textwidth]{dH-A5mu0001.pdf}
\includegraphics[width=0.45\textwidth]{dH-A5mu0.pdf}
\end{center}
\caption{History of Hamiltonian changes of molecular dynamics trajectories
in HMC simulations on $32^3\times64$ lattices. The reweighting method
(left) is more stable than the standard method (right). }
\label{fig:dH}
\end{figure}
Using the reweighting method, we have calculated pion correlators and
compared them with the pion correlators measured on the ensembles
generated in the standard way. Practically one first measures the
correlator in the standard way on the ensemble generated with the
reweighting parameter $\mu$; we refer to this correlator as the
`partially quenched' one. Then one `corrects' it configuration by
configuration with the reweighting factor (\ref{eq:W}). In
Fig. \ref{fig:pioncorr}, we plot the partially quenched pion correlator
(gray) and the reweighted one (red) and compare them with the
pion correlator obtained from the standard simulation (blue).
They are compatible within the uncertainties, and have comparable
statistical errors.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.65\textwidth]{pion-A5.pdf}
\end{center}
\caption{Pion correlator on the $32^3\times64$ lattice, with and without reweighting.
}
\label{fig:pioncorr}
\end{figure}
\section{Conclusion}
We have implemented a proposal by Palombi and L\"uscher to perform a
simulation with a quark determinant that is protected from the
fluctuations of the low-lying modes. We were able to find a
reweighting parameter for which the stability of the DDHMC algorithm
is significantly improved, and at the same time the reweighting factor
is under control. As a test observable, we found that the correlator
of the pseudoscalar density is well behaved under the reweighting
(\ref{eq:reweight}). The behavior of other observables, such as the
nucleon correlator, remains to be explored.
In addition to studying the performance of the reweighted algorithm over more
than 1000 trajectories, it would be worth trying out other versions of
reweighting, where the net benefit in stability is potentially even larger.
Palombi and L\"uscher for instance made a proposal in this
direction~\cite{Luscher:2008tw} in which the UV modes are only affected at
order $\mu^4$.
\acknowledgments{We thank F.\ Palombi and M.\ L\"uscher for sharing part of
their code to compute the reweighting factor. We also thank our colleagues G.\
von Hippel, D.\ Djukanovic, B.\ Brandt and B.\ J\"ager in Mainz for helpful
discussions. The work of HBM is supported by the \emph{Center for
Computational Sciences} in Mainz. The numerical simulations are performed on
the "Wilson" cluster at the Institute for Nuclear Physics of universit\"at
Mainz}
\bibliographystyle{JHEP}
|
\section{Introduction}
\subsection{Hypotheses}
We consider the following reaction-diffusion equation in $(0,+\infty) \times \mathbb{R}$:
\begin{equation}\label{eqn:eqRD}
\partial_t u = \partial_{xx} u + f(x,u).
\end{equation}
We assume that $f=f(x,u)$ is locally Lipschitz-continuous in $u$ and of class $\mathcal{C}^1$ in the neighborhood of $u=0$ uniformly with respect to $x$, so that we can define
$$\mu (x) := f'_u (x,0).$$ Moreover, $f$ is of the KPP type, that is
$$ f(x,0) =0, \ f(x,1) \leq 0, \ \mu (x) >0 \ \hbox{ and } f(x,u) \leq \mu (x) u \ \hbox{ for all } (x,u) \in \mathbb{R} \times (0,1).$$
A typical $f$ which satisfies these hypotheses is $f(x,u) = \mu (x) u(1-u)$, where $\mu$ is a continuous, positive and bounded function.
The very specific hypothesis we make on $f$ in this paper is the following: there exist $\mu_0\in \mathcal{C}^0(\mathbb{R})$ and $\phi\in \mathcal{C}^1 (\mathbb{R})$ such that
\begin{equation}\label{eqn:hypmu}
\left\{
\begin{array}{l}
\displaystyle \mu (x) = \mu_0 (\phi (x)) \hbox{ for all } x\in\mathbb{R}, \\[0.2cm]
\displaystyle 0 < \min_{[0,1]} \mu_0 < \max_{[0,1]} \mu_0 \ \mbox{ and } \ \mu_0 \mbox{ is 1-periodic}, \vspace{3pt} \\
\displaystyle \phi '(x) > 0, \ \lim_{x\rightarrow +\infty} \phi (x) = +\infty \ \mbox{ and } \lim_{x \rightarrow +\infty} \phi ' (x) =0 .
\end{array}
\right.
\end{equation}
That is, our reaction-diffusion equation is strictly heterogeneous (it is not even almost periodic or ergodic), which means that it can provide useful information on both efficiency of recently developed tools and properties of the general heterogeneous problem. But it also satisfies some periodicity properties with a growing period near $+\infty$.
We aim to look at the influence of the varying period $L(x) := x/\phi (x)$ on the propagation of the solutions.
Note that we do not assume here that there exists a positive stationary solution of~$(\ref{eqn:eqRD})$. We require several assumptions that involve the linearization of $f$
near $u=0$ but our only assumption which is related to the behavior of $ f= f(x,u)$ with respect to $u>0$ is that $f(x,1)\leq 0$, that is, $1$ is a supersolution of~$(\ref{eqn:eqRD})$ (it is clear that, up to some change of variables,
$1$ could be replaced by any positive constant in this inequality). It is possible to prove that there exists a minimal and stable positive stationary solution of $(\ref{eqn:eqRD})$
by using this hypothesis and the fact that $\mu_0$ is positive~\cite{BHR07}, but we will not discuss this problem since this is not the main topic of this paper.
\subsection{Definitions of the spreading speeds and earlier works}
For any compactly supported initial condition $u_0$ with $0\leq u_0 \leq 1$ and $u_0 \not \equiv 0$, we define the {\em minimal and maximal spreading speeds} as:
\begin{eqnarray*}
w_* & = & \sup \{ c>0 \ | \ \liminf \; \inf_{x \in [0,ct]} u(t,x) > 0 \mbox{ as } t \rightarrow +\infty \}, \vspace{3pt}\\
w^* &= & \inf \{ c>0 \ | \ \sup_{x \in [ct,+\infty)} u(t,x) \rightarrow 0 \mbox{ as } t \rightarrow +\infty\}.
\end{eqnarray*}
Note that it is clear, from the strong maximum principle, that for any $t >0$ and $x \in \mathbb{R}$, one has $0 < u(t,x) < 1$. One can also easily derive from the homogeneous case~\cite{AronsonWeinberger} that
$$2\sqrt{\min \mu_0} \leq w_* \leq w^* \leq 2\sqrt{\max \mu_0}.$$
The reader could also remark that we just require $\liminf_{t\rightarrow +\infty} u(t,x+ct) > 0 $ in the definition of $w_*$. This is because we did not assume the existence of a
positive stationary solution. Hence, we just require $u$ to ``take off'' from the unstable steady state $0$.
The aim of this paper is to determine if some of these inequalities are indeed equalities.\\
\smallskip
The first result on spreading speeds is due to Aronson and Weinberger \cite{AronsonWeinberger}. They proved that
$w^*=w_* = 2\sqrt{f'(0)}$ in the case where $f$ does not depend on $x$. More generally, even if $f$ does not satisfy $f(u) \leq f'(0) u$ for all $u\in [0,1]$, then
$w^*=w_*$ is the minimal speed of existence of traveling fronts \cite{AronsonWeinberger}. However, because of the numerous applications in various fields of natural sciences, the role of heterogeneity has become an important topic in the mathematical analysis.
When $f$ is periodic in $x$, Freidlin and Gartner
\cite{GartnerFreidlin} and Freidlin \cite{Freidlin}
proved that $w_*= w^*$ using probabilistic techniques. In this case, the spreading speed is
characterized using periodic principal eigenvalues. Namely, assume that $f$ is $1$-periodic in $x$, set $\mu_0 (x):= f_u'(x,0)$ and define for all~$p\in\mathbb{R}$ the elliptic operator
\begin{equation} \mathcal{L}_p \varphi := \varphi'' - 2p\varphi ' + (p^2 + \mu_0 (x))\varphi.\end{equation}
It is known from the Krein-Rutman theory that this operator admits a unique periodic principal eigenvalue $\lambda_p (\mu_0)$, defined by the existence of a positive $1$-periodic
function $\varphi_p\in\mathcal{C}^2 (\mathbb{R})$ so that $\mathcal{L}_p \varphi_p = \lambda_p (\mu_0)\varphi_p$. The characterization of the spreading speed \cite{GartnerFreidlin} reads
\begin{equation} \label{eq:characterizationw} w_* = w^* =\min_{p>0} \frac{\lambda_p (\mu_0)}{p}.\end{equation}
Such a formula is very useful to investigate the dependence between the spreading speed and the growth rate $\mu_0$. Several alternative proofs of this characterization,
based on different techniques, have been given in \cite{BHNa, Weinberger}. The spreading speed $w_*=w^*$ has also been identified later
as the minimal speed of existence of pulsating traveling fronts, which is the appropriate generalization of the notion of traveling fronts to periodic media
\cite{BerestyckiHamel}. Let us mention, without getting into details, that the equality $w_*=w^*$ and the characterization~(\ref{eq:characterizationw}) have been extended
when the heterogeneity is transverse~\cite{MallordyRoquejoffre}, space-time periodic or compactly supported~\cite{BHNa}, or random stationary ergodic
\cite{GartnerFreidlin, NolenXinrandom}. In this last case one has to use Lyapounov exponents instead of principal eigenvalues.
In all these cases (except in the random one), the operator $\mathcal{L}_p$ is compact and thus principal eigenvalues are well-defined. When the dependence of $f$ with respect to $x$
is more general, then classical principal eigenvalues are not always defined, which makes the computation of the spreading speeds much more difficult.
Moreover, in general heterogeneous media, it may happen that $w_*<w^*$. No example of such phenomenon has been given in space heterogeneous media, but there exist
examples in time heterogeneous media \cite{BerestyckiNadin} or when the initial datum is not compactly supported \cite{HamelNadin}.
Spreading properties in general heterogeneous media have recently been investigated by Berestycki, Hamel and the third author in \cite{BHNa}. These authors
clarified the links between the different notions of spreading speeds and gave some estimates on the spreading speeds.
More recently, Berestycki and the third author gave sharper bounds using the notion of generalized principal eigenvalues \cite{BerestyckiNadin}.
These estimates are optimal when the nonlinearity is periodic, almost periodic or random stationary ergodic. In these cases, one gets
$w_*=w^*$ and this spreading speed can be characterized through a formula which is similar to (\ref{eq:characterizationw}),
involving generalized principal eigenvalues instead of periodic principal eigenvalues.
\section{Statement of the results}
Before enouncing our results, let us first roughly describe the situation. As $\phi'(x)\rightarrow 0$ as $x\rightarrow +\infty$, the function $\phi$ is sublinear at infinity and thus $\mu (x) = \mu_0 (\phi (x))$ stays near its extremal values
$\max \mu_0$ or $\min \mu_0$ on larger and larger intervals. If these intervals are sufficiently large, that is, if $\phi$ increases sufficiently slowly,
the solution $u$ of (\ref{eqn:eqRD}) should propagate alternately at speeds close to $2\sqrt{\max \mu_0}$ and $2\sqrt{\min \mu_0}$.
Hence, we expect in such a case that $w^*=2\sqrt{\max \mu_0}$ and $w_*=2\sqrt{\min \mu_0}$.
On the other hand, if one writes $\phi (x) = x/ L(x)$, then the reaction-term locally looks like an $L(x)$-periodic function. Since $L(x) \rightarrow +\infty$, as clearly follows from the fact that $\phi'(x) \to0$ as $x\rightarrow +\infty$,
one might expect to find a link between the spreading speeds and the limit of the spreading speed $w_L$ associated with the $L$-periodic
growth rate $\mu_L (x):= \mu_0 (x/L)$ when $L\rightarrow +\infty$.
This limit has recently been computed by Hamel, Roques and the third author~\cite{HamelNadinRoques}.
As $\mu_L$ is periodic, $w_L$ is characterized by (\ref{eq:characterizationw}) and one can compute the limit of~$w_L$ by computing the limit of $\lambda_p (\mu_L)$ for
all $p$. This is how the authors of \cite{HamelNadinRoques} proved that
\begin{equation}
\lim_{L\rightarrow +\infty} w_L = \min_{k \geq M}\frac{k}{j(k)},
\end{equation}
where $M := \max_{x\in\mathbb{R}} \mu_0 (x) >0 $ and $j: [M,+\infty) \rightarrow [j(M),+\infty)$ is defined for all $k\geq M$ by
\begin{equation} \label{eq:defj} j(k):= \int_0^1 \sqrt{k- \mu_0 (x)}dx.\end{equation}
If $\phi$ increases rapidly, that is, the period $L(x)$ increases slowly, then we expect to recover this type of behavior. More precisely, we expect that $w^*= w_* = \min_{k \geq M}k/j(k)$.
We are now in position to state our results.
\subsection{Slowly increasing $\phi$}
We first consider the case when $\phi$ converges very slowly to $+\infty$ as $x \rightarrow +\infty$. As expected, we prove in this case that
$w_* < w^*$.
\begin{thm}\label{th:gencase}
\begin{enumerate}
\item Assume that $\displaystyle\frac{1}{x \phi '(x)} \rightarrow +\infty$ as $x \rightarrow +\infty$. Then $$w_*= 2\sqrt{\min \mu_0} < w^* = 2\sqrt{\max \mu_0} .$$
\item Assume that $\displaystyle\frac{1}{x \phi '(x)}\rightarrow C$ as $x \rightarrow +\infty$. If $C$ is large enough (depending on $\mu_0$), then $$ w_* < w^* .$$
\end{enumerate}
\end{thm}
This is the first example, as far as we know, of a space heterogeneous nonlinearity $f(x,u)$ for which the spreading speeds $w_*$ and $w^*$
associated with compactly supported initial data are not equal.
In order to prove this Theorem, we will first consider
the particular case when $\mu_0$ is discontinuous and only takes two values (see Proposition \ref{th:2values} below). In this case, we are able to construct sub- and super-solutions on each
interval where $\mu$ is constant, and to conclude under some hypotheses on the length of those intervals. Then, in the general continuous case, our hypotheses on
\textbf{$\displaystyle(x \phi '(x))^{-1}$} allow us to bound $\mu$ from below and above by some two values functions, and our results then follow from the preliminary case.
\begin{rmq}Note that such a two values case is not continuous, so that our Theorem holds under more general hypotheses. In fact, one would only need that $\mu_0$ is continuous on two points such that $\mu_0$ attains its maximum and minimum there, so that, from the asymptotics of $\phi (x)$, the function $\mu (x)=\mu_0 (\phi(x))$ will be close to its maximum and minimum on very large intervals as $x \rightarrow +\infty$.
\end{rmq}
\subsection{Rapidly increasing $\phi$}
We remind the reader that $M := \max_{x\in\mathbb{R}} \mu_0 (x) >0 $ and $j: [M,+\infty) \rightarrow [j(M),+\infty)$ is defined by (\ref{eq:defj}). We expect to characterize the spreading speeds
$w_*$ and $w^*$ using these quantities, as in \cite{HamelNadinRoques}.
Note that $j (M)> 0$ since $\min \mu_0 < M$. The function $j$ is clearly a bijection and thus one can define
\begin{equation}
w_\infty := \min_{\lambda \geq j(M)}\frac{j^{-1}(\lambda)}{\lambda}=\min_{k \geq M}\frac{k}{j(k)}.
\end{equation}
We need in this section an additional mild hypothesis on $f$:
\begin{equation} \label{hyp:fHolder} \exists\, C>0, \gamma>0 \hbox{ such that } \ f(x,u)\geq f_u'(x,0)u -Cu^{1+\gamma} \ \hbox{ for all } (x,u)\in \mathbb{R}\times(0,+\infty).\end{equation}
\begin{thm} \label{thm-vitspread}
Under the additional assumptions $(\ref{hyp:fHolder})$, $\phi\in \mathcal{C}^3 (\mathbb{R})$ and
\begin{equation}\label{hyp:phi2}
\phi'' (x)/\phi' (x)^2 \rightarrow 0, \ \hbox{ and}\ \phi''' (x)/\phi' (x)^2 \rightarrow 0, \hbox{ as } x\rightarrow+\infty ,
\end{equation}
one has
$$w_*= w^* = w_\infty.$$
\end{thm}
Note that (\ref{hyp:phi2}) implies $\displaystyle(x \phi '(x))^{-1}\rightarrow 0$ as $x\rightarrow +\infty$. Hence, this result is somehow complementary to Theorem \ref{th:gencase}. However, this is not optimal as this does not cover all cases. An interesting and open question would be to refine those results to get more precise necessary and sufficient conditions for the equality $w_* = w^*$ to be satisfied. This could provide some insight on the general heterogeneous case, where the establishment of such criteria is an important issue.
This result will mainly be derived from Theorem 2.1 of \cite{BerestyckiNadin}. We first construct some appropriate test-functions using the asymptotic problem associated with
$\mu_L (x) = \mu_0 (x/L)$ as $L\rightarrow +\infty.$ This will enable us to compute the generalized principal eigenvalues and the
computation of the spreading speeds will follow from
Theorem 2.1 in \cite{BerestyckiNadin}.
\subsection{Examples}
We end the statement of our results with some examples which illustrate the different possible behaviors.
\bigskip
\noindent {\bf Example 1:} $\phi (x)=\beta (\ln x)^\alpha,$ with $\alpha, \beta >0$. This function clearly satisfies the hypotheses in (\ref{eqn:hypmu}).
\begin{itemize}
\item If $\alpha \in (0,1)$, one has
$1/(x\phi'(x)) = (\ln x)^{1-\alpha} / (\beta \alpha) \rightarrow +\infty$ as $x\rightarrow +\infty$.
Hence, the assumptions of case $1$ in Theorem \ref{th:gencase} are satisfied and one has $w_*= 2\sqrt{\min\mu_0}$ and $w^* = 2\sqrt{\max \mu_0}.$
\item If $\alpha =1$, then $x\phi'(x) = \beta$ for all $x$ and thus we are in the framework of case $2$ in Theorem \ref{th:gencase},
which means that we can conclude that $w_* < w^*$ provided that $\beta$ is small enough.
\item Lastly, if $\alpha >1$, then straightforward computations give
$$\phi''(x)/\phi'(x)^2 \sim -\frac{1}{\beta \alpha} (\ln x)^{1-\alpha} \rightarrow 0 \hbox{ as } x\rightarrow +\infty,$$
$$\phi'''(x)/\phi'(x)^2 \sim \frac{2}{\beta \alpha x} (\ln x)^{1-\alpha} \rightarrow 0 \hbox{ as } x\rightarrow +\infty.$$
Hence, the assumptions of Theorem \ref{thm-vitspread} are satisfied and there exists a unique spreading speed: $w_*= w^* = w_\infty.$
\end{itemize}
\bigskip
\noindent {\bf Example 2: $\phi (x)=x^\alpha, \alpha \in (0,1)$.} This function clearly satisfies the hypotheses in (\ref{eqn:hypmu}) since $\alpha<1$. One has
$\phi'' (x)/\phi'(x)^2 = \frac{\alpha -1}{\alpha x^\alpha} \rightarrow 0$ and $\phi''' (x)/\phi'(x)^2 = \frac{(\alpha -1)(\alpha -2)}{\alpha x^{1+\alpha}} \rightarrow 0$ as $x\rightarrow +\infty$.
Thus, the assumptions of Theorem \ref{thm-vitspread} are satisfied and $w_*= w^* = w_\infty.$
\bigskip
\noindent {\bf Example 3: $\phi (x)=x/(\ln x)^\alpha, \alpha >0$.} This function satisfies (\ref{eqn:hypmu}) and one has
$$\phi '(x) = \frac{1}{(\ln x)^\alpha} -\frac{\alpha } {(\ln x)^{\alpha+1}},$$
$$\phi ''(x) = \frac{-\alpha}{x(\ln x)^{1+\alpha}} +\frac{\alpha (\alpha+1) } {x(\ln x)^{\alpha+2}},$$
$$\phi '''(x) = \frac{\alpha}{x^2(\ln x)^{1+\alpha}} -\frac{\alpha (\alpha+1)(\alpha +2) } {x^2 (\ln x)^{\alpha+3}}.$$
It follows that $\phi''(x)/\phi'(x)^2 \rightarrow 0$ and $\phi'''(x)/\phi'(x)^2 \rightarrow 0$ as $x\rightarrow +\infty$ since the terms in $x$ will decrease faster than the terms in $\ln x$.
Thus, the assumptions of Theorem \ref{thm-vitspread} are satisfied and $w_*= w^* = w_\infty.$
\bigskip
\noindent {\bf Organization of the paper:} Theorem \ref{th:gencase} will be proved in Section \ref{sec:cont}. As a first step to prove this Theorem, we will
investigate in Section \ref{sec:twovalues} the case where $\mu_0$ is not continuous anymore but only takes two values $\mu_+$ and $\mu_-$. Lastly, Section \ref{sec:unique}
is dedicated to the proof of Theorem \ref{thm-vitspread}.
\bigskip
\noindent {\bf Acknowledgements:} The authors would like to thank Fran\c cois Hamel and Lionel Roques for having drawn their attention to the problems investigated in
this paper. This article was completed while the third author
was visiting the Department of mathematical sciences of Bath whose hospitality is gratefully acknowledged.
\section{The two values case}\label{sec:twovalues}
We assume first that $\mu$ is discontinuous and only takes two distinct values $\mu_- ,$~$\mu_+ \in (0,+\infty)$.
Moreover, we assume that
there exist two increasing sequences $(x_n)_n $ and $(y_n)_n$ such that $x_{n+1} \geq y_n \geq x_n$ for all~$n$, $\lim_{n\rightarrow +\infty} x_n = +\infty$ and
\begin{equation}\label{eqn:hypseq}
\mu (x) = \left\{ \begin{array}{l}
\mu_+ \mbox{ if } x \in (x_n,y_n) ,\vspace{3pt}\\
\mu_- \mbox{ if } x \in (y_n,x_{n+1}).
\end{array}\right.
\end{equation}
\begin{prop}\label{th:2values} We have:
\begin{enumerate}
\item If $y_n/x_n \rightarrow +\infty$, then $w^*= 2\sqrt{\mu_+}$.
\item If $x_{n+1}/y_n \rightarrow +\infty$, then $w_*= 2\sqrt{\mu_-}$.
\item If $y_n/x_n \rightarrow K>1$, then $w^* \geq \displaystyle 2\sqrt{\mu_+}\frac{K}{(K-1)+ \sqrt{\mu_+ / \mu_-}}$.
\item If $x_{n+1}/y_n \rightarrow K>1$, then $w_* \leq 2\sqrt{\mu_-}\displaystyle\frac{K+\sqrt{\mu_+/\mu_-}}{K+\sqrt{\mu_-/\mu_+}}$.
\end{enumerate}
\end{prop}
It is clear in part 3 (resp. 4) of Proposition~\ref{th:2values} that the lower bound on $w^*$ (resp. upper bound on $w_*$) goes to $2\sqrt{\mu_+}$ (resp. $2\sqrt{\mu_-}$) as $K \rightarrow +\infty.$ Hence, for $K$ large enough, we get the wanted result $w_* < w^*$.
\subsection{Maximal speed: proof of parts 1 and 3 of Proposition \ref{th:2values}}
1. We first look for a subsolution of equation (\ref{eqn:eqRD}) going at some speed $c$ close to $2\sqrt{\mu_-}$. Let~$\phi_R$ be a solution of the principal eigenvalue problem:
\begin{equation}\label{eq-eigenphiR}
\left\{
\begin{array}{ll}
\displaystyle \partial_{xx} \phi_R = \lambda_{R} \phi_R & \mbox{ in } B_R,\vspace{3pt}\\
\displaystyle \phi_R=0 & \mbox{ on } \partial B_R,\vspace{3pt}\\
\displaystyle \phi_R>0 & \mbox{ in } B_R.
\end{array}
\right.
\end{equation}
We normalize $\phi_R$ by $\|\phi_R\|_\infty=1$.
We know that $\lambda_{R} \rightarrow 0$ as $R \rightarrow +\infty$. Let $c < 2\sqrt{\mu_-}$ and~$R$ large enough so that
$-\lambda_R < \mu_- - c^2/4$.
Then $\displaystyle v(x)=e^{\frac{-cx}{2}} \phi_R (x)$ satisfies:
$$\partial_{xx} v + c \partial_x v + \mu_- v = \Big( \mu_- - \frac{c^2}{4} + \lambda_R \Big) v >0 \hbox{ in } B_R.$$
By extending $\phi_R$ by~0 outside $B_R$, by regularity of $f$ and since $f'_u (x,0) \geq \mu_-$ for any $x\in \mathbb{R}$, for some small $\kappa$, we also have in $(0,+\infty)\times \mathbb{R}$:
$$\partial_{xx} \kappa v + c \partial_x \kappa v + f(x+ct, \kappa v) \geq 0.$$
Hence, $w (t,x) :=\kappa v (x-ct)$ is a subsolution of (\ref{eqn:eqRD}).
Without loss of generality, we can assume that $u(1,x) \geq w (1,x)$, thus for any $t\geq 1$, $u(t,x) \geq w (t,x)$.
That is, for any speed $c < 2\sqrt{\mu_-}$, we have bounded $u$ from below by a subsolution of (\ref{eqn:eqRD}) with speed c.
In particular,
$$\hbox{ let } \displaystyle t_n := \frac{x_n +R}{c}, \hbox{ then } u(t_n,x) \geq w(t_n,x) \hbox{ for all } x\in\mathbb{R},$$
which is positive on a ball of radius $R$ around $x_n +R$.\\
\smallskip
\noindent 2. Take an arbitrary $c'<2\sqrt{\mu_+}$ and let $\phi_{R'}$ a solution of the principal eigenvalue problem~(\ref{eq-eigenphiR}) with $R'$ such that
$-\lambda_{R'} < \mu_+ - c'^2/4$.
As above, there exists $\displaystyle \widetilde{v} (x)=\kappa' e^{\frac{-c' x}{2}} \phi_{R'} (x)$ compactly supported such that
\begin{equation}
\partial_{xx} \widetilde{v} + c' \partial_x \widetilde{v} + f(x+x_n+R+c't, \widetilde{v}) \geq 0,
\end{equation}
as long as $\widetilde{v}=0$ where $f'_u (x+x_n+R+c't,0) \neq \mu_+$, that is $$(-R'+x_n+R+c't,R'+x_n+R+c't) \subset (x_n,y_n),$$ which is true for
$R>R'$ and
$$0\leq t \leq \frac{ y_n -x_n - R -R'}{c'}.$$
As $R$ could be chosen arbitrarily large, we can assume that the condition $R>R'$ is indeed satisfied.
Moreover, as $\liminf_{n\rightarrow +\infty} y_n/x_n >1$ and $\lim_{n \rightarrow +\infty} x_n = +\infty$, we can assume that~$n$ is large enough so that $y_n-x_n >2R$ and thus the
second condition is also satisfied.
Hence, $\widetilde{w} (t_n+t,x):=\widetilde{v} (x-x_n-R -c't)$
is a subsolution of (\ref{eqn:eqRD}) for $t~\in~(0,\frac{ y_n -x_n - R -R'}{c'})$ and $x\in \mathbb{R}$.
We can take $\kappa'$ small enough so that
\begin{equation}
\kappa \min_{y\in B(0,R')}\phi_R (y)>\kappa' e^{\frac{|c-c'|}{2}R'}.
\end{equation}
For all $x\in B(ct_n,R')$, one has:
\begin{equation}
\begin{array}{rcl} w(t_n,x)=\kappa e^{\frac{-c(x-ct_n)}{2}} \phi_R (x-ct_n)
&\geq &\displaystyle \kappa \Big( \min_{y\in B(0,R')}\phi_R (y) \Big)\ e^{\frac{-(c-c')}{2}(x-ct_n)}e^{\frac{-c'}{2}(x-ct_n)}\\[0.2cm]
&\geq &\displaystyle \kappa \Big( \min_{y\in B(0,R')}\phi_R (y)\Big) \ e^{\frac{-|c-c'|}{2}R'}e^{\frac{-c'}{2}(x-ct_n)} \\[0.2cm]
&\geq &\displaystyle \kappa' e^{\frac{-c'}{2}(x-ct_n)} \\[0.2cm]
&\geq &\displaystyle \kappa' e^{\frac{-c' (x-ct_n)}{2}} \phi_{R'} (x-ct_n)=\widetilde{w} (t_n,x), \end{array}
\end{equation}
since $ct_n = x_n+R$ by definition.
Moreover, $u(t_n,x)\geq w(t_n,x)$ for all $x\in\mathbb{R}$. The parabolic maximum principle thus gives
$$u(t_n+t,x)\geq \widetilde{w} (t_n+t,x) \ \hbox{ for all } t \in \Big(0,\frac{ y_n -x_n - R -R'}{c'}\Big) \hbox{ and } \ x\in\mathbb{R}.$$
\smallskip
\noindent 3. We can now conclude. Indeed, for $n$ large enough one has:
$$u\left(t_n+\frac{ y_n -x_n - R -R'}{c'},y_n -R'\right) \geq \widetilde{w} \left(t_n+\frac{ y_n -x_n - R -R'}{c'},y_n-R'\right)= \widetilde{v} (0).$$
Since the construction of $\widetilde{v}$ did not depend on $n,$ the above inequality holds independently of $n,$ which implies that:
$$\inf_n u\left(t_n+\frac{ y_n -x_n - R -R'}{c'},y_n -R'\right) > 0.$$
If $y_n/x_n \rightarrow +\infty$, we have
$$\frac{y_n - R'}{t_n+\frac{ y_n -x_n - R -R'}{c'}}=\frac{y_n - R'}{\frac{x_n}{c}+\frac{ y_n -x_n - R -R'}{c'}} \rightarrow c' \ \hbox{ as } n \rightarrow +\infty.$$
It follows that $w^* \geq c'$ for any $c' < 2\sqrt{\mu_+}$. The proof of part 1 of Proposition~\ref{th:2values} is completed.
If $y_n/x_n \rightarrow K$, we have
$$\frac{y_n - R'}{t_n+\frac{ y_n -x_n - R -R'}{c'}} \rightarrow \frac{ K}{\frac{1}{c}+\frac{K-1}{c'}} \ \hbox{ as } n \rightarrow +\infty.$$
As this is true for any $c' < 2\sqrt{\mu_+}$ and $c < 2\sqrt{\mu_-}$, this concludes the proof of part 3 of Proposition \ref{th:2values}.
\hfill$\Box$
\subsection{Minimal speed: proof of parts 2 and 4 of Proposition \ref{th:2values}}
Let $\lambda_+ = \sqrt{\mu_+}$ be the solution of $\lambda_+^2 - 2\sqrt{\mu_+} \lambda_+ = - \mu_+$.
One can then easily check, from the KPP hypothesis, that the function
$$v(t,x):=\min \left(1, \kappa e^{-\lambda_+ (x-2\sqrt{\mu_+}t)} \right)$$
is a supersolution of equation (\ref{eqn:eqRD}) going at the speed $2\sqrt{\mu_+}$, for any $\kappa >0$. Since $u_0$ is compactly supported,
we can choose $\kappa$ such that $v(0,\cdot)\geq u_0$ in $\mathbb{R}$. Thus, for any $t\geq 0$ and $x \in \mathbb{R}$, $u(t,x) \leq v(t,x)$.
In particular, the inequality holds for $t=t_n$ the smallest time such that $v(t,y_n)=1$. Note that $t_n=y_n/(2\sqrt{\mu_+})+C$ where $C$ is a
constant independent of $n$.
Then for all $x\in \mathbb{R}$,
$$u(t_n,y_n+x) \leq v(t_n,y_n+x)= \min \left(1, e^{-\lambda_+ x} \right).$$
We now look for a supersolution moving with speed $2\sqrt{\mu_-}$ locally in time around $t_n$. Let us define
$$w(t_n + t,y_n +x):=\min \left(v(t_n +t,y_n+x), e^{-\lambda_- (x-2\sqrt{\mu_-}t)} \right)$$
where $\lambda_- = \sqrt{\mu_-}$. Note that $\lambda_- < \lambda_+$, thus $u(t_n,y_n+x) \leq v(t_n,y_n+x)= w(t_n,y_n+x)$.
We now check that $w$ is indeed a supersolution of equation (\ref{eqn:eqRD}).
We already know that $v$ is a supersolution and it can easily be seen as above from the KPP hypothesis that $(t,x)\mapsto e^{-\lambda_- (x-2\sqrt{\mu_-}t)}$ is a supersolution only where $f'_u (\cdot,0) =\mu_-$. Thus, we want the inequality
$v(t_n + t,y_n+x) \leq e^{-\lambda_- (x-2\sqrt{\mu_-}t)}$ to be satisfied if $ y_n +x \not\in (y_n,x_{n+1}).$ Recall that
$v(t_n+t,y_n+x) = \min \left(1, e^{-\lambda_+ (x-2\sqrt{\mu_+}t)} \right)$ for all $t>0$ and $x\in\mathbb{R}.$
Thus, the inequality is satisfied if $t\geq 0$ and $x\leq 0$ or if
$$x \geq 2 \frac{\lambda_+ \sqrt{\mu_+} - \lambda_- \sqrt{\mu_-}}{\lambda_+ - \lambda_-} t =2(\lambda_++\lambda_-)t.$$
It follows that $w(t_n+t,y_n+x)$ is indeed a supersolution of equation (\ref{eqn:eqRD}) in $\mathbb{R}$ as long as
\begin{equation}\label{eq:tsursol}
0 \leq 2(\lambda_++\lambda_-)t \leq x_{n+1} - y_n,\end{equation}
and that $u(t_n+t,y_n+x) \leq w(t_n+t,y_n+x)$ for any $t$ verifying the above inequality.
To conclude, let now $2\sqrt{\mu_+}>c > 2\sqrt{\mu_-}$, and $t'_n$ the largest $t$ satisfying (\ref{eq:tsursol}), i.e.
$$t_n'= \frac{x_{n+1}-y_n}{2(\lambda_++\lambda_-)}.$$
The sequence $(t_n')_n$ tends to~$+\infty$ as~$n \rightarrow +\infty$ since $\liminf_{n\rightarrow +\infty} x_{n+1}/y_n >1$ and $\lim_{n \rightarrow +\infty} y_n = +\infty$.
Moreover, one has
$$u(t_n+t'_n,y_n+ct'_n) \leq w(t_n+t'_n,y_n+ct'_n) \rightarrow 0\hbox{ as } n \rightarrow +\infty$$
since $c>2\sqrt{\mu_-}$.
If $x_{n+1}/y_n \rightarrow +\infty$ as $n\rightarrow +\infty$, as $t_n=y_n/(2\sqrt{\mu_+})+C$, one gets
$t_n'/t_n\rightarrow +\infty$ as~$n\rightarrow +\infty$.
Hence,
$$\frac{y_n + ct'_n}{t_n+t'_n} \rightarrow c \hbox{ as } n\rightarrow +\infty.$$
It follows that $w_* \leq c$ for any $c > 2\sqrt{\mu_-}$. This proves part 2 of Proposition \ref{th:2values}.
If $x_{n+1}/y_n \rightarrow K$ as $n\rightarrow +\infty$, we compute
$$\frac{y_n + ct'_n}{t_n+t'_n} \rightarrow \displaystyle\frac{1+ \frac{c}{2(\lambda_++\lambda_-)}(K-1)}{\frac{1}{2\sqrt{\mu_+}}+\frac{K-1}{2(\lambda_++\lambda_-)}}\hbox{ as } n\rightarrow +\infty.$$
Hence, $w_*$ is smaller than the right hand-side.
As $c\in (2\sqrt{\mu_-},2\sqrt{\mu_+})$ is arbitrary, $\lambda_-= \sqrt{\mu_-}$ and $\lambda_+= \sqrt{\mu_+}$, we eventually get
$$w_* \leq 2\sqrt{\mu_-}\displaystyle\frac{K+\sqrt{\mu_+/\mu_-}}{K+\sqrt{\mu_-/\mu_+}},$$
which concludes the proof of part 4 of Proposition \ref{th:2values}.
\hfill$\Box$
\section{The continuous case}\label{sec:cont}
\subsection{Proof of part 1 of Theorem \ref{th:gencase}}
We assume that $\mu_0$ is a continuous and 1-periodic function. Let now $\varepsilon$ be a small positive constant and define $\mu_- < \mu_+$ by:
$$\left\{
\begin{array}{rcl}
\mu_+ & := & \max \mu_0 - \varepsilon,\vspace{3pt} \\
\mu_- & := & \min \mu_0.\\
\end{array}
\right.
$$
We want to bound $\mu$ from below by a function taking only the values $\mu_-$ and $\mu_+$, in order to apply Theorem \ref{th:2values}.
Note first that there exist $x_{-1} \in (0,1)$ and $\delta \in (0,1)$ such that $\mu_0 (x) > \mu_+ $ for any $x \in (x_{-1},x_{-1}+\delta)$.
We now let the two sequences $(x_n)_{n \in \mathbb{N}}$ and $(y_n)_{n\in \mathbb{N}}$ defined for any $n$ by:
$$\left\{
\begin{array}{rcl}
\phi (x_n) & = & x_{-1} + n,\vspace{3pt}\\
\phi (y_n) & = & x_{-1} + n + \delta.\\
\end{array}
\right.
$$
Note that since $\phi$ is strictly increasing and $\phi(+\infty)=+\infty$,
then those sequences indeed exist, tend to $+\infty$ as $n \rightarrow +\infty$, and satisfy for any $n$, $x_n < y_n < x_{n+1}$.
It also immediately follows from their definition that for all $x\in\mathbb{R},$
$$
\mu (x) \geq \widetilde\mu(x) \ \hbox{ where }\widetilde\mu(x):= \left\{ \begin{array}{l}
\mu_+ \mbox{ if } x \in (x_n,y_n) ,\vspace{3pt}\\
\mu_- \mbox{ if } x \in (y_n,x_{n+1}) .\\
\end{array}\right.
$$
We now have to estimate the ratio $y_n/x_n$ in order to apply Proposition~\ref{th:2values}. Note that:
\begin{equation}\label{eq:div1}
\delta = \phi (y_n) - \phi (x_n) = \int_{x_n}^{y_n} \phi ' (x) dx.
\end{equation}
Moreover, under the hypothesis $x \phi '(x) \rightarrow 0$ as $x \rightarrow +\infty$, and since $(x_n)_n$, $(y_n)_n$ tend to~$+\infty$ as $n \rightarrow +\infty$:
\begin{eqnarray}\label{eq:div2}
\int_{x_n}^{y_n} \phi ' (x) dx = \int_{x_n}^{y_n} \left( x \phi ' (x) \ \times \ \frac{1}{x} \right)dx= o \left( \ln \left(\frac{y_n}{x_n}\right) \right) \mbox{ as } n \rightarrow +\infty .
\end{eqnarray}
From (\ref{eq:div1}) and (\ref{eq:div2}), we have that $\frac{y_n}{x_n} \rightarrow + \infty$. To conclude, we use the parabolic maximum principle and part 1
of Proposition~\ref{th:2values} applied to problem (\ref{eqn:eqRD}) with a reaction term $\widetilde{f} \leq f$ such that
$$\widetilde{f}'_u (x,0) = \widetilde\mu(x) \ \hbox{ for all }x\in\mathbb{R}.
$$
It immediately follows that $w^* \geq 2 \sqrt{\max \mu_0 - \varepsilon}$. Since this inequality holds for any $\varepsilon >0$, we get $w^* = 2 \sqrt{\max \mu_0}$.
We omit the details of the proof of $w_* = 2 \sqrt{\min \mu_0}$ since it follows from the same method.
Indeed, one only have to choose $y_{-1}'$ and $\delta '$ in $(0,1)$ such that $\mu_0 (x) < \min \mu_0 + \varepsilon$ for any $x \in (y_{-1}',y_{-1}'+\delta ')$
and let two sequences such that
$$\left\{
\begin{array}{rcl}
\phi (y_n') & = & y_{-1}' + n, \vspace{3pt}\\
\phi (x_{n+1}') & = &y_{-1}' + n + \delta '.\\
\end{array}
\right.
$$
One can then easily conclude as above using part 2 of Proposition~\ref{th:2values}.
\hfill$\Box$
\subsection{Proof of part 2 of Theorem \ref{th:gencase}}
As before, we bound $\mu_0$ from below by a two values function, that is, for all $x \in \mathbb{R}$,
$$
\mu (x) \geq \widetilde\mu(x) \ \hbox{ where }\widetilde\mu(x):= \left\{ \begin{array}{l}
\mu_+ =\max \mu_0 - \varepsilon \; \mbox{ if } x \in (x_n,y_n) ,\vspace{3pt}\\
\mu_- = \min \mu_0 \; \mbox{ if } x \in (y_n,x_{n+1}) ,\\
\end{array}\right.
$$
where $\varepsilon$ a small positive constant and the two sequences $(x_n)_n$ and $(y_n)_n$ satisfy for any $n$:
$$\left\{
\begin{array}{l}
x_n < y_n < x_{n+1},\vspace{3pt}\\
\phi (x_n) = x_{-1} + n,\vspace{3pt}\\
\phi (y_n) = x_{-1} + n + \delta (\varepsilon) \mbox{ for some } \delta (\varepsilon) >0,\vspace{3pt}\\
x_n \rightarrow +\infty \mbox{ and } y_n \rightarrow +\infty.
\end{array}
\right.
$$
Here, under the assumption that $x\phi ' (x) \rightarrow 1/C$, we get
\begin{eqnarray*}
\delta (\varepsilon) = \phi (y_n) -\phi (x_n) & =& \int_{x_n}^{y_n} \phi' (x)dx\\
& = & \int_{x_n}^{y_n} \left( x \phi ' (x) \ \times \ \frac{1}{x} \right)dx \\
& = & \frac{1}{C} \ln \left(\frac{y_n}{x_n}\right) + o \left( \ln \left(\frac{y_n}{x_n}\right) \right)\mbox{ as } n \rightarrow +\infty .
\end{eqnarray*}
Hence,
$$\frac{y_n}{x_n} \rightarrow e^{\delta (\varepsilon) C} \mbox{ as } n \rightarrow +\infty.$$
We can now apply the parabolic maximum principle and part~3 of Proposition~\ref{th:2values} to get
\begin{equation}\label{C0-above}
w^* \geq 2\sqrt{\max \mu_0 - \varepsilon}\; \frac{e^{\delta (\varepsilon) C}}{(e^{\delta (\varepsilon) C}-1)+ \sqrt{(\max \mu_0 - \varepsilon) / \min \mu_0}}.
\end{equation}
Notice that the dependence of $\delta$ on $\varepsilon$ prevents us from passing to the limit as $\varepsilon \rightarrow 0$ as we did to prove part~1 of Theorem~\ref{th:gencase}. However, for any fixed $\varepsilon >0$, one can easily check that the right-hand side in the inequation (\ref{C0-above}) converges as $C\rightarrow +\infty$ to $2 \sqrt{\max \mu_0 - \varepsilon}$.
One can proceed similarly to get an upper bound on $w_*$, that is:
\begin{equation}\label{C0-below}
w_* \leq 2 \sqrt{\min \mu_0 + \varepsilon}\; \frac{e^{\delta' (\varepsilon) C} + \sqrt{\max \mu_0 / (\min \mu_0 + \varepsilon)}}{e^{\delta' (\varepsilon) C} + \sqrt{(\min \mu_0 + \varepsilon)/\max \mu_0}},
\end{equation}
where $\varepsilon$ can be chosen arbitrary small and $\delta ' (\varepsilon)$ is such that $\mu_0 (x) \leq \min \mu_0 + \varepsilon$ on some interval of length $\delta' (\varepsilon)$. It is clear that the right-hand side of (\ref{C0-below}) converges to $2 \sqrt{\min \mu_0 + \varepsilon}$ as $C \rightarrow +\infty$.
Therefore, by choosing $\varepsilon < (\max \mu_0 - \min \mu_0) /2$, one easily gets from (\ref{C0-above}) and (\ref{C0-below}) that for $C$ large enough, $w_* < w^*$. This concludes the proof of part~2 of Theorem~\ref{th:gencase}. Moreover, note that the choice of~$C$ to get this strict inequality depends only on the function~$\mu_0$, by the intermediate of the functions $\delta (\varepsilon)$ and $\delta' (\varepsilon)$.
\hfill$\Box$
\section{The unique spreading speed case} \label{sec:unique}
We begin with some preliminary work that will be needed to estimate the spreading speeds.
The proof of Theorem~\ref{thm-vitspread} is then separated into two parts: the first part (Section~\ref{vitspread-p1}) is devoted to the proof that $w^*\leq w_\infty$, while in the second part (Section~\ref{vitspread-p2}) we prove that $w_* \geq w_\infty$.
\subsection{Construction of the approximated eigenfunctions}\label{subsec:approxeigen}
For all $p\in\mathbb{R}$, we define
\begin{equation} \label{eq:defH} H(p):= \left\{ \begin{array}{ccl}
j^{-1} (|p|) &\hbox{ if }& |p |\geq j(M),\\
M &\hbox{ if }& |p |< j(M).\\
\end{array}\right.\end{equation}
The fundamental property of this function is given by the following result.
\begin{prop}\label{prop:exv} (Propositions $3.1$ and $3.2$ in \cite{HamelNadinRoques})
For all $p\in\mathbb{R}$, $H(p)$ is the unique real number such that there exists a continuous $1$-periodic viscosity solution~$v$ of
\begin{equation}\label{HJ-eigen}
(v'(y)-p)^2+\mu_0 (y)= H(p) \hbox{ over } \mathbb{R} .
\end{equation}
\end{prop}
Next, we will need, as a first step of our proof, the function $v$ given by Proposition \ref{prop:exv} to be piecewise $\mathcal{C}^2$.
This is true under some non-degeneracy hypothesis on $\mu_0$. We will check below in the second part of the proof of Theorem~\ref{thm-vitspread} that
it is always possible to assume that this hypothesis is satisfied by approximation.
\begin{lem} \label{lem:regv}
Assume that $\mu_0 \in\mathcal{C}^2 (\mathbb{R})$ and that
\begin{equation} \label{hyp:nondegmu}
\hbox{if } \mu_0 (x_0)=\max_\mathbb{R} \mu_0, \hbox{ then } \mu_0''(x_0)<0.
\end{equation}
Then for all $p\in\mathbb{R}$, equation (\ref{HJ-eigen}) admits a $1$-periodic solution $v_p\in W^{2,\infty}(\mathbb{R})$ which is piecewise $\mathcal{C}^2 (\mathbb{R})$.
\end{lem}
{\bf Proof.}
The proof relies on the explicit formulation of $v_p$. Assume first that
$p > j( M) = j\big(\|\mu_0 \|_\infty\big)$. Then it is easy to check (see \cite{HamelNadinRoques}) that
\begin{equation} \label{explicitv} v_p(x):= p x -\int_0^x \sqrt{H(p)-\mu_0 (y)}dy\end{equation}
satisfies (\ref{HJ-eigen}). Then, the definition of $j$ implies that $v_p$ is $1$-periodic and, as $\mu_0\in \mathcal{C}^1 (\mathbb{R})$
and $H(p) > \mu_0 (y)$ for all $y\in\mathbb{R}$, the function $v_p$ is $\mathcal{C}^2 (\mathbb{R})$.
The case $p < -j( M)$ is treated similarly.
Next, if $|p| \leq j( M)$, let $F$ define for all $Y\in [0,1]$ by:
$$F(Y):= p+\int_Y^1 \sqrt{M-\mu_0 (y)}dy -\int_0^Y \sqrt{M-\mu_0 (y)}dy.$$
Then $F$ is continuous and, as $|p| \leq j( M)$,
$$F(0) = p+\int_0^1 \sqrt{M-\mu_0 (y)}dy = p + j(M)\geq 0.$$
Similarly, $F(1) = p-j(M) \leq 0$. Thus, there exists $X\in [0,1]$ so that $F(X)=0$. We now define:
\begin{equation} v_p (x) = \left\{
\begin{array}{lcl}
\displaystyle px -\int_0^x \sqrt{M-\mu_0 (y)}dy &\hbox{ for all }& x\in [0,X], \\[0.3cm]
\displaystyle px -\int_0^X \sqrt{M-\mu_0 (y)}dy +\int_X^x \sqrt{M-\mu_0 (y)}dy &\hbox{ for all }& x\in [X,1].
\end{array} \right. \end{equation}
From the definition of $X$, the function $v_p$ is $1$-periodic. It is continuous and derivable at any point $x\in [0,1)\backslash \{X\}$ with
$$ v_p'(x) = \left\{ \begin{array}{lcl}
p-\sqrt{M-\mu_0 (x)} &\hbox{ for all }& x\in [0,X),\\
p+\sqrt{M-\mu_0 (x)} &\hbox{ for all }& x\in (X,1).\\
\end{array}\right. $$
Hence, it satisfies (\ref{HJ-eigen}) in the sense of viscosity solutions.
Lastly, for all $x\in (0,X)$ so that $\mu_0 (x) \neq M$, one has
$$v_p''(x) = \frac{\mu_0'(x)}{2\sqrt{M -\mu_0 (x)}}.$$
If $\mu_0 (x_M)=M$, then (\ref{hyp:nondegmu}) implies that $\mu_0 (x) <M$ for all $x\neq x_M$ close to $x_M$ and a Taylor expansion gives
$$\lim_{x\rightarrow x_M, x\neq x_M}v_p''(x) = \sqrt{-\mu_0''(x_M)/2}.$$
Hence, $v_p''$ can be extended to a continuous function over $(0,X)$. Similarly, it can be extended over $(X,1)$. It follows that $v_p''$ is bounded over $[0,1]$
and that it is piecewise $\mathcal{C}^2 (\mathbb{R})$. \hfill $\Box$
\bigskip
For any $p \in \mathbb{R}$, define the elliptic operator:
$$L_p \varphi := \varphi'' - 2p\varphi ' + (p^2 + \mu_0 (\phi (x)))\varphi. $$
\begin{lem}\label{lem-vepapprox}
For all $p\in\mathbb{R}$, let
\begin{equation}
\varphi_p (x):= \exp \Big( \frac{v_p(\phi (x))}{\phi'(x)}\Big).
\end{equation}
If (\ref{hyp:nondegmu}) holds and $\mu_0 \in \mathcal{C}^2 (\mathbb{R})$, then $\varphi_p$ is piecewise $\mathcal{C}^2 (\mathbb{R})$ and one has
\begin{equation}
\frac{L_p \varphi_p(x) -H(p) \varphi_p(x)}{\varphi_p(x)} \rightarrow 0 \hbox{ as } x\rightarrow +\infty.
\end{equation}
\end{lem}
{\bf Proof.} The function $\varphi_p$ is piecewise $\mathcal{C}^2 (\mathbb{R})$ since $v_p$ is piecewise $\mathcal{C}^2 (\mathbb{R})$.
For all $x$ so that~$v_p$ is $\mathcal{C}^2$ in $x$, we can compute
\[
\begin{array}{rcl}
\displaystyle\varphi_p'(x) & = & \left( \displaystyle v'_p(\phi(x))-\frac{\phi''(x)}{(\phi'(x))^2}v_p(\phi(x)) \right) \varphi_p(x),\\
\displaystyle\varphi_p''(x)& = & \left( \displaystyle\phi'(x)v''_p(\phi(x))
- \displaystyle\frac{\phi''(x)}{\phi'(x)}v'_p(\phi(x))
+ \displaystyle\lp2\frac{(\phi''(x))^2}{(\phi'(x))^3}-\frac{\phi'''(x)}{(\phi'(x))^2}\right) v_p(\phi(x)) \right) \varphi_p(x)\\
& & + \left( \displaystyle v'_p(\phi(x))- \frac{\phi''(x)}{(\phi'(x))^2}v_p(\phi(x)) \right) ^2
\varphi_p(x).
\end{array}
\]
This gives
\[
\begin{array}{rcl}
\displaystyle \frac{L_p\varphi_p(x)-H(p)\varphi_p(x)}{\varphi_p(x)} & = &
\displaystyle\phi'(x)v''_p(\phi(x))
- \displaystyle\frac{\phi''(x)}{\phi'(x)}v'_p(\phi(x))
+ \displaystyle\lp2\frac{(\phi''(x))^2}{(\phi'(x))^3}-\frac{\phi'''(x)}{(\phi'(x))^2}\right) v_p(\phi(x))\\[0.3cm]
&& \displaystyle-2\frac{\phi''(x)}{(\phi'(x))^2}v_p(\phi(x))v'_p(\phi(x))
+ \left( \displaystyle\frac{\phi''(x)}{(\phi'(x))^2}v_p(\phi(x))\right) ^2+ v'_p(\phi(x))^2\\[0.3cm]
&& \displaystyle -2p \left( \displaystyle v'_p(\phi(x))-\frac{\phi''(x)}{(\phi'(x))^2}v_p(\phi(x)) \right) +p^2+\mu_0 (\phi(x))-H(p),\\[0.5cm]
&=& \displaystyle\phi'(x)v''_p(\phi(x))
- \displaystyle\frac{\phi''(x)}{\phi'(x)}v'_p(\phi(x))
+ \displaystyle\lp2\frac{(\phi''(x))^2}{(\phi'(x))^3}-\frac{\phi'''(x)}{(\phi'(x))^2}\right) v_p(\phi(x))\\[0.3cm]
&& \displaystyle-2\frac{\phi''(x)}{(\phi'(x))^2}v_p(\phi(x))v'_p(\phi(x))
+ \left( \displaystyle\frac{\phi''(x)}{(\phi'(x))^2}v_p(\phi(x))\right) ^2 \\[0.3cm]
&& \displaystyle+ 2p \frac{\phi''(x)}{(\phi'(x))^2}v_p(\phi(x)) .
\end{array}
\]
As $v_p$ is periodic and $W^{2,\infty}$, $v''_p$ is bounded. It follows from \eqref{hyp:phi2} that
\[
\displaystyle \frac{L_p\varphi_p(x)-H(p) \varphi_p(x)}{\varphi_p(x)} \rightarrow 0 \ \hbox{ as } x\rightarrow+\infty.
\]
\hfill $\Box$
\begin{lem}\label{lem:cvv}
Define $\varphi_p$ as in Lemma \ref{lem-vepapprox}.
Then
\begin{equation}\label{eq:cvv}
\frac{\ln \varphi_p (x)}{x} \rightarrow 0 \hbox{ as } x\rightarrow +\infty.
\end{equation}
\end{lem}
{\bf Proof.}
One has
\begin{equation}
\frac{\ln \varphi_p (x)}{x}=\frac{v_p (\phi (x))}{\phi'(x)x} \ \hbox{ for all }x\in\mathbb{R}.
\end{equation}
The function $x\mapsto v_p(\phi (x))$ is clearly bounded since $v_p$ is periodic. Hence, (\ref{hyp:phi2}) gives the conclusion.
\hfill$\Box$
\subsection{Upper bound for the spreading speed}\label{vitspread-p1}
\noindent {\bf Proof of part 1 of Theorem \ref{thm-vitspread}.} We first assume that $\mu_0 \in \mathcal{C}^2 (\mathbb{R})$. Let us now show that $w^* \leq w_\infty$. Let $c>w_\infty$ and $c_1 \in (w_\infty,c)$. We know
that there exists $p\geq j(M)>0$ such that
$$w_\infty=\min_{p' \geq j(M) }H(p')/p'= H(p)/p.$$
Let $k\geq M$ so that $p=j(k)>0,$ and $\varphi_p$ defined as in Lemma~\ref{lem-vepapprox}.
We know from Lemma~\ref{lem-vepapprox} that there exists $X>0$ such that:
\begin{equation}
|L_{p} \varphi_p(x) -k \varphi_p(x)| \leq (c_1-w_\infty) j(k)\varphi_p(x) \hbox{ for all } x>X.
\end{equation}
Let $\overline{u}$ be defined for all $(t,x) \in [0,+\infty) \times \mathbb{R}$ by:
$$\overline{u}(t,x):= \min \{1, \varphi_p (x)e^{-j(k) (x-h-c_1t)}\},$$
where $h\in\mathbb{R}$ is large enough so that $u_0 (x) \leq \overline{u}(0,x)$ for all $x\in \mathbb{R} $ (this is always possible since $u_0$ is compactly supported).
Moreover, $\overline{u}(t,x)<1$ if and only if $\varphi_p (x)<e^{j(k) (x-h-c_1t)}$, which is equivalent to
$ x-v_p (\phi (x))/(j(k)\phi '(x))> h+c_1t \geq h$. Lemma \ref{lem:cvv} yields that the left hand-side of this inequality goes to $+\infty$ as $x\rightarrow +\infty$. Hence, we
can always take $h$ large enough so that $\overline{u}(t,x) <1$ implies $x>X$.
It follows that for all $(t,x) \in [0,+\infty) \times \mathbb{R}$ such that $\overline{u}(t,x)<1$, one has
$$ \begin{array}{rcl}
\partial_t \overline{u} -\partial_{xx} \overline{u} -f(x,\overline{u}) &\geq & \partial_t \overline{u} -\partial_{xx} \overline{u} -\mu (x)\overline{u}\\[0.2cm]
&\geq & j(k) c_1 \overline{u} - L_{j(k)} \big(\varphi_p \big)(x) e^{-j(k) (x-h-c_1t)}\\[0.2cm]
&\geq & j(k) c_1 \overline{u} - k\overline{u} -(c_1-w_\infty) j(k)\overline{u}\\[0.2cm]
&\geq & \big( j(k) w_\infty - k\big)\overline{u} =0.
\end{array}$$
It follows from the parabolic maximum principle that $\overline{u }(t,x) \geq u(t,x) $ for all $(t,x) \in [0,+\infty) \times \mathbb{R}$.
Hence, for all given $x\in\mathbb{R}$,
$$ u(t,x) \leq \varphi_p (x) e^{-j(k) (x-h-c_1t)} \ \hbox{ for all } t>0.$$
Let $\varepsilon>0$ so that $\varepsilon < j(k) ( c-c_1)/c$. Lemma~\ref{lem:cvv} yields that there exists $R>0$ so that for all $x>R, \ln \big(\varphi_p (x) \big) \leq \varepsilon x$.
Let $T= R/c$ and take $t\geq T$ and $x\geq c t$. One has
$$\begin{array}{rcl}
\ln \Big(\varphi_p (x) e^{-j(k) (x-h-c_1t)}\Big)&=& \ln \big(\varphi_p (x) \big) -j(k) (x-h-c_1t)\\[0.2cm]
&\leq & (\varepsilon-j(k)) x +j(k)(h+c_1 t)\\[0.2cm]
&\leq & \big(\varepsilon c+j(k)(c_1-c) \big) t +j(k)h\\[0.2cm]
& \rightarrow & -\infty \hbox{ as } t\rightarrow +\infty\\
\end{array}$$
since $\varepsilon c<j(k)(c-c_1)$.
Hence,
$$\lim_{t\rightarrow +\infty} \max_{x\geq ct} \Big(\varphi_p (x) e^{-j(k) (x-h-c_1 t)}\Big)= 0 \hbox{ as } t\rightarrow +\infty,$$
which ends the proof in the case $\mu_0 \in \mathcal{C}^2 (\mathbb{R})$.
Lastly, if $\mu_0\in\mathcal{C}^0(\mathbb{R})$ is an arbitrary $1$-periodic function, then one easily concludes by smoothing $\mu_0$ from above. Indeed, one can find a sequence $(\mu_0^n )_n \in \mathcal{C}^2 (\mathbb{R})^{\mathbb{N}}$ converging uniformly to $\mu_0$, and such that for all $n\in \mathbb{N}$ and $x \in \mathbb{R}$, $\mu_0 (x) \leq \mu_0^n (x)$.
It follows from the maximum principle that
$$\lim_{t\rightarrow +\infty} \max_{x\geq ct}u(t,x)=0 \ \hbox{ for all } w > \min_{k\geq M} k/j^n(k),$$
where
$ \displaystyle j^n (k) = \int_0^1 \sqrt{ k-\mu_0^n (x)}dx \geq j (k) >0.$
Letting $n\rightarrow +\infty$, one gets
$$\lim_{t\rightarrow +\infty} \max_{x\geq ct}u(t,x)=0 \ \hbox{ for all } w > w_\infty,$$
which concludes the proof. \hfill $\Box$
\subsection{Lower bound on the spreading speed}\label{vitspread-p2}
\noindent {\bf Proof of part 2 of Theorem \ref{thm-vitspread}.}
First, assume that $\mu_0\in\mathcal{C}^2(\mathbb{R})$ satisfies (\ref{hyp:nondegmu}). Let $\varphi_p$ as in Lemma \ref{lem-vepapprox}. For all $\delta>0$,
take $R$ large enough so that
$L_p \varphi_p \geq (H(p) -\delta) \varphi_p$ at any point of $(R,+\infty)$ where $\varphi_p$ is piecewise $\mathcal{C}^2$.
It is easy to derive from the proof of Lemma \ref{lem-vepapprox} that $\varphi_p'/\varphi_p$ is bounded and uniformly continuous. Take
$C>0$ so that $|\varphi_p' (x)| \leq C \varphi_p (x)$ for all $x\in\mathbb{R}$.
We need more regularity in order to apply the results of~\cite{BerestyckiNadin}. Consider a compactly supported nonnegative
mollifier $\chi \in \mathcal{C}^\infty (\mathbb{R})$ so that $\int_\mathbb{R} \chi = 1$ and define the convoled function
$\psi_p := \exp \big(\chi \star \ln\varphi_p\big) \in\mathcal{C}^2 (\mathbb{R})$.
One has
$\psi_p'/\psi_p = \chi \star \Big(\varphi_p'/\varphi_p\Big). $
Hence, $|\psi_p' (x)| \leq C \psi_p (x)$ for all $x\in\mathbb{R}$ and, as $\varphi_p'/\varphi_p$ and $\big( \varphi_p'/\varphi_p \big)^2$ are uniformly continuous,
up to some rescaling of $\chi$, we can assume that
$$\Big\|\big|\chi \star \Big(\frac{\varphi_p'}{\varphi_p}\Big)\big|^2 -\big|\frac{\varphi_p'}{\varphi_p}\big|^2 \Big\|_\infty \leq \delta,
\ \Big\|\chi \star \Big( \big|\frac{\varphi_p'}{\varphi_p}\big|^2\Big)-\big|\frac{\varphi_p'}{\varphi_p}\big|^2\Big\|_\infty\leq \delta
\hbox{ and } \|\chi\star \mu -\mu \|_\infty \leq \delta.$$
We now compute
$$\frac{ \psi_p''}{\psi_p}=\Big|\frac{\psi_p'}{\psi_p}\Big|^2 +\chi \star \Big(\frac{\varphi_p''}{\varphi_p}-\Big|\frac{\varphi_p'}{\varphi_p}\Big|^2\Big)
=\Big|\chi \star \Big(\frac{\varphi_p'}{\varphi_p}\Big)\Big|^2 -\chi \star \Big( \Big|\frac{\varphi_p'}{\varphi_p}\Big|^2\Big)+\chi \star \Big(\frac{\varphi_p''}{\varphi_p}\Big)
\geq -2\delta +\chi \star \Big(\frac{\varphi_p''}{\varphi_p}\Big).$$
It follows that
$$ \begin{array}{rcl} \displaystyle \frac{L_p \psi_p}{\psi_p} = \frac{\psi_p''-2p \psi_p' + \mu (x)\psi_p}{\psi_p}&\geq&
-2\delta+\chi \star \Big( \displaystyle\frac{\varphi_p''}{\varphi_p}-2p \displaystyle\frac{\varphi_p'}{\varphi_p}\Big)+\mu (x)\\
&\geq & -2\delta+\chi \star \Big( H(p)-\mu-\delta \Big)+\mu (x)\\
&\geq & -3\delta + H(p)+\mu (x)-\chi \star \mu (x)\\
&\geq & -4\delta + H(p)\\
\end{array}$$
in $(R,+\infty)$.
On the other hand, Lemma~\ref{lem:cvv} yields $\psi_p \in\mathcal{A}_R$, where $\mathcal{A}_R$ is the set of admissible test-functions (in the sense of \cite{BerestyckiNadin}) over $(R,\infty)$:
\begin{equation} \label{defA} \begin{array}{rll}
\mathcal{A}_R:= \big\{&\psi\in \mathcal{C}^0([R,\infty) )\cap\mathcal{C}^{2}((R,\infty) ),&\\ &\psi'/\psi \in L^\infty ((R,\infty)), \
\psi>0 \hbox{ in } [R,\infty), \ \lim_{x\rightarrow +\infty} \frac{1}{x}\ln\psi (x) =0 & \big\}.\\ \end{array}
\end{equation}
Thus, one has $\underline{\lambda_1} (L_p ,(R,+\infty)) \geq H(p)-4\delta$, where the principal eigenvalue $\underline{\lambda_1}$ is defined by
\begin{equation}\label{deflambda1'}
\underline{\lambda_1}(L_p,(R,\infty) ):=\sup\{\lambda\ |\
\exists\phi\in \mathcal{A}_R \hbox{ such that } L_p\phi\geq\lambda\phi \hbox{ in }(R,\infty) \},
\end{equation}
Hence, $\lim_{R\rightarrow +\infty} \underline{\lambda_1} (L_p ,(R,+\infty)) \geq H(p)$ for all $p>0$.
In order to use Theorem 2.1 of \cite{BerestyckiNadin}, we need the nonlinearity to have two steady states and to be positive between these two steady states.
It is not the case here but we will bound $f$ from below by such a nonlinearity.
As $\min_\mathbb{R}\mu_0>0$ and $f$ is of class $\mathcal{C}^1$ in the neighborhood of $u=0$, we know that there exists $\theta \in (0,1)$ so that
$$f(x,u) >0 \hbox{ for all } x\in\mathbb{R} \ \hbox{ and } \ u\in (0,\theta).$$
Let $\zeta=\zeta (u)$ a smooth function so that
$$0< \zeta (u)\leq 1 \hbox{ for all } u\in (0,\theta), \hspace{0.2cm} \zeta (u)=0 \hbox{ for all } u\geq \theta \ \hbox{ and } \ \zeta (u)=1\hbox{ for all }u\in (0,\frac{\theta}{2}).$$
Define $\underline{f} (x,u) := \zeta (u) f(x,u)$ for all $(x,u) \in \mathbb{R}\times [0,1]$. Then
$$\underline{f} \leq f \hbox{ in } \mathbb{R}\times [0,1]\ \hbox{ and } \ \underline{f}_u' (x,0)= f_u'(x,0)=\mu_0 (\phi (x))\hbox{ for all } x\in\mathbb{R}.$$ Let $\underline{u}$ the solution of (\ref{eqn:eqRD}) with nonlinearity $\underline{f}$ instead of $f$ and initial datum $u_0$.
The parabolic maximum principle yields $u\geq \underline{u}$.
Since the function $\underline{f}$ satisfies the hypotheses of Theorem 2.1 in \cite{BerestyckiNadin}, we conclude that
$$\lim_{t\rightarrow +\infty} \min_{x\in [0, wt]} \underline{u} (t,x) =1 \hbox{ for all } w\in \Big( 0, \min_{p>0} \displaystyle \frac{H(p)}{p}\Big).$$
It follows that
$$ w_* \geq \min_{p>0} \displaystyle \frac{H(p)}{p} = \min_{k \geq M} \frac{k}{j(k)}.$$
Next, assume that $\mu_0\in\mathcal{C}^2(\mathbb{R})$ does not satisfy (\ref{hyp:nondegmu}).
Let $\overline{y}\in\mathbb{R}$ so that $\mu_0 (\overline{y}) = \displaystyle\max_{y\in\mathbb{R}} \mu_0 (y)$.
Take a $1$-periodic function $\chi \in\mathcal{C}^2 (\mathbb{R})$ so that $\chi (0)=0$, $\chi (y) >0$ for all $y\neq 0$ and $\chi ''(0) >0$. Define for all $n\in\mathbb{N}$ and~$x\in\mathbb{R}$:
$$\mu_0^n (y) := \mu_0 (y) - \frac{1}{n} \chi (y-\overline{y}).$$
This $1$-periodic function satisfies (\ref{hyp:nondegmu}) for all $n$ and one has $0< \mu_0^n \leq \mu_0$ for~$n$ large enough. It follows from the maximum principle that
$$\liminf_{t\rightarrow +\infty} \min_{0\leq x\leq wt}u(t,x) >0 \hbox{ for all } w\in \Big(0,\displaystyle \min_{k\geq M}{ \frac{k}{j^n(k)}}\Big),$$
where
$\displaystyle j^n (k) = \int_0^1 \sqrt{ k-\mu_0^n (x)}dx \geq j (k) > 0$ for all $k \geq M$.
Letting $n\rightarrow +\infty$, one has $\mu_0^n (y) \rightarrow \mu_0 (y)$ uniformly in~$y\in\mathbb{R}$ and thus
$$\liminf_{t\rightarrow +\infty} \min_{0\leq x\leq wt}u(t,x) >0 \hbox{ for all } w\in (0, w_\infty),$$
which concludes the proof in this case.
\smallskip
Lastly, if $\mu_0\in\mathcal{C}^0(\mathbb{R})$ is an arbitrary $1$-periodic function, then one easily concludes by smoothing $\mu_0$ as in the previous step.
\hfill $\Box$
|
\section{Introduction}
By the end of the last century, the Big Bang Model had been worked
out. It contained a huge amount of unobserved, hypothesized "matter"
of a new kind - dark matter. This was postulated as long back as the
1930s to explain the fact that the velocity curves of the stars in
the galaxies did not fall off, as they should. Instead they
flattened out, suggesting that the galaxies contained some
undetected and therefore non-luminous or {\bf dark matter}. The
identity of this dark matter has been a matter of guess work,
though. It could consist of Weakly Interacting Massive Particles
(WIMPS) or Super Symmetric partners of existing particles. Or heavy
neutrinos or monopoles or unobserved brown dwarf stars and so on.\\
In fact Prof. Abdus Salam speculated some two decades ago
\cite{salamnap} "And now we come upon the question of dark matter
which is one of the open problems of cosmology. This is a problem
which was speculated upon by Zwicky fifty years ago. He showed that
visible matter of the mass of the galaxies in the Coma cluster was
inadequate to keep the galactic cluster bound. Oort claimed that the
mass necessary to keep our own galaxy together was at least three
times that concentrated into observable stars. And this in turn has
emerged as a central problem
of cosmology.\\
"You see there is the matter which we see in our galaxy. This is
what we suspect from the spiral character of the galaxy keeping it
together. And there is dark matter which is not seen at all by any
means whatsoever. Now the question is what does the dark matter
consist of? This is what we suspect should be there to keep the
galaxy bound. And so three times the mass of the matter here in our
galaxy should be around in the form of the invisible matter. This is
one of the
speculations."\\
The universe in this picture, contained enough of the mysterious
dark matter to halt the expansion and eventually trigger the next
collapse. It must be mentioned that the latest WMAP survey
\cite{science2}, in a model dependent result indicates that as much
as twenty three percent of the Universe is made up of dark matter,
though there is no definite observational confirmation
of its existence.\\
That is, the Universe would expand
up to a point and then collapse.\\
There still were several subtler problems to be addressed. One was
the famous {\bf horizon problem}. To put it simply, the Big Bang was
an uncontrolled or random event and so, different parts of the
Universe in different directions were disconnected at the very
earliest stage and even today, light would not have had enough time
to connect them. So they need not be the same. Observation however
shows that the Universe is by and large uniform, rather like people
in different countries showing the same habits or dress. That would
not be possible without some form of faster than light
intercommunication which would violate Einstein's Special
Theory of Relativity.\\
The next problem was that according to Einstein, due to the material
content in the Universe, space should be curved whereas the Universe
appears to be {\bf flat}.\\
There were other problems as well. For example astronomers predicted
that there should be {\bf monopoles} that is, simply put, either
only North magnetic poles or only South magnetic poles, unlike the
North South combined magnetic poles we encounter. Such monopoles
have
failed to show up even after seventy five years.\\
Some of these problems as we noted, were sought to be explained by
what has been called {\bf inflationary cosmology} whereby, early on,
just after
the Big Bang the explosion was super fast \cite{zee,lindepl82}.\\
What would happen in this case is, that different parts of the
Universe, which could not be accessible by light, would now get
connected. At the same time, the super fast expansion in the initial
stages would smoothen out any distortion or curvature effects in
space,
leading to a flat Universe and in the process also eliminate the monopoles.\\
Nevertheless, inflation theory has its problems. It does not seem to
explain the cosmological constant observed since. Further, this
theory seems to imply that the fluctuations it produces should
continue to indefinite distances. Observation seems to imply the
contrary.\\
One other feature that has been studied in detail over the past few
decades is that of {\bf structure formation} in the Universe. To put
it simply, why is the Universe not a uniform spread of matter and
radiation? On the contrary it is very lumpy with planets, stars,
galaxies and so on, with a lot of space separating these objects.
This has been explained in terms of fluctuations in density, that
is, accidentally more matter being present in a given region.
Gravitation would then draw in even more matter and so on. These
fluctuations would also cause the cosmic background radiation to be
non uniform or anisotropic. Such anisotropies are in fact being
observed. But this is not the end of the story. The galaxies seem to
be arranged along two dimensional structures and filaments with huge
separating
voids.\\
From 1997, the conventional wisdom of cosmology that had concretized
from the mid sixties onwards, began to be challenged. It had been
believed that the density of the Universe is near its critical
value, separating eternal expansion and ultimate contraction, while
the nuances of the dark matter theories were being fine tuned. But
that year, the author proposed a contra view, which we will examine.
\section{Cosmology}
To proceed, as there are $N \sim 10^{80}$ such particles in the
Universe, we get, consistently,
\begin{equation}
Nm = M\label{3e1}
\end{equation}
where $M$ is the mass of the Universe. It must be remembered that
the energy of gravitational interaction between the particles is
very much insignificant compared to the above electromagnetic considerations.\\
In the following we will use $N$ as the sole cosmological parameter.\\
We next invoke the well known relation
\cite{bgsfluc,nottalefractal,hayakawa}
\begin{equation}
R \approx \frac{GM}{c^2}\label{3e2}
\end{equation}
where $M$ can be obtained from (\ref{3e1}). We can arrive at
(\ref{3e2}) in different ways. For example, in a uniformly expanding
Friedman Universe, we have
$$\dot{R}^{2} = 8 \pi G\rho R^2/3$$
In the above if we substitute $\dot{R} = c$ at $R$, the radius of
the universe, we
get (\ref{3e2}). Another proof can also be given \\
We now use the fact that given $N$ particles, the
(Gaussian)fluctuation in the particle number is of the order
$\sqrt{N}$\cite{hayakawa,huang,ijmpa,ijtp,bgsfqp,bgsmg8}, while a
typical time interval for the fluctuations is $\sim \hbar/mc^2$, the
Compton time, the fuzzy interval within which there is no meaningful
physics as argued by Dirac and in greater detail by Wigner and
Salecker. So particles are created and destroyed - but the ultimate
result is that $\sqrt{N}$ particles are created just as this is the
nett displacement in a random walk of unit step. So we have,
\begin{equation}
\frac{dN}{dt} = \frac{\sqrt{N}}{\tau}\label{3ex}
\end{equation}
whence on integration we get, (remembering that we are almost in the
continuum region that is, $\tau \sim 10^{-23}sec \approx 0$),
\begin{equation}
T = \frac{\hbar}{mc^2} \sqrt{N}\label{3e3}
\end{equation}
We can easily verify that the equation (\ref{3e3}) is indeed
satisfied where $T$ is the age of the Universe. Next by
differentiating (\ref{3e2}) with respect to $t$ we get
\begin{equation}
\frac{dR}{dt} \approx HR\label{3e4}
\end{equation}
where $H$ in (\ref{3e4}) can be identified with the Hubble Constant,
and using (\ref{3e2}) is given by,
\begin{equation}
H = \frac{Gm^3c}{\hbar^2}\label{3e5}
\end{equation}
Already this shows an exponential inflationary behaviour with accelaration while Equation (\ref{3e1}), (\ref{3e2}) and (\ref{3e3}) show that in this
formulation, the correct mass, radius, Hubble constant and age of
the Universe can be deduced given $N$, the number of particles, as
the sole cosmological or large scale parameter. We observe that at
this stage we are not invoking any particular dynamics - the
expansion is due to the random creation of particles from the ZPF
background. Equation (\ref{3e5}) can be written as
\begin{equation}
m \approx \left(\frac{H\hbar^2}{Gc}\right)^{\frac{1}{3}}\label{3e6}
\end{equation}
Equation (\ref{3e6}) has been empirically known as an "accidental"
or "mysterious" relation. As observed by Weinberg \cite{weinberggc},
this is unexplained: it relates a single cosmological parameter $H$
to constants from microphysics. We will touch upon this micro-macro
nexus again. In our formulation, equation (\ref{3e6}) is no longer a
mysterious coincidence but
rather a consequence of the theory.\\
As (\ref{3e5}) and (\ref{3e4}) are not exact equations but rather,
order of magnitude relations, it follows, on differentiating
(\ref{3e4}) that a small cosmological constant $\wedge$ is allowed
such that
$$\wedge \leq 0 (H^2)$$
This is consistent with observation and shows that $\wedge$ is very
small $--$ this has been a puzzle, the so called cosmological
constant problem alluded to, because in conventional theory, it
turns out to be huge \cite{weinbergprl}. But it poses no problem in
this formulation. This is because of the characterization of the ZPF
as independent and primary in our formulation this being the
mysterious dark energy. Otherwise we would encounter the
cosmological constant problem of Weinberg: a $\wedge$ that is some
$10^{120}$ orders of magnitude of observable values!\\
To proceed we observe that because of the fluctuation of $\sim
\sqrt{N}$ (due to the ZPF), there is an excess electrical potential
energy of the electron, which in fact we identify as its inertial
energy. That is \cite{ijmpa,hayakawa},
$$\sqrt{N} e^2/R \approx mc^2.$$
On using (\ref{3e2}) in the above, we recover the well known
Gravitation-Electromagnetism ratio viz.,
\begin{equation}
e^2/Gm^2 \sim \sqrt{N} \approx 10^{40}\label{3e7}
\end{equation}
or without using (\ref{3e2}), we get, instead, the well known so
called Weyl-Eddington formula,
\begin{equation}
R = \sqrt{N}l\label{3e8}
\end{equation}
(It appears that (\ref{3e8}) was first noticed by H. Weyl
\cite{singh}). Infact (\ref{3e8}) is the spatial counterpart of
(\ref{3e3}). If we combine (\ref{3e8}) and (\ref{3e2}), we get,
\begin{equation}
\frac{Gm}{lc^2} = \frac{1}{\sqrt{N}} \propto T^{-1}\label{3e9}
\end{equation}
where in (\ref{3e9}), we have used (\ref{3e3}). Following Dirac
(cf.also \cite{melnikov}) we treat $G$ as the variable, rather than
the quantities $m, l, c \,\mbox{and}\, \hbar$ which we will call
micro physical constants because of their central role
in atomic (and sub atomic) physics.\\
Next if we use $G$ from (\ref{3e9}) in (\ref{3e5}), we can see that
\begin{equation}
H = \frac{c}{l} \quad \frac{1}{\sqrt{N}}\label{3e10}
\end{equation}
Thus apart from the fact that $H$ has the same inverse time
dependance on $T$ as $G$, (\ref{3e10}) shows that given the
microphysical constants, and
$N$, we can deduce the Hubble Constant also, as from (\ref{3e10}) or (\ref{3e5}).\\
Using (\ref{3e1}) and (\ref{3e2}), we can now deduce that
\begin{equation}
\rho \approx \frac{m}{l^3} \quad \frac{1}{\sqrt{N}}\label{3e11}
\end{equation}
Next (\ref{3e8}) and (\ref{3e3}) give,
\begin{equation}
R = cT\label{3e12}
\end{equation}
Equations (\ref{3e11}) and (\ref{3e12}) are consistent with observation.\\
Finally, we observe that using $M,G \mbox{and} H$ from the above, we
get
$$M = \frac{c^3}{GH}$$
This relation is required in the Friedman model of the expanding
Universe (and the Steady State model too). In fact if we use in this
relation, the expression,
$$H = c/R$$
which follows from (\ref{3e10}) and (\ref{3e8}), then we recover
(\ref{3e2}). We will be repeatedly using these relations in the sequel.\\
As we saw the above model predicts a dark
energy driven ever expanding and accelerating Universe with a small
cosmological constant while the density keeps decreasing. Moreover
mysterious large number relations like (\ref{3e5}), (\ref{3e11}) or
(\ref{3e8}) which were considered to be miraculous accidents now
follow from the underlying theory. This seemed to go against the
accepted idea that the density of the Universe equalled the critical
density required for closure and that aided by dark matter, the
Universe was decelerating.\\
However, as noted, from 1998 onwards, following the work of {\bf
Perlmutter}, {\bf Schmidt} and {\bf Riess}, these otherwise apparently heretic conclusions have been vindicated.\\
It may be mentioned that the observational evidence for an
accelerating Universe was the American Association for Advancement
of Science's Breakthrough of the Year, 1998 while the evidence for
nearly seventy five percent of the Universe being Dark Energy, based
on the Wilkinson Microwave Anisotropy Probe (WMAP) and the Sloan Sky
Digital Survey was the Breakthrough of the Year, 2003
\cite{science1,science2}. The trio got the 2011 Nobel for Physics.
\section{Discussion}
1. We observe that in the above scheme if the Compton time $\tau \to
\tau_P$, we recover the Prigogine Cosmology \cite{prig,tryon}. In
this case there is a {\bf phase transition} in the background ZPF or
Quantum Vacuum or Dark Energy and Planck scale particles are
produced.\\
On the other hand if $\tau \to 0$ (that is we return to point
spacetime), we recover the Standard Big Bang picture. But it must be
emphasized that in neither of these two special cases can we recover
the various so called Large Number coincidences for example
Equations like (\ref{3e3}) or (\ref{3e5}) or (\ref{3e7}) or
(\ref{3e8}).\\
2. The above ideas lead to an important characterization of
gravitation. This also explains why it has not been possible to
unify gravitation with other interactions, despite nearly a century
of effort.\\
Gravitation is the only interaction that could not be satisfactorily
unified with the other fundamental interactions. The starting point
has been a diffusion equation
$$| \Delta x|^2 = < \Delta x^2 > = \nu \cdot \Delta t$$
\begin{equation}
\nu = \hbar/m, \nu \approx l v\label{2e3}
\end{equation}
This way we could explain a process similar to the formation of
Benard cells \cite{tduniv,prig} -- there would be sudden formation
of the ``cells" from the background dark energy, each at the Planck
Scale, which is the smallest physical scale. These in turn would be
the underpinning for
spacetime.\\
We could consider an array of $N$ such Planckian cells
\cite{bgsijtp}. This would be described by
\begin{equation}
r = \sqrt{N \Delta x^2}\label{4De1d}
\end{equation}
\begin{equation}
ka^2 \equiv k \Delta x^2 = \frac{1}{2} k_B T\label{4De2d}
\end{equation}
where $k_B$ is the Boltzmann constant, $T$ the temperature, $r$ the
extent and $k$ is the spring constant given by
\begin{equation}
\omega_0^2 = \frac{k}{m}\label{4De3d}
\end{equation}
\begin{equation}
\omega = \left(\frac{k}{m}a^2\right)^{\frac{1}{2}} \frac{1}{r} =
\omega_0 \frac{a}{r}\label{4De4d}
\end{equation}
We now identify the particles or cells with \index{Planck}Planck
\index{mass}masses and set $\Delta x \equiv a = l_P$, the
\index{Planck}Planck length. It may be immediately observed that use
of (\ref{4De3d}) and (\ref{4De2d}) gives $k_B T \sim m_P c^2$, which
ofcourse agrees with the temperature of a \index{black hole}black
hole of \index{Planck}Planck \index{mass}mass. Indeed, Rosen
\cite{rosen} had shown that a \index{Planck}Planck \index{mass}mass
particle at the \index{Planck scale}Planck scale can be considered
to be a \index{Universe}Universe in itself with a Schwarzchild
radius equalling the Planck length. We also use the fact alluded to
that a typical elementary particle like the \index{pion}pion can be
considered to be the result of $n \sim 10^{40}$ \index{Planck}Planck
\index{mass}masses.\\
Using this in (\ref{4De1d}), we get $r \sim l$, the \index{pion}pion
\index{Compton wavelength}Compton wavelength as required. Whence the
pion mass is given by
$$m = m_P/\sqrt{n}$$
which of course is correct, with the choice of $n$. This can be
described by
\begin{equation}
l = \sqrt{n} l_P, \, \tau = \sqrt{n} \tau_P,\label{3e31}
\end{equation}
$$l^2_P = \frac{\hbar}{m_P} \tau_P$$
The last equation is the analogue of the diffusion process seen,
which is in fact the underpinning for particles, except that this
time we have the same Brownian process operating from the Planck
scale to the Compton scale (Cf. also \cite{bgsfpl152002,cu}).\\
We now use the well known result alluded to that the individual
minimal oscillators are black holes or mini Universes as shown by
Rosen \cite{rosen}. So using the Beckenstein temperature formula for
these primordial black holes \cite{ruffinizang}, that is
$$kT = \frac{\hbar c^3}{8\pi Gm}$$
we can show that
\begin{equation}
Gm^2 \sim \hbar c\label{4e4}
\end{equation}
We can easily verify that (\ref{4e4}) leads to the value $m \sim
10^{-5}gms$. In deducing (\ref{4e4}) we have used the typical
expressions for the frequency as the inverse of the time - the
Compton time in this case and similarly the expression for the
Compton length. However it must be reiterated that no specific
values
for $l$ or $m$ were considered in the deduction of (\ref{4e4}).\\
We now make two interesting comments. Cercignani and co-workers have
shown \cite{cer1,cer2} that when the gravitational energy becomes of
the order of the electromagnetic energy in the case of the Zero
Point oscillators, that is
\begin{equation}
\frac{G\hbar^2 \omega^3}{c^5} \sim \hbar \omega\label{4e5}
\end{equation}
then this defines a threshold frequency $\omega_{max}$ above which
the oscillations become chaotic. In other words, for meaningful
physics we require that
$$\omega \leq \omega_{max}.$$
Secondly as we can see from the parallel but unrelated theory of
phonons \cite{huang,reif}, which are also bosonic oscillators, we
deduce a maximal frequency given by
\begin{equation}
\omega^2_{max} = \frac{c^2}{l^2}\label{4e6}
\end{equation}
In (\ref{4e6}) $c$ is, in the particular case of phonons, the
velocity of propagation, that is the velocity of sound, whereas in
our case this velocity is that of light. Frequencies greater than
$\omega_{max}$ in (\ref{4e6}) are again meaningless.
We can easily verify that using (\ref{4e5}) in (\ref{4e6}) gives back (\ref{4e4}).\\
In other words, gravitation shows up as the residual energy from the
formation of the particles in the universe via Planck scales (Benard
like) cells.\\
3. It has been mentioned that despite nearly 75 years of search,
Dark Matter has not been found. More recently there is evidence
against the existence of Dark Matter or its previous models. The
latest LHC results for example seem to rule out SUSY.\\
On the other hand our formulation obviates the need for Dark Matter.
This follows from an equation like (\ref{3e9}) which shows a
gravitational constant decreasing with time. Starting from here it
is possible to deduce not just the anomalous rotation curves of
galaxies which was the starting point for Dark Matter; but also we
could deduce all the known standard results of General Relativity
like the precession of the perihelion of mercury, the bending of
light, the progressive shortening of the time period of binary
pulsars and so on (Cf.ref.\cite{tduniv}).\\
4. {\bf Epilogue}: The idea of a perfect vacuum began to get frayed
in the 19th century. In the 20th century with the advent of Quantum
Theory the concept of a Quantum Vacuum came into being. This Quantum
Vacuum is seething with energy and activity, and it is there
everywhere. With this background we can see the following:\\
Around 1997 I had put forward a radically different model. In this,
there wasn't any \index{Big Bang}Big Bang, with matter and energy
being created instantaneously. Rather the universe is permeated by
an energy field of a kind familiar to modern physicists. The point
is, that according to Quantum theory which is undoubtedly one of the
great intellectual triumphs of the twentieth century, all our
measurements, and that includes measurements of energy, are at best
approximate. There is always a residual error. This leads to what
physicists call a ubiquitous \index{Zero Point Field}Zero Point
Field or \index{Quantum Vacuum}Quantum Vacuum. We will return to
this ``\index{Dark Energy}Dark Energy" soon. Out of such a ghost
background or all pervading energy field, particles are created in a
totally random manner, a process that keeps continuing. However,
much of the matter was created in a fraction of a second. There is
no ``Big Bang" singularity, though, which had posed Wheeler's
greatest problem of physics. The contents of this paper went
diametrically opposite to accepted ideas, that the universe,
dominated by dark matter was actually decelarating. Rather, driven
by dark energy, the universe would be expanding and accelerating,
though slowly. I was quite sure that this paper would be rejected
outright by any reputable scientific journal. So I presented these
ideas at the prestigious \index{Marcell Grossmann meet}Marcell
Grossmann meet in Jerusalem and another International Conference on
Quantum Physics. But, not giving into pessimism, I shot off the
paper to a standard International journal, anyway. To my great
surprise,
it was accepted immediately!\\
There is a further cosmic foot print of this model: a residual
miniscule energy in the \index{Cosmic Microwave Background}Cosmic
Microwave Background, less than a billion billion billion billionth
of the energy of an electron. Latest data has confirmed the presence
of such an energy. All this is in the spirit of the manifest
universe springing out of an unmanifest background, as described in
the \index{Bhagvad Gita}Bhagvad Gita. There are
several interesting consequences.\\
Firstly it is possible to theoretically estimate the size and age of
the universe and also deduce a number of very interesting
interrelationships between several physical quantities like the
charge of the electron, the mass of elementary particles, the
gravitational constant, the number of particles in the universe and
so on. One such, connecting the gravitational constant and the mass
of an elementary particle with the expansion of the universe was
dubbed as inexplicable by Nobel Laureate \index{Steven
Weinberg}Steven Weinberg. But on the whole these intriguing
interrelationships have been considered by
most scientists to be miraculous coincidences.\\
With one exception. The well known Nobel Prize winning physicist
\index{Paul Dirac}Paul Dirac sought to find an underlying reason to
explain what would otherwise pass off as a series of inexplicable
accidents. In this model, there is a departure from previous
theories including the fact that some supposedly constant quantities
like the universal constant of gravitation are actually varying very
slowly with time. Interestingly latest observations seem to point
the finger in this direction.\\
However my model is somewhat different and deduces these mysterious
relations. Further, it sticks its neck out in predicting that the
universe is not only expanding, but also accelerating as it does so.
This went against all known wisdom. Shortly thereafter from 1998
astronomers like \index{Perlmutter}Perlmutter and
\index{Kirshner}Kirshner began to publish observations which
confirmed exactly such a behavior. These shocking results have since
been reconfirmed. The universe had taken
a U Turn.\\
When questioned several astronomers in 1998 confided to me that the
observations were wrong! After the expansion was reconfirmed, some
became cautious. Let us wait and see. At the same time, some rushed
back to their desks and tried to rework their calculations. The
other matter was, what force could cause the accelerated expansion?
The answer would be, some new and inexplicable form of energy, as
suggested by me. \index{Dark Energy}Dark Energy. Later the presence
of dark energy was confirmed by the Wilkinson Microwave Probe
(\index{WMAP}WMAP) and the \index{Sloane Digital Sky Survey}Sloane
Digital Sky Survey. Both these findings were declared by the
prestigious journal Science as breakthroughs of the respective years.\\
The accelerated expansion of the universe and the possibility that
supposedly eternally constant quantities could vary, has been the
new paradigm gifted to science, a parting gift by the departing
millennium.\\
A 2000 article in the \index{Scientific American}Scientific American
observed, ``In recent years the field of cosmology has gone through
a radical upheaval. New discoveries have challenged long held
theories about the evolution of the Universe... Now that observers
have made a strong case for cosmic acceleration, theorists must
explain it.... If the recent turmoil is anything to go by, we had
better keep our options
open."\\
On the other hand, an article in \index{Physics World}Physics World
in the same year noted , ``A revolution is taking place in
cosmology. New ideas are usurping traditional notions about the
composition of the Universe, the relationship between geometry and
destiny, and Einstein's greatest blunder."\\
It is this greatest blunder of Einstein which got the Nobel Prize
for Physics in 2011 for three US astronomers, Perlmutter, Reiss and
Schmidt who observed the accelerated expansion of the universe in
1988.
|
\section{Introduction}
No-scale supergravity models~\cite{Cremmer:1983bf,Ellis:1983sf,Ellis:1983ei}
are a specific set of supergravity models,
in which the vanishing of the tree-level potential in the
hidden sector direction can be automatic for an appropriately chosen form of the K\"ahler potential.
Moreover, the value of the gravitino mass
$m_{3/2}$ can be fixed dynamically by (non-gravitational) radiative correction stabilization, and is related
to other soft SUSY-breaking parameters.
This no-scale mechanism has been known for a long time, but the
complexity of a full minimization of the effective potential lead in the early days to consider only
specific approximations. More recently the strict no-scale boundary conditions $m_0=A_0=0$
have often been studied for their phenomenological consequences but without specifying a precise link
with the above-mentioned scalar potential minimization.
Modern MSSM spectrum calculation tools allow
to incorporate the full one-loop as well as dominant two-loop contributions to the effective scalar potential
and other important radiative corrections.
We will take advantage of this to go further in the study of no-scale models~\cite{Benhenni:2011yt},
implementing the minimization mechanism within SuSpect \cite{Djouadi:2002ze}. In a generalized no-scale inspired
framework, it first requires the definition of the soft parameters at the GUT scale
\begin{eqnarray}
B_0= b_0 \ m_{1/2},\;\; m_0= x_0 \ m_{1/2},\;\; A_0=a_0 \ m_{1/2};
\label{nsbc}
\end{eqnarray}
where the gaugino mass is the unique scale parameter, and the strict no-scale corresponds to
$b_0=x_0=a_0=0$.
Notice that the usual $\tan \beta$ input is replaced by $B_0$, the former
being consistently derived at the electroweak scale.
In addition to this usual parameter set, we will have to consider a new one,
in the form of a boundary condition $\eta_0$ for a vacuum energy term, following \cite{Kounnas:1994fr}.
\section{Renormalization Group invariant effective potential}
The vacuum energy term $\eta_0$ finds its roots in renormalization group (RG) invariance properties
of the effective potential~\cite{VRGinv}.
Adding one-loop contributions to the tree-level potential already ensures a more stable physical spectrum,
but without the vacuum energy contribution the effective potential is not RG-invariant.
In particular in the no-scale approach one is interested in the overall shape of the potential, and
obviously a meaningful minimum is expected to be scale-independent, as much as possible
perturbatively. The (one-loop) RG-invariant potential reads:
\begin{eqnarray}
V_{full} \equiv V_{tree}(Q) +V_{1-loop}(Q) +\tilde{\eta}(Q) m^4_{1/2}
\label{Vfull}
\end{eqnarray}
where as usual the one-loop contribution is expressed in terms of (field dependent) eigenmasses as
\begin{eqnarray}
V_{1-loop}(Q) = \frac {1}{64\pi ^2}
\sum_{all n} (-1)^{2n} M_n^4(H_u,H_d)(\ln \frac{ M_n^2(H_u,H_d)}{Q^2}-\frac 32)
\end{eqnarray}
and in Eq.~(\ref{Vfull}) the vacuum energy term is conveniently scaled by $m_{1/2}$
without much loss of generality.
$\eta(Q)$ runs from $\eta_0$ at GUT scale to $\eta_{EW}$ at EW scale. We have checked that
this term is not only crucial for RG-invariance and stability of the potential and corresponding
physical minimum, but its contribution is also strongly correlated with the position of the minimum,
and the corresponding value of the gaugino mass $m_{1/2}$.
\section{The minimization procedure}
The generalized no-scale electroweak minimization implies, in addition to the two usual EW minimizations
$ \frac{\partial V_{full}}{\partial v_i}=0$, $i=u,d$, an extra minimization in the gaugino direction:
$\frac{\partial V_{full}}{\partial m_{1/2}} = 0 $. This will dynamically determine the soft
parameters if all related to $m_{1/2}$ as in Eq.~(\ref{nsbc}).
Assuming furthermore $\mu \sim m_{1/2}$, the latter minimization takes the convenient form~\cite{Kounnas:1994fr,Benhenni:2011yt}
\begin{eqnarray}
V_{full}(m_{1/2}) +\frac{1}{128\pi^2} \sum_n (-1)^{2n} M^4_n(m_{1/2})
+ \frac{1}{4} m_{1/2}^5 \frac{d \tilde{\eta}_0}{d m_{1/2}} =0
\label{dVm12}
\end{eqnarray}
where the last term is non-vanishing only in case $\tilde \eta_0$ may be a
non-trivial function of $m_{1/2}$.
Upon minimizing the potential, care is to be taken when handling some of the physical constraints.
Typically the right $m_Z$ mass constraint, $m^2_Z ={ v^2 \over 2 } (g'^2+g^2)$,
and the pole-to-running mass relations
(mostly in the top quark sector due to strong dependence on the top quark Yukawa coupling),
$m^{pole}_{top} = Y_t(Q) v_u(Q) (1+\delta^{RC}_y(Q)+\cdots) \; ,
\label{Ythresh}$
should be imposed only {\em after} the global minimum in the three directions $v_u, v_d, m_{1/2}$ has been
found. This implies deviations of a few percent
in the $m_{1/2}$ minima values when taken into account properly~\cite{Benhenni:2011yt}.
\section{No-scale favored regions}
On phenomenological grounds, the no-scale mechanism generally favours a charged (mostly $\tilde \tau$) LSP
for $m_0=0$ or small enough,.
This is not a problem as it is natural to consider the gravitino
as the true LSP within this framework. Current sparticle mass limits from the
LHC~\cite{LHClimits} exclude small
$m_{1/2} \raisebox{-0.13cm}{~\shortstack{$<$ \\[-0.07cm] $\sim$}}~ 300-350$ GeV values. Other indirect constraints, such as $B\to s\gamma$ measurements,
LEP Higgs mass bounds, etc, can be accomodated for sufficiently large $m_{1/2}$.
In our case, $m_{1/2}$ limits translate into bounds on $\eta_0$ values, favouring lower values $\eta_0\raisebox{-0.13cm}{~\shortstack{$<$ \\[-0.07cm] $\sim$}}~ 8-10$
(depending on other parameters, $B_0$ etc)~\cite{Benhenni:2011yt}. But it is
still possible to have viable parameter regions with non-trivial $m_{1/2}$ minima, including
even a decoupled supersymmetric spectrum with a light SM-like Higgs, when $\eta_0 \simeq 0$.
\section{Gravitino dark-matter}
For scenarios with a gravitino LSP (with stau as NLSP), all supersymmetric particles decay to the
NLSP well before the latter has decayed to a gravitino, because all interactions to the gravitino
are suppressed by the Planck mass. We first compute, using micrOMEGAs 2.0 \cite{Belanger:2006is},
the relic density $\Omega_{\rm NLSP}h^2$ the NLSP would have if it did not decay to the gravitino. Then
assuming that each NLSP with mass $m_{\rm NLSP}$ decays to one gravitino, leads to the non-thermal
contribution to the gravitino relic density
\begin{eqnarray}
\Omega_{3/2}^{\rm NTP}h^2={m_{3/2} \over m_{\rm NLSP}} \Omega_{\rm NLSP}h^2
\label{omntp}
\end{eqnarray}
with $h=0.73^{+0.04}_{-0.03}$ the Hubble constant.
\begin{figure}[h!]
\centering
\includegraphics[width=6cm]{NoScale_mhalf_tb_m0_0_mgrav_01mhalf_NewScan2_tb_mhalf_gravrelic}
\includegraphics[width=6cm]{NoScale_mhalf_tb_m0eq05m12_mgrav_01mhalf_NewScan2_tb_mhalf_gravrelic}
\caption{Gravitino LSP and relic density}
\end{figure}
The gravitino can also be produced during reheating after inflation.
The gravitino relic
density from such thermal production, $\Omega_{3/2}^{\rm TP}h^2$, is essentially
controlled by the reheat temperature $T_R$ (see e.g \cite{Pradler:2006qh}).
Comparing the total gravitino relic density $\Omega_{3/2}^{\rm TP}h^2
+\Omega_{3/2}^{\rm NTP}h^2$,
to WMAP constraints\cite{Spergel:2006hy}, will constrain $T_R$ together with the no-scale parameter space.
We illustrate two representative cases in the
($\tan \beta$, $m_{1/2}$) plane, one for the strict no-scale scenario ($m_0=A_0=0$), and another less
stringent scenario where the neutralino has some room as the LSP (though only for rather low $m_{1/2}$).
One recovers consistency with the WMAP relic density constraint in a large part
of the parameter space, provided that $T_R$ is sufficiently large, $T_R\raisebox{-0.13cm}{~\shortstack{$>$ \\[-0.07cm] $\sim$}}~ 10^6$ GeV.
In particular even for the strict no-scale model $B_0=m_0=A_0=0$
there is a range for $m_{1/2}\sim 400-800$ GeV, $\tan\beta\sim 20-25$ compatible
with all present constraints, provided that the reheat temperature is
$10^8-10^9$ GeV.
|
\section{Introduction}
Zeros and factorisations of lacunary polynomials, that is,
polynomials of high degree with relatively small number of
non-zero coefficients, has always been a subject of active
investigation, see~\cite{CTV,FGS,Len1,Len2,Schin} and references
therein. We say that a polynomial $f$ over a field ${\mathbb K}$ is
$k$-lacunary if it has at most $k+1$ non-zero coefficients,
including a non-zero constant term, that is, if $f(0) \ne 0$ and
\begin{equation}
\label{eq:LacPoly}
f(X) = a_0 + a_1X^{t_1} + \ldots + a_kX^{t_k} \in {\mathbb K}[X]
\end{equation}
for some positive integers $t_1 < \ldots <t_k$.
For example, a classical result of Descartes asserts that a
$k$-lacunary polynomial $f\in {\mathbb R}[X]$ may have at most $2k$ real
roots. Furthermore, Lenstra~\cite{Len2} has shown that for an
algebraic number field ${\mathbb K}$ of degree $m$ over ${\mathbb Q}$ and a
$k$-lacunary polynomial $f\in {\mathbb K}[X]$, the product $g$ of all
irreducible divisors $h\mid f$ of degree at most $\deg h \le d$
is of degree
$$
\deg g = O\(k^2 2^{md} md \log (2mdk)\).
$$
Schinzel~\cite{Schin} has obtained a series of statistical results
about the number of $k$-lacunary irreducible polynomials with
prescribed coefficients. In particular,
by~\cite[Corollary~2]{Schin}, for any algebraic numbers $a_0,
\ldots,a_k$ there are at most $O\(T^{\fl{(k+1)/2}}\)$ $k$-tuples
of integers
\begin{equation}
\label{eq:k tuples}
\vec{t} = (t_1,\ldots, t_k),
\qquad 1 \le t_1 < \ldots <t_k,
\end{equation}
with $t_k \le T$ and such that the largest non-cyclotomic factor
(that is, a factor which does not have roots that are roots of
unity) of the $k$-lacunary polynomial~\eqref{eq:LacPoly} is
reducible over ${\mathbb K} = {\mathbb Q}\(a_1/a_0, \ldots, a_k/a_0\)$.
Here we consider a related question about estimating the number
$N_k(p,t)$ of $k$-tuples~\eqref{eq:k tuples} such that there is a
$k$-lacunary polynomial of the form~\eqref{eq:LacPoly} of degree $t_k = t$ over the finite field
${\mathbb K} ={\mathbb F}_p$ of $p$ elements, where $p$ is a prime,
that fully splits over ${\mathbb F}_p$.
\begin{theorem}
\label{thm:split p} If a positive integer $k$ is fixed then for
any prime $p$ and positive integer $t < p$, we have,
$$
N_k(p,t) \le t^{k - k\rf{(k-3)/2}-1} p^{(k-1)\rf{(k-3)/2}+o(1)}
$$
as $p\to \infty$.
\end{theorem}
Clearly, Theorem~\ref{thm:split p} is nontrivial only for $k > 3$
and for
\begin{equation}
\label{eq:large t}
t > p^{1-1/k +\varepsilon},
\end{equation}
with some fixed $\varepsilon>0$. Furthermore, for $t \gg p$ we
obtain the bound
$$
N_k(p,t) \le t^{\rf{k/2} +1 +o(1)}.
$$
Our result is based on a rather unusual combination of two techniques:
a bound on the number of zeros of lacunary polynomials
(see Section~\ref{sec:poly})
and a bound on the so-called domination number of a graph
(see Section~\ref{sec:dom}).
Throughout the paper,
the implied constants in the symbols `$O$', `$\ll$'
and `$\gg$' may depend on $k$ (we recall that
the notations $U \ll V$ and $V \gg U$ is
equivalent to $U = O(V)$).
\section{Zeros of Lacunary Polynomials}
\label{sec:poly}
We need the following estimate from~\cite{CFKLLS} on the number of zeros of lacunary
polynomials over ${\mathbb F}_p$.
\begin{lemma}
\label{lem:SprEq Zeros}
For
$k +1\ge 2$
elements $a_0, a_1, \ldots\,, a_k \in {\mathbb F}_p^*$ and
integers $0 = t_0 < t_1 < \ldots < t_k<p$,
the number of solutions $Q$ to the equation
$$
\sum_{i=0}^k a_ix^{t_i} = 0, \qquad x \in {\mathbb F}_p^*,
$$
with $t_0 = 0$, satisfies
$$
Q \le 2 p^{1 - 1/k} D^{1/k} + O(p^{1 - 2/k} D^{2/k}),
$$
where
$$
D = \min_{0 \le i \le k} \max_{j \ne i} \gcd(t_j - t_i, p-1).
$$
\end{lemma}
\begin{lemma}
\label{lem:SprEq Mult}
For
$k+1 \ge 2$
elements $a_0, a_1, \ldots\,, a_k \in {\mathbb F}_p^*$ and
integers $0 = t_0 < t_1 < \ldots < t_k<p$,
the multiplicity of any root $\rho$ of the polynomial
$$
\sum_{i=0}^k a_iX^{t_i} \in {\mathbb F}_p[X]
$$
is at most $k$.
\end{lemma}
\begin{proof} Let
$$
F(X) = \sum_{i=0}^k a_iX^{t_i}.
$$
Then for the $j$ derivative $F^{(j)}(X)$ we have
$$
F^{(j)}(X)X^j = \sum_{i=0}^k\prod_{h=0}^{j-1} (t_i -h) a_iX^{t_i}
$$
(where as usual, we set $F^{(0)}(X)= F(X)$).
Thus, if $r \ne 0$ is a root of multiplicity at least $k+1\le t_k
<p$ in the algebraic closure of ${\mathbb F}_p$, then
$$
F^{(j)}(r) = 0, \qquad j =0, \ldots, k.
$$
Therefore, the homogeneous system of equations
$$
\sum_{i=0}^k \prod_{h=0}^{j-1} (t_i -h) x_i = 0, \qquad j =0, \ldots, k,
$$
has a non-zero solution $x_i = a_ir^{t_i}$, $i =0, \ldots, k$.
This implies
$$
\det\left[ \( \prod_{h=0}^{j-1} (t_{i} -h) \)_{i,j=0, \ldots, k}\right] = 0,
$$
which is impossible for $0 = t_0 < t_1 < \ldots < t_k <p$ as an easy
calculation shows that
$$
\det\left[ \( \prod_{h=0}^{j-1} (t_{i} -h) \)_{i,j=0, \ldots, k}\right]
= \prod_{0 \le i <j\le k}
(t_j - t_i) \ne 0.
$$
The above contradiction implies the desired result.
\end{proof}
\section{Domination Number of a Graph}
\label{sec:dom}
Let $G=(V,E)$ be a simple undirected graph of order $n$.
A {\it dominating set\/}
$S$ of $G$ is a vertex subset such that any vertex of $V\setminus
S$ has a neighbour in $S$. Intuitively, a dominating set of a
graph is a vertex subset whose neighbours, along with themselves,
make up the vertex set of the graph.
The minimum cardinality of a dominating set of $G$ is called the
{\it domination number\/} $\gamma(G)$ of $G$. In other words,
$$\gamma(G)=\min_{S\subseteq V(G)}\left\{|S|~:~V(G)\subseteq
\bigcup_{v\in S} \hat{N}(v)\right\},$$
where $\hat{N}(v)$ denotes the closed neighbourhood of a vertex
$v$.
We denote by $\delta(G)$ the minimum degree of $G$.
When $\delta(G)$ is big enough, there are very good upper bounds
for the domination number of the graph $G$ in terms of $\delta(G)$
and $n$ (see, for example,~\cite{CSSF, HHS2}). However, for small values
of $\delta(G)$ the classical result of Ore~\cite{ORE} is stronger
and provides an upper bound for the domination number of a graph
with no isolated vertices:
\begin{lemma}
\label{lem:Dom Set}
If $G$ is a graph of order $n$ with
$\delta(G)\geq1$, then
$$\gamma(G)\leq \frac{n}{2}.
$$
\end{lemma}
\section{Proof of Theorem~\ref{thm:split p}}
Since $p > t_k$, by
Lemma~\ref{lem:SprEq Mult}
the
multiplicity of each non-zero root of a polynomial of the
form~\eqref{eq:LacPoly} does not exceed $k$.
Hence, if a polynomial $F(X) \in {\mathbb F}_p[X]$ of the form~\eqref{eq:LacPoly} splits
completely over ${\mathbb F}_p$ then the equation
$$
a_0 + a_{1}x^{t_1} + \ldots + a_{n}x^{t_k} = 0, \qquad x \in {\mathbb F}_p^*,
$$
with $1 \le t_1< \ldots < t_k$
has at least $t_k/k$ solutions. Then, from Lemma~\ref{lem:SprEq Zeros} we have
$$
t_k/k = O\(p^{1 -1/k} D_\vec{t}^{1/k}\),
$$
where
$$
D_\vec{t} = \min_{0 \le i \le n} \max_{j \ne i} \gcd(t_j - t_i, p-1).
$$
Thus $D_\vec{t}t \mid p-1$ and, since $k$ is fixed,
\begin{equation}
\label{eq:Dt}
t\ge D_\vec{t} \gg t_k^{k}p^{-(k-1)} = t^{k}p^{-(k-1)}.
\end{equation}
We now fix $D \mid p-1$, and for each $\vec{t} = (t_1,\ldots,
t_k)$ construct a graph $G_\vec{t}(D)$ on $k+1$ vertices $0,
\ldots, k$, connecting $i$ and $j$ if and only if $\gcd(t_i-t_j,
p-1) \ge D$ (where, as before $t_0=0$).
Clearly, if $D_\vec{t}=D$ and $G_\vec{t}(D)=G$ then $\delta(G)\ge
1$.
Now, for a fixed positive integer $D \le t < p$ and a graph $G$
with $k+1$ vertices and $\delta(G)\ge 1$, we estimate the number $M_p(D,G,t)$ of vectors
$\vec{t} = (t_1,\ldots, t_k) \in {\mathbb Z}^k$ with $1 \le t_1 < \ldots
<t_k$ and $t_k=t$ such that $G_\vec{t}(D)=G$.
Summing over all
graphs $G$ (since $k$ is fixed
there are only finitely many graphs) and admissible values of $D$,
that is, with $t\ge D \gg t^{k}p^{-(k-1)}$, see~\eqref{eq:Dt},
leads to the desired estimate.
Given a graph $G$ with $k+1$ vertices and $\delta(G)\ge 1$,
we now fix a dominating set $S$ in $G$ of cardinality $\# S =
\fl{(k+1)/2}$, which exists by Lemma~\ref{lem:Dom Set} (obviously,
we can always add more vertices to $S$ if necessary to guarantee
$\# S = \fl{(k+1)/2}$). So for each $j \not \in S$ with $j \ne 0,
k$, there is $i\in S$ such that $\gcd(t_i-t_j, q-1) \ge D$. So if
$t_i$ is fixed, then $t_j$ can take at most
\begin{equation}
\label{eq:Dom Choice}
\sum_{\substack{d\mid p-1\\ d \ge D}} \frac{t}{d} \ll \frac{t}{D}
\sum_{d\mid p-1}1 = \frac{t}{D} p^{o(1)}
\end{equation}
values, where we have used the known bound on the divisor
function, (see~\cite[Theorem~320]{HardyWright}). Finally, when
$t_k=t$ is fixed, each $t_i$, $i \in S$, can take at most $t$
values.
Furthermore,
if both $0,k \in S$ then there are only
$$\# S - 2 \le \fl{(k+1)/2}-2 = \fl{(k-3)/2}$$
elements $t_i$ with $i \in S\setminus \{0,k\}$ to be chosen.
After all values of $t_i$ with $i \in S$ are fixed, we see
from~\eqref{eq:Dom Choice} that the remaining
$$k+1 - \#S = \rf{(k+1)/2}$$
elements $t_j$, $j \not \in S$, can be chosen in at most
$(tp^{o(1)}/D)^{\rf{(k+1)/2}}$ ways.
So in this case
\begin{equation}
\label{eq:MDt20} M_p(D,G,t) \le t^{\fl{(k-3)/2}}
(t/D)^{{\rf{(k+1)/2}}}p^{o(1)} = t^{k-1}
D^{-\rf{(k+1)/2}}p^{o(1)}.
\end{equation}
If $0 \in S$ but $k\not \in S$, or $0 \not \in S$ but $k \in S$, then the same
argument implies:
\begin{equation}
\label{eq:MDt11} M_p(D,G,t) \le t^{\fl{(k-1)/2}}
(t/D)^{{\rf{(k-1)/2}}}p^{o(1)} = t^{k-1}
D^{-\rf{(k-1)/2}}p^{o(1)}.
\end{equation}
Finally, if both $0,k \not \in S$ then we get
\begin{equation}
\label{eq:MDt02} M_p(D,G,t) \le t^{\fl{(k+1)/2}}
(t/D)^{{\rf{(k-3)/2}}}p^{o(1)} = t^{k-1}
D^{-\rf{(k-3)/2}}p^{o(1)}.
\end{equation}
Clearly, bound~\eqref{eq:MDt02} dominates the
bounds~\eqref{eq:MDt20} and~\eqref{eq:MDt11}. In particular, for
$t \ge D \gg t^{k}p^{-(k-1)}$ we obtain
$$
M_p(D,G,t) \le t^{k-1 - k\rf{(k-3)/2} }
p^{(k-1)\rf{(k-3)/2}+o(1)}.
$$
Since, as we have mentioned, there are only finitely many possibilities for the graphs
$G_\vec{t}(D)$, recalling~\eqref{eq:Dt} and the bound on
the divisor function (see~\cite[Theorem~320]{HardyWright}),
we obtain the desired result.
\section{Comments}
A slight modification of our
approach can easily produce a nontrivial bound for $1\le k \le 3$
as well, however we do not know how to relax the condition~\eqref{eq:large t}.
It is certainly an interesting question to show that almost all
$k$-lacunary polynomials of a large degree are irreducible over
${\mathbb F}_p$. In fact, as a first step one can try to get a lower bound
on the degree over ${\mathbb F}_p$ of the splitting field of a ``random''
$k$-lacunary polynomial.
\section*{Acknowledgements}
The authors would like to thank the referee for the careful reading
of the manuscript and helpful suggestions.
During the preparation of this work the second author was supported in part by
the Australian Research Council Grant~DP1092835.
|
\section{Introduction}
\label{intro}
The interstellar medium (ISM) of our Milky Way is host to a variety of physical mechanisms that regulate and govern
the structure and evolution of the Galaxy. The current understanding of the ISM is that it is a multi-phase environment
composed of a tenuous plasma, consisting of gas and dust, which is both magnetized and highly turbulent (Ferriere 2001, McKee \& Ostriker 2007).
In particular, the awareness of turbulence
as a dominant physical process in the ISM has only happened in the last decade (Elmegreen \& Scalo 2004).
Turbulence plays a critical role in the areas of star formation, magnetic reconnection, magnetic field amplification,
nearly every transport process, cosmic ray acceleration, magnetic dynamo, and the physics in the intercluster medium, to name just a few
(see Lazarian \& Vishniac 1999, Vishniac \& Cho 2001, Elmegreen \& Scalo 2004, Lazarian 2006, Ballesteros-Paredes et al. 2007,
McKee \& Ostriker 2007 and references therein).
Additionally, turbulence has the unique ability to transfer energy over scales ranging from kiloparsecs down to the proton gyroradius.
This is critical for the ISM, as it explains how energy is distributed from large to small spacial scales in the Galaxy.
\begin{figure*}
\center
\includegraphics[scale=.5]{obs1.eps}
\caption{Top: Observational P (left) and $|\nabla \textbf{P}|$ (right) from the SGPS data used in this study.
Bottom: MHD simulation model number 1 in Table \ref{tab:models} with P (left) and $|\nabla \textbf{P}|$ (right).
The simulated map of P has had the mean subtracted from maps of Q and U, similar to the observations. }
\label{fig:RM}
\end{figure*}
Despite the now obvious importance of magnetized turbulence, the situation of understanding ISM physics is no less complicated.
In spite of the big recent advances in understanding of incompressible and compressible MHD turbulence
(see Goldreich \& Sridhar 1995, Cho \& Vishniac 2000, Maron \& Goldreich 2001, Cho, Lazarian \& Vishniac 2002, 2003,
Cho \& Lazarian 2003, Kowal \& Lazarian 2010, Beresnyak \& Lazarian 2010, Beresnyak 2011) the ISM presents a complex
environment with multiple energy injection sources, different phases and various instabilities acting at different scales. The properties of this turbulence
affect key astrophysical processes and obtaining properties of ISM turbulence from observations opens ways of gauging numerical simulations and testing theory.
In light of these complexities, the most fruitful way of studying astrophysical MHD turbulence and the processes it affects
is to use a \textit{synergetic} approach which combines the knowledge of theoretical predictions, numerical studies, and observational efforts.
Observationally there are several ways of studying MHD turbulence. Many of these techniques hinge on
density fluctuations in ionized or neutral media (see Spangler \& Gwinn 1990, Armstrong et al. 1995,
Padoan et al. 2003, Falgarone et al. 2005, Chepurnov \& Lazarian 2010) and are aimed at finding the density
power spectrum of turbulence. More recently, ways to find the magnetization and the Mach number
of turbulence have been explored (Kowal, Lazarian \& Beresnyak 2007, Burkhart et al. 2009, 2010, Esquivel \& Lazarian 2010, Tofflemire, Burkhart \& Lazarian 2011,
Esquivel \& Lazarian 2011). While column density data are
arguable the most common type, they do not contain the full 3D picture and are only passive tracers of the turbulence velocity field.
Various ways to study the interstellar velocity field have been explored. The
tested and theoretically motivated ways to study turbulent velocities in supersonic interstellar turbulence are based on the use of the Velocity Channel Analysis
(VCA) and Velocity Coordinate Spectrum techniques (Lazarian \& Pogosyan 2000, 20004, 2006, 2008). These techniques have been used both for atomic HI and molecular
data (see Stanimirovic \& Lazarian 2001, Padoan et al. 2006, 2009, Chepurnov et al. 2010) to find the spectra of the turbulent velocity fields.
Naturally, studies of turbulence and magnetic fields are of great importance and synergetic value for the ISM.
In this case, a number of techniques have been explored, e.g.
structure functions of the polarization vectors arising
from dust polarized emission (see Falceta-Gon\c{c}alves et al. 2009, Houde et al. 2011).The quantitative study of
synchrotron intensity fluctuations can be traced to works by
Getmansev (1958), while fluctuations of synchrotron polarization\footnote{For a theoretical description
of synchrotron fluctuations for arbitrary index of cosmic rays and realistic models of anisotropic turbulence see Lazarian \& Pogosyan (2011).}
were used, for instance, to evaluate the spectra of magnetic turbulence in Hydra cluster (En{\ss}lin \& Vogt 2006, En{\ss}lin et al. 2010)
More recently, several authors have discussed the prospects of the use of radio polarization maps to study
turbulence (see
Haverkorn \& Heitsch 2004, Fletcher \& Shukurov 2006, 2007, Gaensler et al. 2011).
Faraday
rotation maps of linearly polarized radio signals are especially promising as they provide
very sensitive probes of fluctuations in magnetic field and ionized gas density
(see Gray et al. 1998, Gaensler et al. 2001, and Landecker et al. 2010).
The Faraday rotation can be calculated as:
\begin{equation}
RM= K \int_l^0 n_e(l) \bf{B}(l) d{\bf l}
\end{equation}
\label{RM}
With units of rad m$^{-2}$, $K=0.81$ rad m$^{-2}$ pc$^{-1}$ cm$^{3}$ $\mu G^{-1}$ and $B$, $l$ and $n_e$ are the magnetic field strength in $\mu G $, the
distance along the LOS
in parsecs, and the electron density in $cm^{-3}$ along the LOS, respectively.
Although many objects seen in the polarization/Faraday maps can be
matched to objects seen in other wavelengths (such as supernova remnants), an extended diffuse polarization emission network that is rich in structure
is also present that can not be mirrored in other wavebands or in total intensity (Fletcher \& Shukurov 2007).
The intensity variations seen in maps of Q, U and P are the result of small-scale angular structure in the
Faraday rotation induced by foreground ionized gas and are thus an indirect representation of
turbulent fluctuations in free electron density and magnetic field throughout the ISM.
In this paper we use polarization gradients to study ISM turbulence.
The use of gradients to highlight small rapid fluctuations seen in polarization maps was first discussed
by Gaensler et al. (2011). When the spatial gradient is applied to maps of vector $\textbf{P}= (Q,U)$ a complex web of filamentary
structures is revealed. These filaments (see right column of Figure \ref{fig:RM} for an example)
were interpreted by Gaensler et al. (2011) as rapid fluctuations in $n_e$ and B along the LOS
due to turbulence. In this paper we will further explain the origins of the filamentary structures as they are related
to turbulence and develop quantitative methods that can be applied to this data in order to obtain the Mach numbers of turbulence.
Taking the gradient of rotation measure or linear polarization maps has its advantages and disadvantages.
The primary advantage is that the spatial gradient of $\textbf{P}$ satisfies the property that it has both translational and rotational invariance
in the (Q,U) plane. Quantities such as the polarization amplitude and polarization angle are not preserved under
arbitrary translations and rotations, which can result from one or more of a smooth
distribution of intervening polarized emission, a smooth uniform screen of foreground Faraday
rotation, or the effects of missing large-scale structure in an interferometric data-set. Thus the
magnitude of the gradient is the simplest quantity that is not significantly effected by missing large-scale structure
or excess foreground emission or Faraday rotation.
Taking the gradient allows one to clearly see jumps and discontinuities,
regardless of whether single-dish (total power) measurements are present in the data.
In particular, this will highlight
areas where a sharp change in $n_e$ or $B$ occurs, which is most likely due to turbulent fluctuations or shock fronts in the ISM.
However, one must also keep in mind the disadvantages of using gradients, namely the fact that the gradient may enhance noise.
\begin{figure}[tbh]
\centering
\includegraphics[scale=.5]{subsonic_grads1.eps}
\caption{Examples of LOS maps and their respective gradients relevant to this paper for subsonic turbulence (model 1).
The first column shows column density, LOS magnetic field, and the rotation measure from top to bottom.
The second column shows the gradients of column density, LOS magnetic field, and polarization vector.}
\label{fig:RM1}
\end{figure}
In this work we explore the physical causes of the filaments by taking the gradient of polarization maps of isothermal MHD turbulence.
We investigate the dependency of the sonic Mach number on the structures seen in $|\nabla \textbf{P}|$
and calculate measures of the Probability Distribution Functions (PDFs) as well as the
a measure of the topology called genus. Genus has been extensively used in cosmology studies (Gott et al. 1986)
and has been suggested for ISM studies in Lazarian (1999). Its use for
synthetic column density maps and observations is discussed in the literature
(Lazarian, Pogosyan \& Esquivel 2002, Kowal, Lazarian \& Beresnyak 2007, Kim \& Park 2007, Chepurnov et al. 2008).
The PDFs and genus have been studied on MHD turbulence in the past and
have both shown sensitivity to the sonic Mach number in density, column density and position-postion-velocity data (Padoan et al. 1999, Kowal, Lazarian \& Beresnyak 2007,
Chepurnov et al. 2008, Burkhart et al 2009, 2010, Tofflemire et al. 2011). We view these statistics as part of a set of tools which can be applied to
polarization data in order to determine the sonic Mach number of the turbulence in the ionized ISM. We stress that the use of the whole set provides a synergetic quantity: obtaining the same result with
different techniques substantially increases how trustworthy the result is.
The paper is organized as follows. In \S~\ref{data} we further describe the data sets used, in particular the
Southern Galactic Plane Survey and a set of ideal MHD simulations, and our calculation of the rotation measure, the linear polarization maps and their gradients.
In \S~\ref{origin} we discuss the origins of the observed filaments as they relate to the sonic Mach number.
We describe different statistical measures of the sonic Mach number in
\S~\ref{ms}; in particular the genus and PDFs.
In we discuss our results followed by conclusions in \S~\ref{concl}.
\section{Data and Method}
\label{data}
\subsection{Gradient Technique and relation of $|\nabla \textbf{P}|$ to $|\nabla RM|$}
An important observational quantity connected to interstellar density and magnetic field fluctuations is the Faraday effect. In the presence of magnetic fields
and free electrons, bifringence of circularly polarized orthogonal modes occurs, giving these modes two different propagation velocities.
In the case of pure polarized background emission propagation through a magnetoionized medium, the linearly polarized
radiation will emerge with its polarization position angle rotated by the amount given in Equation \ref{RM}.
Thus the relation between the observed position angle, emitted position angle, and the rotation measure is:
\begin{equation}
\Theta-\Theta_0 = RM \lambda^2
\end{equation}
which has units of radians. Here $\lambda$ is the wavelength in meters.
\begin{figure}[tbh]
\centering
\includegraphics[scale=.5]{supersonic_grads1.eps}
\caption{Examples of LOS maps and their respective gradients relevant to this paper for supersonic turbulence (model 6).
The first column shows column density, LOS magnetic field, and the rotation measure from top to bottom.
The second column shows the gradients of column density, LOS magnetic field, and polarization vector.}
\label{fig:RM2}
\end{figure}
Observational determination of Faraday rotation comes from measurements of the linear polarization vector $\textbf{P} \equiv$ (Q,U)
(which depends on Stokes U and Q as $|P|=\sqrt{Q^2+U^2}$) as a function of $\lambda^2$.
To avoid confusion between vector and scalar P, we use bold notation to denote the vector quantity of the linear polarization map.
We define the gradient of the polarization vector as:
\begin{equation}
|\nabla \textbf{P}|=\sqrt{\left(\frac{\partial Q}{\partial x}\right)^2+\left(\frac{\partial Q}{\partial y}\right)^2+\left(\frac{\partial U}{\partial x}\right)^2
+\left(\frac{\partial U}{\partial y}\right)^2}
\end{equation}
\label{eq:grad}
We note that in the case of vector $\textbf{P}$ we have:
\begin{equation}
\textbf{P} = |P_0| e^{2i(RM\lambda^2+\theta_0)}
\end{equation}
\label{eq:exp}
From this equation, one can derive a relationship between
$|\nabla \textbf{P}|$ and $|\nabla RM|$ for Faraday-thin polarized emission as:
\begin{equation}
|\nabla RM| = |\nabla \textbf{P}| / 2i \lambda ^2 |\textbf{P}|
\label{eq:1}
\end{equation}
Equation 5 only holds for data in which the entire signal is measured (i.e. single-dish data included) and for which the background is uniform.
When $|\textbf{P}|=1$ and $\lambda =1$ (which are the assumptions we use for the simulations), one finds a trivial relation between $|\nabla \textbf{P}|$ and $|\nabla RM|$ as $|\nabla RM|= |\nabla \textbf{P}|/2i$.
However, this relation can only be used to calculate $|\nabla \textbf{P}|$ or $|\nabla RM|$ in the simulations, since the assumption of $|\textbf{P}|=1$
is almost always to simplistic for the observations because the data are missing single-dish information and/or the background $|\nabla \textbf{P}|$ is not zero.
We calculate maps of RM, $\textbf{P}$ and their gradients from density and LOS magnetic field that is perpendicular and parallel to the mean magnetic field in the simulations.
We calculate the RM as per Equation \ref{RM} at every point and then take its spatial gradient, that is, we compute the gradient vector at every pixel of the image using
neighbor pixels. We can calculate $|\textbf{P}|$ by calculating the stokes
vectors as $Q=\cos(2\theta)$ $U=\sin(2\theta)$. These expressions come from applying Equation 2
with assumed values for $\lambda$ and $\theta_0$
We show a subsonic and supersonic case of density ($n$), $\nabla n$, LOS magnetic field (LOS B), $\nabla B$,
RM and $|\nabla P|$ in Figure \ref{fig:RM1} and Figure \ref{fig:RM2}, respectively. A comparison of the SGPS test data and a subsonic case is given in
Figure \ref{fig:RM}. Inspection of maps of $|\nabla \textbf{P}|$ reveals that filaments are created
in both cases and that there is some correlation between gradients of column density, magnetic field, and $|\nabla \textbf{P}|$. We will discuss these further
in Section \ref{origin}.
\subsection{Southern Galactic Plane Survey }
We use a subsection of radio continuum images of an 18-square-degree patch of the Galactic plane, observed with the Australia Telescope Compact Array
(ATCA, see McClure-Griffiths et al. 2001 and Gaensler et al. 2001 for more details).
We examine the 1.4 GHz frequency data averaged over adjoining frequency channels
with simultaneously recorded Stokes \textit{I}, Stokes \textit{Q}, and Stokes \textit{U}
as part of the SGPS test region (Gaensler et al. 2001).
This field consists of 190 mosaicked pointings of the
Australia Telescope Compact Array (ATCA) and covers the range $325.5 < l < 332.5, -0.5 < b < 3.5$.
Complicated extended structure is seen in linear polarization throughout the test region, almost all of which has no correlation with total intensity.
We select a $512\times512$ pixel subregion from this data to match the resolution of the simulations used in our study and display
it in Figure \ref{fig:RM} in the top row.
The SGPS region we select begins at coordinate l=332.3373, b= -0.3138 and is not overly contaminated by bad pixels and contains significant emission.
\subsection{Simulations}
We generate a database of 3D numerical simulations of isothermal compressible (MHD)
turbulence by using the MHD code of Cho \& Lazarian (2003) and varying the input
values for the sonic and Alfv\'enic Mach number.
The sonic Mach numbers are defined as ${\cal M}_s \equiv \langle |{\bf v}|/C_s \rangle$,
where is ${\bf v}$ is the local velocity, $C_s$ is the sound speed,
and the averaging is done over the whole box.
Similarly, the Alfv\'enic Mach number
is ${\cal M}_A\equiv \langle |{\bf v}|/v_A \rangle$, where
$v_A = |{\bf B}|/\sqrt{\rho}$ is the Alfv\'enic velocity,
${\bf B}$ is magnetic field and $\rho$ is density.
We briefly outline the major points of the numerical setup (for more details see Cho \& Lazarian (2003).
The code is a second-order-accurate hybrid essentially
nonoscillatory (ENO) scheme (Cho \& Lazarian 2003) which solves
the ideal MHD equations in a periodic box:
\begin{eqnarray}
\frac{\partial \rho}{\partial t} + \nabla \cdot (\rho {\bf v}) = 0, \\
\frac{\partial \rho {\bf v}}{\partial t} + \nabla \cdot \left[ \rho {\bf v} {\bf v} + \left( p + \frac{B^2}{8 \pi} \right) {\bf I} - \frac{1}{4 \pi}{\bf B}{\bf B} \right] = {\bf f}, \\
\frac{\partial {\bf B}}{\partial t} - \nabla \times ({\bf v} \times{\bf B}) = 0,
\end{eqnarray}
with zero-divergence condition $\nabla \cdot {\bf B} = 0$,
and an isothermal equation of state $p = C_s^2 \rho$, where
$p$ is the gas pressure.
On the right-hand side, the source term $\bf{f}$ is a random
large-scale solenoidal driving force\footnote{${\bf f}= \rho d{\bf v}/dt$}.
The magnetic field consists of the uniform background field and a
fluctuating field: ${\bf B}= {\bf B}_\mathrm{ext} + {\bf b}$. Initially ${\bf b}=0$.
We scale the simulations to physical units, adopting typical
parameters for warm ionized gas.
We assume a pixel size of 0.15 parsecs and density of 0.1 cm$^{-3}$.
The simulations are assumed to be fully ionized and we do not include the effects of
partial ionization.
To make the maps of $|\nabla \textbf{P}|$ we first calculate the LOS rotation measure at each pixel
then we take the take the gradient of this rotation measure map and convert it to $|\nabla \textbf{P}|$ via
Equation \ref{eq:1}. An equally valid way is to calculate the rotation measure at each pixel, then shine a polarized signal through the cube
with our assumed background values of Q = 1, U = 0. Then one can calculate the emergent values of Q and U by applying the simulated rotation measure map
and then calculate the resulting $|\nabla \textbf{P}|$.
Both methods will produce identical maps of $|\nabla \textbf{P}|$. Additional smoothing of the maps of Q and U using a Gaussian kernel can also be
performed to mimic the telescope beam.
\begin{table}
\begin{center}
\caption{Description of the simulations - MHD, 512$^3$
\label{tab:models}}
\begin{tabular}{cccccc}
\hline\hline
Model & $p_{gas}$ & $B_{\rm ext}$ & ${\cal M}_s$ & ${\cal M}_A$ &Description \\
\tableline
1 &2.00 &1.00 &0.5 &0.7 & subsonic \& sub-Alfv\'enic \\
2 &0.70 &1.00 &1.0 &0.7 & transsonic \& sub-Alfv\'enic \\
3 &0.10 &1.00 &2.0 &0.7 & transsonic \& sub-Alfv\'enic \\
4 &0.05 &1.00 &3.0&0.7 & supersonic \& sub-Alfv\'enic \\
5 &0.025 &1.00 &4.4 &0.7 & supersonic \& sub-Alfv\'enic \\
6 &0.0077 &1.00 &8.0 &0.7 & supersonic \& sub-Alfv\'enic \\
7 &0.0049 &1.00 &10 &0.7 & supersonic \& sub-Alfv\'enic \\
8 &2.00 &0.10 &0.5 &2.0 & subsonic \& super-Alfv\'enic \\
9 &0.70 &0.10 &1.0 &2.0 & transsonic \& super-Alfv\'enic \\
10 &0.10 &0.10 &2.0 &2.0 & transsonic \& super-Alfv\'enic \\
11 &0.05 &0.10 &3.0 &2.0 & supersonic \& super-Alfv\'enic \\
12 &0.025 &0.10 &4.4 &2.0 & supersonic \& super-Alfv\'enic \\
13 &0.0077 &0.10 &8.0 &2.0 & supersonic \& super-Alfv\'enic \\
14 &0.0049 &0.10 &10 &2.0 & supersonic \& super-Alfv\'enic \\
\hline\hline
\end{tabular}
\end{center}
\end{table}
\section{The Origin of Filamentary Structures in Polarization and Rotation Measure Gradients}
\label{origin}
A filament traced by $|\nabla \textbf{P}|$ or $|\nabla RM|$ will form as a result of a localized change in either density or magnetic field as a function of position on the sky
as per Equation \ref{RM}.
There are many physical processes that can result in sharp changes in these quantities in the ISM, including gravitational collapse and outflows.
However, high Reynolds fluids in the ISM are expected to be turbulent (see Cho, Lazarian \& Vishniac 2003, Elmegreen \& Scalo 2004, Lazarian et al. 2009) and a more ubiquitous process responsible for fluctuations in $n_e$ or B is due to MHD turbulence, precisely because it expected everywhere in the
ISM\footnote{Because of the large injection scale of the ISM, the Reynolds numbers can typically reach $10^{10}$}, although the
type of turbulent environment can vary (e.g. the compressibility, magnetization, equation of state, etc.).
In the ISM, fluctuations in density and magnetic field will occur as a result of MHD turbulence, which will be visible in polarimetric maps.
In the case of taking gradients of a turbulent field, one would expect to find filamentary structure
created by shock fronts, jumps and discontinuities.
Figure \ref{fig:jumps} shows a cartoon illustrating these three separate cases of a possible profile and its respective derivative.
The cases are:
\begin{enumerate}
\item A H\"{o}lder continuous profile\footnote{H\"{o}lder continuous functions satisfy $|u(x_1, t)-u(x_2,t)|<C|x_1-x_2|^h$. When the exponent h=1, this satisfies the Lipschitz condition.
In the case of turbulent fields $ h=\frac{1}{3}$} that is not differentiable at a given point (e.g. the absolute value function at the origin): Common for all types of MHD turbulence.
\item A jump profile: Weak shocks, strong fluctuations or edges (e.g. a cloud in the foreground which suddenly stops).
\item A spike profile (e.g. delta function): Strong shock regime.
\end{enumerate}
In respect to case one, it is known that the turbulent velocity field in a Kolmogorov-type inertial range both
in hydro and MHD are not differentiable, but only H\"older continuous (Bernard et al. 1998, Eyink 2009).
Another example is that of any fractal function that displays self-similarity but is not differentiable everywhere.
This profile will naturally create discontinuities when one takes its derivative. Therefore, case one can be found in both subsonic and supersonic type turbulence.
Case one type filaments can be seen in Figure \ref{fig:RM1} for N, B, and $\textbf{P}$ in the right column.
Case two creates a structure in the gradient by a shock jump or a large fluctuation in either $n_e$ or B. Here again, this type of enhancement in $|\nabla \textbf{P}|$ could be found in
supersonic and subsonic type turbulence and is due either to large random spatial increases or decreases due to turbulent fluctuations along the LOS or weak shocks.
We expect weak shock turbulence to show a lager amplitude in $|\nabla \textbf{P}|$ then the subsonic case.
Case three is unique to supersonic turbulence in that it represents a very sharp spike in $n_e$ and/or B across a shock front. The difference between this case
and what might be seen in case two is that here we are dealing with interactions of strong shock fronts, which are known to create delta function like distributions
in density (Kim \& Ryu 2005). In this case, the derivative of case three is has a distinctly different profile with respect to case one and two.
Case three shows a 'double jump' profile across the shock front, which can be seen in Figure \ref{fig:RM2}
in the top right panel for LOS density and the bottom right panel for $|\nabla \textbf{P}|$.
This morphological distinction can be used to determine if one is dealing with
turbulence that is in a shock dominated regime (i.e. supersonic) and can provide researchers with a promising new avenue of obtaining the sonic Mach numbers from
polarimetric data. In the case of the Alfv\'enic Mach number, the morphological difference is less clear, however gradients
will tend to align along the field lines in the case of strong field (sub-Alfv\'enic turbulence).
\begin{figure}[h]
\centering
\includegraphics[scale=.28]{jumps1.eps}
\caption{Cartoon example of three possible scenarios for enhancements in a generic image ``n" , where ``n" could be $\textbf{P}|$, $RM$ or $\rho$ (density or EM).
Case one (top row) shows an example of
a H\"{o}lder continuous function that is not differentiable at the origin (applicable to all turbulent fields). Case two (middle row) shows an example
of a jump resulting from strong turbulent fluctuations along the LOS or weak shocks.
case three (bottom row) shows a delta function profile resulting from interactions of strong shocks. In this case,
the derivative gives a 'double jump' profile which produces morphology that is distinctly different from the previous cases.}
\label{fig:jumps}
\end{figure}
Also of interest is the question of which quantity is providing the dominate contribution to the structures in $|\nabla \textbf{P}|$ or $|\nabla RM|$: $\nabla n_e$, $\nabla B_{LOS}$
or both equally? Especially in the case of compressible turbulence, the magnetic
energy is correlated with density, namely, denser
regions contain stronger magnetic fields, which is due to
the compressibility of the gas (Burkhart et al. 2009). This causes the magnetic field to follow
the flow of plasma if the magnetic tension is negligible. The compressed
regions are dense enough to distort the magnetic
field lines, enhance the magnetic field intensity, and effectively
trap the magnetic energy due to the frozen-in condition.
Thus, for the supersonic cases, the intensity of the structures seen in $\nabla \textbf{P}$ are more pronounced then in the subsonic case, which is observed when
comparing Figures \ref{fig:RM1} and \ref{fig:RM2}.
However, in the case of subsonic turbulence, there are no compressive motions. In this case, random fluctuations in density and magnetic field will create structures
in $|\nabla \textbf{P}|$ and $|\nabla RM|$.
Due to these effects, we might expect different trends in the correlation of supersonic and subsonic $|\nabla \textbf{P}|$ with
$\nabla N$ or $\nabla EM$ (the gradient of the emission measure) and $\nabla B$.
We test this by plotting the pixel by pixel correlation coefficient of $|\nabla \textbf{P}|$ with the gradients of EM, N, and LOS B in Figure \ref{fig:corav}.
In the case of subsonic turbulence, $|\nabla \textbf{P}|$ better traces out the fluctuations in $\nabla B$ (blue line), while the supersonic cases
are dominated by density fluctuations. This is because density enhancements are dominate due to shock fronts in the case of supersonic turbulence, while in
subsonic turbulence density is highly incompressible. In this case, the magnetic field will dominate the topology of the rotation measure and $|\nabla \textbf{P}|$.
This behavior is analogous to velocity in neutral hydrogen radio position-position-velocity cubes of turbulence, where density dominates the power spectrum for the case of supersonic
turbulence and velocity dominates the spectrum for subsonic turbulence (see Lazarian \& Pogosyan 2006, Burkhart et al. 2011a).
This difference in correlation provides yet another way of gauging the Mach numbers if one has both the emission measure
and the linear polarization map. Correlated spatial gradients between the two should indicate regions of shocks.
In the next section we will explore the utility of gradients of polarimetric data for the determination of the Mach numbers by investigating two different statistical
measures of looking at the distribution and topology
of the $|\nabla \textbf{P}|$ maps: PDF moments and genus function.
\section{Statistical Determination of the Sonic Mach Number}
\label{ms}
The previous section provided some theoretical discussion for why we expect $|\nabla \textbf{P}|$ data to be useful for determining the sonic Mach number.
In this section we will attempt to statistically quantify the differences seen in both the morphology and the distribution of maps of $|\nabla \textbf{P}|$.
We again note our assumption for the simulations of $|P|=1$ thus giving a trivial scaling relationship between $|\nabla \textbf{P}|$ and $|\nabla RM|$ as:
$|\nabla RM|=|\nabla \textbf{P}|/2 $.
We also provide an observational comparison for both statistics with the SGPS test region shown in Figure \ref{fig:RM}.
\begin{figure}[h]
\centering
\includegraphics[scale=.5]{correlationav_z.eps}
\caption{Correlation coefficient between$|\nabla \textbf{P}|$ and the LOS $\nabla EM$ (black diamonds), $\nabla N$(red asterisk) and $\nabla B$ (blue plus sign).
The left panel is super-Alfv\'enic and the right panel is sub-Alfv\'enic. In the case of strong shocks, strong correlation
is observed between $\nabla EM$ and $\nabla N$ due to enhanced density. Subsonic models show strong correlations
with $\nabla B$. }
\label{fig:corav}
\end{figure}
\subsection{Moments}
\label{pdfs}
A probability distribution function (PDF) is the function describing the frequency of occurrence of values in the distribution of intensities. PDFs and their quantitative descriptors
have been used to study turbulence in a variety of astrophysical context including diffuse ISM turbulence
(Berkhuijsen \& Fletcher 2011), turbulence characterization (Federrath et al. 2010, Esquivel et al. 2010, Audit \& Hennebelle 2010)
solar wind (Burlaga et al. 2007) and molecular ISM (Padoan et al. 1999). Several authors have discussed the use of PDFs in determining
the Mach numbers of ISM turbulence, in particular the sonic Mach number (Vazquez-Semadeni 1994, Padoan et al. 1999,
Kowal, Lazarian \& Beresnyak 2007, Burkhart et al. 2010, Price, Federrath, \& Brunt 2011). However, this technique is almost always used in the context of the column density.
To our knowledge, no one has applied this technique to the rotation measure, polarization maps, or their gradients.
One method of describing PDFs is by using statistical moments to characterize the mean and variance and departures from Gaussianity.
The first and second order statistical moments (mean and variance) used here are defined as follows:
$\mu_{\xi}=\frac{1}{N}\sum_{i=1}^N {\left( \xi_{i}\right)} $
and $ \nu_{\xi}= \frac{1}{N-1} \sum_{i=1}^N {\left( \xi_{i} - \overline{\xi}\right)}^2$, respectively.
The 3rd and 4th order moments, Skewness and kurtosis respectively, are defined as:
\begin{equation}
\gamma_{\xi} = \frac{1}{N} \sum_{i=1}^N{ \left( \frac{\xi_{i} - \overline{\xi}}{\sigma_{\xi}} \right)}^3
\label{eq:skew}
\end{equation}
\begin{equation}
\beta_{\xi}=\frac{1}{N}\sum_{i=1}^N \left(\frac{\xi_{i}-\overline{\xi}}{\sigma_{\xi}}\right)^{4}-3
\label{eq:kurt}
\end{equation}
Where N is the total number of elements and $\xi_{i}$ is the distribution of intensities.
Past works have focused on the relationship between the moments and the density or column density. As the Mach number increases, so does the mean value of density,
as shocks increase density (the sonic Mach number goes as ${\cal M}_s \approx \rho ^{1/2}$). The variance, skewness, and kurtosis
are less obvious quantities and their relation to the sonic Mach number has been derived numerically (Padoan et al. 1999, Kowal, Lazarian \& Beresnyak 2007, Burkhart et al. 2010
Price, Federrath, \& Brunt 2011). Variance has some disadvantages to skewness and kurtosis. One is that
variance is scale dependent, and values will change between different data set normalizations. This makes direct comparison between simulations and
observations difficult. On the other hand, the higher order moments (skewness and kurtosis) describe deviations from Gaussianity and are unit-less numbers, and therefore
are scale free. They also are shown to increase more linearly with the sonic Mach number in the case of column density (Kowal, Lazarian \& Beresnyak 2007, Burkhart et al. 2010).
However, the increase is not so pronounced in the case of subsonic turbulence, making variance a better indicator in this regime.
\begin{figure}[h]
\centering
\includegraphics[scale=.6]{histgradp.eps}
\caption{PDFs of $|\nabla \textbf{P}|$ normalized as $f(x)=(|\nabla \textbf{P}|-<|\nabla \textbf{P}|>)/\sigma(|\nabla \textbf{P}|)$ where $\sigma(|\nabla \textbf{P}|)$
is the standard deviation of $|\nabla \textbf{P}|$.
The supersonic case (represented by + signs) is highly skewed and kurtotic compared with the subsonic case (represented by * signs).}
\label{fig:PDF}
\end{figure}
While the rotation measure is obviously related to the density, the relationship between the moments of $|\nabla \textbf{P}|$ and the Mach number has never been studied.
The polarization gradient may have an advantage over column density/dispersion measure for investigating the Mach numbers of the WIM, as shocks will cause both an increase in
density and magnetic field due to correlations between density and field strength in supersonic type turbulence (see Burkhart et al. 2009). Furthermore, the gradient will highlight interesting
features in the observational data which might be buried by observational effects, such as a DC offset, that otherwise would not be seen in the maps of
column density or linear polarization (see Figures \ref{fig:RM1} and \ref{fig:RM2}).
Figure \ref{fig:PDF} shows an example PDF of a subsonic and supersonic simulation (model 1 and 6 ). Here it is clear that the general relations that are true
for the column density will apply here: as the Mach number increases the distribution becomes more kurtotic and more skewed.
Figure~\ref{fig:moments} shows the moments of $|\nabla \textbf{P}|$ vs. the sonic Mach number for our simulations with LOS taken along the mean magnetic field.
Error bars are created by taking the
standard deviation between different time snapshots of the data. We see that all four moments tend to increase with increasing sonic Mach number.
This is due to the shock fronts creating more discontinuities and sharper gradients. This not only increases the mean value and the variations in
$|\nabla \textbf{P}|$, but also creates very peaked distributions and distributions with tails skewed towards the right.
In terms of the Alfv\'enic Mach number, the main contributor is the LOS direction chosen relative to the ordered field. When we look parallel to the mean field
as shown in in Figure~\ref{fig:moments}, the sub-Alfv\'enic cases show higher mean and variance while the super-Alfv\'enic cases show generally higher skewness and kurtosis.
We plot the mean and variance of the \textit{polarization angle} in Figure \ref{fig:moments_rm}, which confirms the trend seen in the gradients. In the case of
the LOS parallel to the mean field, the mean value of the polarization angle has a strong Alfv\'enic dependency with no sonic Mach number dependency.
The variance shows strong dependency on both sonic and Alfv\'enic Mach numbers. However, these trends are unique for sight-lines parallel to the mean field LOS.
When looking perpendicular to the LOS the effects change. For example, the moments of $|\nabla \textbf{P}|$ show stronger skewness and kurtosis in the case
of the sub-Alfv\'enic models. Thus to gauge the Alfv\'enic Mach number using the distribution, the ordered LOS magnetic field is required.
An additional effect that must be considered is the issue of the telescope resolution. For example, Figure \ref{fig:smoothiimage} shows model number 6
without smoothing (left) and with Gaussian smoothing with FWHM=6 pixels (right). It is clear that smoothing changes the distribution of
maps of $|\nabla \textbf{P}|$.
We plot the moments vs. smoothing FWHM for four models in Figure \ref{fig:momentsm}. As the resolution of maps of Q and U decrease, so do the moments.
Thus one needs to take the smoothing of the data into account when comparing PDFs of $|\nabla \textbf{P}|$.
We investigate the moments of regions of the SGPS data with the most emission removing the bad pixels from our investigation. We find
that the SGPS data has average skewness of 0.825 and kurtosis of 0.928 and mean value and variance of 0.004 and $3.38x10^{-6}$, respectively.
Comparing with our numerical set up, this most closely matches subsonic to transonic values of the moments, with telescope smoothing taken into account.
\subsubsection{Moment Maps}
\label{mommap}
From section 4.1 it is clear that, for a given map of $|\nabla \textbf{P}|$, as the sonic Mach number increases the skewness and kurtosis also increase.
However, these were globally averaged values of both the moments (i.e. the PDF of the entire image) and the sonic Mach number.
If we want to look at smaller scale variations, we can calculate the PDFs of smaller portions of the image.
Using a moving kernel (box, circle etc.), we can create a ``moment map" which is essentially a smoothed map
that calculates the moments at every point with a given box size. Because we are dealing with gradient quantities, we expect
these maps to particularly trace regions where shocks are interacting and along shock fronts, thus tracing areas
where the sonic Mach number is \textit{changing}.
We make skewness and kurtosis moment maps of the $|\nabla \textbf{P}|$ images similar to the method of Burkhart et al. (2010).
This results in a smoothed moment image, which can be compared to the actual image of the LOS local sonic Mach number (LOS average
sonic Mach number for each pixel) with the same smoothing kernel as the $|\nabla \textbf{P}|$ moment map.
In this section, we investigate the relationship between the LOS sonic Mach number and the values of skewness and kurtosis on a pixel-by-pixel basis.
The key questions one must consider when making a moment map are what the kernel shape should be,
what resolution is appropriate for good statistics, and how to compare the values of the higher order moments to
the sonic Mach number? The last point is particularly important. While Burkhart et al. (2010) took a linear fit between the sonic
Mach number and the moments in the case of column density, we will
test how effective the moment maps can distinguish between supersonic and subsonic regimes on a pixel-by-pixel bases.
Thus we stress that we are not trying to calculate the exact sonic Mach number with this method, which is a topic for another work. We only
seek to determine how well a moment map is able to pick out regimes of supersonic and subsonic.
For our initial test, we choose to use a boxcar kernel for both the $|\nabla \textbf{P}|$ maps and for averaging the sonic Mach number,
since a boxcar can handle edges in a square image most effectively.
\begin{figure}[htb]
\centering
\includegraphics[scale=.6]{Prmmoments_syncx.eps}
\caption{Moments of $|\nabla \textbf{P}|$ vs. ${\cal M}_s$. Error bars are created by taking the
standard deviation of the moments between different time snapshots of the well-developed turbulence. We show sub-Alfv\'enic cases in red and super-Alfv\'enic cases in black. }
\label{fig:moments}
\end{figure}
\begin{figure}[htb]
\centering
\includegraphics[scale=.45]{Prmmoments_rmx.eps}
\caption{Mean (top) and variance (bottom) of the polarization angle (or RM$\lambda^2$) \textit{along the direction of the mean magnetic field} vs. sonic Mach number. A strong
Alfv\'enic dependency is observed. In the case of the mean value, the dependency is independent of the sonic Mach number. The trend is different for a different LOS. }
\label{fig:moments_rm}
\end{figure}
\begin{figure*}[htb]
\centering
\includegraphics[scale=.6]{smooth_gradp.eps}
\caption{Maps of $|\nabla \textbf{P}|$ for model 6 with smoothing FWHM=6 pixels (right) and without smoothing (left). The telescope resolution has a strong effect on the distribution of $|\nabla \textbf{P}|$
yet the morphology remains similar.}
\label{fig:smoothiimage}
\end{figure*}
\begin{figure}[htb]
\centering
\includegraphics[scale=.6]{Prmmoments_smoox.eps}
\caption{Moments of $|\nabla \textbf{P}|$ vs. smoothing for four different models. Error bars are created by taking the
standard deviation between different time snapshots of the data. Smoothing FWHM is given in pixels.}
\label{fig:momentsm}
\end{figure}
How well the moment map is able to determine if the gas is supersonic or subsonic depends on three parameters:
the threshold level for skewness ($\gamma_T$, above which the gas is considered supersonic and below it is considered subsonic),
the threshold level for kurtosis ($\beta_T$, above which the gas is considered supersonic and below it is considered subsonic), and the kernel box size chosen.
The values for $\gamma_T$ and $\beta_T$ represent the threshold value between supersonic and subsonic regimes. Any pixels below either $\gamma_T$ and $\beta_T$ is deemed subsonic and
any pixel above both is deemed supersonic. Of course, an intermediate threshold value could also be chosen to probe the transsonic regime, but we omit this here.
Again we should stress that we are not employing this method to get an exact Mach number;
we simply view it as a way of determining whether our $|\nabla \textbf{P}|$ image shows supersonic or subsonic characteristics.
The best fitting parameter values of box size, $\gamma_T$ and $\beta_T$ are those that provide accurate translation between the moments and the LOS Mach number. In order to determine these values
we use a genetic algorithm which searches possible combinations of these three parameters in order to determine the best fit. The genetic algorithm can provide more computationally efficient way of
testing for fitness rather then iterating over every possible combination of parameter space. See the appendix for a detailed description of our algorithm.
Although multiple combinations of parameters showed high confidence levels, we chose $\gamma_T$=1.1 and $\beta_T$=1.58 and box size=64 pixels.
With these values, the supersonic models were able to determine the Mach number regime with 98\% accuracy, while the subsonic cases had an accuracy of 67\%
(see Figure \ref{fig:corprct}). This is not surprising however, since
the moments are known to be a more robust measure of highly supersonic turbulence (Kowal, Lazarian \& Beresnyak 2007).
We plot the LOS Mach number map for model number 1 in the top panel of Figure \ref{fig:mommap_obs} with contours from the kurtosis moment map.
The moments trace areas where the sonic Mach number is changing.
\begin{figure}[htb]
\centering
\includegraphics[scale=.6]{correlation_box.eps}
\caption{The success rate of the moment map. A moment map is constructed with parameters $\gamma_T$=1.1 and $\beta_T$=1.58 and box size (64 pixel), which were
chosen with a genetic algorithm.
If a value in the moment map is above both $\gamma_T$ and $\beta_T$, then that pixel is considered supersonic. If the value is below either, then
the pixel is considered subsonic.
The moment map is compared with the actual LOS sonic Mach number map to determine whether the moment map alone is sufficient for recovering regimes of
supersonic or subsonic on a pixel-by-pixel basis. The moment map is very successful (almost 100\%) in determining if supersonic turbulence is present.
The success rate drops to $\approx$60\% in the case of diagnosing subsonic turbulence. This further confirms the higher order moments utility in the presence of supersonic
flows.}
\label{fig:corprct}
\end{figure}
We apply the moving box method to the SGPS data cuts using the same parameters for $\gamma_T$, $\beta_T$ and box size as were used for the simulations.
For the SGPS data we obtained average skewness and kurtosis values for this moment map of 0.3 and 0.9 for skewness and kurtosis respectively. We over-plot the SGPS data
with the kurtosis moment map contours in the bottom panel of Figure \ref{fig:mommap_obs}.
While these values seem to point in the direction that the SGPS data is subsonic or transonic, we note that this method is less accurate for subsonic
type turbulence. However, it does confirm that the gas in this patch of the sky is not statistically similar to what is expected for supersonic flows.
\begin{figure}[htb]
\centering
\includegraphics[scale=.6]{mommap_imagsubson1.eps}
\includegraphics[scale=.6]{mommap_imagobs1.eps}
\caption{Top: LOS sonic Mach number map for model number 1 with over-plotted contours of the kurtosis moment map with $\gamma_T$=1.1 and $\beta_T$=1.58 and smoothing
kernel= 64 pixels. The moment map generally outlines areas where the sonic Mach number is changing, which is expected for a gradient quantity.
Bottom: Map of the SGPS test region used in the paper with the kurtosis moment map contours over-plotted.}
\label{fig:mommap_obs}
\end{figure}
\subsection{Topology: The Genus Statistic}
\label{genus}
The filaments seen in $|\nabla \textbf{P}|$ show substantially different morphology when comparing maps of subsonic and supersonic turbulence, as was discussed in
Section 3. Thus, a natural
avenue of characterization would be to use topological measures in order to pick out different structures. In this section we will investigate the utility of the
genus statistics in order to characterize the topology of $|\nabla \textbf{P}|$ filaments.
The genus statistics was developed to study the topology and deviations from Gaussianity of the
universe and the distribution of galaxies in three dimensions (Gott
et al. 1986; 1987). The use of genus statistics for the study of HI was first
discussed in Lazarian (1999), and subsequent studies presented the genus curves for the SMC (Lazarian et al. 2002;
Lazarian 2004, Chepurnov et al. 2008) and for MHD simulations (Kowal, Lazarian \& Beresnyak 2007).
Genus is a quantitative measure of topology. It can characterize
both 2D and 3D distributions.
Generally speaking, the genus is used to detect departures for Gaussianity.
When dealing with the ISM one cannot expect deviations from symmetry to be
small, especially in the presence of supersonic flows. In this case, genus can be used to
characterize flows that are supersonic since these show large deviations from Gaussianity.
The 2D genus can be represented
as (Coles 1988; Melott et al. 1989):
\begin{align*}
G \equiv \text{(\# isolated high-density regions)} & \\ - \text{ (\# isolated low-density regions)}
\end{align*}
where low- and high-density regions are selected with respect to
a given contour threshold.
For instance, a uniform circle would have a genus of 0 (one
connected region of high density, i.e., an ``island,'' and one
connected region of low density, while a ring (a donut, for example)
would have a genus of -1 (one
connected region of high density and two connected regions
of low density). Thus the genus can distinguish between
``meatball" and ``swiss cheese" topologies (Gott 1990).
\begin{figure}[tbh]
\centering
\includegraphics[scale=.5]{out_genus.eps}
\caption{Genus curve of $|\nabla \textbf{P}|$ for model number 1 (pictured in the bottom right of Figure \ref{fig:RM}).
The genus curve of a Gaussian distribution is shown in dashed lines.
The $|\nabla \textbf{P}|$ genus deviates strongly from the Gaussian model. Horizontal and vertical solid lines reference the origin.}
\label{fig:genus}
\end{figure}
For a 2D image, the genus is simply a number which corresponds to
a given threshold value. What is considered to be high or low is dependent on the
threshold value, which acts as a free parameter. As a result, for a given 2D image,
the threshold value can be varied to construct a curve with the genus value on the y-axis and the threshold value on the x-axis.
This is known as the genus curve.
We show an example genus curve for $|\nabla \textbf{P}|$ in Figure \ref{fig:genus}. This genus curve is for the subsonic sub-Aflv\'enic (model number 1) simulation.
In the case of a density or column density field, the genus of subsonic turbulence is close to a Gaussian field. However
the $|\nabla \textbf{P}|$ distribution has longer tails and a more pronounced minimum.
The maximum and minimum points of the genus curve correspond to
percolation of the distribution (see Colombi et al. 2000). The fact
that the observed genus falls more slowly at large thresholds than
the Gaussian distribution indicates that the
islands are more discrete and pronounced than for the Gaussian
distribution.
\begin{figure}[tbh]
\centering
\includegraphics[scale=.5]{RMsuper2.eps}
\caption{Genus zero values of simulated $|\nabla \textbf{P}|$ . Negative values imply the topology is clumpy (meatball) while positive values imply
the topology is hole dominated (swisscheese). The genus of $|\nabla \textbf{P}|$ is fairly insensitive to smoothing of maps of Q and U.
The supersonic cases show a hole topology while the subsonic case shows a clump topology, even in the case of smoothing. Transonic cases show
topology that is a mixture of clumpy and hole dominated.}
\label{fig:gradcd}
\end{figure}
We expect that the sign of the genus curve at the mean intensity level of $|\nabla \textbf{P}|$ (i.e. where the curve crosses the x-axis)
does describe the field topology.
A positive genus
will represent a clump-dominated field, while a negative one will
mean the domination of holes. However, it is more convenient to
work with the zero of the genus curve (genus zero) $\nu_0$, because it can be normalized to the field variance.
In other words, consider a map in which the mean value of the map is subtracted off. $\nu_0$ represents
the location where the genus curve crosses the x-axis, with the origin being the mean value of the field, so that $G(\nu_0)=0$.
In this case, for the intensity with the subtracted mean value, a negative $\nu_0$ corresponds
to the clumpy topology, while a positive $\nu_0$ indicates the ``swiss cheese" topology.
\begin{figure}[tbh]
\centering
\includegraphics[scale=.5]{sgps_genus.eps}
\caption{Genus curve for $|\nabla \textbf{P}|$ of the SGPS data set. The genus zero is at -0.07 which indicates a slightly clumpy topology. This is most similar
to the transonic type genus curves (${\cal M}_s \approx 2.0$). Horizontal and vertical solid lines reference the origin.}
\label{fig:sgps_genus}
\end{figure}
We make plots of the genus zero vs. smoothing in Figure \ref{fig:gradcd}. A negative shift indicates a clumpy topology, while a positive shift indicates a hole topology.
The error bars are derived by estimating the variance of the genus distribution. For the error bar creation, we follow the method used by Chepurnov et al. (2008)
in that we generated a set of images
with randomly shifted phases of individual harmonics. This procedure causes the field to take Gaussian statistics, and therefore
approximately the genus zero is at the origin. However, slight deviations from Gaussian allows us to effectively estimate its variance.
The procedure is as follows: we take a fast Fourier transform (FFT) of
the region being studied and assign the phase of each harmonic
to a random variable uniformly distributed. After the inverse
FFT, we calculate the respective $\nu_0$.
After repeating this procedure ten times, we calculate the variance of the $\nu_0$ values.
Figure \ref{fig:gradcd} plots the genus zero vs. smoothing because the telescope resolution is an important
quantity to consider in any statistical analysis of observational data.
As was demonstrated with the PDF moments in the previous subsection, as maps of Q and U are smoothed, the distribution tends towards Gaussian.
As such we include analysis of the genus zero with smoothing to see what effect the telescope resolution has on the topology.
We include the genus zero vs. smoothing plots for maps of $|\nabla \textbf{P}|$ in the supersonic, transsonic and subsonic turbulence regimes.
We see that over a range of smoothing values
the topology between the three regimes is very different. The supersonic case shows a positive genus zero indicating that the topology is more hole like.
The topology of the subsonic case is more clump like. Visual inspection of Figure \ref{fig:RM1} confirms this.
In the case of the transonic turbulence, the topology is more neutral. Interestingly, there is not a strong dependency when we
smooth the maps of Q and U. The topology remains roughly consistent over a range of smoothing.
We show the genus curve for the SGPS $|\nabla \textbf{P}|$ data in Figure \ref{fig:sgps_genus}. The genus zero is at -0.07, which indicates a slightly clumpy topology.
This value more closely matches the transsonic values accounting also for smoothing effects.
\section{Discussion and Conclusions}
\label{concl}
Quantitative analysis (via PDF moments and genus) of the polarization gradient indicates that turbulence in the
ionized ISM in these directions is in the range of subsonic to transonic. Our findings are supported by recent studies
of Balmer-$\alpha$ emission measures, which have similarly found
Mach number of 2 or less (see Hill et al. 2008). For studies of neutral gas, i.e. 21 cm HI gas, the warm phase of the SMC was shown to have
properties of transonic gas (Burkhart et al. 2010). In the case of the cold component, several independent statistical
and direct measurements applied to the spectral lines and the column density indicate this component of the SMC is supersonic (Stanimirovic \& Lazarian 2001, Burkhart et al. 2010).
Furthermore, a recent study by Chepurnov \& Lazarian (2010) added the Wisconsin H-$\alpha$ Mapper (WHAM) emission measure data to the
``Big Power Law" of electron density fluctuations (see Armstrong et al. 1995)
provided further proof for a parsec to AU -5/3 power law, indicating that this phase of the ISM is not highly compressible.
Thus it becomes apparent that different types of data, sampling different phases of the ISM,
can compliment each other and provide more clues towards furthering our understanding of a turbulent ISM.
Observational studies of ISM turbulence are extremely important. From the observational data one would like to obtain the characteristics of turbulence, e.g.
intensity of its driving, importance of magnetic field in the dynamics of the media. These characteristics can be evaluated if we know sonic and Alfv\'en Mach numbers, i.e.
$ {\cal M}_s$ and ${\cal M}_A$. There have been several attempts to develop the techniques to get these numbers from observations using both column density data (see Kowal, Lazarian \& Beresnyak 2007 Burkhart et al. 2009)
as well as spectroscopic data (Burkhart et al. 2011a). An example of successful application of such techniques to the Small Magellanic Cloud (SMC) HI data is provided in Burkhart et al. (2010).
This paper explores the utility of a new observational motivated technique, namely the gradient of the polarization map, to determine ${\cal M}_s$ and ${\cal M}_A$.
The observational advantage of using gradients stems from the fact that these gradients are easily available from interferometric observations. Our study shows that $|\nabla \textbf{P}|$ is a very useful
measure which allows one to study turbulence. We might also expect gradients be useful for turbulence studies
in other types of data sets where shock morphology will be observed (i.e. column density, see Figures \ref{fig:RM1} and \ref{fig:RM2}).
We view this paper as a first taste of the utility of using polarization data (and their gradients) for studies of turbulence. There are many additional avenues
that should be explored from this first step. For instance, what are the effects of multiple screens along the LOS? What are the effects of changing the assumption
of the constant background polarization? What are the effects of other equations of state and the inclusion of partial ionization?
One may wonder whether an additional way of studying turbulence is valuable, if we already have a few other ways to study turbulence, e.g. with column densities and position-position-velocity
(PPV) spectroscopic data cubes. The answer is a sounding yes!. First of all, dealing with as complex media as the ISM, we would like to have as many independent measures as possible.
Second, different measures may be more sensitive to different phases of the ISM (see the list of the idealized phases and their magnetizations in Yan, Lazarian \& Draine 2004). For example, the
rotation measure, linear polarization and their gradients are biased towards ionized parts of ISM.
The Alfv\'en and sonic Mach number are not the only characteristics of the ISM turbulence.
Spectra of turbulence, its intermittencies (see Kowal \& Lazarian 2010) provide other measures which should
be used to study interstellar turbulence in its complexity.
\subsection{Conclusions}
We created maps of the spatial gradient of the polarization vector of isothermal MHD simulations and compared these with observations.
We tested two statistical methods on gradient polarization maps, namely the genus and higher order moments of the distribution to determine if these
statistics were sensitive to the sonic Mach number. We found that:
\begin{itemize}
\item Filamentary structure was created over a range of sonic Mach numbers, including cases where both shocks and subsonic turbulence were present.
\item Filaments showed different morphology for different regimes of sonic Mach number.
\item $|\nabla \textbf{P}|$ maps with high sonic Mach number showed filaments with a double jump profile
that traced shocks, while subsonic cases showed filaments that were due to random fluctuations in the rotation measure along the LOS. Transonic simulations had both
types of filaments present
\item The moments of the $|\nabla \textbf{P}|$ distribution were higher for larger values of sonic Mach number but were also sensitive to the telescope resolution.
\item There is a strong Alfv\'enic dependency in the moments of polarization angle and $|\nabla \textbf{P}|$ for sight-lines parallel to the ordered magnetic field.
\item The skewness and kurtosis moment maps of $|\nabla \textbf{P}|$ were successful at picking out subsonic or supersonic pixels 67\% and 99\% of the time, respectively.
\item The genus of $|\nabla \textbf{P}|$ revealed a ``hole" topology for supersonic cases and a ``clump" topology for subsonic cases. Transsonic cases showed neutral
topology. This trend is generally smoothing independent.
\item We applied the PDF moments and genus to the SGPS test region and found that this area was statistically similar to models of subsonic to transsonic type turbulence.
\end{itemize}
B.B. acknowledges support from the NSF Graduate Research Fellowship and the NASA Wisconsin Space Grant
Institution. B.B. thanks Andrew Schechtman-Rook for valuable discussions.
A.L. thanks both NSF AST 0808118 and the Center for Magnetic Self-Organization in Astrophysical and Laboratory Plasmas for financial support.
This work was completed during the stay of A.L. as
Alexander-von-Humboldt-Preistr\"ager at the Ruhr-University Bochum and the University of Cologne.
B.M.G. acknowledge the support of an Australian Laureate Fellowship awarded by the Australian
Research Council through grant FL100100114. The Australia Telescope Compact Array is funded
by the Commonwealth of Australia for operation as a National Facility managed by CSIRO.
|
\section{Introduction}
Nuclear matter is an idealized infinite system of interacting protons and neutrons. It is the low density and low temperature phase of QCD. In this range of densities and temperatures QCD theory is non-perturbative and the active degrees of freedom are no longer the quarks and the gluons, but color singlet states, the hadrons. For the description of nuclear matter a variety of approaches has been developed during the years, being an important issue for its applications to heavy-ion collisions \cite{natowitz} and astrophysics \cite{demorest}. A powerful tool for this task is represented by Chiral Perturbation Theory, an effective field theory that incorporates spontaneous and explicit chiral symmetry breaking.
The basic idea of our approach relies on the separation between the long-range correlations and the short-distance physics \cite{kaiser1,kaiser2,kaiser3}. While the long- and intermediate-range dynamics in the nuclear medium, described by $ 1\pi$- and $ 2\pi$-exchange with the inclusion of the $ \Delta$-isobar excitation as an explicit degree of freedom, is treated explicitly, the unresolved short-distance physics is encoded in contact terms fixed to reproduce selected known bulk properties of nuclear matter. For example, the contact term associated to the equation of state of isospin-symmetric nuclear matter is fixed imposing that the energy minimum at $ T = 0 $ is $ -16 $ MeV. The correct saturation density, $ \rho_0 \simeq 0.157\,\text{fm}^{-3} $, follows as prediction. For generalization to isospin-asymmetric nuclear matter one has to introduce two more contact terms that are fixed imposing that the asymmetry energy at the saturation point is $ A(\rho_0) \simeq 34 $ MeV \cite{kaiser3}. Many other ground state and single-particle properties and nuclear thermodynamics emerge as predictions.
Explicit $ 1\pi$- and $ 2\pi$-exchange dynamics are so far treated up to three-loop order in the free energy density. Consider symmetric nuclear matter first. The basic ingredient to perform calculations is the in-medium nucleon propagator. At $ T = 0 $, the nucleon propagator in momentum space has the following representation \cite{kaiser1}:
\begin{equation}\label{propagator}
S_N(p) = \left(\, \mkern-6mu \not \mkern-4mu p + M_N \right) \left[ \frac{i}{p^2- M_N^2 + i\epsilon} - 2 \pi \delta (p^2-M^2) \theta (k_F - |\mathbf{p}|) \theta(p_0) \right] \ ,
\end{equation}
where $ k_F $ is the Fermi momentum of nucleons and $ M_N = 939 $ MeV is the nucleon mass. The first term is the known free nucleon propagator, the second term is called medium insertion and takes into account the presence of a filled Fermi sea of nucleons by means of the theta step function. The calculation is then organized according to the number of medium insertions in the diagrams. The relevant many body dynamics come from diagrams with at least two medium insertions. Two-body forces arise from diagrams with two medium insertions; diagrams with three medium insertions generate the important three body forces. Convergence in the series of medium insertions is realized for $ k_F \ll \Lambda_\chi \sim 1 $ GeV as long as four-nucleon forces are not relevant.
\begin{figure}[tbp]
\center
\subfloat[][Isospin-symmetric nuclear matter.]
{\includegraphics[width=0.495\textwidth]{eos_50}} \
\subfloat[][Neutron-rich matter with $ x_p = 0.1 $.]
{\includegraphics[width=0.495\textwidth]{eos_90}}
\caption{Pressure isotherms as a function of nucleon density for symmetric nuclear matter (a) and neutron-rich matter (b). They display a first-order liquid-gas phase transition. The dotted lines at low temperature in (a) show the non-physical behaviour of the equation of state in the phase transition region. The physical pressure is calculated using the Maxwell construction. The dashed lines delimit the boundary of the coexistence region. The dots indicate the critical temperature.}
\label{eos}
\end{figure}
The free energy is written as a sum of convolution integrals \cite{kaiser2}:
\begin{multline} \label{convolution}
\rho \, \bar{F}(\rho, T) = 4 \int\limits_0^\infty \mbox{d}p \, p \,\mathcal{K}_1(p) \, d(p) + \int\limits_0^\infty \mbox{d}p_1 \int\limits_0^\infty \mbox{d}p_2 \, \mathcal{K}_2(p_1,p_2) \, d(p_1) \, d(p_2) \\
+ \int\limits_0^\infty \mbox{d}p_1 \int\limits_0^\infty \mbox{d}p_2 \int\limits_0^\infty \mbox{d}p_3 \, \mathcal{K}_3(p_1,p_2,p_3) \, d(p_1) \, d(p_2) \, d(p_3) + \rho \, \mathcal{\bar{A}}(\rho,T) \ ,
\end{multline}
where $ \bar{F}(\rho, T) $ is the free energy per nucleon, $ \mathcal{K}_1 $, $ \mathcal{K}_2 $, $ \mathcal{K}_3 $ are respectively one-body, two-body and three-body kernels, and
\begin{equation}\label{density}
d(p) = \frac{p}{2\,\pi^2} \left[ 1 + \exp{\frac{p^2/2 M_N - \tilde{\mu}}{T}} \right]^{-1} \ .
\end{equation}
$ \tilde{\mu} $ is the ``one-body'' chemical potential defined through the relation
\begin{equation}
\rho = 4 \int\limits_0^\infty \mbox{d}p \, p \, d(p) \ .
\end{equation}
$ \mathcal{\bar{A}}(\rho,T) $ is the so-called anomalous contribution and vanishes at $ T = 0 $ \cite{kaiser2}.
The extension of this chiral three loop calculation to isospin-asymmetric nuclear matter requires to take into account the different Fermi seas of protons and neutrons. Consequently we distinguish in Eq.~\eqref{propagator} between proton medium insertion and neutron medium insertion, as well as in Eq.~\eqref{convolution} between proton and neutron distribution functions. Each diagram now involves the sum of all possible combinations of proton and neutron medium insertions with their specific isospin factors.
In section \ref{section2} we present and discuss the equation of state of nuclear matter for different temperatures, featuring the first-order liquid-gas phase transition, and the corresponding phase diagram. In section \ref{section3} we use the scheme above exposed to calculate the chiral condensate of symmetric nuclear matter.
\section{Phase diagram of nuclear matter}\label{section2}
\begin{figure}[tbp]
\centering
\includegraphics[width=.5\columnwidth]{groundenergy}
\caption{Energy per particle as a function of nucleon density for different proton fractions $ x_p = Z/A $ at $ T = 0$. The dashed line shows the trajectory of the saturation point as $ x_p $ varies. For $ x_p \lesssim 0.12 $ the energy is always positive.}
\label{sat_point}
\end{figure}
In Fig.~\eqref{eos} we show the equation of state of nuclear matter as a function of nucleon density $ \rho $ for a series of temperatures running from 0 up to 25 MeV. The plot (a) displays the pressure isotherms for isospin-symmetric nuclear matter, while the plot (b) is indicative for neutron-rich matter and illustrates the pressure isotherms of nuclear matter with proton fraction $ x_p = 0.1 $ ($ x_p = Z/A $). The qualitative behaviour of these curves is reminiscent of the van der Waals equation of state with its characteristic liquid-gas first-order phase transition. The dotted lines in Fig. (a) indicate the non-physical behaviour of the equation of state in the liquid-gas coexistence region. This part of the curve is replaced by the physical one (solid line) using the Maxwell construction. The dashed line sets the boundary of the liquid-gas coexistence region, which terminates at the critical point indicated by the dot. For symmetric nuclear matter we find as critical temperature $ T_c \simeq 15.1 $ MeV. With increasing isospin-asymmetry, energy and pressure of nuclear matter at a given density grow in comparison to the symmetric case. At the same time the coexistence region shrinks. In plot (b) one can observe that the critical temperature has diminished to about 6.3 MeV.
In Fig.~\eqref{sat_point} we study the dependence of the saturation point on the isospin-asymmetry. The saturation point is defined as the energy minimum at $ T = 0 $. The plot shows the energy per particle of nuclear matter at zero temperature for decreasing proton fraction. The saturation point shifts toward smaller nucleon densities. The corresponding binding energy reduces with continuity, as shown by the dashed line, and vanishes at $ x_p \simeq 0.12 $. Below this value neutron-rich matter is unbound for any density and temperature.
\begin{figure}[htbp]
\centering
\includegraphics[width=.5\columnwidth]{T_r}
\caption{$ T-\rho $ phase diagram of nuclear matter for different proton fractions. The dashed line shows the evolution of the critical point.}
\label{phase_asy}
\end{figure}
The thermodynamic information is well summarized by the $ T - \rho $ (temperature versus nucleon density) phase diagram in Fig.~\eqref{phase_asy}. It visualizes clearly how the liquid-gas coexistence region shrinks with increasing isospin-asymmetry and disappears at $ x_p \simeq 0.05 $. Neutron-rich matter with $ x_p \lesssim 0.05 $ is an interacting Fermi gas and cannot undergo any phase transition.
\section{Chiral condensate} \label{section3}
The chiral condensate $ \Braket{\bar{q}q} $ is the order parameter of the chiral phase transition. The non-vanishing of the condensate is connected to the spontaneous breaking of the chiral symmetry of the QCD lagrangian with light quarks. It diminishes with increasing temperature and density. The study of the thermodynamics of the chiral condensate is a key issue in order to locate the boundary of the chiral phase transition in the QCD phase diagram that takes place when the condensate vanishes. The melting of the condensate at high temperature and density restores the chiral symmetry and determines the crossover from the Nambu-Goldstone phase to the Wigner-Weyl realization of the symmetry.
Our starting point is the Hellman-Feynman theorem, which allows to relate the in-medium chiral condensate to the derivative with respect to the quark mass of the free energy per particle of nuclear matter. Using the Gell-Mann-Oakes-Renner relation $ (m_\pi f_\pi) ^2 = -m_q \Braket{0 | \bar{q}q | 0} $, one can rewrite the quark mass derivative as a pion mass derivative, finding the following formula \cite{fio}:
\begin{equation}\label{cond}
\frac{\Braket{\bar{q}q}(\rho,T)}{ \Braket{0|\bar{q}q|0}} = 1 - \frac{\rho}{f_\pi^2} \frac{\partial \bar{F}(\rho,T)}{\partial m_\pi^2} \ .
\end{equation}
The quantities $ f_\pi $ and $ \Braket{0|\bar{q}q|0} $ are taken in the chiral limit ($ m_q, m_\pi \rightarrow 0 $).
The term on the left is the ratio between the in-medium condensate and the vacuum condensate, $ \bar{F} $ is the free energy per particle of nuclear matter. In our model the pion mass appears as an explicit parameter of the free energy and it is possible to calculate its pion mass derivative in a direct way. In the one-body kernel $ \mathcal{K}_1 $ the pion mass dependence is implicit via its dependence on the nucleon mass. The operation of derivation leads to the appearance of the pion nucleon sigma term $ \sigma_N = m_\pi^2\, \partial M/\partial m_\pi^2 $. For our calculation we choose the empirical value $ \sigma_N = (45 \pm 8) $ MeV \cite{sigma}. At $ T = 0 $ the chiral condensate assumes a simpler representation \cite{kaiser7}:
\begin{equation}
\frac{\Braket{\bar{q}q}(\rho)}{ \Braket{0|\bar{q}q|0}} = 1 - \frac{\rho}{f_\pi^2} \left\{ \frac{\sigma_N}{m_\pi^2} \left( 1 - \frac{3\,k_F^2}{10 M^2} \right) + \frac{\partial \bar{E}_{int}(k_F)}{\partial m_\pi^2} \right\} \ .
\end{equation}
The term in the round bracket is the contribution of the non-interacting Fermi gas. It is essentially proportional to the nucleon density and is the leading term of the chiral condensate. The second term is the contribution of the interaction driven by the pion exchange dynamics to the condensate ($ \bar{E}_{int} $ is the interaction energy per particle).
\begin{figure}[htbp]
\centering
\includegraphics[width=.5\columnwidth]{condensate_135_pion}
\caption{Ratio of the chiral condensate of symmetric nuclear matter to its vacuum value as a function of the nucleon density $ \rho $ for different temperatures up to 100 MeV. The contribution of the thermal pions is included. In the liquid-gas coexistence region at low temperatures ($ T \lesssim 15 $ MeV) the quark condensate decreases linearly.}
\label{condensate}
\end{figure}
In Fig.~\ref{condensate} we show the behaviour of the condensate ratio \eqref{cond} as a function of nucleon density for different temperatures from 0 up to $ 100 $ MeV. The contribution of the thermal pions is also included \cite{kaiser}, consisting in a further reduction of the temperature dependent condensate. The effect of the thermal pions is appreciable at high temperatures, $ T > 50 $ MeV. At low temperature ( $ T \lesssim 15 $ MeV) one can recognize the trace of the liquid-gas first-order phase transition on the condensate curves. Using the Maxwell construction one finds that the chiral condensate at a given temperature decreases linearly in the liquid-gas coexistence region (see curve at $ T = 0 $). The net effect of the interactions is to counteract the reduction due to the linear density term. A crucial role is played by the $ \Delta$-isobar excitation. Its contribution is large in comparison to that of the other terms and stabilizes the condensate. The bending of the curves at large densities in Fig.~\ref{condensate} is a consequence of the inclusion of the $ \Delta$-isobar excitation as an explicit degree of freedom. Had we taken into account only the pion exchange dynamics without the $ \Delta$-isobar excitation, the system would have reached chiral restoration not far from the saturation point. With increasing temperature the effect of the interaction weakens and the linear behaviour is practically restored. This recovery is the result of a balance between attractive and repulsive correlations and their mass dependence.
From systematics in the variation of the condensate with nucleon density and temperature presented in Fig.~\ref{condensate} no rapid tendency toward chiral symmetry restoration is seen, at least for densities up to $ \rho \lesssim 2\,\rho_0 $ and temperatures $ T \lesssim 100 $ MeV. Being far from the chiral phase transition, our chiral perturbative approach is reliable in this range of densities and temperatures.
\section{Summary}
In the present review we have presented recent calculations of the equation of state of nuclear matter \cite{fio2} and of the chiral condensate \cite{fio} of isospin-symmetric nuclear matter in the framework of in-medium chiral perturbation theory up to three-loop order in the free energy density.
In our approach long- and intermediate-distance correlations driven by pion exchange dynamics are treated explicitly, while the unresolved short-distance interaction is taken into account by fixing a few contact terms in order to reproduce some selected bulk properties of nuclear matter. Many other ground state, single-particle and thermodynamic properties follow as a prediction of the model. Correlations in the nuclear medium include two- and three-body forces. The resulting equation of state features the characteristic first-order liquid-gas phase transition of nuclear matter, whose dependence on isospin-asymmetry is systematically investigated. We find for isospin-symmetric nuclear matter a critical temperature of about 15 MeV.
The chiral condensate of isospin-symmetric nuclear matter is calculated by differentiating the free energy per particle with respect to the pion mass, which appears as an explicit parameter of the model. Correlations due to $ \Delta$-isobar excitation play the important role of stabilizing the condensate. As a result, we find no indication of a chiral phase restoration for densities up to about twice the density of normal nuclear matter and at temperatures below 100 MeV. This is the first calculation of the chiral condensate at finite density and temperature that systematically includes in-medium two-pion exchange interactions.
The purpose of the present work is to set realistic nuclear physics constraints for ongoing discussions of the QCD phase diagram.
|
\section{Introduction}
Red Clump stars (hereafter RC stars) are a remarkable and well known tool to investigate the structure and the kinematics of various Galactic subsystems.
First two data releases by Asiago Red Clump surveys (at high \cite{Val} and intermediate \cite{Sag} resolving power) provide atmospheric parameters(T$_{eff}$, log(g), [M/H]), spectrophotometric distances and radial velocities for a well selected sample of 439 RC stars, mainly located in a torus extending from 200 to 500 pc from the Sun (future ARCS data releases will explore both inward and outward). Table \ref{tab:1} shows the accuracy of the data provided by the survey. \\
We implemented the catalog with ages computed using the code PARAM, developed by L. Girardi \cite{daS}. The uncertainty on age is up to 40\%-80\%, due to the nature of the RC: isochrones of very different age and metallicity lie closely together in this region. For this reason we used PARAM ages only for a first statistical investigation and we considered only stars with an accuracy on age better than 60\%. The age determination for Red Giants (and hence RC stars) will be improved by recent and exciting asteroseismology findings with data provided by CoRoT and Kepler space missions (\cite{Mos}, \cite{Mont}, \cite{Mig}).\\
A typical analysis of the distribution of ARCS stars in the UV velocity space permits the study of the disk kinematics in the Solar Neighborhood . Different local irregularities that comes out as concentrations of stars in the U-V distribution are the moving groups or stellar streams \cite{Sku}, \cite{Zhao}. Several hypotheses are claimed to explain the origin of moving groups: dispersal stellar clusters, accretion or a resonant mechanism. A homogeneity in chemical abundances and age is expected if the moving group is the debris of star forming aggregates or of an infalling object . On the other hand, an absence of such homogeneities argues for a resonant origin (related to the Galactic bar or spiral arms). \\
A 2-D wavelet transform technique is used to identify moving groups in the U-V plane, as performed by Skuljan \cite{Sku}. We used a Mexican-hat-shaped kernel function with a scale parameter of 5 km s$^{-1}$ (see Figure \ref{fig:1}). Most known moving groups are clearly visible in the ARCS sample. The high accuracy of the data provided by ARCS exclude that detected over-densities are artifacts. In Table \ref{tab:1} we summarized the characteristics of the most prominent moving groups and we compared our results with those of Zhao \cite{Zhao}. There is good agreement between the two works and discrepancies between (U, V) coordinates of the moving groups are mainly due to a different U$_\odot$V$_\odot$W$_\odot$ adopted. The large dispersion in metallicity and age argues in favour of a resonant origin of the moving groups of Sirius-UMa, Coma, Hyades, Pleiades and Hercules. In addiction to these moving groups, Fig. \ref{fig:1} shows other overdensities that are discussed in detail elsewhere (Valentini et al. 2011, submitted).
\begin{table}
\centering
\caption{Accuracy of the ARCS high and medium resolution surveys.}
\label{tab:1}
\begin{tabular}{ l l l l l }
\hline\noalign{\smallskip}
& V$_{rad}$ & T$_{eff}$& log(g) & [M/H] \\
\noalign{\smallskip}\hline\noalign{\smallskip}
ARCS high &0.5 km s$^{-1}$ & 55 K & 0.12 dex & 0.11 dex \\
ARCS medium & 1.3 km s$^{-1}$ & 88 K & 0.38 dex & 0.17 dex \\
\noalign{\smallskip}\hline
\end{tabular}
\end{table}
\begin{figure}
\centering
\resizebox{0.57\columnwidth}{!}{
\includegraphics{valentini_m_fig1} }
\caption{2-D wavelet transform with a Mexican-hat kernel of the U-V distribution of ARCS Red Clump stars. The scale parameter of the analyzing wavelet is 5 km s$^{-1}$, the color scale is adjusted in order to emphasize overdensities.}
\label{fig:1}
\end{figure}
\begin{table}
\centering
\caption{Characteristic U and V velocities, metallicity and age of most prominent features detected in the U-V distribution of ARCS Red Clump stars. Results obtained with ARCS catalog are compared with those present in the catalog of moving groups of Zhao \cite{Zhao}. }
\label{tab:2}
\begin{tabular}{ l l l l l l}
\hline\noalign{\smallskip}
ID & \multicolumn{3}{l}{ARCS} & \multicolumn{2}{l}{Zhao et al. (2009)} \\
& (U,V) & [M/H] & Age & (U,V) & [M/H] \\
& km s$^{-1}$ & dex & Gyr & km s$^{-1}$ & dex \\
\noalign{\smallskip}\hline\noalign{\smallskip}
Sirius-UMa & ($+$0,$+$7) & $-$0.11 $\sigma$=0.22 & 1.8 $\sigma$= 3.0 & ($+$10,$-$14) & $-$0.21 $\sigma$=0.15 \\
Coma & ($-$11,$-$20) & $-$0.12 $\sigma$=0.20 & 1.5 $\sigma$= 3.0 & ($-$38,$-$17) & $-$0.09 $\sigma$=0.17 \\
Pleiades & ($-$8,$-$22) & $-$0.18 $\sigma$=0.26 & 3.0 $\sigma$= 1.8 & ($-$15,$-$23) & $-$0.17 $\sigma$=0.17 \\
Hercules & ($-$10,$-$40) & $-$0.10 $\sigma$=0.29 & 5.0 $\sigma$= 3.0 & ($-$35,$-$51) & $-$0.16 $\sigma$=0.20 \\
\noalign{\smallskip}\hline
\end{tabular}
\end{table}
|
\section{Introduction} The following problem was posed to the author by K. Kurdyka, to which we express our gratitude for stimulating discussions.\\
Given $a=(a_{1},\ldots, a_{n})\in \mathbb{R}^{n}$ we define the function $q_{a}:\mathbb{R}^{n}\to \mathbb{R}$ by
$$q_{a}(x)=a_{1}x_{1}^{2}+\cdots +a_{n}x_{n}^{2}$$
where in the case $b\in \mathbb{R}^{k}, k\neq n,$ we mean $q_{b}$ to belong to $C^{\infty}(\mathbb{R}^{k},\mathbb{R})$ and to be defined in the same similar way.\\
Suppose $f\in C^{\infty }(\mathbb{R}^{n},\mathbb{R})$ and $M\subset \mathbb{R}^{n}$ is a submanifold. Is it true that the set
$$A(f,M)=\{a\in \mathbb{R}^{n}\,|\, (f+q_{a})|_{M}\,\,\textrm{is Morse}\}$$
is residual in $\mathbb{R}^{n}?$\\
The answer to this question turns out to be affirmative, but in a subtle way: standard transversality arguments based on dimension counting do not work and we have to prove it directly.
\section{Failure of parametric transversality argument}
We describe here what it is the usual procedure to prove that given a family of functions $f_{a}:M\to\mathbb{R}$ depending smoothly on the parameter $a\in A$ then the set of $a$ such that $f_{a}$ is Morse is residual in $A.$\\
Let $G:A\times M\to N$ be a smooth map and for every $a\in A$ let $g_{a}:M\to N$ be the function defined by $x\mapsto G(a,x).$ Suppose that $Z\subset N$ is a submanifold and that $F$ is transverse to $Z.$ Then from the \emph{parametric transversality theorem} (see \cite{Hirsch}, Theorem 2.7) it follows that $\{a\in A\,|\, g_{a}\,\, \textrm{is transverse to $Z$}\}$ is residual in $A.$\\
In the case we want to get Morse condition consider $N=T^{*}M,$ $G(a,x)=d_{x}f_{a}$ and $Z\subset T^{*}M$ the zero section. Then $f_{a}$ is Morse if and only if $g_{a}$ is transverse to $Z.$\\
In our case, letting $M=\mathbb{R}^{n}$, we are led to define $G:\mathbb{R}^{n}\times \mathbb{R}^{n}\to \mathbb{R}^{n}$ by
$$(a,x)\mapsto (\partial f/\partial x_{1}(x)+a_{1}x_{1},\ldots, \partial f/\partial x_{n}(x)+a_{1}x_{n})$$
and we consider $Z=\{0\}\in \mathbb{R}^{n}.$ Then $f_{a}=f+q_{a}$ is Morse if and only if $g_{a}$ is transversal to $\{0\}.$ A condition that would ensure this (trough the parametric transversality theorem) is that $G$ is transverse to $\{0\}.$ Computing the differential of $G$ at the point $(a,x)$ we have for $(v,w)\in T_{(a,x)}(\mathbb{R}^{n}\times \mathbb{R}^{n})$
$$(d_{(a,x)}G)(v,w)=\textrm{He}(f)(x)v+\textrm{diag}(a_{1},\ldots, a_{n})v+\textrm{diag}(x_{1},\ldots, x_{n})w$$ and we see that in general this condition does not hold (for example let $f\equiv 0,$ then at alle the points $(a,x)=(0,a_{2},\ldots, a_{n},0,\ldots, 0)$ we have $G(a,x)=0$ but $\textrm{rk}(d_{(a,x)}G)<n$).
\section{A direct approach}
First we recall the following Lemma (see \cite{Bott}).
\begin{lemma}Let $f$ be a smooth function on $\mathbb{R}^{n}$ and for $a\in\mathbb{R}^{n}$ define the function $f_{a}$ by $x\mapsto f(x)+a_{1}x_{1}+\ldots +a_{n}x_{n}.$ The set $$\{a\in\mathbb{R}^{n}\,|\, f_{a}\,\,\textrm{is Morse}\}$$ is residual in $\mathbb{R}^{n}.$ \end{lemma}
\begin{proof}
Define the function $g(x)=(\partial f/\partial x_{1},\ldots, \partial f/\partial x_{n})$ and notice that the Hessian of $f$ is precisely the Jacobian of $g$ and that $x$ is a nondegenerate critical point for $f$ if and only if $g(x)=0$ and the Jacobian $J(g)(x)$ of $g$ at $x$ is nonsingular. Then $g_{a}(x)=g(x)+a$ and $J(g_{a})=J(g).$ We have that $x$ is a critical point for $f_{a}$ if and only if $g(x)=-a;$ moreover it is a nondegenerate critical point if and only if we also have $J(g)(x)$ is nonsingular, i.e. $a$ is a regular value of $g.$ The conlusion follows by Sard's lemma.
\end{proof}
We immediately get the following corollary.
\begin{coro}\label{nonz}If $f$ is a smooth function on an open subset $U$ of $\mathbb{R}^{n}$ such that for every $u=(u_{1},\ldots, u_{n})\in U$ we have $u_{i}\neq 0$ for all $i=0,\ldots ,n,$ then
$$A(f,U)=\{a\in \mathbb{R}^{n}\, |\, f+q_{a} \,\,\textrm{is Morse on $U$}\}$$ is a residual subset of $\mathbb{R}^{n}.$
\end{coro}
\begin{proof}The functions $u_{1}^{2},\ldots,u_{n}^{2}$ are coordinates on $U$ by hypothesis; we let $\tilde f$ be the function $f$ in these coordinates (it is defined on a certain open subset $W$ of $\mathbb{R}^{n}$). Then for every $a\in \mathbb{R}^{n}$ we have that (using the above notation) $\tilde f_{a}$ is Morse on $W$ if and only if $f+q_{a}$ is Morse on $U$ and the conclusion follows applying the previous lemma.
\end{proof}
To prove the general statement we need the following.
\begin{lemma}\label{open}Let $f$ be a smooth function on an (arbitrary) open subset $U$ of $\mathbb{R}^n.$ Then the set $A(f,U)$ is residual in $\mathbb{R}^{n}$.
\begin{proof}
For every $I=\{i_1,\ldots, i_j\}\subset \{1,\ldots,n\}$ define $$H_I=U\cap \{ u_{i}=0,\, i\in I\}\cap\{u_{k}\neq 0,\, k\notin I\}.$$
To simplify notations let $I=\{1,\ldots, j\}$. Notice that if $a=(a_1,\ldots,a_n)$ and $a''=(a_{j+1},\ldots,a_n)$ then $(q_a)|_{H_{I}}=(q_{a''})|_{H_{I}}$ where $q_{a''}:\mathbb{R}^{n-j}\to \mathbb{R}$ is defined as above.
By corollary \ref{nonz} the set $$A''(f, H_{I} )=\{a''\in \mathbb{R}^{n-j}\, |\, \textrm{$f|_{H_{I}}+q_{a''}$ is Morse on $H_I$}\}$$ is residual in $\mathbb{R}^{n-j}$.
Let $a=(a',a'')\in \mathbb{R}^n$ such that $a''\in A''(f, H_{I} )$ and suppose $x\in H_I$ is a critical point of $f+q_a;$ then $x$ is also a critical point of $(f+q_a)|_{H_I}=f|_{H_{I}}+q_{a''}.$ Since $a''\in A''(f, H_{I} )$ then $x$ belongs to a countable set, namely the set $C_{a''}$ of critical points of $ f|_{H_{I}}+q_{a''}$ (each of this critical point must be nondegenerate by the choice of $a''$); moreover we have that $$\textrm{He}(f|_{H_{I}}+q_{a''})(x)=\textrm{He}(f|_{H_{I}})(x)+\textrm{diag}(a_{j+1},\ldots,a_n)$$ is nondegenerate.
Notice that the Hessian of $f+q_a$ at $x$ is a block matrix: $$\textrm{He}(f+q_a)(x)= \left(\begin{array}{c|c}
\textrm{diag}(a_1,\ldots,a_j)+B(x) & C(x) \\ \hline C(x)^{T}& \textrm{He}(f|_{H_{I}}+q_{a''})(x)\end{array}\right). $$
Thus for every $a''=(a_{j+1},\ldots,a_{n})\in A''(f, H_{I})$ and for every $x\in C_{a''}$ consider the polynomial $p_{a'',x}\in \mathbb{R}[t_1,\ldots,t_j]$ defined by $$p_{a'',x}(t_1,\ldots,t_j)=\textrm{det}(\textrm{He}(f)(x)+\textrm{diag}(t_1,\ldots,t_j,a_{j+1},\ldots,a_n))$$
Then the term of maximum degree of $p_{a'',x}$ is $$t_{1}\cdots t_{j}\det (\textrm{He}(f|_{H_{I}}+q_{a'})(x))$$ which is nonzero since $\textrm{det}(\textrm{He}(f|_{H_{I}}+q_{a''})(x))\neq 0$ ($x$ is a \emph{nondegenerate} critical point of $f|_{H_{I}}+q_{a''}$). It follows that $p_{a'',x}$ is not identically zero; hence its zero locus is a \emph{proper} algebraic set. Thus for each $a''\in A''(f, H_{I})$ and each $x\in C_{a''}$ the set $A'(a'',x, I)$ defined by $$\{a'\in \mathbb{R}^j\, |\, \textrm{if $x$ is a critical point of $f+q_{(a',a'')}$ on $H_I$ then it is nondegenerate}\}$$ is residual in $\mathbb{R}^{j}$ (it is the complement of a proper algebraic set); it follows that $$A'(a'',I)=\{a'\in \mathbb{R}^j\, |\, \textrm{each critical point of } f+q_{(a',a'')}\, \textrm{on } H_{I}\textrm{ is nondegenerate}\}$$ is residual in $\mathbb{R}^{j}$, since it is a countable intersection of residual sets, i.e. $$A'(a'',I)=\bigcap_{x\in C_{a''}}A'(a'',x,I)$$
Thus the set $$A(f,I)=\{(a',a'')\, |\, a''\in A''(f, H_{I}), \, a'\in A'(a'',I)\}$$ (which coincides with the set of $a=(a',a'')\in \mathbb{R}^n$ such that each critical point of $f+q_a$ on $H_{I}$ is nondegenerate) is residual: is residual in $a'$ for every $a''$ belonging to a residual set.
Finally $$A(f,U)=\bigcap_{I\subset \{1,\ldots,n\}}A(f,I)$$ is a finite intersection of residual sets, hence residual.
\end{proof}
\end{lemma}
\begin{teo}\label{Morse}Let $f$ be a smooth function on $\mathbb{R}^{n}$ and $M\subset \mathbb{R}^{n}$ be a submanifold. Then the set $A(f,M)$ is residual in $\mathbb{R}^{n}$.
\end{teo}
\begin{proof}We basically improve the proof of Proposition 17.18 of \cite{Bott}.\\
Let $u_1,\ldots,u_n:\mathbb{R}^{n}\to\mathbb{R}$ be the coordinates on $\mathbb{R}^n.$ Suppose $M$ is of dimension $m.$
For every point $\overline{x}\in M$ there exists a neighborhood $W$ of $\overline{x}$ in $M$ such that $u_{i_1},\ldots, u_{i_{m}}$ are coordinates for $M$ on $$W\simeq \mathbb{R}^{m},$$ for some $\{i_1,\ldots,i_m\}\subseteq \{1,\ldots,n\};$ since $M$ is second countable, then it can be covered by a countable (finite if $M$ is compact) number of such open sets. For convenience of notations suppose $\{i_1,\ldots,i_m\}=\{1,\ldots,m\}.$\\
Thus $u_{1},\ldots, u_{m}$ are coordinates on $W\simeq \mathbb{R}^{m}$ and $f|_{W},u_{m+1}|_{W},\ldots, u_{n}|_{W}$ are functions of $u_{1}|_{W},\ldots,u_{m}|_{W}.$
Fix $a''=(a_{m+1},\ldots, a_{n})\in\mathbb{R}^{n-m}$ and define $g_{a''}:W\to \mathbb{R}$ by $$g_{a''}=f|_{W}+a_{m+1}u_{m+1}^{2}|_{W}+\cdots+a_{n}u_{m}^{2}|_{W}=(f+a_{m+1}u_{m+1}^{2}+\cdots+a_{n}u_{n}^{2})|_{W}$$
Notice that $g_{a''}$ is not $(f+q_{a})|_{W}$ since we are taking only the last $n-m$ of the $a_{i}'s;$ we still have the freedom of choice $(a_{1},\ldots,a_{m}).$ \\
By lemma \ref{open}, since $u_{1}|_{W},\ldots, u_{m}|_{W}$ are coordinates on $W$, for every $a''\in \mathbb{R}^{n-m}$ the set $$\{a'=(a_1,\ldots,a_m)\in \mathbb{R}^{m}\quad \textrm{s.t.}\quad g_{a''}+a_1u_1^{2}|_{W}+\cdots +a_mu_m^{2}|_{W}\textrm{ is Morse on $W$}\}$$ is residual in $\mathbb{R}^{m}$.
Notice that $ g_{a''}+a_1u_1^{2}|_{W}+\cdots +a_mu_m^{2}|_{W}=(f+q_{(a',a'')})|_{W};$ hence for every $a''$ the set of $a'$ such that $(f+q_{(a',a'')})|_{W}$ is Morse on $W$ is residual. Thus the set of $a\in \mathbb{R}^n$ such that $(f+q_a)|_{W}$ is Morse on $W$ is residual (it is residual in $a'$ for each fixed $a''$ hence it is globally residual). It follows that $A(f,M)$ is a countable intersection of residual set, hence residual.\end{proof}
|
\section{\label{sec:intr} Introduction}
Data assimilation (DA) is a powerful and versatile method for
combining partial, noisy observational data of a system with its
dynamical model, generally numerically implemented, to generate state
estimates of nonlinear, chaotic systems. A variety of data
assimilation methods, broadly separated into deterministic or
probabilistic methods, have been developed over the past few decades
and used mainly in the earth sciences.\cite{book:lewisvarahan06,
book:park08, Kalnay03, book:malanotte96, Ben01, book:evensen,
PhysicaD:Ide} On the other hand, when two or more chaotic systems
are coupled, they may adjust the properties of their motion due to
coupling in such a way that their evolution becomes
\emph{synchronized}. Several types of coupling, unidirectional,
mutual, common driving, etc., and their effects in the form of a
variety of synchronizations such as identical, generalized, phase,
lag, amplitude, etc., as well as applications of synchronizations have
been and continue to be a very active area of
research.\cite{prl:pecora, book:pikovosky}
We will be interested in this paper with \emph{sequential data
assimilation}, which incorporates observations sequentially in
time. This method can be thought of as synchronization through
unidirectional coupling, with observations acting as the master system
and the numerical model acting as the slave system.\cite{prl:pecora,
atmos:yang, npg:Duane, JGR:Szendro} In fact, this is exactly the
nudging method for assimilation, suggested first in
meteorology,\cite{mwr:hoke} and later used for a number of different
studies.\cite{jgr:verron, tellus:stauffer, mwr:bao, jaot:wang,
ngp:aurox} Recent studies show that such techniques can also be used
for assimilating data in order to track complex spatio-temporal
dynamics of excitable media,\cite{chaos:berg} and forecast the state
of a time-delayed high-dimensional system, e.g., in chaotic
communication.\cite{prl:cohen} Synchronization based DA method may be
useful for small or mesoscale predictions, specially when observations
are frequent and coupled to a small number of
variables.\cite{atmos:yang,ngp:aurox} This method is relatively fast
and robust, and can be implemented easily.\cite{Oceanography:Stammer}
Two of the main characteristics of the observational data used in
earth sciences applications is that they are discrete in time and are
noisy. The main aim of this paper is to study synchronization with
these characteristics in mind, as we explain below.
\begin{enumerate}
\item It has been observed recently\cite{DynAtmOcean:ide,
jhydro:walker, waterres:walker, tellus:apte} that the time period
between two observations is one of the important factors which
significantly affect conventional DA schemes. Though frequent
updates may produce better results depending on the quality of the
model as well as the observations, they increase the burden of
computational cost.\cite{advspaceres:Riishojgaard} Since the choice
of observational frequency depends on many factors such as cost and
ease of observations, it is important to understand in detail how
the time period between observations affects the accuracy of
predictions which in turn depends on the accuracy of the data
assimilation.
For these reasons, we investigate the effect of varying
observational frequency, i.e., the gaps between the observations, on
discrete time synchronization -- this discussion forms the first
part of the paper. As expected, we find that observations which are
farther apart lead to poorer synchronization and consequently low
accuracy of prediction. We also find that there is a certain
threshold, which in the system we study is roughly of the order of
the inverse of Lyapunov exponent of the system, such that using
observations with a time period greater than that threshold leads to
no synchronization.
\item It is of course well known that the observational noise plays
crucial role in data assimilation as well as in
synchronization.\cite{PRL:Neiman, PRL:Teramae, PRE:Guan} Thus we
also carry out the study described in the above paragraph with noisy
observations with different noise levels, in order to investigate
the effect of noise on synchronization. The discussion of these
results forms the second part of the paper. Again as expected, the
increase in noise leads to poorer synchronization and prediction.
\end{enumerate}
We present these results in detail in section~\ref{sec:results} and a
discussion of these results along with directions for further studies
are presented in section~\ref{sec:conclude}.
In order to investigate the effects of these two aspects, we perform
the so-called ``identical twin experiments''\cite{tellus:talagrand81}
as follows. We generate observations using a known trajectory, called
the ``true'' trajectory of a numerical model. We use these
observations (master) to try to synchronize the same numerical model
starting with a different initial condition (slave). We quantify the
``degree of synchronization'' by calculating the root mean square
(RMS) error of the slave trajectory with respect to the true
trajectory, leaving out the first half of the time-span, in order to
get rid of transients. We assess the effect on predictions by
calculating the RMS error over certain time-spans after the
synchronization is stopped and the slave model is let to run by
itself. Currently we are investigating the effect of these two
factors, the time period between observations and the noise level, on
data assimilation methods such as ensemble Kalman filter
(ENKF).\cite{book:evensen}
The dynamical model we use is that of a chaotic Chua
circuit.\cite{IJBC:chua, pre:singla} One of the main reason for this
choice is that in future we would like to use the actual data from the
circuit experiments being currently performed by two of the
authors.\cite{pre:singla} This model is described in detail in the
following section.
\section{Chua circuit model and synchronization}
\label{sec:model}
To study synchronization we have used the Chua model used in
Ref.~\onlinecite{pre:singla}. The model of the circuit is made
dimensionless by substituting $\tau=\beta t$, where
$\beta=(R_1C_1)^{-1}$, and $R_1$ and $C_1$ are the resistance and
capacitor used in the circuit respectively. The dimensionless model is
given by.
\begin{eqnarray}
\frac{dx}{d\tau}&=&\frac{1200}{R} (y-x)- g(x) \,,\nonumber \\
\frac{dy}{d\tau}&=&\sigma\left[\frac{1200}{R} (y-x) +z\right] \,,\nonumber\\
\frac{dz}{d\tau}& =&-c [y +r'z] \,,\label{eqn:master}
\end{eqnarray}
where
$$ g(x) = \left\{ \begin{array}{ll}
-x &\mbox{ $|x|<1$} \\
-[1+b(|x| -1)]sign(x) &\mbox{$1<|x|\le10$}\\
10[(|x| -10)-(9b+1)]sign(x) &\mbox{$10<|x|$}
\end{array} \right. $$
and $\sigma, c, r',$ and $b$ are the parameters whose values depend on
the capacitors, resistances and inductor used in the
circuit.\cite{pre:singla} The resistance $R$ is the control parameter
of the system. Depending on R, the system shows chaotic or regular
behavior. In the present study, we have chosen $\sigma = 4.6/69
\approx 0.067$, $c = (1200 \times 3300 \times 4.6/8.5) \times 10^{-6}
\approx 0.779$, $r' = 85/1200 \approx 0.071$ and $b = 1 - 1200/3300
\approx 0.636$. The model was integrated using the fourth order
Runge-Kutta numerical scheme with integration step $\Delta \tau =
0.01$, and the output of the model has been expressed in terms of real
time $t = \tau/\beta$ rather than in terms of the dimensionless time
$\tau$.
\begin{figure}
\includegraphics[width=8.5 cm]{fig1}
\caption{Two Chua attractors for the system in
Eqs.\eqref{eqn:master}. Inset shows the large attractor, which
represents the diverging dynamics, along with the above two
attractors.}
\label{fig:attractors}
\end{figure}
The model equations show three attractors depending upon the initial
conditions as shown in Fig.~\ref{fig:attractors}(inset), and zoomed
version of the smaller attractors by black and gray plots. As the
basin of attractors is not known, it is not possible to tell \emph{a
priory} which initial condition will lead to a particular
attractor. Generally, for high values of initial condition
(approximately greater than $1$ in absolute value), system goes to the
large limit cycle attractor [inset of Fig.~\ref{fig:attractors}], and
for smaller initial conditions, system stays on any of the two small
attractors [Fig.~\ref{fig:attractors}]. As the larger attractor is not
of experimental interest, the dynamics on this attractor is considered
as divergent dynamics.
In the present simulation, we have used $R = 1245 \Omega$ for which
the system shows chaotic behavior. The ``true'' (or master) initial
condition is chosen such that the trajectory stays on any one of the
small attractors depending upon the initial conditions. The largest
Lyapunov exponent ($\lambda$) is positive ($\approx 4.04$ bits/msec),
and corresponding Lyapunov time is $t_{\lambda} = 3.6 \times 10^{-4}
sec$. $\lambda$ was estimated using Rosenstein
techniques~\cite{physicaD:Rosenstein} and is consistent with the
earlier results.\cite{JCktSystComp:parlitz}
In order to generate any one set of observations, we first simulated a
long trajectory on one of the smaller attractors. We refer to such a
trajectory as ``truth'' in keeping with other data assimilation
literature. Next, we sub-sampled this trajectory to extract several
sets of discrete-time equispaced observations with varying time
interval between the observations, which will henceforth be denoted by
$T_{\textrm{obs}}$. For the present study, observations were
generated with four different values of $T_{\textrm{obs}}$, namely,
$2.0\times10^{-5}$, $5.0\times10^{-5}$, $7.0\times10^{-5}$, and
$1.0\times10^{-4}$ sec. Noisy observations were created by adding
noise of particular intensity ($D$) to these discrete-time
observations, where $D$ was chosen to be a specified fraction of the
standard deviation of the attractors. The results presented in
Sec.~\ref{sec:results} are for one specific set of observations, but
the qualitative features of these results were found to be identical
when we used observational sets generated from many different initial
conditions.
\subsection{Master-slave coupling}
The model (slave) was unidirectionally coupled with $N$ discrete-time
observations (master) at times $\tau_1,\dots,\tau_N$ as follows.
\begin{widetext}
\begin{eqnarray}
\frac{dx}{d\tau}&=&\frac{1200}{R} (y-x)- g(x) -
k_x\sum_{i=1}^N \left[x(\tau) - x_{obs}(\tau_i)\right]\delta(\tau-\tau_i) \nonumber \\
\frac{dy}{d\tau}&=&\sigma\left[\frac{1200}{R} (y-x) +z\right] -
k_y\sum_{i=1}^N \left[y(\tau) - y_{obs}(\tau_i)\right]\delta(\tau-\tau_i) \nonumber \\
\frac{dz}{d\tau}& =&-c [y +r'z] -
k_z\sum_{i=1}^N \left[z(\tau) - z_{obs}(\tau_i)\right]\delta(\tau-\tau_i)
\label{eqn:sync}\end{eqnarray}
\end{widetext}
where, $k_x, k_y$, and $k_z$ represent the $x$, $y$, and $z$
coupling. In the case of continuous time observations which is also
discussed below, the slave is coupled to the master directly, by
replacing the summations of delta functions in the above equations by
terms such as $k_x[x(\tau) - x_{\textrm{obs}}(\tau)]$. We have quantified
synchronization using RMS error of the difference of the synchronized
and true trajectory over the last half of the trajectory, and the
quality of prediction is again quantified by the RMS error with
respect to truth over varying time periods, as discussed in detail
below. The results about synchronization and prediction for different
$T_{\textrm{obs}}$ and noise level ($D$), and a comparison of these
results with the case of continuous time synchronization are presented
in the next section.
\section{Results}
\label{sec:results}
The main aim of this section is to discuss the effects of the
observational time period $\tobs$ and the observational noise variance
$D$ on the root mean square of the difference between the slave system
and the master, both during the time period in which the slave is
coupled to the master (synchronization phase) and after the coupling
is switched off (the prediction phase).
\subsection{Synchronization with observations without noise ($D=0$)}
\label{subsec:obs_time}
\begin{figure}[t]
\includegraphics[width=8.5 cm]{fig2}
\caption{RMS errors in the $x$-component (top) and the total RMS error
(bottom) when $x$-component was observed. In both these plots as
well as in Fig.~\ref{fig:rms_y}, thick solid line is for continuous
time synchronization while dashed, thin solid, and gray lines are
for cases with $T_{\textrm{obs}}=2.0\times10^{-5}$,
$5.0\times10^{-5}$, and $7.0\times10^{-5}$ sec respectively. The
$y$- and $z$-components also show similar qualitative behavior. Note
that RMS error above approximately $0.5$ indicates loss of
synchronization.}
\label{fig:rms_x}
\end{figure}
In order to investigate the performance of the synchronization, we
have first studied the case of coupling only the $x$-component,
i.e. setting $k_y = 0 = k_z$ in Eqns.~\ref{eqn:sync} while $k_x$ is
varied from zero to higher values. We quantify synchronization of each
components by calculating the RMS error of the component-wise difference of
master and slave trajectories, and also the total RMS error. For
continuous case, the slave model synchronizes with the master within a
very short time. Solid lines in Fig~\ref{fig:rms_x} (a) and (b) show
the RMS errors in $x$-component and total errors respectively. They
show that the system started synchronizing around $k_x\approx 0.005$,
and achieved full synchronization around $k_x\approx 0.03$ as the RMS
errors are almost zero. The system remained in synchronized state even
at higher values of $k_x$.
\begin{figure}
\includegraphics[width=8.5 cm] {fig3}
\caption{First and second columns represent the $x$ and $y$ components
of the slave trajectories (solid line) and model trajectories (gray
line) at $k_x = 1.74$ for the observations taken at
$T_{\textrm{obs}}=2.0\times10^{-5}$ (top) and $5.0\times10^{-5}$
(bottom). Note that the jumps in $x$-component and in $y$-derivative
are because of the $\delta$-functions in Eqs.~\eqref{eqn:sync}. }
\label{fig:time_sr}
\end{figure}
Synchronization with $x$-component observations at discrete time is
shown in Fig.~\ref{fig:rms_x}. In this figure, the dashed-lines show
the RMS errors when $x$-observation were recorded with
$T_{\textrm{obs}}=2.0\times10^{-5}$ sec, which is approximately 18
times less than $t_{\lambda}$. The figure also shows that the model
started synchronizing around $k_x=0.18$ with the observations and
achieved full synchronization around $k_x>0.5$. For $k_x>2.5$ model
diverges from the observations. So the range of the coupling $k_x$ for
which there is synchrony has become finite when observations were
taken at a finite interval. Synchrony is also clear from the plots of
$x$ and $y$ time series of the observational trajectories and slave
trajectories shown in Fig.~\ref{fig:time_sr}. They show the model
(gray lines) gets perfectly synchronized with the observational
trajectory (black lines).
Note that RMS error above approx. $0.5$ indicates a loss of
synchronization, hence effectively synchronization is obtained only
for the first two of these cases, i.e., continuous time observations
and with $\tobs = 2.0 \times 10^{-5}$ sec.
Black thin solid lines [Fig~\ref{fig:rms_x}] show the RMS errors, when the
observations were taken with $T_{\textrm{obs}}=5.0\times10^{-5}$ sec.~
($ \approx \frac{1}{7} t_{\lambda}$). In this case errors decreased slowly with
$k_x$ and showed partial synchronization only in a very small range of
$k_x$ around $k_x \approx 1.7$. The time series at $k_x=1.74$ in plots
(c) and (d) of Fig.~\ref{fig:time_sr} show such partial synchrony
between the observed trajectory (gray lines) and the slave model
(black line) for $\tobs = 5.0\times10^{-5}$ sec.~The behaviour of the
$z$-component was similar to that of the $y$. For
$T_{\textrm{obs}}=7.0\times10^{-5}$ sec, the RMS error is least around
$k_x=1.19$ but even in this case, it is far from being
synchronized. These results show that $T_{\textrm{obs}}$ is an
important factor in synchronization that must be considered in DA
experiments.
\begin{figure}
\includegraphics[width=8.5 cm]{fig4}
\caption{RMS errors in the $x$-component (top) and the total RMS error
(bottom) when $x$-component was observed. The lines are for same
$\tobs$ as described in Fig.~\ref{fig:rms_x}. Comparing the RMS
error for $\tobs = 5.0 \times 10^{-5}$ sec (thin solid lines) in
this figure and in Fig.~\ref{fig:rms_x}, we see that
$y$-observations are more informative than the $x$-observations.}
\label{fig:rms_y}
\end{figure}
When $y$-component was observed, i.e., when we set $k_x = 0 = k_z$ and
vary $k_y$, the model showed synchrony for different range of
$k_y$ than that was for $x$ observations. In Fig.~\ref{fig:rms_y}, the
solid lines show the RMS errors in (a) $x$-component and (b) total RMS
error for continuous observation case. Around $k_y=0.2$ the model
achieved synchrony and remains in this state even at higher values of
$k_y$. For $T_{\textrm{obs}}=2.0\times10^{-5}$ sec (dashed lines),
slave system achieved full synchronization around $k_y=0.53$, and
remains synchronized for $0.53<k_y<2.22$. When
$T_{\textrm{obs}}=5.0\times10^{-5} sec$ (thin solid), the
synchronization region shifts to $1.14<k_y<2.72$. Similar to
$x$-observation case, there is no synchronization for the
$y$-observation with $T_{\textrm{obs}}=7.0\times10^{-5} sec$ (gray
line) and above. Thus we see that the $y$-component is a more dominant
dynamical variable than the $x$-component.
\begin{figure}
\includegraphics[width=8.5 cm]{fig5}
\caption{(a) RMS errors in $x$ (solid), $y$ (dashed) and $z$-component
(dotted line) when $z$-component was observed continuously in time,
showing that there is no synchronization when $z$ is observed. Time
series of slave (gray) and master (black line) at $k_z= 1.2$ is
shown in (b) $x$, (c) $y$ and and (d) $z$. The $x$- and
$y$-components of the slave have been multiplied by factors of 20
and 0.5 respectively after subtracting mean, in order to see the comparison clearly.}
\label{fig:rms_z}
\end{figure}
\begin{figure}
\includegraphics[width=8.5 cm]{fig6}
\caption{Attractors of the master (solid black) and slave system
(dotted line) when $z$ was observed. The structures of these
attractors are clearly different from each other.}
\label{fig:z_attract}
\end{figure}
When model was coupled with master through $z$-observations ($k_x = 0
= k_y$), there was no synchronization even for the continuous
observations case. Panel (a) of Fig~\ref{fig:rms_z} shows the RMS
errors in $x$ (solid line), $y$ (dashed line), and $z$ (dotted line)
components as $k_z$ is varied. This is also clear from the plots of
$x$, $y$, and $z$ time series of the master (gray) and slave (solid)
shown in panels (b), (c), and (d), respectively. The difference
between the master (solid line) and slave (dotted line) is also clear
from the attractors shown in Fig~\ref{fig:z_attract}. A possible
explanation for not having synchronization with $z$ coupling is that
this variable may not contain sufficient information about the
dynamics. Hence in the current regime of the Chua system $x$ and $y$
contain useful information for synchronization but not the
$z$-component.\cite{atmos:yang} Finally as the model did not show any
synchrony for continuous case, obviously we did not get any synchrony
with discrete time observations either.
\subsection{Synchronization with noisy observations}
\label{subsec:noisy_obs}
\begin{figure}
\includegraphics[width=8.5 cm]{fig7}
\caption{RMS error of synchronization when noisy $x$-component
observations with $T_{\textrm{obs}}=2.0\times10^{-5}$ are coupled to
the slave model: (a) RMS error in $x$-component and (b) total RMS
error. The solid, dashed, dotted, dashed-dotted, and gray lines
represent the RMS errors for noise levels $D=2\%, 10\%, 20\%, 30\%$,
and $40\%$, respectively.}
\label{fig:rms_nz}
\end{figure}
\begin{figure}
\includegraphics[width=8.5 cm]{fig8}
\caption{Time series of the observed trajectory (thin black), and the
slave trajectories coupled to observations with $D=10\%$ (thick
gray) and $40\%$ (thick black) noise with
$T_{\textrm{obs}}=2.0\times10^{-5} sec$. The noisy observations with
$D=40\%$ are also shown (stars).}
\label{fig:time_sr_nz}
\end{figure}
To study the effect of noise on synchronization, we added noise to the
$x$-component of observation with $T_{\textrm{obs}}=2.0\times10^{-5}
sec$, which is used in Sec.~\ref{subsec:obs_time}, and repeated the
same experiments. We have not performed any experiments with noisy
continuous time observations, since we are only considering
deterministic models.
In Fig.~\ref{fig:rms_nz} (a) and (b), the solid lines represent the
RMS errors in $x$-component and total error respectively when added
noise intensity is $D=2\%$. In this case, we got synchrony in the
range of $0.66<k_x<2.1$, and the model started desynchronizing around
$k_x=2.1$ as shown in Figs.~\ref{fig:rms_nz}. So the noisy
observations squeeze the k range further. With increase in $D$ overall
range remained almost same. In Figs.~\ref{fig:rms_nz} solid, dashed,
dotted, dashed-dotted, and gray lines are the RMS errors when
$D=10\%$, $20\%$, $30\%$ and $40\%$ respectively.
Effect of the noise is also clear from the time series
plots. Fig.~\ref{fig:time_sr_nz} shows the time series of the
observation (thin solid line) and slave trajectories for $D=10\%$ (thick gray)
and $40\%$ (thick solid line). They show that the deviation of the
slave trajectories from the observed one increases with increasing
noise level. But $z$-component deviate lesser than $x$ and $y$-component (not shown).
It is also observed that the RMS errors increases almost linearly
(not shown here) with increases in $D$.
\subsection{Prediction with truths and noisy observation}
\begin{figure}[t]
\includegraphics[width=8.5 cm]{fig9}
\caption{Prediction RMS errors in the $x$-component for different
$t_{fh}$ when $x$ was observed without noise with
$T_{\textrm{obs}}=2.0\times10^{-5}$ (top) and
$T_{\textrm{obs}}=5.0\times10^{-5}$ (bottom). Plots in the right
column are the zoomed version of the boxed regions of the left
plots. In each plot, the thick solid, gray, dotted, dash-dot, and
thin solid lines show the prediction RMS errors when $t_{fh}=
\frac{1}{10} t_\lambda$, $t_\lambda$, $5t_\lambda$, $10t_\lambda$
and $50t_\lambda$ respectively.}
\label{fig:rms_x_predict}
\end{figure}
\begin{figure}[t]
\includegraphics[width=8.5 cm]{fig10}
\caption{Prediction RMS errors in the $x$-component for different
$t_{fh}$ when $x$ was observed with noise levels $D=6\%$ (top) and
$D=10\%$ (bottom) with $T_{\textrm{obs}}=2.0\times10^{-5}$. Plots in
the right column are the zoomed version of the boxed regions of the
left plots. In each plot, the thick solid, gray, dotted, dash-dot,
and thin solid lines show the prediction RMS errors when $t_{fh}=
\frac{1}{10} t_\lambda$, $t_\lambda$, $5t_\lambda$, $10t_\lambda$
and $50t_\lambda$ respectively.}
\label{fig:predict_noise}
\end{figure}
\begin{figure}[t]
\includegraphics[width=8.5 cm]{fig11}
\caption{Differnce between the observed and the predicted
trajectories, when the coupling is switched off at $t = 10$
msec. Top panel (a) is for the case of observations with no noise,
and the bottom panels (b) and (c) are for the case with noisy
observations with intensity $D = 6\%$ and $D = 10\%$ respectively.}
\label{fig:predict_traj}
\end{figure}
In this section we consider the use of the slave model for
prediction. This is done in the following manner. We couple the
observations with the slave system through Eqs.~\eqref{eqn:sync} for a
long time period, in particular, until $t_{coup} = 10$ msec in our
numerical experiments. Then we switch off the coupling, i.e., use
Eqs.~\eqref{eqn:master} for the slave beyond this time. The RMS errors
described below are calculated over varying time periods, which are
called the forecast horizons $t_{fh}$. We present the results for
$x$-observations with $T_{\textrm{obs}}=2.0\times10^{-5}$ and
$T_{\textrm{obs}}=5.0\times10^{-5}$ since larger $\tobs$ do not show
any synchronization.
In Fig.~\ref{fig:rms_x_predict} the prediction RMS errors in
$x$-component has been shown for the range of $0<k_x<5$. The $y$- and
$z$- components follow the same qualitative trend. Thick-solid, gray, dotted, dash-dotted and thin solid lines
in Fig.~\ref{fig:rms_x_predict} are the RMS errors for
$t_{fh}=\frac{1}{10} t_\lambda, t_\lambda, 5t_\lambda, 10t_\lambda$,
and $50t_\lambda$ respectively, where boxed regions of
Fig.~\ref{fig:rms_x_predict} (a) has been zoomed in
Fig.~\ref{fig:rms_x_predict} (a$'$). We see that for
$t_{fh}=\frac{1}{10} t_\lambda, t_\lambda, 5t_\lambda, 10t_\lambda$
slave system predicts well, i.e., the RMS errors are small, but when
$t_{fh}>10t_\lambda$ prediction diverges from true
trajectory. Figs~\ref{fig:rms_x_predict} (a$'$) [thin solid line]
show that RMS error at $t_{fh}= 50t_\lambda$ is significantly
large. Fig.~\ref{fig:predict_traj} (a) show that slave model becomes
completely uncorrelated with the master around $t_{fh}= 40t_\lambda$.
Another important observations from our numerical experiments is that
the forecast horizon for which the slave shows small RMS error with
respect to the master is highly dependent on the state of the system
at which the coupling is turned off. This is because the uncertainty
growth rate is different on different parts of the
attractor.~\cite{pla:Ziehmann}
When noisy observations are coupled with the slave system, the
divergence of the slave from the master is rapid once the coupling is
switched off. Fig.~\ref{fig:predict_noise} shows errors in
$x$-components when noise with intensity $D=6\%$ (top panels) and
$10\%$ (bottom panels) is added to the observations. We have seen that
for $t_{fh} \le \frac{1}{5}t_\lambda$ (not shown in figure), there is
a range of coupling constants for which the prediction RMS error is
small enough to give good prediction. For larger $t_{fh}$, the slave
model become uncorrelated with the master and predictions cannot be
obtained. This is the case for both cases of the observations with
$6\%$ and $10\%$ noise. Divergence is also clear from the difference
of $x_{obs}-x$ time series of with $6\%$ of noise
[Fig.~\ref{fig:predict_traj} (b)] and $10\%$ of noise
[Fig.~\ref{fig:predict_traj} (c)] respectively.
\section{Conclusion}
\label{sec:conclude}
This paper is devoted to the study of discrete-time unidirectional
synchronization, with specific emphasis on aspects which are of
relevance to data assimilation problems. In particular we consider the
observations of a system, in this case the chaotic Chua system of
Eqs.~\eqref{eqn:master}, to be the master coupled unidirectionally to
the slave system of Eqs.~\eqref{eqn:sync} and examine how the slave
system synchronizes with the master as we change various
characteristics of the observations. The main discussion is focused
on studying the effects of changing observational period, i.e. the
time between the observations $\tobs$, and observational noise levels
$D$. Some of the main results as well as directions for future
research which we are currently pursuing are discussed below.
We obtained synchronization with continuous time observations of the
$x$- and $y$-components of the system, but not with $z$-component
observations. When discrete observations with finite
$T_{\textrm{obs}}$ were used, Chua model showed synchrony for a finite
range of coupling constants, and this range decreases with increase in
$T_{\textrm{obs}}$ and noisy observations. The study also shows that
the system shows synchrony only when the observations were taken with
$T_{\textrm{obs}}$ much less than $t_{\lambda}$ (almost
$\frac{1}{7}t_{\lambda}$) while for larger $\tobs$ there is no
synchronization between the master and the slave.
In many cases of practical interest in earth sciences, data
assimilation is used as a tool to improve prediction of the observed
as well as unobserved components of the system. With this in mind, we
also study the prediction errors over different forecast horizons
after the coupling between the observations and the slave is turned
off. We see that in the cases when synchronization errors are small,
the prediction errors even for forecast horizons far beyond the
Lyapunov timescale are small. This happens only for the case of
observations which are not noisy. When using observations even with a
very small noise, the slightly higher synchronization errors coupled
with chaotic nature of the system lead to loss of predictive power
over forecast horizons which are a fraction of the Lyapunov timescale.
These results indicates that the association of the Lyapunov exponent
with predictability horizons has limited use.\cite{pla:Ziehmann} These
results also show the limitations of discrete-time synchronization
with noisy observations and the importance of the study of other data
assimilation techniques which systematically take into account the
noise in the observations.
In synchronization choice of coupling constant $k$ is merely ad hoc
and depends on a particular application. In case of data assimilation
using similar techniques, e.g. nudging\cite{ngp:aurox}, the choice of
$k$ can be interpreted in terms of variational approach. In
particular, Ref.~\onlinecite{ngp:aurox} shows, using the variational
approach, that the optimal coupling constant is inversely proportional
to observational noise covariance: $k \sim R^{-1}$, where $R$ is the
covariance matrix of observational errors. As the range of $k$ on
which model shows synchrony depends mainly on observed component and
noise, we are investigating the possibility that a similar formulation
may also be possible for synchronization to choose the appropriate
$k$. We will also be comparing the synchronization based methods with
data assimilation techniques such as the ensemble Kalman
filter\cite{book:evensen} and Bayesian methods.\cite{physd:apte07}
Another ongoing investigation is to study the performance of these
methods in the chaotic Chua model using real data from the
circuit\cite{pre:singla}. This study will give us an insight about the
choice of coupling constants and observational time period, the
effects of observational noise as well as of the errors in the
numerical model.
\section*{Acknowledgments}
AA and MN would like to thank the Indian National Centre for Ocean
Information Services for financial support through Project
No. INCOIS/93/2007.
\nocite{*}
\bibliographystyle{aipnum4-1}
|
\section{Introduction}
Hierarchical Bayesian models are a mainstay of the machine learning
and statistics communities. Exact posterior inference in such models
is rarely tractable, however, and so researchers and practitioners
must usually resort to approximate statistical inference methods.
Deterministic approximate inference algorithms (for example, those
reviewed by \citet{Wainwright:2008}) can be efficient, but introduce
bias and can be difficult to apply to some models. Rather than
computing a deterministic approximation to a target posterior (or
other) distribution, Markov chain Monte Carlo (MCMC) methods offer
schemes for drawing a series of correlated samples that will converge
in distribution to the target distribution \citep{Neal:1993}. MCMC
methods are sometimes less efficient than their deterministic
counterparts, but are more generally applicable and are asymptotically
unbiased.
Not all MCMC algorithms are created equal. For complicated models with
many parameters, simple methods such as random-walk Metropolis
\citep{Metropolis:1953} and Gibbs sampling \citep{Geman:1984} may
require an unacceptably long time to converge to the target
distribution. This is in large part due to the tendency of these
methods to explore parameter space via inefficient random walks
\citep{Neal:1993}. When model parameters are continuous rather than
discrete, Hamiltonian Monte Carlo (HMC), also known as hybrid Monte
Carlo, is able to suppress such random walk behavior by means of a
clever auxiliary variable scheme that transforms the problem of
sampling from a target distribution into the problem of simulating
Hamiltonian dynamics \citep{Neal:2011}. The cost of HMC per
independent sample from a target distribution of dimension $D$ is
roughly $O(D^{5/4})$, which stands in sharp contrast with the $O(D^2)$
cost of random-walk Metropolis \citep{Creutz:1988}.
HMC's increased efficiency comes at a price. First, HMC requires the
gradient of the log-posterior. Computing the gradient for a complex
model is at best tedious and at worst impossible, but this requirement
can be made less onerous by using automatic differentiation
\citep{Griewank:2008}. Second, HMC requires that the user specify at
least two parameters: a step size $\epsilon$ and a number of steps $L$
for which to run a simulated Hamiltonian system. A poor choice of
either of these parameters will result in a dramatic drop in HMC's
efficiency. Methods from the adaptive MCMC literature (see
\citet{Andrieu:2008} for a review) can be used to tune $\epsilon$ on
the fly, but setting $L$ typically requires one or more costly tuning
runs, as well as the expertise to interpret the results of those
tuning runs. This hurdle limits the more widespread use of HMC, and
makes it challenging to incorporate HMC into a general-purpose
inference engine such as BUGS \citep{Gilks:1992}, JAGS
(http://mcmc-jags.sourceforge.net), Infer.NET \citep{infer.net},
HBC \citep{Daume:2007}, or PyMC \citep{Patil:2010}.
The main contribution of this paper is the No-U-Turn Sampler (NUTS),
an MCMC algorithm that closely resembles HMC, but eliminates the need
to choose the problematic number-of-steps parameter $L$. We also
provide a new dual averaging \citep{Nesterov:2009} scheme for
automatically tuning the step size parameter $\epsilon$ in both HMC
and NUTS, making it possible to run NUTS with no hand-tuning at
all. We will show that the tuning-free version of NUTS samples as
efficiently as (and sometimes more efficiently than) HMC, even
ignoring the cost of finding optimal tuning parameters for HMC. Thus,
NUTS brings the efficiency of HMC to users (and generic inference
systems) that are unable or disinclined to spend time tweaking an MCMC
algorithm.
\section{Hamiltonian Monte Carlo}
In Hamiltonian Monte Carlo (HMC) \citep{Neal:2011, Neal:1993,
Duane:1987}, we introduce an auxiliary momentum variable $r_d$ for
each model variable $\theta_d$. In the usual implementation, these
momentum variables are drawn independently from the standard normal
distribution, yielding the (unnormalized) joint density
\begin{equation}
\textstyle
p(\theta, r) \propto \exp\{\mathcal{L}(\theta) - \frac{1}{2}r \cdot r\},
\end{equation}
where $\mathcal{L}$ is the logarithm of the joint density of the variables of
interest $\theta$ (up to a normalizing constant) and $x\cdot y$
denotes the inner product of the vectors $x$ and $y$. We can
interpret this augmented model in physical terms as a fictitious
Hamiltonian system where $\theta$ denotes a particle's position in
$D$-dimensional space, $r_d$ denotes the momentum of that particle in
the $d$th dimension, $\mathcal{L}$ is a position-dependent negative potential
energy function, $\frac{1}{2}r \cdot r$ is the kinetic energy of the particle,
and $\log p(\theta, r)$ is the negative energy of the particle. We can
simulate the evolution over time of the Hamiltonian dynamics of this
system via the ``leapfrog'' integrator, which proceeds according to
the updates
\begin{equation}
\label{eq:leapfrog}
r^{t + \epsilon/2} = r^t + (\epsilon/2) \nabla_\theta \mathcal{L}(\theta^t);
\quad \theta^{t + \epsilon} = \theta^t + \epsilon r^{t+\epsilon/2};
\quad r^{t + \epsilon} = r^{t+\epsilon/2} + (\epsilon/2)\nabla_\theta
\mathcal{L}(\theta^{t+\epsilon}),
\end{equation}
where $r^t$ and $\theta^t$ denote the values of the momentum and
position variables $r$ and $\theta$ at time $t$ and $\nabla_\theta$
denotes the gradient with respect to $\theta$. Since the update for
each coordinate depends only on the other coordinates, the leapfrog
updates are volume-preserving---that is, the volume of a region
remains unchanged after mapping each point in that region to a new
point via the leapfrog integrator.
\begin{algorithm}[tb]
\caption{Hamiltonian Monte Carlo}
\label{alg:HMC}
\begin{algorithmic}
\STATE Given $\theta^0$, $\epsilon$, $L$, $\mathcal{L}, M$:
\FOR{$m=1$ to $M$}
\STATE Sample $r^0\sim\mathcal{N}(0, I)$.
\STATE Set $\theta^m\leftarrow \theta^{m-1}, \tilde\theta\leftarrow \theta^{m-1}, \tilde r\leftarrow r^0$.
\FOR{$i=1$ to $L$}
\STATE Set $\tilde\theta, \tilde r \leftarrow \mathrm{Leapfrog}(\tilde\theta, \tilde r, \epsilon)$.
\ENDFOR
\STATE With probability $\alpha = \min\left\{1, \frac{\exp\{\mathcal{L}(\tilde\theta)-\frac{1}{2}\tilde r\cdot\tilde r\}}
{\exp\{\mathcal{L}(\theta^{m-1}) - \frac{1}{2} r^{0}\cdot r^{0}\}}\right\},$
set $\theta^m\leftarrow\tilde \theta$, $r^m\leftarrow -\tilde r$.
\ENDFOR
\STATE
\STATE {\bf function} $\mathrm{Leapfrog}(\theta, r, \epsilon)$
\STATE Set $\tilde r \leftarrow r +
(\epsilon/2)\nabla_\theta\mathcal{L}(\theta)$.
\STATE Set $\tilde \theta \leftarrow \theta +
\epsilon \tilde r$.
\STATE Set $\tilde r \leftarrow \tilde r +
(\epsilon/2)\nabla_\theta\mathcal{L}(\tilde\theta)$.
\RETURN $\tilde \theta, \tilde r$.
\end{algorithmic}
\end{algorithm}
A standard procedure for drawing $M$ samples via Hamiltonian Monte
Carlo is described in Algorithm \ref{alg:HMC}. $I$ denotes the
identity matrix and $\mathcal{N}(\mu, \Sigma)$ denotes a multivariate normal
distribution with mean $\mu$ and covariance matrix $\Sigma$. For each
sample $m$, we first resample the momentum variables from a standard
multivariate normal, which can be inetpreted as a Gibbs sampling
update. We then apply $L$ leapfrog updates to the position and
momentum variables $\theta$ and $r$, generating a proposal
position-momentum pair $\tilde\theta, \tilde r$. We propose setting
$\theta^m=\tilde\theta$ and $r^m=-\tilde r$, and accept or reject this
proposal according to the Metropolis algorithm
\citep{Metropolis:1953}. This is a valid Metropolis proposal because
it is time-reversible and the leapfrog integrator is
volume-preserving; using an algorithm for simulating Hamiltonian
dynamics that did not preserve volume would seriously complicate the
computation of the Metropolis acceptance probability. The negation of
$\tilde r$ in the proposal is theoretically necessary to produce
time-reversibility, but can be omitted in practice if one is only
interested in sampling from $p(\theta)$. The algorithm's original
name, ``Hybrid Monte Carlo,'' refers to the hybrid approach of
alternating between updating $\theta$ and $r$ via Hamiltonian
simulation and updating $r$ via Gibbs sampling.
The term $\log \frac{p(\tilde\theta, \tilde r)}{p(\theta, r)}$, on
which the acceptance probability $\alpha$ depends, is the negative
change in energy of the simulated Hamiltonian system from time 0 to
time $\epsilon L$. If we could simulate the Hamiltonian dynamics
exactly, then $\alpha$ would always be 1, since energy is conserved in
Hamiltonian systems. The error introduced by using a discrete-time
simulation depends on the step size parameter
$\epsilon$---specifically, the change in energy $|\log
\frac{p(\tilde\theta, \tilde r)}{p(\theta, r)}|$ is proportional to
$\epsilon^2$ for large $L$, or $\epsilon^3$ if $L=1$
\citep{Leimkuhler:2004}. In theory the error can grow without bound as
a function of $L$, but in practice it typically does not when using
the leapfrog discretization. This allows us to run HMC with many
leapfrog steps, generating proposals for $\theta$ that have high
probability of acceptance even though they are distant from the
previous sample.
The performance of HMC depends strongly on choosing suitable values
for $\epsilon$ and $L$.
If $\epsilon$ is too large, then the simulation will be inaccurate and
yield low acceptance rates. If $\epsilon$ is too small, then
computation will be wasted taking many small steps. If $L$ is too
small, then successive samples will be close to one another, resulting
in undesirable random walk behavior and slow mixing. If $L$ is too
large, then HMC will generate trajectories that loop back and retrace
their steps. This is doubly wasteful, since work is being done to
bring the proposal $\tilde \theta$ {\it closer} to the initial
position $\theta^{m-1}$. Worse, if $L$ is chosen so that the
parameters jump from one side of the space to the other each
iteration, then the Markov chain may not even be ergodic
\citep{Neal:2011}. More realistically, an unfortunate choice of $L$
may result in a chain that is ergodic but slow to move between regions
of low and high density.
\section{Eliminating the Need to Hand-Tune HMC}
HMC is a powerful algorithm, but its usefulness is limited by the need
to tune the step size parameter $\epsilon$
and number of steps $L$. Tuning these parameters for any particular
problem requires some expertise, and usually one or more preliminary
runs. Selecting $L$ is particularly problematic; it is difficult to
find a simple metric for when a trajectory is too short, too long, or
``just right,'' and so practitioners commonly rely on heuristics based
on autocorrelation statistics from preliminary runs \citep{Neal:2011}.
Below, we present the No-U-Turn Sampler (NUTS), an extension of HMC
that eliminates the need to specify a fixed value of $L$. In section
\ref{sec:sa} we present schemes for setting $\epsilon$ based on the
dual averaging algorithm of \citet{Nesterov:2009}.
\subsection{No-U-Turn Hamiltonian Monte Carlo}
\label{sec:NUTS}
Our first goal is to devise an MCMC sampler that retains HMC's ability
to suppress random walk behavior without the need to set the number
$L$ of leapfrog steps that the algorithm takes to generate a proposal.
We need some criterion to tell us when we have simulated the dynamics
for ``long enough,'' i.e., when running the simulation for more steps
would no longer increase the distance between the proposal
$\tilde\theta$ and the initial value of $\theta$. We use a convenient
criterion based on the dot product between $\tilde r$ (the current
momentum) and $\tilde \theta- \theta$ (the vector from our initial
position to our current position), which is the derivative with
respect to time (in the Hamiltonian system) of half the squared
distance between the initial position $\theta$ and the current
position $\tilde \theta$:
\begin{equation}
\label{eq:timederiv}
\frac{d}{dt} \frac{(\tilde\theta - \theta)\cdot(\tilde\theta - \theta)}{2}
= (\tilde\theta - \theta) \cdot \frac{d}{dt} (\tilde\theta - \theta)
= (\tilde\theta - \theta) \cdot \tilde r.
\end{equation}
In other words, if we were to run the simulation for an infinitesimal
amount of additional time, then this quantity is proportional to the
progress we would make away from our starting point $\theta$.
This suggests an algorithm in which one runs leapfrog steps until the
quantity in equation \ref{eq:timederiv} becomes less than 0; such an
approach would simulate the system's dynamics until the proposal
location $\tilde\theta$ started to move back towards
$\theta$. Unfortunately this algorithm does not guarantee time
reversibility, and is therefore not guaranteed to converge to the
correct distribution. NUTS overcomes this issue by means of a
recursive algorithm reminiscent of the doubling procedure devised by
\citet{Neal:2003} for slice sampling.
\begin{figure}[t]
\begin{center}
\centerline{\includegraphics[width=1\columnwidth]{doubling3}}
\end{center}
\vskip -0.3in
\caption{Example of building a binary tree via repeated doubling.
Each doubling proceeds by choosing a direction (forwards or
backwards in time) uniformly at random, then simulating Hamiltonian
dynamics for $2^j$ leapfrog steps in that direction, where $j$ is
the number of previous doublings (and the height of the binary
tree). The figures at top show a trajectory in two dimensions (with
corresponding binary tree in dashed lines) as it evolves over four
doublings, and the figures below show the evolution of the binary
tree. In this example, the directions chosen were forward (light
orange node), backward (yellow nodes), backward (blue nodes), and
forward (green nodes).}
\vskip -0.1in
\label{fig:doubling}
\end{figure}
\begin{figure}[t]
\begin{center}
\vskip -0.3in
\centerline{\includegraphics[width=1.125\columnwidth]{nuts_stopping3}}
\vskip -0.3in
\end{center}
\vskip -0.3in
\caption{Example of a trajectory generated during one iteration of
NUTS. The blue ellipse is a contour of the target distribution, the
black open circles are the positions $\theta$ traced out by the
leapfrog integrator and associated with elements of the set of
visited states $\mathcal{B}$, the black solid circle is the starting
position, the red solid circles are positions associated with states
that must be excluded from the set $\mathcal{C}$ of possible next samples
because their joint probability is below the slice variable $u$, and
the positions with a red ``x'' through them correspond to states
that must be excluded from $\mathcal{C}$ to satisfy detailed balance. The
blue arrow is the vector from the positions associated with the
leftmost to the rightmost leaf nodes in the rightmost height-3
subtree, and the magenta arrow is the (normalized) momentum vector
at the final state in the trajectory. The doubling process stops
here, since the blue and magenta arrows make an angle of more than
90 degrees. The crossed-out nodes with a red ``x'' are in the right
half-tree, and must be ignored when choosing the next sample.}
\vskip -0.1in
\label{fig:stopping}
\end{figure}
NUTS begins by introducing a slice variable $u$ with conditional
distribution $p(u|\theta,
r)=\dunif(u;[0,\exp\{\mathcal{L}(\theta)-\frac{1}{2}r\cdot r\}])$, which
renders the conditional distribution
$p(\theta,r|u)=\dunif(\theta,r;\{\theta',
r'|\exp\{\mathcal{L}(\theta)-\frac{1}{2}r\cdot r\}\ge u\})$.
This slice sampling step is not strictly necessary, but it simplifies
both the derivation and the implementation of NUTS. In addition to
being more complicated, the analogous algorithm that eliminates the
slice variable seems empirically to be slightly less efficient than
the algorithm presented in this paper.
At a high level, after resampling $u|\theta, r$, NUTS uses the
leapfrog integrator to trace out a path forwards and backwards in
fictitious time, first running forwards or backwards 1 step, then
forwards or backwards 2 steps, then forwards or backwards 4 steps,
etc. This doubling process implicitly builds a balanced binary tree
whose leaf nodes correspond to position-momentum states, as
illustrated in Figure \ref{fig:doubling}. The doubling is halted when
the subtrajectory from the leftmost to the rightmost nodes of any
balanced subtree of the overall binary tree starts to double back on
itself (i.e., the fictional particle starts to make a ``U-turn'').
At this point NUTS stops the simulation and samples from among the set
of points computed during the simulation, taking care to preserve
detailed balance. Figure \ref{fig:stopping} illustrates an example of
a trajectory computed during an iteration of NUTS.
Pseudocode implementing a efficient version of NUTS is provided in
Algorithm \ref{alg:NUTS}. A detailed derivation follows below, along
with a simplified version of the algorithm that motivates and builds
intuition about Algorithm \ref{alg:NUTS} (but uses much more memory
and makes smaller jumps).
\subsubsection{Derivation of simplified NUTS algorithm}
NUTS further augments the model
$p(\theta,r)\propto\exp\{\mathcal{L}(\theta)-\frac{1}{2}r\cdot r\}$ with a
slice variable $u$ \citep{Neal:2003}. The joint probability of
$\theta, r,$ and $u$ is
\begin{equation}
\label{eq:augmented}
\textstyle
p(\theta, r, u) \propto \mathbb{I}[u\in[0,
\exp\{\mathcal{L}(\theta) - \frac{1}{2}r\cdot r
\}]],
\end{equation}
where $\mathbb{I}[\cdot]$ is 1 if the expression in brackets is true
and 0 if it is false. The (unnormalized) marginal probability of
$\theta$ and $r$ (integrating over $u$) is
\begin{equation}
\textstyle p(\theta, r) \propto \exp\{\mathcal{L}(\theta) - \frac{1}{2}r\cdot r\},
\end{equation}
as in standard HMC. The conditional probabilities $p(u|\theta, r)$ and
$p(\theta, r|u)$ are each uniform, so long as the condition
$u\le\exp\{\mathcal{L}(\theta)-\frac{1}{2}r\cdot r\}$ is satisfied.
We also add a finite set $\mathcal{C}$ of candidate position-momentum
states and another finite set $\mathcal{B}\supseteq\mathcal{C}$ to the
model. $\mathcal{B}$ will be the set of all position-momentum states that the
leapfrog integrator traces out during a given NUTS iteration, and
$\mathcal{C}$ will be the subset of those states to which we can transition
without violating detailed balance. $\mathcal{B}$ will be built up by randomly
taking forward and backward leapfrog steps, and $\mathcal{C}$ will selected
deterministically from $\mathcal{B}$. The random procedure for building $\mathcal{B}$
and $\mathcal{C}$ given $\theta,$ $r,$ $u,$ and $\epsilon$ will define a
conditional distribution $p(\mathcal{B}, \mathcal{C} | \theta, r, u, \epsilon)$, upon
which we place the following conditions:
\begin{enumerate}
\renewcommand{\labelenumi}{C.\arabic{enumi}:}
\item All elements of $\mathcal{C}$ must be chosen in a way that preserves
volume. That is, any deterministic transformations of $\theta, r$
used to add a state $\theta', r'$ to $\mathcal{C}$ must
have a Jacobian with
unit determinant.
\item $p((\theta, r)\in \mathcal{C} | \theta, r, u, \epsilon)=1$.
\item $p(u \le \exp\{\mathcal{L}(\theta') - \frac{1}{2}r'\cdot r' \} |
(\theta', r') \in \mathcal{C}) = 1$.
\item If $(\theta, r)\in \mathcal{C}$ and $(\theta', r')\in
\mathcal{C}$ then for any $\mathcal{B}$, $p(\mathcal{B}, \mathcal{C} |
\theta, r, u, \epsilon)= p(\mathcal{B}, \mathcal{C} | \theta', r',
u, \epsilon)$.
\end{enumerate}
C.1 ensures that $p(\theta, r|(\theta, r)\in\mathcal{C})\propto p(\theta, r)$,
i.e. if we restrict our attention to the elements of $\mathcal{C}$ then we can
treat the unnormalized probability density of a particular element of
$\mathcal{C}$ as an unnormalized probability mass. C.2 says that the current
state $\theta, r$ must be included in $\mathcal{C}$. C.3 requires that
any state in $\mathcal{C}$ be in the slice defined by $u$, i.e., that
any state $(\theta', r')\in\mathcal{C}$ must have equal (and positive)
conditional probability density $p(\theta', r'|u)$. C.4 states that
$\mathcal{B}$ and $\mathcal{C}$ must have equal probability of being
selected regardless of the current state $\theta, r$ as long as
$(\theta, r)\in\mathcal{C}$ (which it must be by C.2).
Deferring for the moment the question of how to construct and sample
from a distribution $p(\mathcal{B},\mathcal{C}|\theta, r, u, \epsilon)$ that satisfies
these conditions, we will now show that the the following procedure
leaves the joint distribution $p(\theta, r, u, \mathcal{B},
\mathcal{C}|\epsilon)$ invariant:
\begin{enumerate}
\item sample $r\sim\mathcal{N}(0, I)$,
\item sample $u\sim\dunif([0, \exp\{\mathcal{L}(\theta^t) -
\frac{1}{2}r\cdot r \}])$,
\item sample $\mathcal{B}, \mathcal{C}$ from their
conditional distribution $p(\mathcal{B},\mathcal{C} | \theta^t,
r, u, \epsilon)$,
\item sample $\theta^{t+1}, r\sim T(\theta^t, r, \mathcal{C})$,
\end{enumerate}
where $T(\theta', r' | \theta, r, \mathcal{C})$ is a
transition kernel that leaves the uniform distribution over
$\mathcal{C}$ invariant, i.e., $T$ must satisfy
\begin{equation}
\frac{1}{|\mathcal{C}|}\sum_{(\theta, r)\in\mathcal{C}} T(\theta',
r'|\theta, r, \mathcal{C}) = \frac{\mathbb{I}[(\theta', r')\in\mathcal{C}]}
{|\mathcal{C}|}
\end{equation}
for any $\theta', r'$. The notation $\theta^{t+1}, r\sim T(\theta^t,
r, \mathcal{C})$ denotes that we are resampling $r$ in a way that
depends on its current value.
Steps 1, 2, and 3 resample $r$, $u$, $\mathcal{B}$, and $\mathcal{C}$
from their conditional joint distribution given $\theta^t$, and
therefore together constitute a valid Gibbs sampling update. Step 4 is
valid because the joint distribution of $\theta$ and $r$ given $u,
\mathcal{B}, \mathcal{C}$, and $\epsilon$ is uniform on the elements
of $\mathcal{C}$:
\begin{equation}
\begin{split}
\label{eq:pthetargivenuc}
p(\theta, r | u, \mathcal{B}, \mathcal{C}, \epsilon)
& \propto p(\mathcal{B}, \mathcal{C} | \theta, r, u, \epsilon) p(\theta, r | u) \\
& \propto p(\mathcal{B}, \mathcal{C} | \theta, r, u, \epsilon)
\textstyle \mathbb{I}[u\le\exp\{\mathcal{L}(\theta) - \frac{1}{2}r\cdot r\}] \\
& \propto \mathbb{I}[(\theta, r)\in\mathcal{C}].
\end{split}
\end{equation}
Condition C.1 allows us to treat the unnormalized conditional density
$p(\theta, r | u)\propto\mathbb{I}[u\le\exp\{\mathcal{L}(\theta) -
\frac{1}{2}r\cdot r\}]$ as an unnormalized conditional probability
mass function. Conditions C.2 and C.4 ensure that $p(\mathcal{B}, \mathcal{C}|\theta,
r, u, \epsilon)\propto\mathbb{I}[(\theta, r)\in\mathcal{C}]$ because by C.2
$(\theta, r)$ must be in $\mathcal{C}$, and by C.4 for any $\mathcal{B}, \mathcal{C}$ pair
$p(\mathcal{B}, \mathcal{C}|\theta, r, u, \epsilon)$ is constant as a function of
$\theta$ and $r$ as long as $(\theta, r)\in\mathcal{C}$. Condition C.3 ensures
that $(\theta, r)\in\mathcal{C} \Rightarrow u\le\exp\{\mathcal{L}(\theta) -
\frac{1}{2}r\cdot r\}$ (so the $p(\theta, r | u, \epsilon)$ term is
redundant).
Thus, equation \ref{eq:pthetargivenuc} implies that the joint
distribution of $\theta$ and $r$ given $u$ and $\mathcal{C}$ is
uniform on the elements of $\mathcal{C}$, and we are free to choose a
new $\theta^{t+1}, r^{t+1}$ from any transition kernel that leaves
this uniform distribution on $\mathcal{C}$ invariant.
We now turn our attention to the specific form for $p(\mathcal{B}, \mathcal{C} |
\theta, r, u, \epsilon)$ used by NUTS. Conceptually, the generative
process for building $\mathcal{B}$ proceeds by repeatedly doubling the size of
a binary tree whose leaves correspond to position-momentum
states. These states will constitute the elements of $\mathcal{B}$. The
initial tree has a single node corresponding to the initial
state. Doubling proceeds by choosing a random direction
$v_j\sim\dunif(\{-1,1\})$ and taking $2^j$ leapfrog steps of size
$v_j\epsilon$ (i.e., forwards in fictional time if $v_j=1$ and
backwards in fictional time if $v_j=-1$), where $j$ is the current
height of the tree. (The initial single-node tree is defined to have
height 0.) For example, if $v_j=1$, the left half of the new tree is
the old tree and the right half of the new tree is a balanced binary
tree of height $j$ whose leaf nodes correspond to the $2^j$
position-momentum states visited by the new leapfrog trajectory. This
doubling process is illustrated in Figure \ref{fig:doubling}. Given
the initial state $\theta, r$ and the step size $\epsilon$, there
are $2^j$ possible trees of height $j$ that can be built according to
this procedure, each of which is equally likely. Conversely, the
probability of reconstructing a particular tree of height $j$ starting
from any leaf node of that tree is $2^{-j}$ regardless of which leaf
node we start from.
We cannot keep expanding the tree forever, of course. We want to
continue expanding $\mathcal{B}$ until one end of the trajectory we are
simulating makes a ``U-turn'' and begins to loop back towards another
position on the trajectory. At that point continuing the simulation is
likely to be wasteful, since the trajectory will retrace its steps and
visit locations in parameter space close to those we have already
visited. We also want to stop expanding $\mathcal{B}$ if the error in the
simulation becomes extremely large, indicating that any states
discovered by continuing the simulation longer are likely to have
astronomically low probability. (This may happen if we use a step size
$\epsilon$ that is too large, or if the target distribution includes
hard constraints that make the log-density $\mathcal{L}$ go to $-\infty$ in
some regions.)
The second rule is easy to formalize---we simply stop doubling if the
tree includes a leaf node whose state $\theta, r$ satisfies
\begin{equation}
\label{eq:stoperror}
\mathcal{L}(\theta)-\frac{1}{2}r\cdot r - \log u < -\Delta_\mathrm{max}
\end{equation}
for some nonnegative $\Delta_\mathrm{max}$. We recommend setting
$\Delta_\mathrm{max}$ to a large value like 1000 so that it does not
interfere with the algorithm so long as the simulation is even
moderately accurate.
We must be careful when defining the first rule so that we can build a
sampler that neither violates detailed balance nor introduces
excessive computational overhead. To determine whether to stop
doubling the tree at height $j$, NUTS considers the $2^j-1$ balanced
binary subtrees of the height-$j$ tree that have height greater than
0. NUTS stops the doubling process when for one of these subtrees the
states $\theta^-, r^-$ and $\theta^+, r^+$ associated with the
leftmost and rightmost leaves of that subtree satisfies
\begin{equation}
\label{eq:stopangle}
(\theta^+-\theta^-)\cdot r^- < 0 \quad\mathrm{or}\quad
(\theta^+-\theta^-)\cdot r^+ < 0.
\end{equation}
That is, we stop if continuing the simulation an infinitesimal amount
either forward or backward in time would reduce the distance between
the position vectors $\theta^-$ and $\theta^+$. Evaluating the
condition in equation \ref{eq:stopangle} for each balanced subtree of
a tree of height $j$ requires $2^{j+1}-2$ inner products, which is
comparable to the number of inner products required by the $2^j-1$
leapfrog steps needed to compute the trajectory. Except for very
simple models with very little data, the cost of these inner products
is usually negligible compared to the cost of computing gradients.
This doubling process defines a distribution $p(\mathcal{B}|\theta, r, u,
\epsilon)$. We now define a deterministic process for deciding which
elements of $\mathcal{B}$ go in the candidate set $\mathcal{C}$, taking care to
satisfy conditions C.1--C.4 on $p(\mathcal{B},\mathcal{C}|\theta, r, u, \epsilon)$
laid out above. C.1 is automatically satisfied, since leapfrog steps
are volume preserving and any element of $\mathcal{C}$ must be within some
number of leapfrog steps of every other element of $\mathcal{C}$. C.2 is
satisfied as long as we include the initial state $\theta, r$ in
$\mathcal{C}$, and C.3 is satisfied if we exclude any element $\theta',
r'$ of $\mathcal{B}$ for which $\exp\{\mathcal{L}(\theta')-\frac{1}{2}r'\cdot r'\} <
u$. To satisfy condition C.4, we must ensure that $p(\mathcal{B},\mathcal{C}|\theta,
r, u, \epsilon)=p(\mathcal{B},\mathcal{C}|\theta', r', u, \epsilon)$ for any
$(\theta', r')\in\mathcal{C}$. For any start state $(\theta', r')\in\mathcal{B}$,
there is at most one series of directions $\{v_0,\ldots,v_j\}$ for
which the doubling process will reproduce $\mathcal{B}$, so as long as we
choose $\mathcal{C}$ deterministically given $\mathcal{B}$ either $p(\mathcal{B},\mathcal{C}|\theta',
r', u, \epsilon)=2^{-j}=p(\mathcal{B},\mathcal{C}|\theta, r, u, \epsilon)$ or
$p(\mathcal{B},\mathcal{C}|\theta', r', u, \epsilon)=0$. Thus, condition C.4 will be
satisfied as long as we exclude from $\mathcal{C}$ any state $\theta', r'$
that could not have generated $\mathcal{B}$. The only way such a state can
arise is if starting from $\theta', r'$ results in the stopping
conditions in equations \ref{eq:stoperror} or \ref{eq:stopangle} being
satisfied before the entire tree has been built, causing the doubling
process to stop too early. There are two cases to consider:
\begin{enumerate}
\item The doubling procedure was stopped because either equation
\ref{eq:stoperror} or equation \ref{eq:stopangle} was satisfied by a
state or subtree added during the final doubling iteration. In this
case we must exclude from $\mathcal{C}$ any element of $\mathcal{B}$ that was added
during this final doubling iteration, since starting the doubling
process from one of these would lead to a stopping condition being
satisfied before the full tree corresponding to $\mathcal{B}$ has been
built.
\item The doubling procedure was stopped because equation
\ref{eq:stopangle} was satisfied for the leftmost and rightmost
leaves of the full tree corresponding to $\mathcal{B}$. In this case no
stopping condition was met by any state or subtree until $\mathcal{B}$ had
been completed, and condition C.4 is automatically satisfied.
\end{enumerate}
\begin{algorithm}[t]
\caption{Naive No-U-Turn Sampler}
\label{alg:naive-NUTS}
\begin{algorithmic}
\small
\STATE Given $\theta^0$, $\epsilon$, $\mathcal{L}$, $M$:
\FOR{$m=1$ to $M$}
\STATE Resample $r^0\sim\mathcal{N}(0, I)$.
\STATE Resample $u \sim \dunif([0, \exp\{\mathcal{L}(\theta^{m-1} - \frac{1}{2}r^0\cdot r^0\}])$
\STATE Initialize $\theta^-= \theta^{m-1}$, $\theta^+= \theta^{m-1}$,
$r^-= r^0$, $r^+= r^0$, $j= 0$,
$\mathcal{C}= \{(\theta^{m-1}, r^0)\}, s= 1$.
\WHILE {$s=1$}
\STATE Choose a direction $v_j \sim \dunif(\{-1, 1\})$.
\IF {$v_j = -1$}
\STATE $\theta^-, r^-, -, -, \mathcal{C}', s' \leftarrow
\mathrm{BuildTree}(\theta^-, r^-, u, v_j, j, \epsilon)$.
\ELSE
\STATE $-, -, \theta^+, r^+, \mathcal{C}', s' \leftarrow
\mathrm{BuildTree}(\theta^+, r^+, u, v_j, j, \epsilon)$.
\ENDIF
\IF {$s' = 1$}
\STATE $\mathcal{C} \leftarrow \mathcal{C} \cup \mathcal{C}'$.
\ENDIF
\STATE $s\leftarrow s' \mathbb{I}[(\theta^+-\theta^-)\cdot r^- \ge 0]
\mathbb{I}[(\theta^+-\theta^-)\cdot r^+ \ge 0]$.
\STATE $j\leftarrow j+1$.
\ENDWHILE
\STATE Sample $\theta^{m}, r$ uniformly at random from
$\mathcal{C}$.
\ENDFOR
\STATE
\STATE {\bf function} $\mathrm{BuildTree}(\theta, r, u, v, j,
\epsilon)$
\IF {$j=0$}
\STATE {\it Base case---take one leapfrog step in the direction $v$.}
\STATE $\theta', r'\leftarrow\mathrm{Leapfrog}(\theta, r, v\epsilon)$.
\STATE $\mathcal{C}' \leftarrow \left\{\begin{array}{ll}
\{(\theta', r')\} & \mbox{if $u\le\exp\{\mathcal{L}(\theta')-\frac{1}{2}r'\cdot r'\}$}
\\ \emptyset & \mbox{else} \end{array} \right.$
\STATE $s' \leftarrow \mathbb{I}[u < \exp\{\Delta_\mathrm{max} + \mathcal{L}(\theta')-\frac{1}{2}r'\cdot r'\}]$.
\STATE {\bf return} $\theta', r', \theta', r', \mathcal{C}', s'$.
\ELSE
\STATE {\it Recursion---build the left and right subtrees.}
\STATE $\theta^-, r^-, \theta^+, r^+, \mathcal{C}', s' \leftarrow
\mathrm{BuildTree}(\theta, r, u, v, j-1, \epsilon)$.
\IF {$v = -1$}
\STATE $\theta^-, r^-, -, -, \mathcal{C}'', s'' \leftarrow
\mathrm{BuildTree}(\theta^-, r^-, u, v, j-1, \epsilon)$.
\ELSE
\STATE $-, -, \theta^+, r^+, \mathcal{C}'', s'' \leftarrow
\mathrm{BuildTree}(\theta^+, r^+, u, v, j-1, \epsilon)$.
\ENDIF
\STATE $s' \leftarrow s's'' \mathbb{I}[(\theta^+-\theta^-)\cdot r^- \ge 0]
\mathbb{I}[(\theta^+-\theta^-)\cdot r^+ \ge 0]$.
\STATE $\mathcal{C}' \leftarrow \mathcal{C}' \cup \mathcal{C}''$.
\STATE {\bf return} $\theta^-, r^-, \theta^+, r^+, \mathcal{C}', s'$.
\ENDIF
\end{algorithmic}
\end{algorithm}
Algorithm \ref{alg:naive-NUTS} shows how to construct $\mathcal{C}$
incrementally while building $\mathcal{B}$. After resampling the initial
momentum and slice variables, it uses a recursive procedure resembling
a depth-first search that eliminates the need to explicitly store the
tree used by the doubling procedure. The $\mathrm{BuildTree}()$ function
takes as input an initial position $\theta$ and momentum $r$, a slice
variable $u$, a direction $v\in\{-1,1\}$, a depth $j$, and a step size
$\epsilon$. It takes $2^j$ leapfrog steps of size $v\epsilon$
(i.e. forwards in time if $v=1$ and backwards in time if $v=-1$), and
returns
\begin{enumerate}
\item the backwardmost and forwardmost position-momentum states
$\theta^-, r^-$ and $\theta^+, r^+$ among the $2^j$ new states
visited;
\item a set $\mathcal{C}'$ of position-momentum states containing each
newly visited state $\theta', r'$ for which
$\exp\{\mathcal{L}(\theta')-\frac{1}{2}r'\cdot r'\}>u$; and
\item an indicator variable $s$; $s=0$ indicates that a stopping
criterion was met by some state or subtree of the subtree
corresponding to the $2^j$ new states visited by
$\mathrm{BuildTree}()$.
\end{enumerate}
At the top level, NUTS repeatedly calls $\mathrm{BuildTree}()$ to
double the number of points that have been considered until either
$\mathrm{BuildTree}()$ returns $s=0$ (in which case doubling stops and
the new set $\mathcal{C}'$ that was just returned must be ignored) or equation
\ref{eq:stopangle} is satisfied for the new backwardmost and
forwardmost position-momentum states $\theta^-, r^-$ and $\theta^+,
r^+$ yet considered (in which case doubling stops but we can use the
new set $\mathcal{C}'$). Finally, we select the next position and momentum
$\theta^{m}, r$ uniformly at random from $\mathcal{C}$, the union of
all of the valid sets $\mathcal{C}'$ that have been returned, which clearly
leaves the uniform distribution over $\mathcal{C}$ invariant.
To summarize, Algorithm \ref{alg:naive-NUTS} defines a transition
kernel that leaves $p(\theta, r, u, \mathcal{B}, \mathcal{C}|\epsilon)$ invariant, and
therefore leaves the target distribution
$p(\theta)\propto\exp\{\mathcal{L}(\theta)\}$ invariant. It does so by
resampling the momentum and slice variables $r$ and $u$, simulating a
Hamiltonian trajectory forwards and backwards in time until that
trajectory either begins retracing its steps or encounters a state
with very low probability, carefully selecting a subset $\mathcal{C}$ of the
states encountered on that trajectory that lie within the slice
defined by the slice variable $u$, and finally choosing the next
position and momentum variables $\theta^{m}$ and $r$ uniformly at
random from $\mathcal{C}$. Figure \ref{fig:stopping} shows an example of a
trajectory generated by an iteration of NUTS where equation
\ref{eq:stopangle} is satisfied by the height-3 subtree at the end of
the trajectory. Below, we will introduce some improvements to
algorithm \ref{alg:naive-NUTS} that boost the algorithm's memory
efficiency and allow it to make larger jumps on average.
\subsubsection{Efficient NUTS}
Algorithm \ref{alg:naive-NUTS} requires $2^{j}-1$ evaluations of
$\mathcal{L}(\theta)$ and its gradient (where $j$ is the number of times
$\mathrm{BuildTree}()$ is called), and $O(2^j)$ additional operations to
determine when to stop doubling. In practice, for all but the smallest
problems the cost of computing $\mathcal{L}$ and its gradient still dominates
the overhead costs, so the computational cost of algorithm
\ref{alg:naive-NUTS} per leapfrog step is comparable to that of a
standard HMC algorithm. However, Algorithm \ref{alg:naive-NUTS} also
requires that we store $2^j$ position and momentum vectors, which may
require an unacceptably large amount of memory. Furthermore, there are
alternative transition kernels that satisfy detailed balance with
respect to the uniform distribution on $\mathcal{C}$ that produce larger jumps
on average than simple uniform sampling. Finally, if a stopping
criterion is satisfied in the middle of the final doubling iteration
then there is no point in wasting computation to build up a set $\mathcal{C}'$
that will never be used.
The third issue is easily addressed---if we break out of the recursion
as soon as we encounter a zero value for the stop indicator $s$ then
the correctness of the algorithm is unaffected and we save some
computation. We can address the second issue by using a more
sophisticated transition kernel to move from one state $(\theta,
r)\in\mathcal{C}$ to another state $(\theta', r')\in\mathcal{C}$ while leaving the
uniform distribution over $\mathcal{C}$ invariant. This kernel admits a
memory-efficient implementation that only requires that we store
$O(j)$ position and momentum vectors, rather than $O(2^j)$.
Consider the transition kernel
\begin{equation}
T(w'|w,\mathcal{C}) = \left\{ \begin{array}{ll}
\frac{\mathbb{I}[w'\in\mathcal{C}^\textrm{\tiny new}]}{|\mathcal{C}^\textrm{\tiny new}|} &
\mbox{if $|\mathcal{C}^\textrm{\tiny new}|>|\mathcal{C}^\textrm{\tiny old}|$}, \\
\frac{|\mathcal{C}^\textrm{\tiny new}|}{|\mathcal{C}^\textrm{\tiny old}|}\frac{\mathbb{I}[w'\in\mathcal{C}^\textrm{\tiny new}]}{|\mathcal{C}^\textrm{\tiny new}|}
+ \left(1-\frac{|\mathcal{C}^\textrm{\tiny new}|}{|\mathcal{C}^\textrm{\tiny old}|}\right)\mathbb{I}[w'=w] &
\mbox{if $|\mathcal{C}^\textrm{\tiny new}|\le|\mathcal{C}^\textrm{\tiny old}|$}
\end{array} \right. ,
\end{equation}
where $w$ and $w'$ are shorthands for position-momentum states
$(\theta, r)$, $\mathcal{C}^\textrm{\tiny new}$ and $\mathcal{C}^\textrm{\tiny old}$ are disjoint subsets of $\mathcal{C}$
such that $\mathcal{C}^\textrm{\tiny new}\cup\mathcal{C}^\textrm{\tiny old}=\mathcal{C}$, and $w\in\mathcal{C}^\textrm{\tiny old}$. In English,
$T$ proposes a move from $\mathcal{C}^\textrm{\tiny old}$ to a random state in $\mathcal{C}^\textrm{\tiny new}$
and accepts the move with probability
$\frac{|\mathcal{C}^\textrm{\tiny new}|}{|\mathcal{C}^\textrm{\tiny old}|}$. This is equivalent to a
Metropolis-Hastings kernel with proposal distribution $q(w',
{\mathcal{C}^\textrm{\tiny old}}', {\mathcal{C}^\textrm{\tiny new}}' |w, \mathcal{C}^\textrm{\tiny old},
\mathcal{C}^\textrm{\tiny new})\propto\mathbb{I}[w'\in\mathcal{C}^\textrm{\tiny new}]\mathbb{I}[{\mathcal{C}^\textrm{\tiny old}}'=\mathcal{C}^\textrm{\tiny new}]
\mathbb{I}[{\mathcal{C}^\textrm{\tiny new}}'=\mathcal{C}^\textrm{\tiny old}]$, and it is straightforward to show
that it satisfies detailed balance with respect to the uniform
distribution on $\mathcal{C}$, i.e.
\begin{equation}
p(w|\mathcal{C})T(w'|w,\mathcal{C}) = p(w'|\mathcal{C})T(w|w',\mathcal{C}),
\end{equation}
and that $T$ therefore leaves the uniform distribution over $\mathcal{C}$
invariant. If we let $\mathcal{C}^\textrm{\tiny new}$ be the (possibly empty) set of
elements added to $\mathcal{C}$ during the final iteration of the doubling
(i.e. those returned by the final call to $\mathrm{BuildTree}()$ and
$\mathcal{C}^\textrm{\tiny old}$ be the older elements of $\mathcal{C}$, then we can replace the
uniform sampling of $\mathcal{C}$ at the end of Algorithm \ref{alg:naive-NUTS}
with a draw from $T(\theta^t, r^t, \mathcal{C})$ and leave the uniform
distribution on $\mathcal{C}$ invariant. In fact, we can apply $T$ after {\it
every} doubling, proposing a move to each new half-tree in
turn. Doing so leaves the uniform distribution on each partially built
$\mathcal{C}$ invariant, and therefore does no harm to the invariance of the
uniform distribution on the fully built set $\mathcal{C}$. Repeatedly
applying $T$ in this way increases the probability that we will jump
to a state $\theta^{t+1}$ far from the initial state $\theta^t$;
considering the process in reverse, it is as though we first tried to
jump to the other side of $\mathcal{C}$, then if that failed tried to make a
more modest jump, and so on. This transition kernel is thus akin to
delayed-rejection MCMC methods \citep{Tierney:1999}, but in this
setting we can avoid the usual costs associated with evaluating new
proposals.
The transition kernel above still requires that we be able to sample
uniformly from the set $\mathcal{C}'$ returned by $\mathrm{BuildTree}()$,
which may contain as many as $2^{j-1}$ elements. In fact, we can
sample from $\mathcal{C}'$ without maintaining the full set $\mathcal{C}'$ in memory
by exploiting the binary tree structure in Figure
\ref{fig:doubling}. Consider a subtree of the tree explored in a call
to $\mathrm{BuildTree}()$, and let $\mathcal{C}_\mathrm{subtree}$ denote the
set of its leaf states that are in $\mathcal{C}'$: we can factorize the
probability that a state $(\theta, r)\in\mathcal{C}_\mathrm{subtree}$ will
be chosen uniformly at random from $\mathcal{C}'$ as
\begin{gather}
p(\theta, r|\mathcal{C}') = \frac{1}{|\mathcal{C}'|} =
\frac{|\mathcal{C}_\mathrm{subtree}|}{|\mathcal{C}'|}\frac{1}{|\mathcal{C}_\mathrm{subtree}|}
\\ \nonumber
= p((\theta, r)\in\mathcal{C}_\mathrm{subtree} | \mathcal{C})
p(\theta, r |
(\theta, r)\in\mathcal{C}_\mathrm{subtree}, \mathcal{C}).
\end{gather}
That is, $p(\theta, r|\mathcal{C}')$ is the product of the probability of
choosing some node from the subtree multiplied by the probability of
choosing $\theta, r$ uniformly at random from
$\mathcal{C}_\mathrm{subtree}$. We use this observation to sample from $\mathcal{C}'$
incrementally as we build up the tree. Each subtree above the bottom
layer is built of two smaller subtrees. For each of these smaller
subtrees, we sample a $\theta, r$ pair from $p(\theta, r |
(\theta, r)\in\mathcal{C}_\mathrm{subtree})$ to represent that subtree. We
then choose between these two pairs, giving the pair representing each
subtree weight proportional to how many elements of $\mathcal{C}'$ are in that
subtree. This continues until we have completed the subtree associated
with $\mathcal{C}'$ and we have returned a sample $\theta'$ from $\mathcal{C}'$ and an
integer weight $n'$ encoding the size of $\mathcal{C}'$, which is all we need
to apply $T$. This procedure only requires that we store $O(j)$
position and momentum vectors in memory, rather than $O(2^j)$, and
requires that we generate $O(2^j)$ extra random numbers (a cost that
again is usually very small compared with the $2^{j}-1$ gradient
computations needed to run the leapfrog algorithm).
Algorithm \ref{alg:NUTS} implements all of the above improvements in
pseudocode. Matlab code implementing the algorithm is also available
at \url{http://www.cs.princeton.edu/~mdhoffma}, and a C++
implementation will also be available as part of the
soon-to-be-released Stan inference package.
\subsection{Adaptively Tuning $\epsilon$}
\label{sec:sa}
Having addressed the issue of how to choose the number of steps $L$,
we now turn our attention to the step size parameter $\epsilon$. To
set $\epsilon$ for both NUTS and HMC, we propose using stochastic
optimization with vanishing adaptation \citep{Andrieu:2008},
specifically an adaptation of the primal-dual algorithm of
\citet{Nesterov:2009}.
\begin{algorithm}[t!]
\caption{Efficient No-U-Turn Sampler}
\label{alg:NUTS}
\begin{algorithmic}
\small
\STATE Given $\theta^0$, $\epsilon$, $\mathcal{L}$, $M$:
\FOR{$m=1$ to $M$}
\STATE Resample $r^0\sim\mathcal{N}(0, I)$.
\STATE Resample $u \sim \dunif([0, \exp\{\mathcal{L}(\theta^{m-1} - \frac{1}{2}r^0\cdot r^0\}])$
\STATE Initialize $\theta^-=\theta^{m-1}$,
$\theta^+=\theta^{m-1}$, $r^-=r^0, r^+=r^0, j=0,
\theta^{m}=\theta^{m-1}, n=1, s=1$.
\WHILE {$s=1$}
\STATE Choose a direction $v_j \sim \dunif(\{-1, 1\})$.
\IF {$v_j = -1$}
\STATE $\theta^-, r^-, -, -, \theta', n', s' \leftarrow
\mathrm{BuildTree}(\theta^-, r^-, u, v_j, j, \epsilon)$.
\ELSE
\STATE $-, -, \theta^+, r^+, \theta', n', s' \leftarrow
\mathrm{BuildTree}(\theta^+, r^+, u, v_j, j, \epsilon)$.
\ENDIF
\IF {$s' = 1$}
\STATE With probability $\min\{1, \frac{n'}{n}\}$, set
$\theta^{m}\leftarrow \theta'$.
\ENDIF
\STATE $n\leftarrow n + n'$.
\STATE $s\leftarrow s' \mathbb{I}[(\theta^+-\theta^-)\cdot r^- \ge 0]
\mathbb{I}[(\theta^+-\theta^-)\cdot r^+ \ge 0]$.
\STATE $j\leftarrow j+1$.
\ENDWHILE
\ENDFOR
\STATE
\STATE {\bf function} $\mathrm{BuildTree}(\theta, r, u, v, j,
\epsilon)$
\IF {$j=0$}
\STATE {\it Base case---take one leapfrog step in the direction $v$.}
\STATE $\theta', r'\leftarrow\mathrm{Leapfrog}(\theta, r, v\epsilon)$.
\STATE $n' \leftarrow\mathbb{I}[u\le\exp\{\mathcal{L}(\theta')-\frac{1}{2}r'\cdot r'\}]$.
\STATE $s' \leftarrow \mathbb{I}[u < \exp\{\Delta_\mathrm{max} + \mathcal{L}(\theta')-\frac{1}{2}r'\cdot r'\}]$.
\STATE {\bf return} $\theta', r', \theta', r', \theta', n', s'$.
\ELSE
\STATE {\it Recursion---implicitly build the left and right subtrees.}
\STATE $\theta^-, r^-, \theta^+, r^+, \theta', n', s' \leftarrow
\mathrm{BuildTree}(\theta, r, u, v, j-1, \epsilon)$.
\IF {$s' = 1$}
\IF {$v = -1$}
\STATE $\theta^-, r^-, -, -, \theta'', n'', s'' \leftarrow
\mathrm{BuildTree}(\theta^-, r^-, u, v, j-1, \epsilon)$.
\ELSE
\STATE $-, -, \theta^+, r^+, \theta'', n'', s'' \leftarrow
\mathrm{BuildTree}(\theta^+, r^+, u, v, j-1, \epsilon)$.
\ENDIF
\STATE With probability $\frac{n''}{n'+n''}$,
set $\theta'\leftarrow\theta''$.
\STATE $s' \leftarrow s'' \mathbb{I}[(\theta^+-\theta^-)\cdot r^- \ge 0]
\mathbb{I}[(\theta^+-\theta^-)\cdot r^+ \ge 0]$
\STATE $n' \leftarrow n' + n''$
\ENDIF
\STATE {\bf return} $\theta^-, r^-, \theta^+, r^+, \theta', n', s'$.
\ENDIF
\end{algorithmic}
\end{algorithm}
Perhaps the most commonly used vanishing adaptation algorithm in MCMC
is the stochastic approximation method of
\citet{Robbins:1951}. Suppose we have a statistic $H_t$ that describes
some aspect of the behavior of an MCMC algorithm at iteration $t\ge
1$, and define its expectation $h( x)$ as
\begin{equation}
h( x)\equiv\mathbb{E}_t[H_t| x]\equiv
\lim_{T\rightarrow\infty}\frac{1}{T} \sum_{t=1}^T \mathbb{E}[H_t| x],
\end{equation}
where $ x\in\mathbb{R}$ is a tunable parameter to the MCMC
algorithm. For example, if $\alpha_t$ is the Metropolis acceptance
probability for iteration $t$, we might define $H_t =
\delta-\alpha_t$, where $\delta$ is the desired average acceptance
probability. If $h$ is a nondecreasing function of $ x$ and a few
other conditions such as boundedness of the iterates $x_t$ are met
(see \citet{Andrieu:2008} for details), the update
\begin{equation}
x_{t+1}\leftarrow x_t - \eta_t H_t
\end{equation}
is guaranteed to cause $h( x_t)$ to converge to 0 as long as the
step size schedule defined by $\eta_t$ satisfies the conditions
\begin{equation}
\label{eq:sasteps}
\sum_t\eta_t=\infty;\quad \sum_t\eta_t^2<\infty.
\end{equation}
These conditions are satisfied by schedules of the form $\eta_t\equiv
t^{-\kappa}$ for $\kappa\in(0.5, 1]$. As long as the per-iteration
impact of the adaptation goes to 0 (as it will if $\eta_t\equiv
t^{-\kappa}$ and $\kappa>0$) the asymptotic behavior of the sampler
is unchanged. That said, in practice $ x$ often gets ``close
enough'' to an optimal value well before the step size $\eta$ has
gotten close enough to 0 to avoid disturbing the Markov chain's
stationary distribution. A common practice is therefore to adapt any
tunable MCMC parameters during the burn-in phase, and freeze the
tunable parameters afterwards (e.g., \citep{Gelman:2004}).
\paragraph{Dual averaging:}
The optimal values of the parameters to an MCMC algorithm during the
burn-in phase and the stationary phase are often quite different.
Ideally those parameters would therefore adapt quickly as we shift
from the sampler's initial, transient regime to its stationary
regime. However, the diminishing step sizes of Robbins-Monro give
disproportionate weight to the {\it early} iterations, which is the
opposite of what we want.
Similar issues motivate the dual averaging scheme of
\citet{Nesterov:2009}, an algorithm for nonsmooth and stochastic
convex optimization. Since solving an unconstrained convex
optimization problem is equivalent to finding a zero of a
nondecreasing function (i.e., the (sub)gradient of the cost function),
it is straightforward to adapt dual averaging to the problem of MCMC
adaptation by replacing stochastic gradients with the statistics
$H_t$. Again assuming that we want to find a setting of a parameter $
x\in\mathbb{R}$ such that $h( x)\equiv\mathbb{E}_t[H_t| x]=0$, we can apply
the updates
\begin{equation}
\label{eq:daupdates}
x_{t+1}\leftarrow \mu -
\frac{\sqrt{t}}{\gamma}\frac{1}{t+t_0}\sum_{i=1}^t H_i;\quad
\bar x_{t+1}\leftarrow \eta_t x_{t+1} + (1-\eta_t)\bar x_t,
\end{equation}
where $\mu$ is a freely chosen point that the iterates $x_t$ are
shrunk towards, $\gamma>0$ is a free parameter that controls the
amount of shrinkage towards $ \mu$, $t_0\ge 0$ is a free parameter
that stabilizes the initial iterations of the algorithm, $\eta_t\equiv
t^{-\kappa}$ is a step size schedule obeying the conditions in
equation \ref{eq:sasteps}, and we define $\bar x_1 = x_1$. As in
Robbins-Monro, the per-iteration impact of these updates on $x$ goes
to 0 as $t$ goes to infinity. Specifically, for large $t$ we have
\begin{equation}
x_{t+1}-x_t = O(-H_t t^{-0.5}),
\end{equation}
which clearly goes to 0 as long as the statistic $H_t$ is bounded.
The sequence of averaged iterates $\bar x_t$ is guaranteed to converge
to a value such that $h(\bar x_t)$ converges to 0.
The update scheme in equation \ref{eq:daupdates} is slightly more
elaborate than the update scheme of \citet{Nesterov:2009}, which
implicitly has $t_0\equiv0$ and $\kappa\equiv 1$. Introducing these
parameters addresses issues that are more important in MCMC adaptation
than in more conventional stochastic convex optimization
settings. Setting $t_0>0$ improves the stability of the algorithm in
early iterations, which prevents us from wasting computation by trying
out extreme values. This is particularly important for NUTS, and for
HMC when simulation lengths are specified in terms of the overall
simulation length $\epsilon L$ instead of a fixed number of steps
$L$. In both of these cases, lower values of $\epsilon$ result in more
work being done per sample, so we want to avoid casually trying out
extremely low values of $\epsilon$. Setting the parameter $\kappa<1$
allows us to give higher weight to more recent iterates and more
quickly forget the iterates produced during the early burn-in stages.
The benefits of introducing these parameters are less apparent in the
settings originally considered by Nesterov, where the cost of a
stochastic gradient computation is assumed to be constant and the
stochastic gradients are assumed to be drawn i.i.d. given the
parameter $x$.
Allowing $t_0>0$ and $\kappa\in(0.5,1]$ does not affect the asymptotic
convergence of the dual averaging algorithm. For any $\kappa\in(0.5,
1]$, $\bar x_t$ will eventually converge to the same value
$\frac{1}{t}\sum_{i=1}^t x_t$. We can rewrite the term
$\frac{\sqrt{t}}{\gamma}\frac{1}{t+t_0}$ as
$\frac{t\sqrt{t}}{\gamma (t+t_0)}\frac{1}{t}$;
$\frac{t\sqrt{t}}{\gamma (t+t_0)}$ is still $O(\sqrt{t})$, which
is the only feature needed to guarantee convergence.
We used the values $\gamma=0.05, t_0=10,$ and $\kappa=0.75$ for all
our experiments. We arrived at these values by trying a few settings
for each parameter by hand with NUTS and HMC (with simulation lengths
specified in terms of $\epsilon L$) on the stochastic volatility model
described below and choosing a value for each parameter that seemed to
produce reasonable behavior. Better results might be obtained with
further tweaking, but these default parameters seem to work
consistently well for both NUTS and HMC for all of the models that we
tested. It is entirely possible that these parameter settings may not
work as well for other sampling algorithms or for $H$ statistics other
than the ones described below.
\paragraph{Setting $\epsilon$ in HMC:}
In HMC we want to find a value for the step size $\epsilon$ that is
neither too small (which would waste computation by taking needlessly
tiny steps) nor too large (which would waste computation by causing
high rejection rates). A standard approach is to tune $\epsilon$ so
that HMC's average Metropolis acceptance probability is equal to some
value $\delta$. Indeed, it has been shown that (under fairly strong
assumptions) the optimal value of $\epsilon$ for a given simulation
length $\epsilon L$ is the one that produces an average Metropolis
acceptance probability of approximately 0.65 \citep{Beskos:2010,
Neal:2011}. For HMC, we define a criterion
$h^{\mathrm{HMC}}(\epsilon)$ so that
\begin{equation}
H^{\mathrm{HMC}}_t \equiv \min\left\{1, \frac{p(\tilde \theta^t,
\tilde r^t)}{p(\theta^{t-1}, r^{t,0})}\right\};
\quad h^{\mathrm{HMC}}(\epsilon) \equiv \mathbb{E}_t[H_t^{\mathrm{HMC}}|\epsilon],
\end{equation}
where $\tilde \theta^t$ and $\tilde r^t$ are the proposed position and
momentum at the $t$th iteration of the Markov chain, $\theta^{t-1}$
and $r^{t,0}$ are the initial position and (resampled) momentum for
the $t$th iteration of the Markov chain, $H^{\mathrm{HMC}}_t$ is the
acceptance probability of this $t$th HMC proposal and
$h^{\mathrm{HMC}}$ is the expected average acceptance probability of
the chain in equilibrium for a fixed $\epsilon$. Assuming that
$h^{\mathrm{HMC}}$ is nonincreasing as a function of $\epsilon$, we
can apply the updates in equation \ref{eq:daupdates} with $H_t\equiv
\delta-H^{\mathrm{HMC}}_t$ and $x\equiv\log \epsilon$ to coerce
$h^{\mathrm{HMC}}=\delta$ for any $\delta\in(0,1)$.
\paragraph{Setting $\epsilon$ in NUTS:}
Since there is no single accept/reject step in NUTS we must define an
alternative statistic to Metropolis acceptance probability. For each
iteration we define the statistic $H^{\mathrm{NUTS}}_t$ and its
expectation when the chain has reached equilibrium as
\begin{equation}
H^{\mathrm{NUTS}}_t \equiv
\frac{1}{|\mathcal{B}_t^\mathrm{final}|}
\sum_{\theta, r\in\mathcal{B}_t^\mathrm{final}}
\min\left\{1,\frac{p(\theta,r)}{p(\theta^{t-1},r^{t,0})}\right\}
;\quad h^\mathrm{NUTS} \equiv \mathbb{E}_t[H_t^\mathrm{NUTS}],
\end{equation}
where $\mathcal{B}_t^\mathrm{final}$ is the set of all states explored during
the final doubling of iteration $t$ of the Markov chain and
$\theta^{t-1}$ and $r^{t,0}$ are the initial position and (resampled)
momentum for the $t$th iteration of the Markov
chain. $H^\mathrm{NUTS}$ can be understood as the average acceptance
probability that HMC would give to the position-momentum states
explored during the final doubling iteration. As above, assuming that
$H^\mathrm{NUTS}$ is nonincreasing in $\epsilon$, we can apply the
updates in equation \ref{eq:daupdates} with $H_t\equiv
\delta-H^\mathrm{NUTS}$ and $x\equiv\log \epsilon$ to coerce
$h^\mathrm{NUTS}=\delta$ for any $\delta\in(0,1)$.
\begin{algorithm}[tb]
\caption{Heuristic for choosing an initial value of $\epsilon$}
\label{alg:initialepsilon}
\begin{algorithmic}
\small
\STATE {\bf function} $\mathrm{FindReasonableEpsilon}(\theta)$
\STATE Initialize $\epsilon=1$, $r\sim\mathcal{N}(0,I)$.
\STATE Set $\theta', r'\leftarrow \mathrm{Leapfrog}(\theta, r, \epsilon)$.
\STATE $a\leftarrow 2\mathbb{I}\left[\frac{p(\theta', r')}{p(\theta, r)} > 0.5\right]-1.$
\WHILE {$\left(\frac{p(\theta', r')}{p(\theta, r)}\right)^a > 2^{-a}$}
\STATE $\epsilon\leftarrow 2^a \epsilon$.
\STATE Set $\theta', r'\leftarrow \mathrm{Leapfrog}(\theta, r, \epsilon)$.
\ENDWHILE
\RETURN $\epsilon$.
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[t!]
\caption{Hamiltonian Monte Carlo with Dual Averaging}
\label{alg:dacdhmc}
\begin{algorithmic}
\small
\STATE Given $\theta^0$, $\delta$, $\lambda$, $\mathcal{L}, M, M^\mathrm{adapt}$:
\STATE Set $\epsilon_0=\mathrm{FindReasonableEpsilon}(\theta), \mu=\log(10\epsilon_0), \bar\epsilon_0=1, \bar H_0=0,
\gamma=0.05, t_0=10, \kappa=0.75.$
\FOR{$m=1$ to $M$}
\STATE Reample $r^0\sim\mathcal{N}(0, I)$.
\STATE Set $\theta^m\leftarrow \theta^{m-1}, \tilde\theta\leftarrow \theta^{m-1}, \tilde r\leftarrow r^0,
L_m=\max\{1,\mathrm{Round}(\lambda/\epsilon_{m-1})\}$.
\FOR{$i=1$ to $L_m$}
\STATE Set $\tilde\theta, \tilde r \leftarrow \mathrm{Leapfrog}(\tilde\theta, \tilde r, \epsilon_{m-1})$.
\ENDFOR
\STATE With probability $\alpha = \min\left\{1, \frac{\exp\{\mathcal{L}(\tilde\theta)-\frac{1}{2}\tilde r\cdot\tilde r\}}
{\exp\{\mathcal{L}(\theta^{m-1}) - \frac{1}{2} r^{0}\cdot r^{0}\}}\right\},$
set $\theta^m\leftarrow\tilde \theta, r^m\leftarrow -\tilde r$.
\IF {$m \le M^\mathrm{adapt}$}
\STATE Set $\bar H_m = \left(1-\frac{1}{m+t_0}\right)\bar H_{m-1} + \frac{1}{m+t_0}(\delta-\alpha)$.
\STATE Set $\log\epsilon_{m} = \mu - \frac{\sqrt{m}}{\gamma}\bar H_m,
\log\bar\epsilon_{m} = m^{-\kappa}\log\epsilon_{m} + (1-m^{-\kappa})\log\bar\epsilon_{m-1}.$
\ELSE
\STATE Set $\epsilon_{m} = \bar\epsilon_{M^\mathrm{adapt}}$.
\ENDIF
\ENDFOR
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[]
\caption{No-U-Turn Sampler with Dual Averaging}
\label{alg:danuts}
\begin{algorithmic}
\footnotesize
\STATE Given $\theta^0$, $\delta$, $\mathcal{L}, M, M^\mathrm{adapt}$:
\STATE Set $\epsilon_0=\mathrm{FindReasonableEpsilon}(\theta), \mu=\log(10\epsilon_0), \bar\epsilon_0=1, \bar H_0=0,
\gamma=0.05, t_0=10, \kappa=0.75.$
\FOR{$m=1$ to $M$}
\STATE Sample $r^0\sim\mathcal{N}(0, I)$.
\STATE Resample $u \sim \dunif([0, \exp\{\mathcal{L}(\theta^{m-1} - \frac{1}{2}r^{0}\cdot r^{0}\}])$
\STATE Initialize $\theta^-=\theta^{m-1}$,
$\theta^+=\theta^{m-1}$, $r^-=r^{0}, r^+=r^{0}, j=0,
\theta^{m}=\theta^{m-1}, n=1, s=1$.
\WHILE {$s=1$}
\STATE Choose a direction $v_j \sim \dunif(\{-1, 1\})$.
\IF {$v_j = -1$}
\STATE $\theta^-, r^-, -, -, \theta', n', s', \alpha, n_\alpha \leftarrow
\mathrm{BuildTree}(\theta^-, r^-, u, v_j, j, \epsilon_{m-1} \theta^{m-1}, r^{0})$.
\ELSE
\STATE $-, -, \theta^+, r^+, \theta', n', s', \alpha, n_\alpha \leftarrow
\mathrm{BuildTree}(\theta^+, r^+, u, v_j, j, \epsilon_{m-1}, \theta^{m-1}, r^{0})$.
\ENDIF
\IF {$s' = 1$}
\STATE With probability $\min\{1, \frac{n'}{n}\}$, set
$\theta^{m}\leftarrow \theta'$.
\ENDIF
\STATE $n\leftarrow n + n'$.
\STATE $s\leftarrow s' \mathbb{I}[(\theta^+-\theta^-)\cdot r^- \ge 0]
\mathbb{I}[(\theta^+-\theta^-)\cdot r^+ \ge 0]$.
\STATE $j\leftarrow j+1$.
\ENDWHILE
\IF {$m \le M^\mathrm{adapt}$}
\STATE Set $\bar H_m = \left(1-\frac{1}{m+t_0}\right)\bar H_{m-1} + \frac{1}{m+t_0}(\delta-\frac{\alpha}{n_\alpha})$.
\STATE Set $\log\epsilon_{m} = \mu - \frac{\sqrt{m}}{\gamma}\bar H_m,
\log\bar\epsilon_{m} = m^{-\kappa}\log\epsilon_{m} + (1-m^{-\kappa})\log\bar\epsilon_{m-1}.$
\ELSE
\STATE Set $\epsilon_{m} = \bar\epsilon_{M^\mathrm{adapt}}$.
\ENDIF
\ENDFOR
\STATE
\STATE {\bf function} $\mathrm{BuildTree}(\theta, r, u, v, j,
\epsilon, \theta^0, r^0)$
\IF {$j=0$}
\STATE {\it Base case---take one leapfrog step in the direction $v$.}
\STATE $\theta', r'\leftarrow\mathrm{Leapfrog}(\theta, r, v\epsilon)$.
\STATE $n' \leftarrow\mathbb{I}[u\le\exp\{\mathcal{L}(\theta')-\frac{1}{2}r'\cdot r'\}]$.
\STATE $s' \leftarrow \mathbb{I}[u < \exp\{\Delta_\mathrm{max} + \mathcal{L}(\theta')-\frac{1}{2}r'\cdot r'\}]$.
\STATE {\bf return} $\theta', r', \theta', r', \theta', n', s',
\min\{1,\exp\{\mathcal{L}(\theta')-\frac{1}{2}r'\cdot r' - \mathcal{L}(\theta^0) + \frac{1}{2}r^0\cdot r^0\}\}, 1$.
\ELSE
\STATE {\it Recursion---implicitly build the left and right subtrees.}
\STATE $\theta^-, r^-, \theta^+, r^+, \theta', n', s', \alpha', n_\alpha' \leftarrow
\mathrm{BuildTree}(\theta, r, u, v, j-1, \epsilon, \theta^0, r^0)$.
\IF {$s' = 1$}
\IF {$v = -1$}
\STATE $\theta^-, r^-, -, -, \theta'', n'', s'', \alpha'', n_\alpha'' \leftarrow
\mathrm{BuildTree}(\theta^-, r^-, u, v, j-1, \epsilon, \theta^0, r^0)$.
\ELSE
\STATE $-, -, \theta^+, r^+, \theta'', n'', s'', \alpha'', n_\alpha'' \leftarrow
\mathrm{BuildTree}(\theta^+, r^+, u, v, j-1, \epsilon, \theta^0, r^0)$.
\ENDIF
\STATE With probability $\frac{n''}{n'+n''}$,
set $\theta'\leftarrow\theta''$.
\STATE Set $\alpha'\leftarrow\alpha' + \alpha''$, $n_\alpha'\leftarrow n_\alpha' + n_\alpha''$.
\STATE $s' \leftarrow s'' \mathbb{I}[(\theta^+-\theta^-)\cdot r^- \ge 0]
\mathbb{I}[(\theta^+-\theta^-)\cdot r^+ \ge 0]$
\STATE $n' \leftarrow n' + n''$
\ENDIF
\STATE {\bf return} $\theta^-, r^-, \theta^+, r^+, \theta', n', s', \alpha', n_\alpha'$.
\ENDIF
\end{algorithmic}
\end{algorithm}
\paragraph{Finding a good initial value of $\epsilon$:}
The dual averaging scheme outlined above should work for any initial
value $\epsilon_1$ and any setting of the shrinkage target
$\mu$. However, convergence will be faster if we start from a
reasonable setting of these parameters. We recommend choosing an
initial value $\epsilon_1$ according to the simple heuristic described
in Algorithm \ref{alg:initialepsilon}. In English, this heuristic
repeatedly doubles or halves the value of $\epsilon_1$ until the
acceptance probability of the Langevin proposal with step size
$\epsilon_1$ crosses 0.5. The resulting value of $\epsilon_1$ will
typically be small enough to produce reasonably accurate simulations
but large enough to avoid wasting large amounts of computation. We
recommend setting $\mu=\log(10\epsilon_1)$, since this gives the dual
averaging algorithm a preference for testing values of $\epsilon$ that
are larger than the initial value $\epsilon_1$. Large values of
$\epsilon$ cost less to evaluate than small values of $\epsilon$, and
so erring on the side of trying large values can save computation.
Algorithms \ref{alg:dacdhmc} and \ref{alg:danuts} show how to
implement HMC (with simulation length specified in terms of $\epsilon
L$ rather than $L$) and NUTS while incorporating the dual averaging
algorithm derived in this section, with the above initialization
scheme. Algorithm \ref{alg:dacdhmc} requires as input a target
simulation length $\lambda\approx\epsilon L$, a target mean acceptance
probability $\delta$, and a number of iterations $M^\mathrm{adapt}$
after which to stop the adaptation. Algorithm \ref{alg:danuts}
requires only a target mean acceptance probability $\delta$ and a
number of iterations $M^\mathrm{adapt}$. Matlab code implementing both
algorithms can be found at
\url{http://www.cs.princeton.edu/~mdhoffma}, and C++ implementations
will be available as part of the Stan inference package.
\section{Empirical Evaluation}
\label{sec:evaluation}
In this section we examine the effectiveness of the dual averaging
algorithm outlined in section \ref{sec:sa}, examine what values of the
target $\delta$ in the dual averaging algorithm yield efficient
samplers, and compare the efficiency of NUTS and HMC.
For each target distribution, we ran HMC (as implemented in algorithm
\ref{alg:dacdhmc}) and NUTS (as implemented in algorithm
\ref{alg:danuts}) with four target distributions for 2000 iterations,
allowing the step size $\epsilon$ to adapt via the dual averaging
updates described in section \ref{sec:sa} for the first 1000
iterations. In all experiments the dual averaging parameters were set
to $\gamma=0.05, t_0=10,$ and $\kappa=0.75$. We evaluated HMC with 10
logarithmically spaced target simulation lengths $\lambda$ per target
distribution. For each target distribution the largest value of
$\lambda$ that we tested was 40 times the smallest value of $\lambda$
that we tested, meaning that each successive $\lambda$ is
$40^{1/9}\approx 1.5$ times larger than the previous $\lambda$. We
tried 15 evenly spaced values of the dual averaging target $\delta$
between 0.25 and 0.95 for NUTS and 8 evenly spaced values of the dual
averaging target $\delta$ between 0.25 and 0.95 for HMC. For each
sampler-simulation length-$\delta$-target distribution combination we
ran 10 iterations with different random seeds. In total, we ran 3,200
experiments with HMC and 600 experiments with NUTS.
We measure the efficiency of each algorithm in terms of effective
sample size (ESS) normalized by the number of gradient evaluations
used by each algorithm. The ESS of a set of $M$ correlated samples
$\theta^{1:M}$ with respect to some function $f(\theta)$ is the number
of independent draws from the target distribution $p(\theta)$ that
would give a Monte Carlo estimate of the mean under $p$ of $f(\theta)$
with the same level of precision as the estimate given by the mean of
$f$ for the correlated samples $\theta^{1:M}$. That is, the ESS of a
sample is a measure of how many independent samples a set of
correlated samples is worth for the purposes of estimating the mean of
some function; a more efficient sampler will give a larger ESS for
less computation. We use the number of gradient evaluations performed
by an algorithm as a proxy for the total amount of computation
performed; in all of the models and distributions we tested the
computational overhead of both HMC and NUTS is dominated by the cost
of computing gradients. Details of the method we use to estimate ESS
are provided in appendix \ref{app:ess}. In each experiment, we
discarded the first 1000 samples as burn-in when estimating ESS.
ESS is inherently a univariate statistic, but all of the distributions
we test HMC and NUTS on are multivariate. Following
\citet{Girolami:2011} we compute ESS separately for each dimension and
report the minimum ESS across all dimensions, since we want our
samplers to effectively explore all dimensions of the target
distribution. For each dimension we compute ESS in terms of the
variance of the estimator of that dimension's mean and second central
moment (where the estimate of the mean used to compute the second
central moment is taken from a separate long run of 50,000 iterations
of NUTS with $\delta=0.5$), reporting whichever statistic has a lower
effective sample size. We include the second central moment as well as
the mean because for simulation lengths $\epsilon L$ that hit a
resonance of the target distribution HMC can produce samples that are
{\it anti-}correlated. These samples yield low-variance estimators of
parameter means, but very high-variance estimators of parameter
variances, so computing ESS only in terms of the mean of $\theta$ can
be misleading.
\subsection{Models and Datasets}
\label{sec:models}
To evaluate NUTS and HMC, we used the two algorithms to sample from
four target distributions, one of which was synthetic and the other
three of which are posterior distributions arising from real datasets.
\paragraph{250-dimensional multivariate normal (MVN):}
In these experiments the target distribution was a zero-mean
250-dimensional multivariate normal with known precision matrix $A$,
i.e.,
\begin{equation}
\textstyle
p(\theta)\propto\exp\{-\frac{1}{2}\theta^T A \theta\}.
\end{equation}
The matrix $A$ was generated from a Wishart distribution with identity
scale matrix and 250 degrees of freedom. This yields a target
distribution with many strong correlations. The same matrix $A$ was
used in all experiments.
\paragraph{Bayesian logistic regression (LR):}
In these experiments the target distribution is the posterior of a
Bayesian logistic regression model fit to the German credit dataset
(available from the UCI repository \citep{Frank:2010}). The target
distribution is
\begin{equation}
\begin{split}
\textstyle
p(\alpha, \beta|x, y) & \propto p(y|x, \alpha, \beta)p(\alpha)p(\beta) \\
& \textstyle \propto \exp\{-\sum_i \log(1 + \exp\{-y_i (\alpha + x_i\cdot \beta\})
- \frac{1}{2\sigma^2}\alpha^2 - \frac{1}{2\sigma^2}\beta\cdot \beta\},
\end{split}
\end{equation}
where $x_i$ is a 24-dimensional vector of numerical predictors
associated with a customer $i$, $y_i$ is $-1$ if customer $i$ should
be denied credit and 1 if that customer should receive credit,
$\alpha$ is an intercept term, and $\beta$ is a vector of 24
regression coefficients. All predictors are normalized to have zero
mean and unit variance. $\alpha$ and each element of $\beta$ are given
weak zero-mean normal priors with variance $\sigma^2=100$. The dataset
contains predictor and response data for 1000 customers.
\label{sec:hconvergence}
\begin{figure}[t!]
\vskip-0.3in
\begin{center}
\centerline{\includegraphics[width=1\columnwidth]{normalhconvergence}}
\vskip-0.1in
\centerline{\includegraphics[width=1\columnwidth]{lrhconvergence}}
\vskip-0.1in
\centerline{\includegraphics[width=1\columnwidth]{hlrhconvergence}}
\vskip-0.1in
\centerline{\includegraphics[width=1\columnwidth]{sv2hconvergence}}
\end{center}
\vskip -0.3in
\caption{Discrepancies between the realized average acceptance
probability statistic $h$ and its target $\delta$ for the
multivariate normal, logistic regression, hierarchical logistic
regression, and stochastic volatility models. Each point's distance
from the x-axis shows how effectively the dual averaging algorithm
tuned the step size $\epsilon$ for a single experiment. Leftmost
plots show experiments run with NUTS, other plots show experiments
run with HMC with a different setting of $\epsilon L$.}
\vskip -0.2in
\label{fig:hconvergence}
\end{figure}
\paragraph{Hierarchical Bayesian logistic regression (HLR):}
In these experiments the target distribution is again the posterior of
a Bayesian logistic regression model fit to the German credit dataset,
but this time the variance parameter in the prior on $\alpha$ and
$\beta$ is given an exponential prior and estimated as well. Also, we
expand the predictor vectors by including two-way interactions,
resulting in ${24 \choose 2} + 24 = 300$-dimensional vectors of
predictors $x$ and a 300-dimensional vector of coefficients
$\beta$. These elaborations on the model make for a more challenging
problem; the posterior is in higher dimensions, and the variance term
$\sigma^2$ interacts strongly with the remaining 301 variables. The
target distribution for this problem is
\begin{equation}
\begin{split}
\textstyle
p(\alpha, \beta, \sigma^2|x, y) &\propto p(y|x, \alpha, \beta)p(\beta|\sigma^2)p(\alpha|\sigma^2)p(\sigma^2)
\\ &\textstyle\propto \exp\{-\sum_i \log(1 + \exp\{-y_i x_i\cdot \beta\})
- \frac{1}{2\sigma^2}\alpha^2 - \frac{1}{2\sigma^2}\beta\cdot \beta
- \frac{N}{2}\log\sigma^2 - \lambda\sigma^2\},
\end{split}
\end{equation}
where $N=1000$ is the number of customers and $\lambda$ is the rate
parameter to the prior on $\sigma^2$. We set $\lambda=0.01$, yielding
a weak exponential prior distribution on $\sigma^2$ whose mean and
standard deviation are 100.
\paragraph{Stochastic volatility (SV):}
In the final set of experiments the target distribution is the
posterior of a relatively simple stochastic volatility model fit to
3000 days of returns from the S\&P 500 index. The model assumes that
the observed values of the index are generated by the following
generative process:
\begin{gather}
\nonumber \tau\sim \Exp(100);\quad \nu\sim\Exp(100); \quad s_1\sim\Exp(100); \\
\textstyle \log s_{i>1}\sim \dnormal(\log s_{i-1}, \tau^{-1}); \quad
\frac{\log y_i - \log y_{i-1}}{s_i} \sim \dt_\nu,
\end{gather}
where $s_{i>1}$ refers to a scale parameter $s_i$ where $i>1$. We
integrate out the precision parameter $\tau$ to speed mixing, leading
to the 3001-dimensional target distribution
\begin{multline}
\textstyle p(s, \nu|y)\propto e^{-0.01\nu}e^{-0.01s_1}
(\prod_{i=1}^{3000}\dt_\nu(s_i^{-1}(\log y_i - \log y_{i-1}))) \times \\
\textstyle
(0.01 + 0.5 \sum_{i=2}^{3000} (\log s_i - \log s_{i-1})^2)^{-\frac{3001}{2}}.
\end{multline}
\subsection{Convergence of Dual Averaging}
\begin{figure}[t]
\vskip-0.3in
\begin{center}
\centerline{\includegraphics[width=1.025\columnwidth]{epsilonconvergence}}
\end{center}
\vskip -0.3in
\caption{Plots of the convergence of $\bar\epsilon$ as a function of
the number of iterations of NUTS with dual averaging with
$\delta=0.65$ applied to the multivariate normal (MVN), logistic
regression (LR), hierarchical logistic regression (HLR), and
stochastic volatility (SV) models. Each trace is from an independent
run. The y-axis shows the value of $\bar\epsilon$, divided by one of
the final values of $\bar\epsilon$ so that the scale of the traces
for each problem can be readily compared.}
\vskip -0.1in
\label{fig:epsilonconvergence}
\end{figure}
Figure \ref{fig:hconvergence} plots the realized versus target values
of the statistics $h^{\mathrm{HMC}}$ and $h^{\mathrm{NUTS}}$. The $h$
statistics were computed from the 1000 post-burn-in samples. The dual
averaging algorithm of section \ref{sec:sa} usually does a good job of
coercing the statistic $h$ to its desired value $\delta$. It performs
somewhat worse for the stochastic volatility model, which we attribute
to the longer burn-in period needed for this model; since it takes
more samples to reach the stationary regime for the stochastic
volatility model, the adaptation algorithm has less time to tune
$\epsilon$ to be appropriate for the stationary distribution. This is
particularly true for HMC with small values of $\delta$, since the
overly high rejection rates caused by setting $\delta$ too small lead
to slower convergence.
\begin{figure}[t!]
\vskip-0.3in
\begin{center}
\centerline{\includegraphics[width=1\columnwidth]{nfevals-hist2}}
\end{center}
\vskip -0.3in
\caption{Histograms of the trajectory lengths generated by NUTS with
various acceptance rate targets $\delta$ for the multivariate normal
(MVN), logistic regression (LR), hierarchical logistic regression
(HLR), and stochastic volatility (SV) models.}
\vskip -0.1in
\label{fig:nutstraj}
\end{figure}
Figure \ref{fig:epsilonconvergence} plots the convergence of the
averaged iterates $\bar\epsilon_m$ as a function of the number of dual
averaging updates for NUTS with $\delta=0.65$. Except for the
stochastic volatility model, which requires longer to burn in,
$\bar\epsilon$ roughly converges within a few hundred iterations.
\subsection{NUTS Trajectory Lengths}
\label{sec:nutstraj}
Figure \ref{fig:nutstraj} shows histograms of the trajectory lengths
generated by NUTS. Most of the trajectory lengths are integer powers
of two, indicating that the U-turn criterion in equation
\ref{eq:stopangle} is usually satisfied only after a doubling is
complete and not by one of the intermediate subtrees generated during
the doubling process. This behavior is desirable insofar as it means
that we only occasionally have to throw out entire half-trajectories
to satisfy detailed balance.
\begin{figure}[t!]
\vskip-0.3in
\begin{center}
\centerline{\includegraphics[width=1.0125\columnwidth]{normaless}}
\vskip-0.1in
\centerline{\includegraphics[width=1.0125\columnwidth]{lress}}
\vskip-0.1in
\centerline{\includegraphics[width=1.0125\columnwidth]{hlress}}
\vskip-0.1in
\centerline{\includegraphics[width=1.0125\columnwidth]{sv2ess}}
\end{center}
\vskip -0.3in
\caption{Effective sample size (ESS) as a function of $\delta$ and
(for HMC) simulation length $\epsilon L$ for the multivariate
normal, logistic regression, hierarchical logistic regression, and
stochastic volatility models. Each point shows the ESS divided by
the number of gradient evaluations for a separate experiment; lines
denote the average of the points' y-values for a particular
$\delta$. Leftmost plots are NUTS's performance, each other plot
shows HMC's performance for a different setting of $\epsilon L$.}
\vskip -0.1in
\label{fig:ess}
\end{figure}
The trajectory length (measured in number of states visited) grows as
the acceptance rate target $\delta$ grows, which is to be expected
since a higher $\delta$ will lead to a smaller step size $\epsilon$,
which in turn will mean that more leapfrog steps are necessary before
the trajectory doubles back on itself and satisfies equation
\ref{eq:stopangle}.
\subsection{Comparing the Efficiency of HMC and NUTS}
\label{sec:hmcvnuts}
Figure \ref{fig:ess} compares the efficiency of HMC (with various
simulation lengths $\lambda\approx\epsilon L$) and NUTS (which chooses
simulation lengths automatically). The x-axis in each plot is the
target $\delta$ used by the dual averaging algorithm from section
\ref{sec:sa} to automatically tune the step size $\epsilon$. The
y-axis is the effective sample size (ESS) generated by each sampler,
normalized by the number of gradient evaluations used in generating
the samples. HMC's best performance seems to occur around
$\delta=0.65$, suggesting that this is indeed a reasonable default
value for a variety of problems. NUTS's best performance seems to
occur around $\delta=0.6$, but does not seem to depend strongly on
$\delta$ within the range $\delta\in[0.45, 0.65]$. $\delta=0.6$
therefore seems like a reasonable default value for NUTS.
On the two logistic regression problems NUTS is able to produce
effectively independent samples about as efficiently as HMC can. On
the multivariate normal and stochastic volatility problems, NUTS with
$\delta=0.6$ outperforms HMC's best ESS by about a factor of three.
As expected, HMC's performance degrades if an inappropriate simulation
length is chosen. Across the four target distributions we tested, the
best simulation lengths $\lambda$ for HMC varied by about a factor of
100, with the longest optimal $\lambda$ being 17.62 (for the
multivariate normal) and the shortest optimal $\lambda$ being 0.17
(for the simple logistic regression). In practice, finding a good
simulation length for HMC will usually require some number of
preliminary runs. The results in Figure \ref{fig:ess} suggest that
NUTS can generate samples at least as efficiently as HMC, even
discounting the cost of any preliminary runs needed to tune HMC's
simulation length.
\subsection{Qualitative Comparison of NUTS, Random-Walk Metropolis, and Gibbs}
\begin{figure}
\begin{center}
\vspace*{-8pt}
\noindent\makebox[\textwidth]{
\includegraphics[width=1.05\columnwidth]{nuts-v-gibbs-v-rwm.pdf}
}
\end{center}
\vspace*{-12pt}
\caption{{\small\it Samples generated by random-walk Metropolis, Gibbs
sampling, and NUTS. The plots compare 1,000 independent draws
from a highly correlated 250-dimensional distribution (right) with
1,000,000 samples (thinned to 1,000 samples for display) generated
by random-walk Metropolis (left), 1,000,000 samples (thinned to
1,000 samples for display) generated by Gibbs sampling (second
from left), and 1,000 samples generated by NUTS (second from
right). Only the first two dimensions are shown here.
}}\label{fig:nuts-v-gibbs-rwm}
\end{figure}
In section \ref{sec:hmcvnuts}, we compared the efficiency of NUTS and
HMC. In this section, we informally demonstrate the advantages of NUTS
over the popular random-walk Metropolis (RWM) and Gibbs sampling
algorithms. We ran NUTS, RWM, and Gibbs sampling on the
250-dimensional multivariate normal distribution described in section
\ref{sec:models}. NUTS was run with $\delta=0.5$ for 2,000
iterations, with the first 1,000 iterations being used as burn-in and
to adapt $\epsilon$. This required about 1,000,000 gradient and
likelihood evaluations in total. We ran RWM for 1,000,000 iterations
with an isotropic normal proposal distribution whose variance was
selected beforehand to produce the theoretically optimal acceptance
rate of 0.234 \citep{Gelman:1996}. The cost per iteration of RWM is
effectively identical to the cost per gradient evaluation of NUTS, and
the two algorithms ran for about the same amount of time. We ran Gibbs
sampling for 1,000,000 sweeps over the 250 parameters. This took
longer to run than NUTS and RWM, since for the multivariate normal
each Gibbs sweep costs more than a single gradient evaluation; we
chose to nonetheless run the same number of Gibbs sweeps as RWM
iterations, since for some other models Gibbs sweeps can be done more
efficiently.
Figure \ref{fig:nuts-v-gibbs-rwm} visually compares independent
samples (projected onto the first two dimensions) from the target
distribution with samples generated by the three MCMC algorithms. RWM
has barely begun to explore the space. Gibbs does better, but still
has left parts of the space unexplored. NUTS, on the other hand, is
able to generate many effectively independent samples.
We use this simple example to visualize the relative performance of
NUTS, Gibbs, and RWM on a moderately high-dimensional distribution
exhibiting strong correlations. For the multivariate normal, Gibbs or
RWM would of course work much better after an appropriate rotation of
the parameter space. But finding and applying an appropriate rotation
can be expensive when the number of parameters $D$ gets large, and RWM
and Gibbs both require $O(D^2)$ operations per effectively independent
sample even under the highly optimistic assumption that a
transformation can be found that renders all parameters i.i.d. and can
be applied cheaply (e.g. in $O(D)$ rather than the usual $O(D^2)$ cost
of matrix-vector multiplication and the $O(D^3)$ cost of matrix
inversion). This is shown for RWM by \citet{Creutz:1988}, and for
Gibbs is the result of needing to apply a transformation requiring
$O(D)$ operations $D$ times per Gibbs sweep. For complicated models,
even more expensive transformations often cannot render the parameters
sufficiently independent to make RWM and Gibbs run efficiently. NUTS,
on the other hand, is able to efficiently sample from high-dimensional
target distributions without needing to be tuned to the shape of those
distributions.
\section{Discussion}
We have presented the No-U-Turn Sampler (NUTS), a variant of the
powerful Hamiltonian Monte Carlo (HMC) Markov chain Monte Carlo (MCMC)
algorithm that eliminates HMC's dependence on a number-of-steps
parameter $L$ but retains (and in some cases improves upon) HMC's
ability to generate effectively independent samples efficiently. We
also developed a method for automatically adapting the step size
parameter $\epsilon$ shared by NUTS and HMC via an adaptation of the
dual averaging algorithm of \citet{Nesterov:2009}, making it possible
to run NUTS with no hand tuning at all. The dual averaging approach
we developed in this paper could also be applied to other MCMC
algorithms in place of more traditional adaptive MCMC approaches based
on the Robbins-Monro stochastic approximation algorithm
\citep{Andrieu:2008, Robbins:1951}.
In this paper we have only compared NUTS with the basic HMC algorithm,
and not its extensions, several of which are reviewed by
\citet{Neal:2011}. We only considered simple kinetic energy functions
of the form $\frac{1}{2}r\cdot r$, but both NUTS and HMC can benefit
from introducing a ``mass'' matrix $M$ and using the kinetic energy
function $\frac{1}{2}r^T M^{-1} r$. If $M^{-1}$ approximates the
covariance matrix of $p(\theta)$, then this kinetic energy function
will reduce the negative impacts strong correlations and bad scaling
have on the efficiency of both NUTS and HMC. Another extension of HMC
introduced by \citet{Neal:1994} considers windows of proposed states
rather than simply the state at the end of the trajectory to allow for
larger step sizes without sacrificing acceptance rates (at the expense
of introducing a window size parameter that must be tuned). The
effectiveness of the windowed HMC algorithm suggests that NUTS's lack
of a single accept/reject step may be responsible for some of its
performance gains over vanilla HMC.
\citet{Girolami:2011} recently introduced Riemannian Manifold
Hamiltonian Monte Carlo (RMHMC), a variant on HMC that simulates
Hamiltonian dynamics in Riemannian rather than Euclidean spaces,
effectively allowing for position-dependent mass matrices. Although
the worst-case $O(D^3)$ matrix inversion costs associated with this
algorithm often make it expensive to apply in high dimensions, when
these costs are not too onerous RMHMC's ability to adapt its kinetic
energy function makes it very efficient. There are no technical
obstacles that stand in the way of combining NUTS's ability to adapt
its trajectory lengths with RMHMC's ability to adapt its mass
matrices; exploring such a hybrid algorithm seems like a natural
direction for future research.
Like HMC, NUTS can only be used to resample unconstrained
continuous-valued variables with respect to which the target
distribution is differentiable almost everywhere. HMC and NUTS can
deal with simple constraints such as nonnegativity or restriction to
the simplex by an appropriate change of variable, but discrete
variables must either be summed out or handled by other algorithms
such as Gibbs sampling. In models with discrete variables, NUTS's
ability to automatically choose a trajectory length may make it more
effective than HMC when discrete variables are present, since it is
not tied to a single simulation length that may be appropriate for one
setting of the discrete variables but not for others.
Some models include hard constraints that are too complex to eliminate
by a simple change of variables. Such models will have regions of the
parameter space with 0 posterior probability. When HMC encounters such
a region, the best it can do is stop short and restart with a new
momentum vector, wasting any work done before violating the
constraints \citep{Neal:2011}. By contrast, when NUTS encounters a
0-probability region it stops short and samples from the set of points
visited up to that point, making at least some progress.
NUTS with dual averaging makes it possible for Bayesian data analysts
to obtain the efficiency of HMC without spending time and effort
hand-tuning HMC's parameters. This is desirable even for those
practitioners who have experience using and tuning HMC, but it is
especially valuable for those who lack this experience. In particular,
NUTS's ability to operate efficiently without user intervention makes
it well suited for use in generic inference engines in the mold of
BUGS \citep{Gilks:1992}, which until now have largely relied on much
less efficient algorithms such as Gibbs sampling. We are currently
developing an automatic Bayesian inference system called Stan, which
uses NUTS as its core inference algorithm for continuous-valued
parameters. Stan promises to be able to generate effectively
independent samples from complex models' posteriors orders of
magnitude faster than previous systems such as BUGS and JAGS.
In summary, NUTS makes it possible to efficiently perform Bayesian
posterior inference on a large class of complex, high-dimensional
models with minimal human intervention. It is our hope that NUTS will
allow researchers and data analysts to spend more time developing and
testing models and less time worrying about how to fit those models to
data.
\acks{This work was partially supported by Institute of Education
Sciences grant ED-GRANTS-032309-005, Department of Energy grant
DE-SC0002099, National Science Foundation grant ATM-0934516, and
National Science Foundation grant SES-1023189.}
|
\section{O VI and Trends with Star Formation}
\noindent{\Large \bf S5 \,\, O VI and Trends with Star Formation} \\
We find a sharp difference in the detection rate and typical strength of O VI absorption surrounding galaxies that are star-forming and galaxies that are passively evolving. This trend is clearly evident in Figures 2 and 3 of the main text, where we use specific star formation rate as the key variable, because it cleanly divides those galaxies where we have detected star formation from those where we have not, at sSFR $\approx 10^{-11}$ yr$^{-1}$. For the 30 galaxies in which we have detected star formation (those with sSFR $>10^{-11}$ yr$^{-1}$), there are 27 detections of O VI (binomial probability $P = 0.90\pm0.07$ at 95\% confidence), while for the 12 galaxies with lower sSFR there are 4 O VI detections ($P = 0.33 \pm 0.27$ at 95\% confidence).
We can also test the statistical significance of the apparent correlation of O VI with sSFR. We use a generalized Kendall's tau statistic that can handle doubly censored data \cite{Isobe:86:490}. First, we establish that the correlation of O VI with star formation (sSFR) is strong, when the passive galaxies are included. For all 42 sample galaxies, we obtain $\tau = 0.617$ (significance $P = 1.32\times10^{-5}$) and so can reject at $>99.99$\% confidence the null hypothesis that there is no correlation between $N_{\rm OVI}$ and sSFR. This no-correlation null hypothesis is rejected at $>98$\% in the inner two annular bins of $R$ (Figure 1 in the main text), and at 93\% over $R = 100-160$ kpc. Thus there is strong evidence that the correlation with the presence or absence of star formation holds at all radii in the detected halos.
The well-known correlation of galaxy star formation with stellar mass is potentially a confusing influence on the O VI, one which hinders our ability to conclusively identify star formation as the key influence on halo O VI. However, even though the size of our sample is relatively small, we have evidence that star formation itself, and not simply stellar mass, has a direct relationship with the observed O~VI. To assess this case we consider the star-forming vs. passive comparison in the range of stellar mass where we have both types of galaxies. We take subsamples with $\log M_* > 10.5$, where there are 11 star-forming galaxies (8 detections) and 12 passive galaxies (4 detections).
We first consider how the quantity of observed O VI differs between star-forming and passive galaxies in the same mass range. This comparison appears in Figure S2, where we repeat the $N_{\rm OVI}$ versus sSFR plot shown in the upper panel of Figure 3 in the main text but this time including only those galaxies with $\log M_* > 10.5$. Here we see that the strong difference in hit rates between star-forming and passive galaxies is attributable to the very different column density distributions that they follow. The detections of O~VI surrounding passive galaxies are generally weaker than the detections in the star-forming sample, while the non-detections surrounding passive galaxies imply weaker absorption than almost all the detections in either category. If we compare the star-forming and passive galaxies in terms of $N_{\rm OVI}$ using a two-sided Kolmogorov-Smirnov test, we find that we can reject at $>99$\% confidence the null hypothesis that the two column density subsamples are drawn from the same parent distribution (KS statistic $D = 0.644$, probability $0.00866$).
As a consequence of the lower typical O VI column density surrounding passive galaxies, their hit rate is lower than for the star-forming galaxies. We can use the Wilson score interval to estimate the distribution of the underlying binomial probability of detection, given an observed sample size and hit rate. For the star-forming galaxies, the underlying binomial probability of a detection lies in the range $0.36 - 0.93$ at 99\% confidence with a hit rate of 0.73. For passive galaxies, the 99\% confidence interval is $0.36 - 0.68$ with a hit rate of 0.33. Thus, the hit rate for each category lies just outside the 99\% confidence interval on the binomial probability given by the other, showing that the two rates are different at $\sim 2.6 \sigma$. Because of the lower column densities and hit rates for passive galaxies, even when we control for stellar mass, we conclude that the dichotomy of O~VI is related to star formation, not solely to mass. We still lack the statistics necessary to identify what fraction of the variation in O~VI between the blue cloud and the red sequence is related to star formation and what fraction owes to stellar mass.
\clearpage
\begin{figure*}[!ht]
\begin{center}
\includegraphics[width=4.5in]{figs2}
\label{fig_colden_ssfr_mcut}
\end{center}
\end{figure*}
{\bf Figure S2:} O~VI vs. sSFR for only those galaxies with stellar masses $\log M_* > 10.5$. This figure includes all 12 passive galaxies and 11 of the star-forming galaxies from the main sample. The column density distributions of O~VI are clearly different even over the stellar mass range that star-forming and passive galaxies have in common.
\vspace{0.3in}
From inspection of Figure~3 in the main text and Figures~S2 and S3 here, it is evident that much of the statistical power of the full sample resides in the red galaxies with no detected star formation and little O VI. We can also test whether the O VI is influenced by the quantity of star formation, or only its presence or absence. To do this we omit the passive galaxies from the sample and use only the 30 galaxies for which star formation was detected.
Removing the SFR upper limits weakens these correlations significantly. For the three 50 kpc annular bins the no-correlation null hypothesis can only be rejected at $\sim 75 - 80$~\% confidence (Spearman's $\rho \sim 0.37-0.39$, $P \sim 0.21 - 0.26$). However, for the whole dataset over $R = 0 - 160$ kpc, we still see an indication ($\rho = 0.352$, $P = 94$\% confidence with $N = 30$) of a correlation of O VI with sSFR. Though these correlations are less statistically significant than those for the whole sample including passive galaxies, there is an indication that the quantity of O VI is at least weakly dependent on the quantity of nearby star formation and not simply to its presence or absence.
\clearpage
\begin{figure*}[!ht]
\begin{center}
\includegraphics[width=4.5in]{figs3}
\label{fig_colden_ssfr_rcuts}
\end{center}
\end{figure*}
{\bf Figure S3:} O VI vs. SFR over all $R$ in the sample out to 160 kpc.
\vspace{0.3in}
Previous characterizations of gas in galaxy halos at low redshift have not found a strong correlation between halo gas and galaxy bimodality. The behavior of O~VI in halos contrasts with that of the cold gas clouds traced by singly-ionized magnesium, Mg II ({\it 30}), which is rarely detected at $R > 100$ kpc, has a steeper radial dependence within that radius, and has a positive correlation with galaxy luminosity that is opposite to our findings for O~VI. Our findings on O~VI are in better accord with the $140$ kpc size of the triply-ionized carbon (C IV) envelopes surrounding galaxies as reported by \cite{Chen:01:158}, but that study did not detect a correlation with star-formation (as proxied by morphological type). Thus the O~VI, C IV, and Mg II may trace distinct phases of the circumgalactic medium with possibly different origins. Studies of star-forming ``Lyman-break'' galaxies at $z \sim 2-3$ have found a connection between star-formation and outflowing material traced by Ly$\alpha$ and low and intermediate ions ({\it 6,7}), and total masses of cold gas comparable to that in the galaxies themselves. \\
\noindent{\Large \bf S6 \,\, Physical Conditions, Distribution, and Mass of the O VI-bearing Halo Gas} \\
The narrow range of physical conditions in diffuse gas that contains significant O VI allows us to obtain robust lower limits to the mass of oxygen in the detected halo gas. O VI is not the dominant ion of oxygen in astrophysical conditions, so the ionization correction applied to derive the total oxygen density (or mass) from the O VI density (or mass) is always substantial (Figure 4 of the main text). To estimate the correction, we have considered the ionization state of O VI over a wide range of temperature as computed by the CLOUDY photoionization code \cite{Ferland:98:761} assuming ionization equilibrium and including both collisional ionization and photo-ionization with the $z=0.2$ radiation background from \cite{Haardt:01}. Under plausible CGM physical conditions, the ionization timescales are usually $10^7-10^8$ years or less, compared to halo dynamical timescales of $\sim 10^9$ years, so it seems unlikely that a large fraction of the CGM gas could be far from ionization equilibrium.
To calculate the mass implied by our detection of strong O VI surrounding star-forming galaxies, we can exploit the fact that column densities are not sensitive to the exact distribution of densities along the line of sight. We can therefore apply basic geometry to deduce the total mass of O VI in a volume projecting a circular region of $R = 150$ kpc on the sky with a typical column density $\log N_{\rm OVI} = 14.5$. The total mass of O VI in solar masses is:
\begin{equation}
M_{\rm OVI} = \pi R^2 N_{\rm OVI} (16 m_H).
\end{equation}
Scaling to typical values for our sample and including the O~VI ionization correction, we obtain the total oxygen mass:
\begin{equation}
M_{\rm O} = 1.2\times10^7 M_{\odot} \left( \frac{R}{150 \,{\rm kpc}} \right)^2 \left( \frac{N_{\rm OVI}}{10^{14.5 }\, {\rm cm}^{-2}} \right) \left( \frac{0.2}{f_{\rm OVI}} \right),
\end{equation}
which has the same functional form and scalings as the relation in the main text. As in the main text we also correct the coefficient from 1.4 down to $1.2$ according to the measured covering fraction around star-forming galaxies in three radial bins: 11/11 detections with $R < 50$ kpc, 9 of 10 with $50 < R < 100$, and 7/9 with $100 < R < 150$. Taking the mean column density separately in each of these three bins gives $\log \langle N_{\rm OVI} \rangle =$ 14.7, 14.6, and 14.5 respectively. Adopting the $\langle N_{\rm OVI} \rangle$ and $f_{hit}$ for the individual bins gives a mass coefficient $1.4 \times 10^7 M_{\odot}$, but to impose a firm lower limit for the main text we have adopted the lower envelope of the star-forming detections at $\log \langle N_{\rm OVI} \rangle = 14.5$ to avoid being skewed by the few high values at small $R$.
The interstellar gas masses cited in the text were derived from the individual stellar masses by applying the mean gas fraction versus stellar mass relation as tabulated by ({\it 27}). The typical interstellar oxygen masses in the galaxies were then obtained by applying the metallicities, $Z = 0.5 - 1.2 Z_{\odot}$, obtained from the measurements of nebular emission lines. These values define the green band in Figure 4 of the main text.
\clearpage
\begin{figure*}[!t]
\begin{center}
\includegraphics[width=4.5in]{figs4}
\end{center}
\label{oxy_frac_fig}
\end{figure*} \vspace{-1.5in}
{\bf Figure S4:} The ratio between our inferred minimum CGM oxygen mass ($M_{\rm O} = 1.2 \times 10^7 M_{\odot}$, see Equation 2) and the typical ISM oxygen mass $M_{\rm ISM}^{\rm O}$, as a function of galaxy stellar mass, and with the conservative O VI ionization correction $f_{\rm OVI} = 0.2$. The blue curve and band correspond to the mean observed MZR and its systematic uncertainties. The CGM oxygen mass $M_{\rm O}$ range from 70\% of $M_{\rm ISM}^{\rm O}$ at the lower end of the mass range in our sample, $10^{9.5} M_{\odot}$, down to 10\% at upper mass limit of $10^{11} M_{\odot}$. Because the CGM oxygen mass increases in inverse proportion to $f_{\rm OVI}$, the blue band expresses a lower limit to the CGM-to-ISM ratio. The heavy dashed line shows the ratio between the inferred minimum CGM mass and the oxygen mass produced by these galaxies over the last 1.5 Gyr.
\vspace{0.3in}
We have also calculated $M_{ISM}^{\rm O}$ for typical star-forming galaxies, using a mean gas fraction and mean mass-metallicity relation for SDSS galaxies following the analysis of ({\it 27}). These results are shown in Figures 4 and S4. Here we have derived the gas fractions of galaxies using the total budget of cold gas (atomic plus molecular). The mean ISM oxygen mass as a function of galaxy stellar mass is then derived and compared with the minimum halo oxygen mass, $1.2 \times 10^7 M_{\odot} \times (0.2 / f_{\rm OVI})$. The widths of these hashed regions express the systematic uncertainty in the mass-metallicity relation. For the mean mass-metallicity relation adopting the total cold gas budget, we find that $M_{\rm O} / M_{\rm ISM}^{\rm O}$ = 0.7 at $\log M_* = 9.5$ and 0.1 at $\log M_* = 11$ (Figure~S4). This ratio increases as $f_{\rm OVI}$ declines from its maximal value of 0.2. Notably, the ratio between CGM and ISM oxygen masses increase as galaxy stellar mass declines, if $f_{\rm OVI}$ is held fixed.
We can also consider the ratio between the measured CGM oxygen mass $M_{\rm O}$ and the oxygen produced by star formation over some timescale. In this comparison we adopt the mean SDSS SFRs as a function of stellar mass, the oxygen yield of 0.014 $M_{\odot}$ per $M_{\odot}$ of star formation, and a timescale of 1.5 Gyr, the travel time to 150 kpc of a cloud moving at 100 km s$^{-1}$. The minimum ratio between the detected CGM oxygen mass and this ``produced'' mass is given by the heavy dashed line in Figure S4. As stated in the main text, the detected CGM mass is a notable fraction of the total amount of oxygen produced by these galaxies over a significant timescale.
Because we measure column densities, many possible arrangements of gas along the line of sight are possible. However, we can estimate the total line-of-sight extent of the gas using simple scalings. For a solar ($Z=Z_\odot$) oxygen abundance ({\it 26}), the O VI column density is
\begin{equation}
N_{\rm O VI} = L \left({n_{\rm O}\over n_{\rm H}}\right)\left({\rho \over m_{\rm H}} \right) \left({ Z \over Z_{\odot} } \right)
= 10^{14} L_{\rm kpc} \left({\rho/\bar{\rho}\over 1000}\right)
\left({ f_{\rm O VI} \over 0.2 }\right) \left({ Z \over Z_{\odot} } \right) \,{\rm cm}^{-2}~,
\label{a}
\end{equation}
where $L_{\rm kpc}$ is the path length in kpc and $\bar{\rho}$ is the mean cosmic density at $z = 0.2$ (particle density of $n_H = 3.26 \times 10^{-7}$ cm$^{-3}$). Thus halo clouds with density $n_H = 10^{-4}\,{\rm cm}^{-3}$ and $f_{\rm OVI}=0.2$ can give $\log N_{\rm OVI} = 14.5$ if they extend 10 kpc along the line of sight, but the required path length rises to 200 kpc if the ionization fraction of O VI is only 1\% rather than its peak value of 20\% (Figure S5). Except at very low densities, where the cloud sizes become too large to fit within galaxy halos and still give the column densities we observe for star-forming galaxies, the ionization fraction of O VI is never more than $f_{\rm OVI} = 0.2$.
The cited oxygen masses for the Mg II halos of galaxies were obtained by taking the total baryonic mass estimate for that medium from ({\it 30}), correcting down by 0.1 solar metallicity as argued in that paper, and then applying the solar relative abundances of magnesium and oxygen. The cited mass estimate for Galactic HVCs is taken from the estimate by ({\it 28}) for the more conservative scenario that locates all the HVCs within 60 kpc of the Milky Way disk.
To calculate the fraction of cosmic oxygen represented by the ionized halos of star forming galaxies, we apply the generic $R = 150$ kpc halo with $N_{\rm O VI} = 10^{14.5}$ cm$^{-2}$ to all star-forming galaxies with $M_* > 10^{9.5} \, M_{\odot}$, the range covered by our survey. Integrating over the representative population (specified by the K-selected late-type luminosity function of \cite{Bell:03:289}, with $\alpha=-0.94$), we obtain $\Omega _{\rm OVI}^{\rm halos} = 8.26/h \times 10^{-8} \times (1/ f_{\rm OVI})$. The cosmic density of oxygen is $\Omega _{\rm O} = y\Omega_* = 0.014 \times 0.00197/h = 2.76/h \times 10^{-5}$. The fraction of cosmic metals in star-forming halos like those in our sample is then $\Omega_{\rm OVI}^{\rm halos} / \Omega_{\rm O} = 0.015 \times (0.2/ f_{\rm OVI}) $, as given in the text. If we correct this cosmic mass budget to the total gas mass implied in these halos, scaling to the Milky Way HVC abundances, we obtain $\Omega _{b} ^{\rm halos} = 177 (Z_{\odot} / Z) \Omega_{\rm OVI}^{\rm halos} = 7.31 / h \times 10^{-4} \times (0.1Z_\odot / Z) \times (0.2 / f_{\rm OVI})$. Dividing by the cosmic baryon fraction \cite{Dunkley:09:306}, we obtain $\Omega _{b} ^{\rm halos} = 0.023 (Z_\odot / Z) \times (0.2 / f_{\rm OVI})$. Thus the ionized gas in the extended halos of our star forming sample represent at least 0.2\% of cosmic baryons if they have $f_{\rm OVI} = 0.2$ and solar metallicity, 2\% if conditions are not optimized for O~VI ($f_{\rm OVI} \sim 0.02$), or if the metallicity is $0.1 Z_{\odot}$, or up to 20\% if they have both low metallicity and low $f_{\rm O VI}$. The contribution of these halos to the baryon budget is similar to or exceeds that from atomic and molecular phases of the ISM \cite{Fukugita:98:503}, which also make $1-2$\% contributions, but only if conditions are not optimal for O VI and/or the gas metallicity is low.
Figure S5 suggests that it is just possible to explain our observed column densities with solar metallicity, photo-ionized gas at $T \leq 10^5\,$K, provided that this gas fills enough of the halo to accumulate path lengths $\geq 100\,$kpc. However, for $Z=0.1 Z_\odot$ the required path lengths become much larger than the virial radii of the galaxy halos, expected to be $200-350\,$kpc for galaxies in our observed stellar mass range. Furthermore, if halo gas traces the dark matter, then typical overdensities at radii $\sim 100\,$kpc are several hundred or more, in which case photo-ionization is ineffective. These arguments favor collisionally ionized gas, which requires finely tuned temperatures to achieve high $f_{\rm OVI}$ (see Figure~4 and discussion in the main text). There are two significant caveats to the argument for collisional ionization. First, the gas responsible for the OVI absorption could have lower overdensity than the halo dark matter, with $(\rho/\bar{\rho})_{\rm gas} \sim 50-100$ at $R \sim 100\,$kpc. Large path lengths like those shown in Figure S5 would still be required, so this solution is only viable if the metallicity is close to solar. Second, the UV background at the 10 Rydberg ionization threshold of O VI could be a factor of several higher than the value we have adopted \cite{Haardt:01}, in which case photo-ionization becomes important at an overdensity higher by the same factor. To give a concrete example, if the UV background intensity is a factor of four higher, then $f_{\rm OVI} \approx 0.05$ at $T=10^{4.5}\,$K and $\rho/\bar{\rho}=400$ (see the $\rho/\bar{\rho}=100$ curve in Figure~4), and the required path length for $N_{\rm OVI}=10^{14.5}\,{\rm cm}^{-2}$ drops from $100(Z_\odot/Z)\,$kpc (Figure S5, red curve) to $25(Z_\odot/Z)\,$kpc. With our current data, we cannot determine the metallicity of the absorbing gas or decide between collisionally ionized and photo-ionized scenarios (or a mix of the two). In any case, however, achieving our high covering factors and observed column densities requires that favorable combinations of physical conditions arise in the typical halos of star-forming galaxies.
\clearpage
\begin{figure*}[!ht]
\begin{center}
\includegraphics[width=4.5in]{figs5}
\end{center}
\label{ovifracfig}
\end{figure*}
\vspace{-0.2in}
{\bf Figure S5:} Path length required to obtain $N_{\rm OVI} = 10^{14.5}\,{\rm cm}^{-2}$ for solar metallicity gas with the indicated overdensity and temperature, using the $f_{\rm OVI}$ values shown in Figure~4 of the main text. These pathlengths scale in inverse proportion to metallicity according to Equation (3), and so are $10\times$ longer for $Z = 0.1 Z_{\odot}$. The overdensity and ionizing background are computed at $z=0.2$, using the UV background model of \cite{Haardt:01}. Gas with $\rho/\bar{\rho} \geq 1000$ is close to collisional ionization equilibrium, so $f_{\rm OVI}$ depends only on temperature and required path lengths are inversely proportional to density. At overdensity $\rho/\bar{\rho} \leq 100$, photo-ionization becomes important for $T < 10^{5.5}\,$K, so $f_{\rm OVI}$ depends on the gas density and the scaling of required path length with density is no longer simple.
\vspace{0.3in}
\bibliographystyle{Science}
|
\section{Introduction}
\label{sec:intro}
Onsager's intuition that purely repulsive rods undergo an entropy-driven transition from an isotropic ($I$) to an orientationally ordered nematic ($N$) phase constitutes one of the major milestones in our understanding of liquid crystals \cite{onsager}. The key ingredient of this phenomenon relies on considering markedly non-spherical particles, which can be modeled as cylindrically symmetric ``rods'' and ``plates''. In the early 1970s Freiser pointed out that a richer phase behavior is expected, if the assumption of cylindrical symmetry is released \cite{freiser}. Besides the usual prolate ($N_+$) and oblate ($N_-$) uniaxial nematic phases, normally developed by uniaxial rods and plates, respectively, a novel nematic phase with an increased orientational order can appear in the phase diagram. Such a liquid crystal phase is characterized by alignment along three directors and, consequently, by the presence of two distinct optical axes, hence the name {\em biaxial nematic} ($N_B$) \cite{tschierske}. Further studies suggested that $N_B$ stability could be interpreted as a balanced competition between rodlike (favoring $N_+$) and platelike (favoring $N_-$) behavior \cite{alben, straley, mulder}.
In more than 40 years since its first theoretical prediction, extensive theoretical \cite{boccara,mulder,taylor,vanroij,martinez2,varga,vanakaras,dematteis,longa,allender,martinez,belli} and simulation \cite{allen,camp,berardi2,bates3,bates2,berardi,cuetos} work has been devoted to identify the conditions under which a stable $N_B$ phase could be observed. The practical limitations in this sense are testified by the fact that, apart from the micellar system studied by Yu and Saupe \cite{yu}, no such state has been observed for more than 30 years. A renewed interest towards the topic has grown due to the first experimental realization of thermotropic $N_B$ liquid crystals in systems of bent-core molecules a few years ago \cite{madsen,acharya}. In lyotropics, a remarkably stable $N_B$ phase was recently discovered in a colloidal suspension of mineral boardlike particles \cite{vandenpol}.
Boardlike particles, that is, particles with the symmetry of a brick, represent the simplest model in which an $N_B$ phase has been predicted \cite{tschierske}. However, the emergence of smectic layering is expected to prevent the realization of this phase, unless the constituent particles are designed with a precision far beyond present-day ability \cite{taylor, vandenpol}. A higher $N_B$ stability can be achieved by considering size-polydisperse systems of boardlike particles, as demonstrated by a recent experiment \cite{vandenpol}. In fact, polydispersity seems to enhance $N_B$ stability through two distinct phenomena: (i) a reduced smectic stability \cite{vanakaras} and (ii) an $N_+$-$N_-$ competition, which manifests itself exclusively in systems of slightly elongated (rodlike) boards \cite{belli}. The first phenomenon does not come as a surprise \cite{vanakaras}, since it is well known that polydispersity renders the establishment of long-distance positional ordering unfavorable \cite{barrat, mcrae, bates}. On the contrary, the reason behind the second phenomenon appears to be more obscure.
In this paper we investigate the effect of a non-adsorbing depletant on the biaxial nematic stability of (monodisperse) boardlike particles. Our understanding of depletion dates back to the pioneering work by Asakura and Osawa \cite{asakura} and Vrij \cite{vrij}, who showed that the addition of small co-solutes (e.g. polymers, surfactants, micelles) to a colloidal suspension gives rise to an effective attraction between colloidal particles. Since then, the concepts related to depletion have been widely applied to various scientific fields \cite{tuinier}: in biology by interpreting phenomena like macromolecular crowding \cite{odijk} and protein crystallization \cite{piazza}; in nanotechnology through e.g. the development of self-assembly processes as key-lock structures \cite{odriozola, sacanna}; in condensed matter physics, furnishing answers to fundamental problems like the condition for gas-liquid phase separation \cite{gast}, the kinetics of crystallization \cite{auer, anderson} and the nature of glassy states \cite{weeks}. More recently, the liquid crystal phase behavior of non-spherical colloids, typically rods \cite{lekkerkerker, buitenhuis, bolhuis, dogic, savenko, jungblut} and plates \cite{bates4, vanderkooij, zhang, kleshchanok}, in presence of a depletant has also been addressed. As a general feature, the addition of a depletant reduces the stability of liquid-crystal phases, leading to a direct isotropic-crystal transition at high enough depletant mole fraction. Moreover, when the size of the depletant particles is big enough, one or more critical points appear in the phase diagram, indicating a liquid-gas separation between phases with same spatial symmetries.
In contrast to the aforementioned work on rods and plates, we focus here on the low depletant density limit, where the stability of the nematic liquid crystal phases developed by the pure system of boardlike particles is preserved.
In the same spirit as the Asakura-Oosawa-Vrij model for spheres \cite{asakura,vrij}, we consider the limit of low depletant density and neglect depletant-depletant interactions. For the sake of convenience, we model the depletants as cubic particles excluded from the surface of the cuboids via a hard-core interaction. A mean-field theory at second virial order \cite{onsager, evans} with restricted orientations (Zwanzig model) \cite{zwanzig} constitutes our theoretical framework. The degrees of freedom of the depletant in the partition function can be systematically integrated out, giving rise to an effective potential between boardlike particles \cite{dijkstra, dijkstra2}, where only two-body interactions are considered. The assumption of ideal depletant allows to determine an explicit expression for such a pairwise depletion potential. We show that, by varying the depletant density, the system develops an $N_+$-$N_-$ competition remarkably similar to that predicted for a polydisperse system of boardlike particles in absence of depletant \cite{belli}. If in Ref. \cite{belli} the origin of this competition is not evident, here it appears to be due to a balance between the hard-core repulsion between boardlike particles, favoring $N_+$ ordering, and the depletion attraction, favoring $N_-$ ordering. As a consequence of this effect, the biaxial nematic phase appears to be stable over a wide range of depletant density. We therefore suggest that the concentration of a non-adsorbing depletant furnishes in practical situations the simplest, though effective, way to control the liquid-crystal phase behavior of boardlike particles and to select states of high biaxial-nematic stability.
The paper is organized as follows. We illustrate in Sec. \ref{sec:theory} our theoretical framework and in Sec. \ref{sec:model} the model describing the boards-depletant mixture. Sec. \ref{sec:results} is devoted to the results, whereas in Sec. \ref{sec:conclusions} we draw our conclusions.
\section{Second-virial density functional theory with restricted orientations}
\label{sec:theory}
\begin{figure}
\includegraphics[scale=0.14]{fig1.pdf}
\caption{\label{fig1} The 6 independent orientations of a boardlike particle within the restricted orientations (Zwanzig) model.}
\end{figure}
We consider a system of $N$ boardlike particles with dimensions $l \times w \times t$ ($l > w > t$ and particle volume $v=l w t$) in a box of volume $V$ at temperature $T$. Accounting for the orientational degrees of freedom at the single-particle level requires a numerically demanding description based on $3$ Euler angles. In order to circumvent this problem while keeping the essential physics of the system, we turn to the so-called Zwanzig model: the only allowed orientations are those with the main particle axes aligned along the axes of a fixed reference frame \cite{zwanzig}. Within this model a boardlike particle can take the $6$ orientations depicted in Fig. \ref{fig1}, and the orientation distribution function (ODF) is a $6$-dimensional vector $\boldsymbol{\psi}$ with components $\psi_i$ ($i=1,...,6$), subject to the normalization condition
\begin{equation}
\sum_{i=1}^6 \psi_i = 1.
\label{eq1}
\end{equation}
Being interested in the low-density phase behavior of the system, where the stable phases are expected to be homogeneous in space, we can neglect spatial modulations in the single-particle density. Under these conditions, the intrinsic free energy $\mathcal{F}$ reads \cite{belli}
\begin{equation}
\frac{\beta \mathcal{F}[\boldsymbol{\psi}]}{N} = \ln(\eta) + \sum_{i=1}^6 \psi_i \ln(\psi_i ) + \frac{\beta \mathcal{F}_{\mathrm{exc}}[\boldsymbol{\psi}]}{N},
\label{eq2}
\end{equation}
where $\beta=(k_B T)^{-1}$, $k_B$ is the Boltzmann constant and $\eta=N v/V$ is the packing fraction. A closed expression for the excess free energy $\mathcal{F}_{\mathrm{exc}}$ in terms of the ODF is not known in general, but for short-range pairwise additive potentials it is possible to write it as a virial series in $\eta$. Let $u_{i i'}(\mathbf{r})$ be the interaction potential between a pair of particles with orientations $i$ and $i'$, respectively, and a separation $\mathbf{r}$ between their centers of mass. At second virial order the excess free energy reads
\begin{equation}
\frac{\beta \mathcal{F}_{\mathrm{exc}}[\boldsymbol{\psi}]}{N} = \frac{\eta}{2 v} \sum_{i,i'=1}^{6} E_{i i'} \psi_{i} \psi_{i'},
\label{eq3}
\end{equation}
where the second-virial coefficients $E_{i i'}/2$ are given by
\begin{equation}
E_{i i'} = -\int_V d \mathbf{r} f_{i i'}(\mathbf{r}),
\label{eq4}
\end{equation}
with the Mayer function
\begin{equation}
f_{i i'}(\mathbf{r}) = \exp \bigl[-\beta u_{ii'}(\mathbf{r}) \bigr] - 1.
\label{eq5}
\end{equation}
More refined approximations than Eq. (\ref{eq3}) have been recently developed in order to take into account higher-order virial terms, which would give a quantitatively more reliable description \cite{martinez}. On the other hand, truncating the virial series at second order allows for an appreciable simplification of the mathematics and the numerics involved, while retaining the essential physics.
Let us indicate with $X_i$ the main axis ($l$, $w$ or $t$) of a particle with orientation $i$ along the $x$ axis of a fixed reference frame, and similarly with $Y_i$ and $Z_i$. Within the Zwanzig model each of the $6$ independent orientations of a particle can be identified by $(X_i,Y_i,Z_i)$, which is one of the $6$ permutations of the three elements $l$, $w$ and $t$. With these definitions one can write the interaction potential between two identical boardlike particles, modeled as hard cuboids, as
\begin{equation}
\beta u_{i i'}(\mathbf{r}) =
\begin{cases}
\infty \hspace{0.8cm} \text{if $|x|< (X_{i}+X_{i'})/2$ } \\
\hspace{1.15 cm} \text {and $|y|<(Y_{i}+Y_{i'})/2$} \\
\vspace{0.15 cm}
\hspace{1.15 cm} \text{and $|z|<(Z_{i}+Z_{i'})/2$;}\\
0 \hspace{1cm} \text{otherwise,}
\end{cases}
\label{eq6}
\end{equation}
from which explicit expressions for $E_{ii'}$ in terms of $l$, $w$ and $t$ directly follow through Eqs. (\ref{eq4})-(\ref{eq5}).
At second-virial order the equilibrium ODF at fixed packing fraction $\eta$ is obtained by minimizing the free energy of Eqs. (\ref{eq2})-(\ref{eq3}) with respect to $\boldsymbol{\psi}$, subject to the normalization condition Eq. (\ref{eq1}). In practice, the minimization problem is performed by numerically solving the system of $6$ non-linear Euler-Lagrange equations
\begin{equation}
\psi_i = C \exp \Biggl[-\frac{\eta}{v} \sum_{i'=1}^{6} E_{i i'} \psi_{i'} \Biggr],
\label{eq7}
\end{equation}
with the proportionality constant $C$ determined by Eq. (\ref{eq1}). The symmetry of the solution of Eq. (\ref{eq7}) allows to identify the stable homogeneous phase. When $\psi_i=1/6$ for every $i=1,...,6$, the phase is isotropic ($I$), whereas in the opposite case, when all the $\psi_i$ assume different values, the ODF describes a biaxial nematic ($N_B$) phase. When the system is characterized by the presence of a single axis of symmetry (uniaxial nematic phase), the coefficients $\psi_i$ are coupled two-by-two. Let us suppose this axis of symmetry to be the vertical axis of Fig. \ref{fig1}. In this case, we distinguish between prolate uniaxial nematic phase ($N_+$), when the most likely configurations of Fig. \ref{fig1} are (1) and (4), and oblate uniaxial nematic phase ($N_-$), when the most likely configurations are (3) and (6).
In the treatment described so far, we assume the system to be homogeneous in space. In order to estimate the limit of validity of this assumption, we adopt bifurcation theory to calculate the minimum packing fraction $\bar{\eta}$, beyond which homogeneous phases are unstable with respect to smectic states \cite{mulder2}. The mathematical details regarding the application of bifurcation theory to the free energy of Eqs. (\ref{eq2})-(\ref{eq3}) can be found in the Supplemental Material of Ref. \cite{belli}, and we report here only the final result. Let us indicate with $Q_{ii'}^{(x)}(q_x)$ the function
\begin{equation}
Q_{ii'}^{(x)}(q_x) = \frac{\eta}{v} \sqrt{\psi_i \psi_{i'}} \int_V d \mathbf{r} \, f_{ii'}(\mathbf{r}) \exp(-i q_x x),
\label{eq8}
\end{equation}
where $\boldsymbol{\psi}$ is the ODF of the equilibrium homogeneous phase at packing fraction $\eta$, and analogously for $Q_{ii'}^{(y)}(q_y)$ and $Q_{ii'}^{(z)}(q_z)$. The bifurcation packing fraction $\bar{\eta}_x$ for smectic fluctuations along the $x$ axis is found as the minimum packing fraction at which the $6 \times 6$ matrix with entries $Q_{ii'}^{(x)}(q_x)$ has an eigenvalue $1$ for some $\bar{q}_x$. Therefore, the smectic bifurcation packing fraction is $\bar{\eta}=\min(\bar{\eta}_x, \bar{\eta}_y, \bar{\eta}_z )$. As a final remark, it is important to notice that the present bifurcation analysis allows only to predict when homogeneous phases are unstable with respect to one-dimensional modulations in the single-particle density. Therefore, nothing ensures the corresponding stable inhomogeneous phase to be characterized by one-dimensional (smectic), rather than two- (columnar) or three-dimensional (crystal) positional ordering.
\section{Effective depletion interaction}
\label{sec:model}
Our aim is to study the influence of a depletant on the phase behavior of a system of boardlike particles. Hence, the system described in Sec. \ref{sec:theory} is modified by the addition of a second species of particles (the depletant), modeled as cubes with dimensions $d \times d \times d$. The binary mixture of boardlike particles and depletant is assumed to be in equilibrium with a reservoir of depletant particles at fixed fugacity $z_D = \exp(\beta \mu_D)/\Lambda_D^3$, where $\mu_D$ is the chemical potential of the depletant and $\Lambda_D$ its thermal wavelength. Following the pioneering approaches to the topic \cite{asakura,vrij}, we neglect interactions between depletants, in which case the fugacity $z_D$ coincides with the density $n_D$ in the reservoir. The ideal-depletant assumption is justified {\em a posteriori} by the low packing fractions $n_D d^3$ considered. Modeling the depletant with cubic particles appears to be rather unrealistic, especially if compared to typical polymeric depletants, usually treated as spheres. However, we claim that our choice contains the essential features of the physical phenomenon, while considerably simplifying the mathematics that follows. In the next section we show that the peculiar phase behavior of our system is due to the asphericity of the depletion volume, which, in turns, is a consequence of the asphericity of boardlike particles. Therefore, we do not expect the specific shape of the depletion region (cuboidal for cubic depletant, spherocuboidal for spherical depletant) to play a major role in our results. Moreover, the relative difference between cuboidal and spherocuboidal depletion volume for the values of the particles dimensions considered here amounts to few percentage points. The interactions in the mixture are given by the cuboid-cuboid potential Eq. (\ref{eq6}) between boardlike particles, and by the cuboid-cube potential between boardlike particles and depletant, given by
\begin{equation}
\beta v_{i}(\mathbf{r}) =
\begin{cases}
\infty \hspace{0.8cm} \text{if $|x|< (X_{i}+d)/2$ } \\
\hspace{1.15 cm} \text {and $|y|<(Y_{i}+d)/2$} \\
\vspace{0.15 cm}
\hspace{1.15 cm} \text{and $|z|<(Z_{i}+d)/2$;}\\
0 \hspace{1cm} \text{otherwise,}
\end{cases}
\label{eq9}
\end{equation}
which explicitly depends on the orientation $i$ of the boardlike particle.
At fixed fugacity $z_D$ the configurational entropy of the depletant is maximized when the total depletion volume, i.e. the region of space forbidden to the depletant due to the presence of boardlike particles, is minimized. As a consequence, an effective attraction between boardlike particles appears. Such a depletion interaction can be explicitly calculated by integrating out the depletant degrees of freedom and must be expressed in general as a sum of many-body interaction terms \cite{dijkstra, dijkstra2}. For the sake of simplicity, we include the effect of the depletant by considering only the effective two-body interaction potential, while neglecting higher order terms. The effective pairwise depletion potential $w_{i i'}(\mathbf{r})$ between cuboids with orientations $i$ and $i'$, respectively, and center-to-center separation $\mathbf{r}$ is given by \cite{gotzelmann}
\begin{equation}
\beta w_{i i'}(\mathbf{r}) = -n_D \mathcal{V}_{i i'}(\mathbf{r}),
\label{eq10}
\end{equation}
with $\mathcal{V}_{i i'}(\mathbf{r})$ the overlap volume of the depletion regions,
\begin{equation}
\mathcal{V}_{i i'}(\mathbf{r})=
\begin{cases}
0 \hspace{1.1cm} \text{if $|x|>(2d+X_{i}+X_{i'})/2$} \\
\hspace{1.3 cm} \text {and $|y|>(2d+Y_{i}+Y_{i'})/2$} \\
\vspace{0.15 cm}
\hspace{1.3 cm} \text{and $|z|>(2d+Z_{i}+Z_{i'})/2$;}\\
\lambda^{(x)}_{i i'} \lambda^{(y)}_{i i'} \lambda^{(z)}_{i i'} \hspace{1cm} \text{otherwise.}
\end{cases}
\label{eq11}
\end{equation}
Here $\lambda^{(x)}_{i i'}$ is defined as
\begin{equation}
\lambda^{(x)}_{i i'}(x)=
\begin{cases}
d + \frac{X_{i}+X_{i'}}{2} - |x| \hspace{0.5cm} \text{if $|x|>\frac{|X_{i}-X_{i'}|}{2}$} \\
\vspace{0.15 cm}
\hspace{3 cm} \text{and $|x|<(d+\frac{X_{i}+X_{i'}}{2})$;} \\
d + \min(X_{i},X_{i'}) \hspace{0.5cm} \text{if $|x|<\frac{|X_{i}-X_{i'}|}{2}$,}
\end{cases}
\label{eq12}
\end{equation}
and analogous definitions hold for $\lambda^{(y)}_{i i'}(y)$ and $\lambda^{(z)}_{i i'}(z)$.
Let us indicate with a tilde the properties obtained by adding the effective two-body depletion potential $w_{i i'}(\mathbf{r})$ to the cuboid-cuboid potential $u_{i i'}(\mathbf{r})$. The Mayer function Eq. (\ref{eq5}) becomes
\begin{equation}
\widetilde{f}_{i i'}(\mathbf{r}) = \exp \bigl[-\beta u_{ii'}(\mathbf{r}) + n_D \mathcal{V}_{i i'}(\mathbf{r}) \bigr] - 1.
\label{eq13}
\end{equation}
The phase behavior of this effective one-component system can then be calculated by following the prescriptions of Sec. \ref{sec:theory}, with the function $f_{i i'}(\mathbf{r})$ substituted by $\widetilde{f}_{i i'}(\mathbf{r})$. Unfortunately, the expression of $\mathcal{V}_{i i'}(\mathbf{r})$ given in Eq. (\ref{eq11}) does not allow for an analytical calculation of the integrals $\widetilde{E}_{ii'}$ and $\widetilde{Q}_{ii'}^{(x)}(q_x)$ in Eqs. (\ref{eq4}) and (\ref{eq8}). However, an analytical expression can be obtained by truncating the Taylor series of the Mayer function Eq. (\ref{eq13}) in $n_D \mathcal{V}_{i i'}(\mathbf{r})$,
\begin{equation}
\widetilde{f}_{ii'}(\mathbf{r}) = f_{ii'}(\mathbf{r}) + \sum_{m=1}^{\infty} \frac{n_D^m}{m!} \bigl (\mathcal{V}_{i i'}(\mathbf{r}) \bigr)^m \exp \bigl [-\beta u_{ii'}(\mathbf{r})\bigr].
\label{eq14}
\end{equation}
By inserting Eq. (\ref{eq14}) into Eq. (\ref{eq4}), one obtains for the effective excluded-volume coefficients
\begin{equation}
\widetilde{E}_{i i'} = E_{i i'} - \sum_{m=1}^{\infty} \frac{n_D^m}{m!} \int_V d \mathbf{r} \bigl (\mathcal{V}_{i i'}(\mathbf{r}) \bigr)^m \exp \bigl [-\beta u_{ii'}(\mathbf{r})\bigr],
\label{eq15}
\end{equation}
where the integrals of the r.h.s. can now be solved analytically for every $m$. Similar considerations hold for the functions $\widetilde{Q}_{ii'}^{(x)}(q_x)$ of Eq. (\ref{eq8}). We verified by comparison with exact numerical calculations of the effective excluded-volume coefficients that quantitative agreement can be obtained by truncating the series of Eq. (\ref{eq15}) at fifth order in $n_D$ for all $n_D$ considered in this paper. For consistency, the Taylor expansion in $n_D$ of the functions $\widetilde{Q}_{ii'}^{(x)}(q_x)$ is truncated at the same order.
\section{Results}
\label{sec:results}
The framework developed in Sec. \ref{sec:model} allows to determine the effective excluded-volume coefficients $\widetilde{E}_{i i'}$ of a system of cuboidal $l \times w \times t$ particles due to the presence of a cubic $d \times d \times d$ depletant at fugacity $z_D$ (and density $n_D=z_D$). The phase behavior of this effective one-component system of boardlike particles is then analyzed by applying the theory described in Sec. \ref{sec:theory}.
\begin{figure}
\includegraphics[scale=0.65]{fig2.pdf}
\caption{\label{fig2} Effective excluded-volume coefficients $\widetilde{E}_{1 i}$ (in units of boardlike particle volume $v=l w t$) for the 6 independent orientational configurations of a pair of boardlike particles in the Zwanzig model as a function of the depletant number density $n_D$. Here the boardlike particles have dimensions $l/t=9.3$, $w/t=3.0$ and are in contact with a reservoir of ideal cubic depletants with side $d/t=1.0$ and number density $n_D$.}
\end{figure}
It is readily understood from Eq. (\ref{eq15}) that adding the depletion attraction Eq. (\ref{eq10}) to the cuboid-cuboid pairwise potential $u_{ii'}({\mathbf r})$ gives rise to a monotonic decrease of the coefficients $\widetilde{E}_{i i'}$ with $n_D$. This effect is depicted in Fig. \ref{fig2}, where we report the 6 independent values of the matrix elements $\widetilde{E}_{1 i}$, corresponding to the 6 two-particle configurations (1,1), (1,2), (1,3), (1,4), (1,5) and (1,6) (cf. Fig. \ref{fig1}), as a function of the reservoir depletant concentration $n_D$. In order to allow for a comparison with previous experimental \cite{vandenpol} and theoretical \cite{belli} work on the subject, the aspect ratios are chosen as $l/t=9.3$ and $w/t=3.0$, while for the cubic depletant we set $d/t=1.0$. At $n_D=0$ the 6 excluded-volume matrix elements are positive definite, but with increasing $n_D$ their value decreases until becoming negative (see $\widetilde{E}_{11}$ in Fig. \ref{fig1}). Such a behavior is well known from the study of systems of spherically symmetric particles with short-range attractive potentials, where one can define a temperature at which the second virial coefficient changes its sign (``Boyle temperature''). The change in sign of the second-virial coefficient is related to a tendency of the system to develop a gas-liquid phase separation. Also in the present case, where the role of the (inverse) temperature is played by the depletant density $n_D$, this change in sign can indicate a tendency towards a phase separation between two homogeneous phases. On the other hand, when the dimension of the depletant is sufficiently small, one expects the gas-liquid phase separation to be metastable with respect to a broad gas-solid coexistence \cite{dijkstra,dijkstra2,bolhuis,savenko}. As we ignore the stability of inhomogeneous phases like smectic, columnar or crystal states, in the present work we limit our investigations to values of $n_D$ small enough as to guarantee a positive value of all the effective excluded-volume matrix elements, and to avoid strong tendency towards a broad phase separation.
\begin{figure}
\includegraphics[scale=0.65]{fig3.pdf}
\caption{\label{fig3} Ratio between the second-virial coefficients corresponding to the two-particle configurations (1,4) and (1,5) of Fig. \ref{fig1} as a function of the depletant density $n_D$ for boardlike particles with $w/t=3.0$ and $l/t=8.7$ ($\nu=l/w-w/t=-0.1$), $l/t=9.0$ ($\nu=0.0$), $l/t=9.3$ ($\nu=0.1$), and $l/t=10.0$ ($\nu=0.33$). The solid circles highlight the value of the depletant density $n_D^*$, defined by the condition $\widetilde{E}_{14}=\widetilde{E}_{15}$.}
\end{figure}
Although the monotonic decrease with $n_D$ is a feature of all the 6 effective excluded-volume coefficients $\widetilde{E}_{1i}$, their rate of change is not the same. Let us focus on the coefficients corresponding to the two-particle configurations (1,4) and (1,5). In absence of depletant ($n_D=0$), $\widetilde{E}_{15}=E_{15}$ is slightly bigger than $\widetilde{E}_{14}=E_{14}$, but its first derivative at $n_D>0$ is smaller. As a consequence, there exists a value of the depletant density $n_D^*$, such that $\widetilde{E}_{14}<\widetilde{E}_{15}$ for $n_D<n_D^*$, and $\widetilde{E}_{14}>\widetilde{E}_{15}$ for $n_D>n_D^*$ (see inset of Fig. \ref{fig2}). This fact, which we will show to have deep consequences for the phase behavior of the system, is more clearly represented by the red solid curve of Fig. \ref{fig3}, representing the ratio $\widetilde{E}_{14}/\widetilde{E}_{15}$ as a function of $n_D$. In the same plot, we report the ratio $\widetilde{E}_{14}/\widetilde{E}_{15}$ relative to boards-cubes mixtures with fixed ratio $w/t=3.0$ and $d/t=1.0$, but different values of $l/t$ ($l/t=8.7$, $9.0$ and $10.0$). In the four cases we observe a monotonically increasing dependence of $\widetilde{E}_{14}/\widetilde{E}_{15}$ on $n_D$, which implies that the existence of $n_D^*$, defined by the condition $\widetilde{E}_{14}=\widetilde{E}_{15}$, is determined by the value of $E_{14}/E_{15}$, which in turn depends only on $l$, $w$ and $t$. In other words, $n_D^*$ exists only if $E_{14} \leq E_{15}$, with $n_D^*=0$ if $E_{14} = E_{15}$. On the contrary, if $E_{14}>E_{15}$ one has $\widetilde{E}_{14}>\widetilde{E}_{15}$ independently of the depletant density $n_D$.
Before addressing the physical consequences of the existence of the density $n_D^*$, it is worth seeing how the relative value of the excluded volume coefficients $E_{14}$ and $E_{15}$ determines the phase behavior of boardlike particles in absence of depletant (i.e. $n_D=0$). It is well-known that monodisperse hard boardlike particles are expected to undergo an $I N$ transition, where the particles spontaneously break the orientational symmetry by aligning along common directions in space \cite{straley}. The nematic phase emerging from the $I$ can be i) {\em uniaxial prolate} $N_+$ with alignment of the long axis $l$; ii) {\em uniaxial oblate} $N_-$ with alignment of the short axis $t$; iii) {\em biaxial} $N_B$ with alignment of the three axes of the particle. Following Onsager \cite{onsager}, the origin of this phase transition can be understood by considering that orientational ordering determines an increase in excluded-volume entropy, which compensates the decrease in orientational entropy. Therefore at fixed orientational entropy, when $E_{14}<E_{15}$ the $N_+$ phase will be thermodynamically favored over the $N_-$, the opposite being the case when $E_{14}>E_{15}$ (cf. Fig. \ref{fig1}). In the intermediate situation, when $E_{14}=E_{15}$, the system undergoes instead a direct second-order $I N_B$ transition. By explicitly calculating $E_{14}$ and $E_{15}$ in terms of $l$, $w$ and $t$, and defining a shape parameter $\nu = l/w-w/t$, one can show that
\begin{equation}
\frac{E_{14}}{E_{15}}
\begin{cases}
<1 \hspace{0.4cm} \Leftrightarrow \nu>0, \\
=1 \hspace{0.4cm} \Leftrightarrow \nu=0,\\
>1 \hspace{0.4cm} \Leftrightarrow \nu<0.
\end{cases} \label{eq16}
\end{equation}
This is consistent with Straley's result that a system of boardlike particles undergoes i) a first-order $I N_+$ transition if $\nu>0$; ii) a first-order $I N_-$ transition if $\nu<0$; iii) a second-order $I N_B$ transition if $\nu=0$ \cite{mulder}.
\begin{figure*}
\includegraphics[scale=0.7]{fig4.pdf}
\caption{\label{fig4} Phase diagrams of boardlike particles with aspect ratios $w/t=3.0$ and (a) $l/t=8.7$ ($\nu=-0.1$), (b) $l/t=9.0$ ($\nu=0.0$), (c) $l/t=9.3$ ($\nu=0.1$), (d) $l/t=10.0$ ($\nu=0.33$) in contact with a reservoir of cubic depletant with side length $d/t=1.0$ at number density $n_D$. The diagrams feature isotropic ($I$, green regions), prolate ($N_+$, red regions) and oblate ($N_-$, oblate regions) uniaxial and biaxial ($N_B$, yellow regions) nematic phases. The black circles highlight the Landau critical points, whereas the dotted lines indicate the limit of stability of nematic phases with respect to smectic ($Sm$) fluctuations along the long (red dotted line) and short (blue dotted line) particle axis respectively.}
\end{figure*}
The relation between the excluded-volume coefficients and the character of the $I N$ transition can be generalized to the case of boardlike particles immersed in a depletant, provided that the coefficients $E_{i i'}$ are substituted with $\widetilde{E}_{i i'}$. According to this interpretation, one is lead to identify the density $n_D^*$, defined by the condition $\widetilde{E}_{14}=\widetilde{E}_{15}$, with a Landau critical point at which the system undergoes a direct second-order $I N_B$ transition. Since the $N_B$ stability is generally due to a balanced competition between rodlike and platelike behavior, the {\em critical depletant density} $n_D^*$ is expected to divide the phase diagram into two distinct regions, one where the stable uniaxial nematic phase is prolate ($N_+$, corresponding to rodlike behavior), the others where the stable uniaxial nematic phase is oblate ($N_-$, corresponding to platelike behavior).
This picture is confirmed by the $(\eta,n_D)$ phase diagrams of Fig. \ref{fig4}, describing the phase behavior of boardlike particles with dimensions $w/t=3.0$ and (a) $l/t=8.7$ ($\nu=-0.1$), (b) $l/t=9.0$ ($\nu=0.0$), (c) $l/t=9.3$ ($\nu=0.1$), and (d) $l/t=10.0$ ($\nu=0.33$) immersed in a cubic depletant with sides $d/t=1.0$ and number density $n_D$. As a general feature, at packing $\eta \approx 0.2-0.3$ the system undergoes a phase transition from an $I$ phase (green region) to $N_+$ (red regions), $N_-$ (blue regions) or $N_B$ (yellow regions) states. The first-order character of the $I N_+$ and $I N_-$ transitions is not visible on the scale of Fig. \ref{fig4}. The dotted lines indicate the limit of stability of the homogeneous phases with respect to one-dimensional (smectic, $Sm$) fluctuations along the long axis $l$ (red dotted lines) or along the short axis $t$ (blue dotted lines). Therefore, the smectic bifurcation analysis confirms that inhomogeneous phases (white regions) preempt nematic states at sufficiently high packing fractions. In particular, the higher the depletant density $n_D$, the lower the stability of homogeneous phases with respect to inhomogeneous one. This result is in agreement with previous studies on the phase behavior of hard rods interacting via an attractive depletion potential \cite{bolhuis, savenko}. In the latter case, in fact, the coexistence regions between different phases increase with the depletant density leading, at sufficiently high $n_D$, to a wide isotropic-crystal coexistence and a consequent disappearance of the liquid crystal phases. We expect similar phenomena at depletant concentrations higher than those considered here, but a description beyond the second virial order would be needed in this case.
As deduced by the analysis of Fig. \ref{fig3}, in the case of ``platelike'' boards with $\nu<0$ (Fig. \ref{fig4}(a)) the absence of a Landau critical point implies that the $I$ phase undergoes a transition towards a $N_-$ phase for every value of $n_D$. If we instead consider a system of boardlike particles with $\nu=0$ (Fig. \ref{fig4}(b)), it is well known that in absence of depletant a direct second-order $I N_B$ transition is expected. In our picture, this corresponds to the presence of a critical depletant density at $n_D^*=0$, beyond which $N_-$ states appear at intermediate packing between the $I$ and the $N_B$ phase. More interesting conclusions can be drawn when considering ``rodlike'' boards with $\nu>0$, in which case the critical depletant density $n_D^*$ assumes non-zero values (Figs. \ref{fig4}(c) and (d)). If the pure system at $n_D=0$ is expected to develop an $I N_+$ transition, the attraction induced by the depletant reduces the $N_+$ stability until determining a direct $I N_B$ transition at $n_D=n_D^*$. Surprisingly, at even higher depletant densities ($n_D>n_D^*$), the stable uniaxial nematic phase in between $I$ and $N_B$ has oblate ordering $N_-$, in sharp contrast with the behavior of the pure rodlike boards system. Moreover, the phase diagrams of Figs. \ref{fig4}(c) and (d) suggest that, when dealing with boardlike particles with $\nu>0$, setting the depletant density at values close to $n_D^*$ allows to select regions of the phase diagram with relatively high $N_B$ stability. This is possible also when the regime of $N_B$ stability of the pure boardlike particle system is small (Fig. \ref{fig4}(c)) or even absent (Fig. \ref{fig4}(d)).
A relevant feature of the present analysis is that a critical depletant density $n_D^*$ exists only for slightly elongated, or rodlike, boards ($\nu > 0$). On the contrary, no Landau critical point is predicted when $\nu<0$, in which case for every value of $n_D$ the system develops a first-order $I N_-$ transition, typical of platelike particles. This fact can be interpreted in the following terms. At low enough depletant density ($n_D<n_D^*$) the role of the depletant is weak and the isotropic-nematic phase transition is driven by the gain in boardlike particles' excluded-volume entropy, leading to $N_+$ ($N_-$) ordering if $\nu>0$ ($\nu<0$). On the other hand, at high enough depletant density ($n_D>n_D^*$) the thermodynamically more favored states are those maximizing the depletant entropy, i.e. states where the overall depletion volume is minimized. It appears clear that, at fixed boardlike particles' orientational entropy and independently of the sign of $\nu$, $N_-$ rather than $N_+$ ordering tends to maximize the overlap between the depletion regions of single boards. Therefore, when $\nu>0$ the Landau critical point at $n_D^*$ appears as a result of a competition between the excluded-volume entropy of boardlike particles and depletant. Instead, when $\nu<0$ this competition does not happen since both entropies are maximized by $N_-$ states, and thus no critical depletant density exists.
\begin{figure}
\includegraphics[scale=0.65]{fig5.pdf}
\caption{\label{fig5} Critical depletant density $n_D^*$ as a function of the side $d$ of the cubic depletant for different boardlike particles with same shape parameter $\nu=0.1$: $l/t=9.3$ and $w/t=3.0$ (solid red line); $l/t=16.4$ and $w/t=4.0$ (solid green line); $l/t=25.5$ and $w/t=5.0$ (solid blue line). The dashed lines represent the approximate analytical dependence given by Eq. (\ref{eq17}). The inset illustrates the same data in terms of the critical depletant packing fraction $n_D^* d^3$.}
\end{figure}
One could wonder how the phase behavior of our system changes by varying its relevant parameters, that is, the dimensions of the boardlike particles $l$, $w$, $t$ and the size of the depletant $d$. The phase diagrams of Fig. \ref{fig4} highlight the monotonic increasing dependence of $n_D^*$ on $\nu$, when the dimensions of the depletant $d/t$ and the aspect ratios of one of the two boards (here $w/t$) are fixed. By numerically solving the equation $\widetilde{E}_{14}=\widetilde{E}_{15}$, we investigate also the role of the depletant dimension on the critical depletant density. In Fig. \ref{fig5} we report the critical depletant density $n_D^*$ as a function of the depletant side $d/t$ for three boardlike particle dimensions with different aspect ratios $l/t$ and $w/t$, but same shape parameter $\nu=l/w-w/t=0.1$. As a general trend, by increasing the size of the depletant, the critical depletant number density $n_D^*$ decreases, in accordance with the intuitive notion that the smaller the depletant, the more one needs to establish enough depletion attraction. Moreover, at fixed $\nu$ the critical depletion density decreases most for the more extreme aspect ratios of the particles. In other words, the bigger the aspect ratios $l/t$ and $w/t$ at fixed $\nu$, the smaller the amount of depletant needed to reach $n_D^*$. If instead of the number density one considers the critical depletant packing fraction $n_D^* d^3$, one sees that this quantity is an increasing function of $d$ (inset of Fig. \ref{fig5}). Therefore, one expects the ideal depletant approximation (cf. Sec. \ref{sec:model}) to be increasingly reliable in the limit of small depletant.
In practical situations, one could be interested in estimating the critical depletant density $n_D^*$, which is defined as the solution of the non-linear equation $\widetilde{E}_{14}=\widetilde{E}_{15}$ and which has then to be calculated numerically. If $n_D^*$ is sufficiently small, one can obtain an approximate expression for this quantity by linearizing both sides of equation $\widetilde{E}_{14}=\widetilde{E}_{15}$ in $n_D$. The approximate critical depletant density is given by the following expression
\begin{equation}
n_D^* = \frac{2 \bigl [t(l+w)^2 - l(w+t)^2 \bigr](l-t)^{-1}}{2(lt-w^2) d^3 + w (lw+tw-2lt) d^2},
\label{eq17}
\end{equation}
and it can be compared (dotted lines of Fig. \ref{fig5}) with the numerical calculation (solid lines), showing good overall agreement, which improves the larger the depletant side $d/t$.
\section{Conclusions}
\label{sec:conclusions}
In the present paper we investigate for the first time the effect of a short-range depletion-induced attraction on the liquid-crystal phase behavior of boardlike particles. To this aim, we make use of classical density functional theory truncated at second virial order, and adopt the Zwanzig model for the description of the orientational degrees of freedom. In close analogy with the Asakura-Oosawa-Vrij model for mixtures of spheres, by neglecting interactions between the cubic depletant particles we can explicitely calculate the effective two-body attractive depletion potential between boardlike particles.
We predict that in systems of slightly elongated boardlike particles ($\nu>0$), there exists a critical depletant density at which the uniaxial nematic phase is substituted by a direct second-order transition from the isotropic to a biaxial nematic phase. At higher depletant concentrations, a large region of oblate uniaxial nematic ordering develops, rendering the system of attractive rodlike boards behaving like a system of hardly repulsive platelike boards. The origin of this phenomenon is due to two competing mechanisms: the maximization of the boardlike particle entropy, favoring $N_+$ ordering, and the maximization of the depletant entropy, favoring $N_-$ ordering.
The phase behavior described in this work shares many similarities with our findings in Ref. \cite{belli}, where we showed that size-polydispersity in a system of hard boardlike particles with the same shape and different volume induces the appearance of a Landau tetracritical point at a specific system composition. This fact is related to a competition between prolate and oblate ordering, which in turn is realized only when the boardlike particles are slightly elongated. In the light of our present findings, we suggest that this $N_+$-$N_-$ competition and the corresponding emergence of a Landau tetracritical point can be understood in terms of a depletion effect. More specifically, when size-polydispersity becomes relevant, $N_-$ rather than $N_+$ ordering determines the higher total entropy due to the minimization of the overall depletion regions of the big particles with respect to the smaller ones. In further analogy with the boards-depletant mixture, no such competition is predicted for platelike boards and consequently no tetracritical point appears in this case.
Besides furnishing an explanation for the results of Ref. \cite{belli}, we suggest that manipulating the attraction induced by a depletant, e.g. a non-adsorbing polymer, furnishes an original and effective way to control the phase behavior of boardlike particles, allowing to stabilize prolate and oblate uniaxial and biaxial nematic states. Moreover, the depletant density is expected to be an easy experimental control parameter.
\section{Acknowledgements}
It is our pleasure to thank R. Kamien for suggesting, during the 8th Liquid Matter Conference in Vienna, to interpret the results of Ref. \cite{belli} as a depletion effect. We would also like to thank R. Ni for stimulating discussions.
This work is financed by an NWO-VICI grant and is part of the research program of the ``Stichting door Fundamenteel Onderzoek der Materie (FOM)'', which is financially supported by the ``Nederlandse Organisatie voor Wetenschappelijk Onderzoek (NWO)''.
|
\section{\label{sec:level1}Introduction}
A solid sphere on an inclined surface rolls down under the action of gravity. A rectangular object however is more likely to slide down. The choice of sliding versus rolling motion is determined by the shape and weight of the body and the frictional forces at the supporting surface. On the other hand, a liquid drop of a given volume on an inclined surface can attain a variety of shapes. The static shape depends on the equilibrium contact angle, solid surface characteristics including the pinning location, interfacial tension between the substrates, gravity and plate inclination, while a moving drop has to contend in addition with viscous and inertial stresses \cite{Book_deGennes}. Moreover, the reaction forces and moments provided by the supporting surface are distributed and depend strongly on the shape, see e.g. \cite{sumesh_2011,sumesh2010}. These differences between a solid and a liquid make the study of the latter complex, but pose an interesting question, viz. whether a liquid drop sitting on an inclined solid surface will roll, slide, or do both.
The motion of the liquid drops on solid surfaces is much studied, especially in the two limiting cases of perfect wetting and zero wetting. When the contact angle is small, the lubrication approximation of the Navier-Stokes equation is a good model \cite{Dussan_1983}, to describe the dynamics, with a prescribed parabolic velocity profile. The sliding motion of the drops and the associated instabilities of the receding front has received much attention theoretically and experimentally, see e.g. \cite{limat_2005, limat_2005b}. At the other limit, when the contact angle approaches $180 ^{\circ}$, a complete rolling of the drop is observed under certain situations \cite{quere_2005}. An anomalous increase in the speed of smaller drops has been observed on super hydrophobic surfaces \cite{quere_1999} and explained based on the scaling arguments in the model of \cite{mahadevan_1999}. The shape of these droplets as they roll down has been studied by \cite{quere_2004}. While these two limits are well studied, intermediate contact angles are studied much less. They are harder to analyze since they do not lend themselves to simplifying approximations. While analytical solutions are practically impossible, numerical solutions pose considerable challenges due to the multiple length scales present and the coupling between the evolving field and the interface shape \cite{ShikhmurzaevBook}. Here we analyze the entire spectrum of shapes for two-dimensional drops, for a range of the relevant non-dimensional parameters. A splitting of the motion into sliding, shear and rolling leads to a better understanding of the dynamics.
The competition between rolling and sliding motion of droplets on hydrophobic surfaces has been investigated experimentally \cite{Sakai_2006, Sakai_2008}. In these studies the effect of surface coating and roughness on the internal fluidity of the droplet was of primary concern. The velocity inside the droplet was obtained by particle image velocimetry, and the contribution to slip versus to roll was evaluated for an accelerating drop on an inclined surface. Rolling motion is clearly observed in molecular dynamics \cite{Muller_2008}, and lattice Boltzmann, simulations \cite{Moradi_2011} on cylindrical drops. However, such studies have concentrated on the total velocity of the droplet and its dependence on the driving force and the contact angle. With a somewhat different goal, we investigate here the motion of cylindrical drops of various contact angles on smooth surfaces, where the effect of fluid properties like density and viscosity, and also of slip are evaluated.
The idea of splitting the motion into shear and roll is standard in fluid mechanics. However, the standard splitting does not distinguish global rotation from local rotation of the fluid element. Such distinction is necessary to understand the global dynamics of a drop, so we distinguish between these, in the manner introduced by \cite{kolar_2007} in a different context. A splitting of the velocity field in a drop into slip and roll was done recently by \cite{yeomans_2010}, although in a different manner than done here. The linear part of the velocity profile on a line passing through the center of the drop was attributed to rotation whereas we will see in section \ref{sec:kolar} that this would overestimate the rotation inside the drop. In addition, the study of \cite{yeomans_2010} is valid for drop sizes smaller than the capillary length, so large deformations from the circular shape were not under consideration, while the present approach has no such restriction. We show that the shape, and hence the size is very important in determining the amount of rolling inside the drop.
Movement of the contact line, by slipping or otherwise, is imperative for a moving drop unless the contact angle is exactly $180 ^{\circ}$ \cite{mahadevan_1999}. Though the exact mechanism of contact line movement is not understood \cite{eggers_2009}, it is generally believed that the macroscopic behavior of the drop is independent of the assumptions at the contact line. However, it is clear that even a small amount of global rotation can manifest itself as tank-treading \cite{Dussan_1974} near the contact line, which means that contact line behavior is not local in nature, and must be consistent with the macroscopic motion. We demonstrate this using the diffuse interface (DI) model. Most earlier studies have concentrated on the rolling motion near the contact line \cite{Clarke_1995, Rame_1996} while we look at the rolling motion in the bulk of the drop. The association between the two, if any, implies the nonlocal hydrodynamic effects of the contact line movement \cite{ShikhmurzaevBook}.
Understanding the kinematics is not just a curiosity but has practical relevance too. For example, rolling droplets play an important role in self-cleaning devices. As they roll, they pick up and remove dirt as observed on hydrophobic surfaces \cite{rothstein_2010}. Hence it is desirable to know under what situations one can maximize the rolling motion inside the drop.
\section{\label{sec:level2}Theory: Diffuse Interface Model and Hydrodynamics}
We use a coupled system of equations describing the hydrodynamics of a conserved order parameter $\psi$ and the conserved momentum density $\rho {\bf u}$, where $\rho$ and ${\bf u}$ are the total density and the local fluid velocity.
\subsection{Landau-Ginzburg theory}
For a binary fluid system consisting of species $I$ and $II$ with local densities $n_I$ and $n_{II}$, the order parameter is defined as the normalized density difference, $\psi = \frac{n_{II} - n_{I}}{n_{II} + n_{I}}$ which quantifies the local composition. The equilibrium thermodynamics of the fluid is described by the Landau free-energy functional \cite{ChaikinBook, RowlinsonBook}
\begin{equation}
F(\psi) = \int(f(\psi)+\frac{K}{2}\left|\nabla\psi|^2\right)d\mathbf{r},
\label{eqn:Feng}
\end{equation}
where $\mathbf{r}$ stands for the spatial dimensions. The first term represents the local free energy density of the bulk fluid, and is approximated as $ f(\psi) = \frac{A}{2}\psi^2 + \frac{B}{4} \psi^4$ with $A<0$ and $B>0$. The three parameters $A$, $B$, and $K$ control the interfacial thickness and interfacial energy of the mixture. The second term of Eq. \ref{eqn:Feng} involving the square gradient gives a free energy cost to any variation in the order parameter, and is related to the interfacial tension between the two fluid phases \cite{Kendon2001}. Two uniform solutions $\psi = \pm \sqrt{A/B}$ can coexist across a fluid interface. For a planar interface; the concentration profile between the two bulk phases is given by
\begin{equation}
\psi(z) = \sqrt{\frac{A}{B}}\tanh{\frac{z}{\xi}},
\label{eqn:tanh}
\end{equation}
where $z$ is the coordinate normal to the interface while $ \xi = \sqrt{\frac{2K}{A}}$ determines the interfacial thickness. The energy associated with this profile in excess of the energy in the bulk, defined per unit area, provides the interfacial tension $ \gamma = \frac{2}{3} \sqrt{\frac{2KA^3}{B^2}}$. The corresponding chemical potential is given by the variational derivative of the free energy with respect to the order parameter $\mu = \delta F/\delta \psi = A \psi + B \psi^3 - K \nabla^2 \psi$. Gradients in the order parameter produce additional stresses, which follow from the relation $\psi\nabla\mu = \nabla \cdot \boldsymbol{\sigma}^{\psi}$ \cite{anderson1998}, including Laplace and Marangoni stresses due to a fluid-fluid interface.
\subsection{Governing equations}
The order parameter is described by a Cahn-Hilliard equation (CHE), which includes advection by fluid flow and relaxation due to chemical potential gradients,
\begin{equation}
\partial_t \psi + \nabla \cdotp \left( \mathbf{u}\psi\right) = \nabla \cdotp \left( M\nabla \mu \right).
\label{eqn:advdiff}
\end{equation}
The mobility $M$ is the constant of proportionality in the linear phenomenological law relating the thermodynamic flux of $\psi$ to the thermodynamic force $\nabla \mu$. The order parameter dynamics is coupled to a Navier-Stokes equation (NSE) \cite{LandauBook} with additional stress densities arising from the order parameter. For an incompressible fluid, the dynamics is governed by
\begin{eqnarray}
\partial_t (\rho \mathbf{u}) + \nabla \cdotp (\rho \mathbf{uu}) = -\nabla p + \eta \nabla^2 \mathbf{u} + \psi \nabla \mu + \mathbf{G}
\label{eqn:ns}
\end{eqnarray}
together with the continuity equation for the density. In the above, $p$ stands for the isotropic contribution of the pressure, $\eta$ is the shear viscosity and $\mathbf{G}$ is the gravitational force density.
\subsection{Numerical Algorithm}
We now briefly review the numerical algorithm used to solve these coupled equations. We use a hybrid algorithm by combining the lattice Boltzmann method for hydrodynamics and method of lines for the order parameter dynamics. The interested reader is referred to \cite{sumeshlb_2011} for a detailed description.
The Lattice Boltzmann (LB) method for solving the Navier-Stokes equations is modified to include force densities such as the divergences of order parameter stresses and gravity \cite{nash2008}. In a standard $DdQn$ LB model where the velocity space is discretized into $n$ components in $d$ dimensional space, the discrete form of the Boltzmann equation reads
\begin{equation}
\partial_t f_i + \mathbf{c}_i \cdot \nabla f_i + [{\bf F}\cdot\nabla_{\bf c}f]_i= - \sum_j L_{ij}( f_j -f_j^0),
\label{eqn:flbe}
\end{equation}
where $\mathbf{F}(\mathbf{x},t)$ is an effective force density. The moments of the single particle distribution function $f_i$, defined at lattice node $\bf{x}$ with velocity $\mathbf{c}_i$ at time $t$, give the fluid mass, momentum and stress densities:
\begin{equation}
\rho = \sum_{i=0}^{n} f_i, \hspace{2mm} \rho\mathbf{v} = \sum_{i=0}^{n} f_i\mathbf{c}_i, \hspace{2mm}S_{\alpha \beta} = \sum_{i=0}^{n}{f}_i Q_{i \alpha \beta},
\label{eqn:momreln}
\end{equation}
where $Q_{i\alpha\beta} = c_{i\alpha}c_{i\beta} - c_s^2 \delta_{\alpha\beta}$. The collision operator $L_{ij}$ is the discrete form of the collision integral. It controls the relaxation of $f_j$ to equilibrium, $f_j^0$, and may be modeled in terms of a single relaxation time $\tau$ as $L_{ij} = \delta_{ij}/\tau$. The viscosity is then obtained as $\eta = \tau c_s^2$ where $c_s = 1/\sqrt{3}$ is the sound speed in LB units.
The spatial discretization of the Cahn-Hilliard equation is based on a finite-volume formulation \cite{RotenBerg2008}. For a given node, the divergence is written as a sum of fluxes defined on the midpoint of the link connecting the node to its neighbors. The resulting equation is temporally integrated using a Runge-Kutta algorithm \cite{sumeshlb_2011}.
A liquid drop will typically be of higher viscosity and density than its surroundings. Secondly, the contact line must be treated with care. In the present work, these features are incorporated into the numerical approach as discussed below.
\subsubsection{Viscosity Ratio}
\begin{figure}
\includegraphics[trim = 0 0 0 0, clip, width=\linewidth]{comparison_realcurve.eps}
\caption{ Comparison of 2D channel flow velocity profiles of an immiscible binary system and its approximations in the DI framework. The thick continuous line is the velocity profile of an immiscible binary fluid. In the simulations, the binary fluid was approximated as a viscosity stratified fluid flow according to Eq. \ref{eqn:vislin} and \ref{eqn:vispow}, the results are shown by red stars and blue circles respectively. The corresponding analytical solutions for viscosity-stratified flow are shown by the red dashed and blue solid lines respectively. While LB perfectly reproduces the flow for a given viscosity stratification, the use of a thin viscosity stratified layer itself is a good approximation for an immiscible binary fluid system.}
\label{fig:viscontrast}
\end{figure}
The differences in properties between the two fluids could be introduced at a molecular level, as done by \cite{Luo2003, asinari_2005}. Our approach is macroscopic, and the simplest way to introduce a viscosity difference across the fluid interface is to prescribe the relaxation time as a function of the order parameter. The underlying assumption is that the molecular structure of two fluids is the same, and is analogous to the introduction of interfacial tension using Cahn-Hilliard theory. To test this, we first design a model problem, of the laminar pressure-driven flow of two fluids in a two-dimensional channel. Fluid I is of higher viscosity $\eta_{I}$, and occupies the lower portion of the channel, while fluid II of lower viscosity $\eta_{II}$ occupies the upper portion.
In the literature we find two different expressions for the relationship between the concentration and the relaxation time.
(i) Effective relaxation time as a polynomial function of order parameter, the simplest being linear \cite{zhang_1999, eggert_1993},
\begin{equation}
\tau = 0.5[\tau_{I}(1-\psi)] + [ \tau_{II}(1+\psi)].
\label{eqn:vislin}
\end{equation}
(ii) As Arrhenius suggested, prescribe an effective relaxation time as a product of relaxation times, but raised to a power proportional to concentration \cite{yeomansvis_2000}
\begin{equation}
\tau = \tau_{I}^{(\frac{1-\psi}{2})} \tau_{II}^{(\frac{1+\psi}{2})}.
\label{eqn:vispow}
\end{equation}
We test both relationships by simulating the two-fluid flow described above, first defining the viscosity ratio simply by $\eta_r = \eta_I/ \eta_{II} = \tau_I/\tau_{II}$. The velocity profile from simulations using the two expressions above are compared to the analytical solution for two immiscible fluids separated by a sharp interface in Fig. \ref{fig:viscontrast}. When the viscosity contrast is small, both match well with the analytical solution. However, when the ratio is large ($\eta_r$ = 10), Eq. \ref{eqn:vispow} is closer to the immiscible result than Eq. \ref{eqn:vislin}. This is because the effective mixed layer where the viscosity varies between $\eta_I$ and $\eta_{II}$ is smaller by the former relationship. This is evidenced by the fact that a corresponding analytical solution for the laminar velocity profile for the parallel flow of two miscible fluids with a thin mixed region between them agrees in each case with the computed result. We have thus shown that this approach is a good one for incorporating viscosity contrasts in LB simulations. Note that no ad-hoc fixes are needed. We have used Eq. \ref{eqn:vispow} in our calculations.
\subsubsection{Gravity}
The hybrid algorithm written using a single particle distribution function is modified to include gravitational effects as follows. Following \cite{greated_2000}, we may write the gravitational force acting on the fluids as,
\begin{equation}
\mathbf{G} = \frac{\rho}{2} [g_{I}(1-\psi)+g_{II}(1+\psi)] \mathbf{1_g},
\label{eqn:grav}
\end{equation}
where $\mathbf{1_g}$ is the direction in which gravity is acting. Therefore, when $\psi=1$, $\mathbf{G} = \rho g_{II} \mathbf{1_g}$ and when $\psi = -1$, $\mathbf{G} = \rho g_{I} \mathbf{1_g}$. At the interface when $\psi=0$, $\mathbf{G} = \frac{\rho}{2}(g_{I}+g_{II})\mathbf{1_g}$. Thus $g_{I}/g_{II}$ determines the density ratio between two fluids. In order to ensure that we are in the incompressible limit, we must have $|\mathbf{G}z| << \rho c_s^2$ where $z$ is the vertical extent of the simulation domain, which means that thermodynamic pressure is large compared to the hydrostatic pressure difference.
In our simulations, wall boundary conditions are applied on two sides of the domain. Thus a computation of a single component fluid, with gravity prescribed along the flow direction, develops a channel flow between the walls. For droplet simulations, we implement the body force only on one fluid, which is equivalent to solving the NSE with the Boussinesq approximation as described below. For small density variations, i.e., $\Delta \rho \equiv g_I - g_{II} << g_{I}$, the Boussinesq approximation provides that we may neglect the density variation everywhere in the NSE except in the buoyancy term, $\mathbf{G}$ of Eq. \ref{eqn:ns}. For simplicity we may further absorb the body force $\rho g_{II}$ into the pressure term by redefining pressure. Density in the interface region is prescribed as a linear function of the order parameter. Note that the validity of our simulations is thus limited to situations where the Boussinesq approximation holds.
\subsubsection{Wetting Boundary Conditions}
\begin{figure}
\includegraphics[width=\linewidth]{cont.eps}
\caption{Verification of the implementation of wetting conditions, Eq. \ref{eqn:walleng}, on the walls by comparing the numerically obtained contact angle with Eq. \ref{eqn:cont}. The agreement is good except at very large and small contact angles.}
\label{fig:cont}
\end{figure}
The algorithm implemented to get the correct contact angle on the wall is based on \cite{Desplat2001, yeomanscont_2002}. The solid-fluid surface tensions are introduced by defining the Landau free energy functional
\begin{equation}
F = F_{bulk}(\psi) + \int f(\psi_s) ds,
\end{equation}
where $\psi_s$ is the value of order parameter at the wall. Minimization of this energy functional near the wall gives a relation between energy gradient and the gradient of the order parameter
\begin{equation}
\frac{df_s}{d\psi_s} = k \nabla \psi \cdot \mathbf{n},
\end{equation}
where $\mathbf{n}$ is normal to the wall. The form $ f_s = \frac{C}{2} \psi_s^2 + H \psi_s$ is known to be sufficient to produce various wetting behavior. By tuning the parameters $C$ and $H$ we can modify the properties of the surface. If $H=0$ we have neutral wetting. Nonzero values of $H$ therefore allows an asymmetry in the surface value of the order parameter and a contact angle different from $90^{\circ}$. Therefore we have,
\begin{equation}
\frac{df_s}{d\psi_s} = C \psi_s + H = K \nabla \psi \cdot \mathbf{n}.
\label{eqn:walleng}
\end{equation}
It is found sufficient to retain only the linear term of the surface energy functional \cite{Desplat2001, yeomanscont_2002}, i.e., to set $C=0$. We use a second order central difference formula to calculate the normal derivative of order parameter at the wall. Thus, $ H = K \frac{\psi_1 - \psi_0}{\Delta x}$, where the subscripts $0$ and $1$ represents the $0^{th}$ and first node respectively, which may be used to obtain the order parameter $\psi_0$ at the boundary node. The wall is placed at the $\frac{1}{2}$ location, as is usual in the bounce back schemes used to represent wall in LB procedures \cite{SucciBook}. Defining a parameter $h\equiv H \sqrt{\frac{2}{kB}}$, the contact angles may be calculated as
\begin{equation}
\cos{\theta} = \frac{1}{2} \left[ (1+h)^{3/2} - (1-h)^{3/2} \right].
\label{eqn:cont}
\end{equation}
In addition $\mu_0 = \mu_1$ is imposed to ensure no order parameter flux into the wall, up to second order accuracy. Also, advection terms have been carefully discretized to preserve mass conservation upto machine accuracy \cite{sumeshlb_2011}. The implementation of the wetting properties of the wall is verified by comparing the static equilibrium angle obtained from the simulation to that from Eq. \ref{eqn:cont} as shown in Fig. \ref{fig:cont}. Deviations are seen at very large and very small contact angles and this appears to be consistent with the literature \cite{Desplat2001, yeomanscont_2002} (both used kinetic schemes for CHE). Both the grid size and the interfacial thickness were doubled separately to ensure that the deviations seen are not due to the contact line pinning. One reason for the deviation may just be the error involved in the derivation of Eq. \ref{eqn:cont} where $h$ is assumed to be small and Taylor series expansions are used. In spite of this, we restrict our simulations as far as possible to the wide range of intermediate contact angles where we are sure the simulations capture the relevant physics in this respect.
\subsection{Measure of Rolling Motion}
\label{sec:kolar}
\begin{figure}
\subfigure[\quad $\psi$]{\includegraphics[trim=50 120 60 130, clip, width=0.3\linewidth]{drop.eps}\label{fig:drop}}
\subfigure[\quad $v_{cm}$]{\includegraphics[trim=50 100 60 120, clip, width=0.3\linewidth]{dropvel.eps}\label{fig:dropvel}}
\subfigure[\quad $\omega$]{\includegraphics[trim=50 120 0 90, clip, width=0.3\linewidth]{tvort.eps}\label{fig:tvort}}\\
\subfigure[\quad $W$]{\includegraphics[trim=50 120 0 80, clip, width=0.3\linewidth]{weiss.eps}\label{fig:weiss}}
\subfigure[\quad $m$]{\includegraphics[trim=50 120 0 80, clip, width=0.3\linewidth]{kvn.eps}\label{fig:kvn}}
\subfigure[\quad $\omega_{res}$]{\includegraphics[trim=50 120 0 80, clip, width=0.3\linewidth]{rvort.eps}\label{fig:rvort}}
\caption{(a) $\psi$, order parameter field in the entire simulation box, two colors for the two different fluids with a thick interface. (b) $v_{cm}$, velocity field in the center of mass frame of the drop, the rolling motion of the drop may be observed. (c) $\omega$, the vorticity field inside the drop, regions concentrated vorticity may be noticed near the contact line. (d) $W$, a measure for the Weiss criterion to distinguish regions of high shear from vortical regions. (e) $m$, the kinematic vorticity number, again a measure of highly vortical regions, (f) $\omega_{res}$, residual vorticity showing the regions associated with solid body rotation.}
\label{fig:vortmeas}
\end{figure}
Before discussing the simulations and results, we take a typical simulated drop and discuss how the slide, shear and roll may be estimated. The droplet is illustrated in \ref{fig:drop}. The rolling motion inside the drop is evident in the corresponding velocity field, plotted in the center of mass reference frame of the moving drop as illustrated in \ref{fig:dropvel}. Our objective is to quantify this rotation. The first quantity to look at would be the vorticity field, shown in \ref{fig:tvort}. The velocity gradient tensor is usually split into a symmetric part $\mathbf{S}$ and an antisymmetric part $\boldsymbol{\Omega}$ as
\begin{equation}
\nabla \mathbf{u} = \mathbf{S} + \boldsymbol{\Omega}.
\label{split}
\end{equation}
$\mathbf{S}(x,y)$ for shear tensor is a measure of the deformation of a small fluid element located at $(x,y)$, while the vorticity tensor $\boldsymbol{\Omega}(x,y)$ identifies the angular velocity of the fluid element, with $(x,y)$ as the center of rotation. There are different measures available in the literature to identify regions of high vorticity relative to shear \cite{Vorticity_book}, and most of them have been derived in the context of turbulence. A commonly used measure is $W = \frac{1}{2}||\boldsymbol{\Omega}||^2 - ||\mathbf{S}||^2$ of the Weiss criterion. This is illustrated in Fig \ref{fig:weiss} for the drop under consideration. Another measure is the kinematic vorticity number, defined as $m=||\boldsymbol{\Omega}||/||\mathbf{S}||$ where $||[.]|| = \text{trace}([.]\cdot [.]^{T})^{1/2}$. This quantity is plotted in Fig \ref{fig:kvn}. Different criteria have been developed later on to define a vortex exactly and many of them reduce in two dimensions to the Weiss criterion \cite{Hussain_1995}. However the main drawback of these measures is that they all estimate vorticity, which does not in general give a direct measure of rolling motion. This is because the vorticity $\boldsymbol{\omega} = \nabla \times \mathbf{u}$ is a local quantity which includes both solid body rotation and shearing motion of a fluid element, and thus cannot distinguish between them. The residual vorticity which we describe below is shown in \ref{fig:rvort}. Although all three measures broadly describe the region of high rotationality in a similar fashion, only the last is good for obtaining a quantitative estimate of solid body rotation.
\begin{figure}
\subfigure[]{\includegraphics[trim = 40 250 30 190, clip, width=\linewidth]{split.eps}\label{fig:split}}\\
\subfigure[]{\includegraphics[trim = 40 250 30 190, clip, width=\linewidth]{convert.eps}\label{fig:convert}}\\
\subfigure[\quad Example I]{\includegraphics[trim = 40 250 30 190, clip, width=\linewidth]{split1.eps}\label{fig:split1}}
\subfigure[\quad Example II]{\includegraphics[trim = 40 250 30 190, clip, width=\linewidth]{split2.eps}\label{fig:split2}}\\
\caption{(a) Standard velocity gradient decomposition into symmetric and antisymmetric parts for a a flow with $u_x=0.1$, $u_y=0.3$, $v_x=-0.2$, $v_y=-0.1$. (b) The strain rate tensor in the principal coordinate system is rotated by $45^{\circ}$ and added to the antisymmetric tensor to generate the same flow in BFR. (c) The same flow is decomposed into a simple shear flow and residual flow, the residual being purely rotational flow. (d) A strain dominated flow, $u_x=0.1$, $u_y=0.2$, $v_x=0.1$, $v_y=-0.1$ is decomposed into simple shear and residual flow, with residual consists of only straining flow.}
\label{fig:tdm}
\end{figure}
To demonstrate that these conventional measures of vorticity are inadequate in describing the global rolling motion inside a drop, let us discuss a simple flow configuration, namely, a shear flow with $u = \dot{\gamma} y$ where $x$ is the flow direction and $y$ is the gradient direction. The velocity gradient tensor $\nabla \mathbf{u}$ splits into its symmetric and antisymmetric parts as follows:
\[
\nabla \mathbf{u} =
\begin{pmatrix}
0 & \dot{\gamma}\\
0 & 0
\end{pmatrix}
=
\begin{pmatrix}
0 & \dot{\gamma}/2\\
\dot{\gamma}/2 & 0
\end{pmatrix}
+
\begin{pmatrix}
0 & -\dot{\gamma}/2\\
\dot{\gamma}/2 & 0
\end{pmatrix}
= \mathbf{S} + \boldsymbol{\Omega}.
\]
Although $\boldsymbol{\Omega}$ is non-zero, there is no rolling motion, which tells us that we need a modified vorticity to characterize rolling motion.
To demarcate a coherent vortex correctly from high shear regions, Kolar \cite{kolar_2007} proposed a scheme to remove the `shear' vorticity from the total vorticity. He proposed a triple decomposition of the relative motion of a fluid element, where the velocity gradient tensor is split into a straining part, a rigid body rotation and a simple shear flow part. For clarity, we illustrate these pictorially in Fig. \ref{fig:tdm}. Given a velocity field in two dimensions, $\mathbf{u}$, the velocity gradient tensor (Fig. \ref{fig:split}) is a $2 \times 2$~matrix
\begin{eqnarray}
\begin{pmatrix}
u_x & u_y \\
v_x & v_y
\end{pmatrix}
=
\begin{pmatrix}
u_x & \frac{u_y+v_x}{2} \\
\frac{u_y+v_x}{2} & v_y
\end{pmatrix}
+
\begin{pmatrix}
0 & \frac{u_y-v_x}{2} \\
\frac{v_x-u_y}{2} & 0
\end{pmatrix}.
\end{eqnarray}
The symmetric part can be diagonalized to give $\left(\begin{smallmatrix} s/2 & 0 \\ 0 & -s/2 \end{smallmatrix}\right)$ where $s=\sqrt{4u_x^2+(u_y+v_x)^2}$. This is the strain rate tensor in the principal coordinates, see Fig. \ref{fig:convert}, which represents the total straining of the fluid element. The rotation tensor being antisymmetric will not change with a rotation of the coordinate system, and remains as $\left(\begin{smallmatrix} 0 & -\omega/2 \\ \omega/2 & 0 \end{smallmatrix}\right)$ where $\omega = v_x-u_y$, the vorticity. Therefore, in the principal axis coordinates, the velocity gradient tensor is, $\left(\begin{smallmatrix} s/2 & -\omega/2 \\ \omega/2 & -s/2 \end{smallmatrix}\right)$. We now rotate the coordinate system further by $\pi/4$ (Fig. \ref{fig:convert}). This frame is called a basic frame of reference (BFR), and the velocity gradient tensor in this frame is
\[
\begin{pmatrix}
0 & (s-\omega)/2\\
(s+ \omega)/2 & 0
\end{pmatrix}
=
\begin{pmatrix}
0 & s/2\\
s/2 & 0
\end{pmatrix}
+
\begin{pmatrix}
0 & -\omega/2\\
\omega/2 & 0
\end{pmatrix}.
\]
In this reference frame, the contribution due to shear is maximized in a triple decomposition of $\nabla \mathbf{u}$. One may write \cite{kolar_2007}
\begin{eqnarray}
\nabla u &=& \nabla u_{shear} + \nabla u_{residual} \nonumber\\
&=& \mathbf{S}_{shear} + \boldsymbol {\Omega}_{shear} + \mathbf{S}_{residual} + \boldsymbol {\Omega}_{residual}. \nonumber
\end{eqnarray}
Only one of the residual terms $\mathbf{S}_{residual}$ or $\boldsymbol {\Omega}_{residual}$ will be non-zero. The residual matrices above are constructed as explained below. The residual vorticity and strain may be written respectively as
\begin{eqnarray}
\omega_{res} &=& 0 \quad \text{if} \quad |s|\ge|\omega| \nonumber\\
& = & \text{sgn}(\omega) \left[|\omega|-|s|\right] \quad \text{if} \quad |s|\le|\omega|,\nonumber\\
s_{res}&=& \text{sgn}(s)\left[|s|-|\omega|\right] \quad \text{if} \quad |s| \ge |\omega|\nonumber \\
&=& 0 \quad \text{if} \quad |s| \le |\omega|.\nonumber
\end{eqnarray}
An example flow which is vorticity dominated, where $|s|<|\omega|$ is illustrated in figure \ref{fig:split1}. For this case we may write
\begin{eqnarray}
\nabla \mathbf{u}_{BFR} =
\begin{pmatrix}
0 & s/2\\
s/2 & 0
\end{pmatrix} +
\begin{pmatrix}
0 & -sgn(\omega)|\frac{s}{2}|\\
sgn(\omega)|\frac{s}{2}| & 0
\end{pmatrix} +\nonumber\\
\begin{pmatrix}
0 & 0\\
0 & 0
\end{pmatrix}+
\begin{pmatrix}
0 & -sgn(\omega)\frac{|\omega|-|s|}{2}\\
sgn(\omega)\frac{|\omega|-|s|}{2} & 0\nonumber
\end{pmatrix},
\end{eqnarray}
where the first two matrices will combine to produce a simple shear flow. The residual straining, shown by the third matrix, is zero. The portion of the velocity gradient tensor contributing to solid body rotation is given by the fourth matrix as illustrated in Fig. \ref{fig:split1}. On the other hand, for a flow which is strain dominated, i.e. $|s|>|\omega|$, an example of which is demonstrated in Fig. \ref{fig:split2}, we write
\begin{eqnarray}
\nabla \mathbf{u}_{BFR} =
\begin{pmatrix}
0 & sgn(s)|\frac{\omega}{2}|\\
sgn(s)|\frac{\omega}{2}| & 0
\end{pmatrix} +
\begin{pmatrix}
0 & \frac{-\omega}{2}\\
\frac{\omega}{2} & 0
\end{pmatrix} + \nonumber\\
\begin{pmatrix}
0 & sgn(s)\frac{|s|-|\omega|}{2}\\
sgn(s)\frac{|s|-|\omega|}{2} & 0
\end{pmatrix} +
\begin{pmatrix}
0 & 0\\
0 & 0\nonumber
\end{pmatrix}.
\end{eqnarray}
Again the first two matrices on the right hand side together produce a simple shear flow. The remainder is seen in the third matrix to be purely straining, as illustrated in Fig. \ref{fig:split2}. We may now see that for the simple shear flow discussed above, $|s|=|\omega|$ and $\omega_{res} = 0$, while for a solid body rotation, $\omega=\omega_{res}$. If $|s|<|\omega|$, the flow is vorticity dominated, and the residual tensor will consist of only rotation, and vice versa. Therefore residual vorticity can characterize the rolling motion inside a drop.
This separation of the shear vorticity from residual is different from the shear and curvature vorticity used by the atmospheric science community. There the shear vorticity is defined as $-\frac{\partial u}{\partial n}$ where $u$ is the local velocity along the streamline and $\mathbf{n}$ is normal to the stream line. The remainder is defined as the `curvature vorticity', which is associated with the curvature of streamlines. Note that these values will not be Galilean invariant. This procedure, when applied to a solid body rotation, predicts equal values for both shear and curvature vorticity, though there is no shear component present in solid body rotation. Hence, we will not benefit here from this procedure. The residual vorticity that we use does not suffer from this drawback.
In contrast, in an irrotational vortex where $u_r =0$ and $u_{\theta} = 1/r$ the residual vorticity is zero because vorticity associated with shear and rotation are equal and of opposite signs. However the curvature vorticity is non-zero showing the swirling motion of fluid elements. We may neglect such a contribution as we do not expect a point vortex like motion inside the drop. In other words, in the strict limit of Stokes flow, the velocity field inside a drop is constituted by growing harmonics alone. Therefore it is reasonable to consider that the residual vorticity gives the correct quantitative measure of rolling motion inside a drop. Note that neither residual vorticity nor curvature vorticity are complete measures of rolling motion. One may use residual vorticity in low Re flows and curvature vorticity in high Re flows.
By splitting the vorticity into two parts, one can identify the regions of shear. This could have also been identified by looking at the shear rate or viscous dissipation. However, such an approach will miss a very important factor to the motion which is solid body rotation which will not produce shear, but is important in determining the dynamics. Hence looking at the actual and residual vorticity plots gives an idea about overall motion, straining regions and rotating regions. As we will see that solid body rotation is indeed an important ingredient in the dynamics.
\section{Results and discussion}
Simulations have been performed in a box of dimensions $512 \times 256 \times 1$ LB units for a cylindrical drop. Wall boundary conditions are applied on two sides and periodic boundary conditions are applied on the other two sides. The simulation is initiated with a semicircular drop of radius $L = 60$ sitting on one wall, which is inclined at an angle $\alpha$ to the horizontal. In response to gravity and surface forces, the drop starts moving on the solid surface. We impose a smooth surface which thus does not display any hysteresis. The simulation is continued till the drop reaches a steady state velocity $\mathbf{V}$. Before compiling all the results into a unified framework, we first examine the effect of varying one physical quantity at a time. We define the Bond number as $Bo \equiv L^2|\mathbf{G}|/\sigma$, the Reynolds number as $Re \equiv L V \rho/\eta$ and the Capillary number as $Ca \equiv \eta V/\sigma$.
\begin{figure*}
\subfigure
\quad $Bo = 0.038, \%R = 38.1, \eta_r = 10, \theta_e = 152^{\circ}, Re=11.8, Ca=10^{-3}$]{\includegraphics[trim=130 460 80 60, clip, width=\linewidth]{218.eps}\label{fig:geodefa}}\\
\subfigure
\quad $Bo = 0.38, \%R = 28.7, \eta_r = 10, \theta_e = 152^{\circ}, Re=124, Ca=10^{-2}$]{\includegraphics[trim=130 460 80 60, clip, width=\linewidth]{220.eps}\label{fig:geodefb}}
\caption{Effect of gravity on the drop shape, streamline patterns, vorticity ($\omega$), residual vorticity ($\omega_{res}$) and residual angular velocity ($v_{res}$) are illustrated in a coordinate frame moving with the center of mass of the drop. A streamline of a given color in the left-most figure is shown in the same color in the three polar plots of $\omega$, $\omega_{res}$ and $v_{res}$. The azimuthal angle, measured from a line parallel to the solid plate, corresponds to that of the streamline, and the radial location at a given polar angle indicates the magnitude of the respective quantities. The magenta lines represent $\psi=0.9, \psi=0$ and $\psi=-0.9$, showing the thick interface. The drop is moving on a surface inclined to the horizontal, and the black dashed line indicates the direction of gravity. The red dashed line is normal to it.}
\label{fig:geodef}
\end{figure*}
The effect of increasing gravity on the steady state drop shape and streamline patterns, total and residual vorticity and angular velocity are illustrated in Fig. \ref{fig:geodef}. A larger driving force means larger deformations, so the drop deviates further from its equilibrium shape at zero plate inclination. The streamlines are plotted in the center of mass frame. Fixing the center as the innermost streamline, vorticity and residual vorticity along different streamlines are plotted as functions of azimuthal angle, thus mapping the entire vorticity field inside the drop. The thick interface and the contact line region are excluded from quantitative consideration as there may be spurious velocities generated due to the LB-DI model \cite{Yeomans2008}. In Fig. \ref{fig:geodefa}, the total vorticity is seen to lie within a small range almost everywhere in the drop, but as discussed above, we may not use this to determine whether there is a solid body rotation. In this case, the residual vorticity indicates that the bulk of the drop is indeed in solid body rotation. The total and residual vorticity values are comparable, showing that there is hardly any shear vorticity. However, in a region near the rear contact line, shear vorticity dominates. Thus the up-down symmetry is broken. Finally the angular velocity based on residual vorticity, $v_{res} = r\omega_{res}$ where $r$ is taken as the radial distance from the center of the innermost streamline, is plotted as a function of azimuthal angle. A perfect solid body rotation, such as that of a solid wheel would have appeared as concentric circles in this plot. At small Bond number, we do have something like this, except for a slight loss of symmetry in that the outer streamlines move faster at the top and slower at the bottom. In the case of large $Bo$, the drop is elongated normal to gravity, with a clear breakdown in left-right symmetry in its shape, as illustrated in Fig. \ref{fig:geodefb}. Except for the very center of the drop there is no resemblance to solid body rotation or even to tank-treading. This is reflected in the the angular velocity plot as well. The residual vorticity is higher in the direction of elongation. Interestingly the residual vorticity is now higher near the rear of the drop, exactly where it was lower at low $Bo$. This is because at higher gravity the rear of the drop has a tendency to lift off the surface. A given fluid element accelerates and decelerates significantly as it moves on a streamline.
If a solid body is rolling on an inclined surface with an angular velocity of $N$, then the corresponding vorticity is $2N$. Therefore, we can find the average residual vorticity inside a drop and calculate a corresponding forward velocity of the drop corresponding to the roll as
\begin{equation}
V_{rolling} = \frac{\text{Average}(\omega_{res})}{2} \times \frac{\text{Height of the drop}}{2}.
\end{equation}
Here we take the radius of the drop to be half of the maximum height. Then a quantity called percentage rotation, denoted by $\%R$, is calculated based on the total translational velocity $V$ of the drop, as
\begin{equation}
\%R=V_{rolling}/V \times 100.
\label{perr}
\end{equation}
This is different from the roll versus slip velocity as defined in \cite{yeomans_2010}, where the velocity profile at a single location is considered and the definition does not distinguish the shear vorticity from residual vorticity. In Fig. \ref{fig:geodef} one may see that increasing gravity increases the translational velocity by an order of magnitude, as seen in the increase in Reynolds number, but the associated deformation reduces the percentage rotation $\%R$ by 10\%.
\begin{figure}
\subfigure
\quad $Bo = 0.027, \theta_e>180^{\circ} (h=-0.7), \% R = 17.8, Re = 0.23, Ca = 0.002$]{\includegraphics[trim=130 420 290 50, clip, width=\linewidth]{282.eps}\label{fig:thetavar_samerea}}
\subfigure
\quad $Bo = 0.028, \theta_e = 138^{\circ}, \% R = 18.5, Re = 0.23, Ca = 0.002$]{\includegraphics[trim=130 420 290 50, clip, width=\linewidth]{284.eps}\label{fig:thetavar_samereb}}\\
\subfigure
\quad$Bo = 0.038, \theta_e = 90^{\circ}, \% R = 8.7, Re = 0.23, Ca = 0.002$]{\includegraphics[trim=130 420 290 50, clip, width=\linewidth]{255.eps}\label{fig:thetavar_samerec}}
\subfigure
\quad $Bo = 0.068, \theta_e = 42^{\circ}, \% R = 1.9, Re = 0.23, Ca = 0.002$]{\includegraphics[trim=130 420 290 50, clip, width=\linewidth]{287.eps}\label{fig:thetavar_samered}}
\caption{Effect of contact angle and thence the geometry on the rolling behavior. The $\% R$ is larger when the drop shape is closer to a circle. In all cases $Re$ and $Ca$ are kept constant by adjusting the $Bo$ and $\eta_r$ is kept as $10$. A larger reduction in $\omega_{res}$ as compared to $\omega$ may be observed as $\theta_e$ decreases.}
\label{fig:thetavar_samere}
\end{figure}
Therefore one may infer that an important property that determines the motion of a drop is its geometrical characteristics. Needless to say, the geometry is in turn determined by the volume, the contact angle, the gravitational force and the plate inclination, apart from the viscosity and density ratios. The present study is thus valid over a wide range of parameters in contrast to \cite{mahadevan_1999} wherein a spherical drop deformed by incremental gravity was studied. The crucial assumption in \cite{mahadevan_1999} was that the deviation of the shape from a sphere is very small. Relevant length scales of the deformation as a response to gravity were thence derived. These scaling arguments break down when $\theta_e \ne 180^{\circ}$ due to a finite contact area as we have in our simulations. Also, since we do not restrict our analysis to small $Bo$, the changes in the surface energy need not scale with that of gravitational potential energy unlike in \cite{mahadevan_1999}.
We now analyze the effect of contact angle alone as illustrated in Fig. \ref{fig:thetavar_samere} for drops of same volume. Here gravity is adjusted so that the drop attains the same terminal settling velocity in all cases and hence the same Reynolds number. We thus ensure that effects of inertia are nullified in this comparison. In Fig. \ref{fig:thetavar_samere}, one may see that no contribution of rotation is present when $\theta_e = 42^{\circ}$. Here the entire vorticity of the fluid elements can be attributed to that associated with shear. As the equilibrium contact angle increases, the percentage rotation increases with the maximum in the case of an almost circular drop. The effect of equilibrium contact angle is thus an intuitive result.
Compared to the case of Fig. \ref{fig:thetavar_samereb}, case \ref{fig:thetavar_samerea} shows a marginal reduction in the percentage rotation despite an increase in the contact angle. This can be attributed to the deformation of the drop near the wall in the latter case. The total vorticity is a function of shape as seen in different cases with a large contribution coming from the shear vorticity.
\begin{figure}
\subfigure
\quad $\alpha= 4, Bo = 0.1, \%R = 5.73, Re = 0.57, Ca=0.005$]{\includegraphics[trim=130 460 290 60, clip, width=0.9\linewidth]{324.eps}}
\subfigure
\quad $\alpha= 94, Bo = 1.5, \%R = 7.3, Re = 8.1, Ca = 0.065$]{\includegraphics[trim=130 460 290 60, clip, width=0.9\linewidth]{327.eps}}
\subfigure
\quad $\alpha= 176, Bo = 0.1, \%R = 10.7, Re = 0.67, Ca = 0.005$]{\includegraphics[trim=130 460 290 60, clip, width=0.9\linewidth]{329.eps}}
\caption{Effect of plate inclination on the shape and rotation behavior of drops. Equilibrium contact angle is $90^{\circ}$ and $\eta_r = 10$. A pendant drop is elongated to almost same size as the radius, producing more solid body rotation in the drop.}
\label{fig:plateinc}
\end{figure}
It is of interest to study how the rolling behavior changes as a function of the tilt angle of the plate. This is because the ratio of the components of gravity normal and tangential to the plate changes. The application of the normal component alone does not produce any movement of the drop, but both components contribute to deciding the shape, and hence the dynamics. The effect of plate inclination on the shape, streamlines and vorticities and their angular dependencies are illustrated in Fig. \ref{fig:plateinc}. This illustration is for an equilibrium contact angle of $90^{\circ}$. As the plate inclination increases the height of the drop increases, and it tends to lift off from the plate at some inclination. In turn the percentage rotation increases, and is highest for a tilt angle of $176^{\circ}$. This is another indication that the drop shape is a very important parameter in determining the kinematics inside the drop. The presence of corners and deformed parts of the drop always increase the shear vorticity locally.
\begin{figure}
\subfigure
\quad $\alpha= 4, Bo = 0.53, \%R = 5.3, Re = 3.3, Ca = 0.03$]{\includegraphics[trim=130 460 290 60, clip, width=\linewidth]{350.eps}}
\subfigure[
\quad $\alpha= 4, Bo = 0.53, \%R = 14.5, Re = 4.2, Ca = 0.03$, Normal component of gravity is $1/10^{th}$ of the above.]{\includegraphics[trim=130 460 290 60, clip, width=\linewidth]{350-10th.eps}}\\
\caption{Effect of normal component of gravity on the shape and rolling behavior of drops. Equilibrium angle is $90^\circ$ and $\eta_r = 10$. The tilt angle is chosen as $4^\circ$. To differentiate the effect of the tangential component of gravity, the normal component of gravity in (b) is artificially suppressed to $1/10^{th}$ of its value. However, the tangential component being maintained the same in both (a) and (b) yields a similar settling velocity. The percentage rotation can clearly be very different even at the same settling velocity, due to the change in shape.}
\label{fig:normvary}
\end{figure}
As the plate inclination changes, not only the ratio of normal to tangential forces changes, but also their magnitudes. In order to study the effect of this ratio alone, the normal force component was artificially varied keeping the tangential force the same. This corresponds to a simultaneous variation in plate inclination and gravity to achieve the same settling velocity. The results are illustrated in Fig. \ref{fig:normvary}. One can clearly see that as the normal component of gravity is slowly reduced, the shape becomes more and more elongated in the direction normal to the plate and this increases the amount of rotation considerably. This is yet another indication that the shape of the drop plays a big role in the rolling behavior of the drop.
\begin{figure}
\includegraphics[trim = 0 0 0 0, clip, width=\linewidth]{qall.eps}
\caption{The variation of percentage rotation with the isoperimetric quotient is illustrated for different sets of simulations. Each color represents a fixed equilibrium contact angle. Within each set, plate inclinations vary from $\alpha = 4^{\circ}$ to $176^{\circ}$. Also, $Bo$ ranges from $5 \times 10^{-3}$ to $1.5$ by varying gravity and surface tension. The slip length and the viscosity ratio are kept fixed. An exponential curve fitted through all data points is also shown. Note that moving drops for a wide variation in physical properties fall on this curve.}
\label{fig:qall}
\end{figure}
In all these cases, we see that the deviation from a circular shape plays an important role in determining the dynamics. One can then suitably define a shape parameter to describe the closeness of the shape to a circle, for example, the isoperimetric quotient, $q$,
\begin{equation}
q = \frac{4 \pi \times \text{Area}}{\text{Perimeter}^2}.
\end{equation}
This ratio is unity for a circle and is less than this value for any other shape, since a circle has the least circumference for a given area. The percentage rotation in all the cases we have computed shows a direct dependence on the isoperimetric quotient, as shown in Fig. \ref{fig:qall}. As expected, the percentage rotation is higher for a shape which is closer to a circle. It is however of interest to note that irrespective of the parameters such as equilibrium contact angle, gravity and plate inclination which determines the shape and the deformation, the percentage of roll is a function only of this quantity. The collapse of data that we obtain here is a strong indication of this. Note that the data in this figure represents results from over $100$ simulations, spanning a wide range of $\theta_e$, $\alpha$ and $Bo$. However the viscosity, the mobility and the viscosity ratio are kept fixed in the simulations shown so far. As we will see below, percentage rotation curve gets shifted in response to changes in these parameters. We have defined the outline of the drop as a contour of $\psi=-0.9$ which can be thought of as the inner limit of the interface. Since we use a combination of LB and DI models, there can be spurious interface velocities \cite{Yeomans2008} and hence the data outside this line is not considered. We have also tried to calculate this shape parameter from $\psi=0$ which is theoretically the interface. However this shape fails to capture the deformations correctly for large contact angles.
\begin{figure}
\includegraphics[trim = 0 0 0 0, clip, width=\linewidth]{shapefactor.eps}
\caption{Isoperimetric quotient of static shapes compared with dynamic drops for the same Bond numbers. The equilibrium contact angle is $90^{\circ}$ and three different plate inclinations, $30^{\circ}, 90^{\circ}$ and $135^{\circ}$, are chosen for comparison. Here `FP' stands for `front pinned', `RP' stands for `rear pinned' and `D' stands for dynamic cases. The shape factor of a dynamic drop lies in between those corresponding to front pinned and back pinned static shapes \cite{sumesh_2011}. This behavior breaks down at large Bond numbers where inertia is higher. In that case, the moving drop is closer to circular than either static shape.}
\label{fig:sfcomp}
\end{figure}
It would be interesting to compare the isoperimetric quotient of static drops with that of dynamic cases and see whether any predictions can be made. This is illustrated in Fig. \ref{fig:sfcomp}. As explained in \cite{sumesh_2011} we obtain minimum energy static shapes of drops with either the front end or the rear end pinned. As illustrated in Fig. \ref{fig:sfcomp}, the isoperimetric quotient of the dynamic drops resides between that of front pinned and back pinned cases. This is interesting because one can make predictions about kinematics inside the drop by the analysis of static drops, but only for small Bond numbers. As $Bo$ increases, such monotonic variations in the shape factor is violated, necessitating the full calculations.
\begin{figure}
\subfigure[]{\includegraphics[trim = 0 0 0 0, clip, width=0.49\linewidth]{widthindeca.eps}}
\subfigure[]{\includegraphics[trim = 0 0 0 0, clip, width=0.48\linewidth]{widthinderotn.eps}}
\caption{A large interfacial thickness is an artifact of the DI method. However, here the $Ca$ and $\% R$ are shown to be independent of $Cn$.In the above plots the symbols represent different sets of simulations; $\bullet$ for $Bo = 0.19$,$\theta_e = 138^{\circ}$, $\eta_r = 1$, $\small\blacksquare$ for $Bo = 0.04$, $\theta_e = 138$, $\eta_r = 1$, $\blacklozenge$ for $Bo = 1.90$, $\theta_e = 90$, $\eta_r = 1$ and $\blacktriangle$ for $Bo = 0.19$, $\theta_e = 138$, $\eta_r = 10$.}
\label{fig:widthinde}
\end{figure}
One of the main drawbacks of the DI models is that it imposes a finite thickness of the interface while for most macroscopic drops, the interface thickness would be negligible compared to any other length scale in the problem. In order to ensure that our results are independent of the assumed interfacial width, simulations with a range of interface thicknesses were conducted. As shown in Fig. \ref{fig:widthinde}, both the Capillary number and the percentage rotation remain insensitive to interfacial width, to within numerical errors. Here $Cn = \xi/L$, the ratio of interfacial thickness to the macroscopic length, is the Cahn number.
\begin{figure}
\includegraphics[trim = 0 0 0 0, clip, width=\linewidth]{cavsslip.eps}
\caption{$Ca$ is plotted as a function of the nondimensional slip length, $S$. In order to obtain a range of $S$, the viscosity was independently varied by three orders of magnitude and mobility by one order of magnitude. The equilibrium contact angle is $152^{\circ}$. Larger slip length at the contact line results in larger translational velocity of the drop.}
\label{fig:cavsslip}
\end{figure}
\begin{figure}
\includegraphics[trim =0 0 0 0, clip, width=\linewidth]{sliplength.eps}
\caption{Percentage roation, \%R, is plotted as a function of non-dimensionalised slip length, S. The simulation parameters are same as in Fig. \ref{fig:cavsslip}. A strong dependence of the rolling behavior on the slip length may be observed.}
\label{fig:sliplength}
\end{figure}
Apart from the solid body rotation, which gives a forward velocity to the entire drop, the contact line moves due to the slip provided by the diffusion of the order parameter \cite{jacqmin_2000}. Balancing the advection and diffusion of order parameter across the interface provides a length scale for this diffusion as $\lambda = \sqrt{\eta M}$. We define a nondimensional slip length as $S = \lambda/L $. This slip length is the same as that used in the slip-induced movement of contact line in sharp interface models \cite{feng_2010}, and is not an artificial parameter. They actually show that this slip length should not be dependent on the interfacial width. Hence we can use $\lambda$ as a measure of slip at the contact line. This means that either mobility or viscosity can be independently or simultaneously varied to change the slip at the contact line. Slip length is here defined using the viscosity of the drop $\eta_I$. In principle, the external fluid plays a very important role and an effective viscosity to define the slip length may need to account for the viscosity ratio $\eta_r$. For example a geometric mean of the two is used in \cite{feng_2010}. However, we refrain from using this relation as it lacks a physical significance.
Both viscosity and mobility are independently varied by at least one order of magnitude in Fig. \ref{fig:cavsslip} to obtain a range of $S$. As the slip length increases, the $Ca$ also increases as illustrated in Fig. \ref{fig:cavsslip}. Intuitively, a slipping drop on an inclined surface will roll less. This is verified in our simulations as shown in Fig. \ref{fig:sliplength} wherein the importance of slip length in determining the amount of rotation inside the drop may be inferred. And this dependence appears to be exponential. Larger percentage rotations than those shown, which would correspond to smaller slip lengths could not be obtained reliably with the present numerical simulations.
\begin{figure}
\subfigure[$Bo = 0.04, M = 0.2, \eta = 0.15 , \%R = 19.7, Re = 0.3, Ca = 0.003$]{\includegraphics[trim=130 420 290 50, clip, width=\linewidth]{215.eps}}
\subfigure[$Bo = 0.03, M = 2 , \eta = 0.015 , \%R = 19.9, Re = 25, Ca = 0.002$]{\includegraphics[trim=130 420 290 50, clip, width=\linewidth]{239.eps}}
\caption{Though viscosity affects the overall dynamics, the contribution of rolling motion to the dynamics remains the same. Here $\alpha = 30^{\circ}, \theta_e = 152^{\circ}$ and $\eta_r = 10$. Since viscosity plays a direct roll in the slip of the contact line, mobility was simultaneously adjusted to obtain same slip length to isolate the effect of viscosity alone.}
\label{fig:viscfree}
\end{figure}
In our simulations, the nondimensional slip length, $S$, varies from $10^{-3}$ to $10^{-2}$. In the light of experimental evidence where slip length varies from $nm$ to $\mu m$ \cite{meinhart_2002} we expect that our observations remain valid for a range of drop sizes. For macroscopic drops smaller than the capillary length, this ratio is very small and hence a larger fraction of rolling motion may be expected than those seen here. It is worth mentioning however that slip lengths of $10-100$ of micron have been reported on patterned surfaces \cite{lohse_2009} or when lubricating gas layers are present \cite{vinogradova_2009} or on super-hydrophobic surfaces \cite{rothstein_2010}. Since we concentrate on the bulk motion of fluid elements, we expect that our simulations are relevant in several practical applications.
It is interesting to note that viscosity does not have an explicit dependence on the percentage contribution of rolling in the total motion. Viscosity enters the problem in two ways, via the Reynolds number, and via the slip at the contact line. The former effect is illustrated in Fig. \ref{fig:viscfree}. In order to effect changes in viscosity in the stress term alone, the slip length was maintained the same in the two cases shown by simultaneously changing the mobility by the appropriate factor. As a consequence of change in dissipation, the settling velocity is different. But the percentage contribution of rolling is the same in the two cases showing that the viscosity does not explicitly affect the rolling motion.
\begin{figure}
\includegraphics[trim = 0 0 0 0, clip, width=\linewidth]{cavsbo.eps}
\caption{The linear relationship between $Ca$ and $Bo$ is illustrated for the data given in Fig \ref{fig:qall}. This data consists of different contact angles, plate inclinations, gravity and surface tension values.}
\label{fig:cavsbo}
\end{figure}
\begin{figure}
\includegraphics[trim =0 0 0 0, clip, width=\linewidth]{mbovsca.eps}
\caption{Both Capillary number, $Ca$ and modified Capillary number, $Ca_M$ are plotted against $Bo$. The relative standard deviation (RSD) is small for $Ca_M$ justifying Eq. \ref{eqn:mcabo}. Both viscosity and mobility are varied by at least one order of magnitude to produce range of slip lengths in these simulations. Equilibrium contact angle is $152^{\circ}$ and $Cn=0.07$. Value of $\beta$ is adjusted in Eq. \ref{eqn:mcabo} to obtain the least RSD, hence verifying the role of slip length. Since, $0.1 < Re < 70$ for these simulations, a wider distribution at large $Bo$ may be related to the unaccounted inertial effects.}
\label{fig:mbovsca}
\end{figure}
Neglecting inertial terms in NSE, one may balance the gravitational driving forces to viscous dissipation to obtain \cite{Kang_2001, limat_2005},
\begin{equation}
Ca \sim Bo - \Delta_{\theta},
\label{eqn:cavsbo}
\end{equation}
where $\Delta_{\theta} = \cos{\theta_r} - \cos{\theta_a}$, which is small for most of our simulations. A plot of $Ca$ vs $Bo$ is shown in Fig. \ref{fig:cavsbo} which corresponds to the data described in Fig. \ref{fig:qall}. This scaling holds good for all our simulations. The small deviations seen are due to inertial effects not accounted for in Eq. \ref{eqn:cavsbo}, as well as by the intercept, as shown below.
Having seen the effect of contact-line slip on the drop dynamics, we now make a connection between the slip velocity and the above scaling relation. For this we use the generalized Navier's boundary condition \cite{Qian2003, eggers_2009},
\begin{equation}
\frac{u^{slip}}{\lambda} = [\partial_n u] + \frac{{\cal L}(\psi)}{\eta} \partial_x \psi
\end{equation}
to make quantitative estimates of the sliding velocity. In the above equation $n$ is a direction normal to the wall and ${\cal L}(\psi) = K \partial_n \psi + \partial \sigma_{wf}/\partial \psi$, $K$ is the coefficient in the free energy functional and $\sigma_{wf}$ is the interfacial free energy per unit area at the fluid-solid boundary. The second term is called the uncompensated Young's stress and one may see that
\begin{equation}
\int_{int} dx [{\cal L}(\psi) \partial_x \psi] = \gamma (\cos{\theta_d} - \cos{\theta_e}),
\end{equation}
where $int$ means `across the interface' and $\theta_d$ is the dynamic contact angle. We can calculate an order of magnitude estimate of $\Delta_{\theta}$ from the above equation as $Ca\frac{\xi}{\lambda}$. Since we are at steady state, both the advancing and receding contact lines move with the same velocity as the center of mass of the drop. We assume that any variation in order parameter and hence the slip is felt over a region of interfacial thickness $\xi$ while the associated slip length $\lambda$ is the same slip length calculated in a DI model. Accounting for a possible pre-factor in the addition of scaling estimates in Eq. \ref{eqn:cavsbo} we obtain,
\begin{equation}
Ca \left[1 + \frac{\xi}{\beta\lambda}\right] \sim Bo,
\label{eqn:mcabo}
\end{equation}
which explicitly includes the role of slip at the contact line. This relationship is verified in Fig. \ref{fig:mbovsca}, where the change in only the slip length affects the linear relationship between $Ca$ and $Bo$. We define $ Ca_M \equiv \left[1 + \xi/(\beta\lambda)\right]$. By choosing a suitable pre-factor $\beta$ one may see that the second term in the above equation explains in part the distribution in $Ca$ at a given $Bo$. In other words, the relative standard deviation, which is the ratio of standard deviation to the mean value expressed as a percentage comes down dramatically when Eq. \ref{eqn:mcabo} is used.
\begin{figure}
\includegraphics[trim=0 0 0 0, clip, width=\linewidth]{cavsviscratio.eps}
\caption{Effect of surrounding fluid viscosity on $Ca$. As the viscosity ratio $\eta_r$ increases, which corresponds to a reduction in the viscosity of the surrounding fluid, the drop translates faster.}
\label{fig:cavsviscratio}
\end{figure}
\begin{figure}
\includegraphics[trim=0 0 0 0, clip, width=\linewidth]{muratio.eps}
\caption{Effect of external fluid viscosity on rolling behavior of the drop. The percentage rotation exhibits a strong dependence on the viscosity ratio, $\eta_r$. As the external fluid viscosity comes down, rolling motion inside the drop increases. Low viscosity drops in a higher viscosity fluid are seen to almost slide on the wall rather than execute a rolling motion. }
\label{fig:muratio}
\end{figure}
\begin{figure}
\subfigure[$Bo = 0.19, \eta_r = 20, \%R = 22.0, Re = 0.3, Ca = 0.01$]{\includegraphics[trim=130 420 290 50, clip, width=\linewidth]{206.eps}}
\subfigure[$Bo = 0.19, \eta_r = 1, \%R = 12.1, Re = 0.1, Ca = 0.004$]{\includegraphics[trim=130 420 290 50, clip, width=\linewidth]{172.eps}}
\caption{Change in the velocity and vorticity fields when the external fluid viscosity is changed keeping the drop viscosity same. Here $\alpha = 30^{\circ}$ and $\theta_e = 138^{\circ}$. As the external viscosity increases, it start affecting the dynamics more as seen in Fig. \ref{fig:muratio}.}
\label{fig:muratio_vel}
\end{figure}
Finally we investigate the role of the viscosity ratio between the drop and the surrounding fluid, in Fig. \ref{fig:cavsviscratio} and \ref{fig:muratio}. When the viscosity of the external fluid is reduced, the settling velocity and hence the $Ca$, as expected, increase. Also the percentage of rolling motion is larger. In line with this, one may expect significant amount of rolling in case of a water-air system where viscosity contrast is large. As the viscosity of the external fluid increases and goes beyond that of the drop, the entire dynamics shifts to the external fluid. The drops slides in that case. This too is consistent with intuition, since a `bubble' will simply slide in a liquid rather than roll when moving on a surface. The changes in the vorticity and residual vorticity fields are illustrated in Fig. \ref{fig:muratio_vel}. One may observe that, despite the geometry remaining similar, the percentage rotation increases when the viscosity ratio increases. Therefore the universal curve obtained in Fig. \ref{fig:qall} will be shifted appropriately by a change in the slip length and viscosity ratio.
\section{Conclusions}
A hybrid simulation method implementing lattice Boltzmann algorithm with diffuse interface model is used to analyze the drop motion on inclined surfaces under gravity. Modification of the model to provide a viscosity contrast between the fluids, wetting boundary conditions and to introduce gravity are discussed. By removing shear vorticity from total vorticity, it is shown that residual vorticity is a good measure to characterize the rolling motion inside the drop. It is shown that the drop shape can be described by a geometrical quantity, isoperimetric quotient, that is primarily responsible for determining the fraction of forward motion accruing from solid body rotation. For a given slip length and viscosity ratio with the outer fluid, this dependence is universal, irrespective of the equilibrium contact angle, plate inclination and $Bo$. The importance of the slip mechanism of the contact line is discussed, not only in relation to the rolling motion inside the drop, but also in modifying the coefficient in the scaling relationship between $Bo$ and $Ca$. The external fluid certainly affects the drop motion with larger rolling motion observed when its viscosity is small compared to the drop viscosity.
\begin{acknowledgments}
We gratefully acknowledge Ignacio Pagonabarraga for the fruitful discussions.
\end{acknowledgments}
|
\section{Introduction}
There is a vivid interest in the cross-section of charmed hadron
production in the proton-antiproton collisions to be measured by
the future $\bar{P}ANDA$ experiment (see, e.g., \cite{panda}). The
amount of produced charmed mesons and baryons is important for
assessing the ability of this experiment to perform
flavour-physics oriented studies, such as the measurement of
charm-anticharm mixing, the search for $CP$-violation in $D$
decays or the studies of $\Lambda_c$ decays. A reliable estimate
of the $p\bar{p}\to ~charm $ cross section is however a very
difficult task. The main problem is that the projected energy
range (with the c.m. energy $\sqrt{s}$ varying from 2.25 GeV to
5.47 GeV), being reasonably high for the proton-antiproton
collisions, is still not far from the threshold of
charm-anticharm production. Several models of charm production at
these energies can be found in the literature
\cite{KaidV,TitovKampfer,Titov:2011vc,Kroll:1988cd,Goritschnig:2009sq,Braaten,Haidenbauer:2009ad,Haidenbauer:2010nx,Kerbikov:1994xx},
their predictions differing by several orders of magnitude, as
emphasized, e.g. in \cite{Haidenbauer:2009ad}. Especially
difficult is to predict the inclusive charm-anticharm cross
section in the situation where not many exclusive channels are
open. Hence, it is more realistic to assess the exclusive
production of baryons or mesons, such as $p\bar{p}\to {\Lambda_c}{\bar{\Lambda}_c},
\bar{D}D$. A successful model of strange-hadron pair production in
$p\bar{p}$ collisions, which was measured at similar energies, can
serve as a useful tool, provided there is a reliable way to
replace the model parameters of strange hadrons by the ones for
charmed hadrons. The key parameters in many hadronic models of
these processes are the strong couplings of strange or charmed
baryons with mesons and nucleons. To relate them, the
$SU(4)_{fl}$-symmetry is frequently used in the literature. Note
however, that it is difficult to justify this symmetry in QCD, due
to the large mass difference of the $c$- and $s$- quarks,
$m_c-m_s\gg \Lambda_{QCD}$.
In this paper we employ the strong baryon-meson
couplings of charmed and strange hadrons calculated
from QCD light-cone sum rules (LCSR),
where finite masses of $c$ and $s$ quark are taken
into account. Recently, we calculated \cite{LcDN} the charmed baryon
strong couplings with a charmed meson and a nucleon. In addition
to these results, here we obtain the corresponding strong couplings
of strange hadrons. The nonperturbative inputs in the LCSR method
are the universal nucleon distribution
amplitudes (DA's).
Hence, the extension of our calculation from charmed to
strange hadrons is straightforward and is reduced to a replacement
of the virtual $c$-quark by an $s$-quark in the
underlying correlation functions. In what follows, we also employ the
ratios of calculated strong couplings which are predicted from LCSR
with smaller uncertainties than the individual couplings.
The results for the strong couplings presented here can be used in
various models of exclusive charm and strange hadron production.
As an application of our results, we use the quark-gluon string
(QGS) model of binary reactions developed by Kaidalov and his
collaborators
\cite{Kaidalov:1981jw,Kaidalov:1981rw,Kaidalov:1999zb,KaidN}. One
version of this model was already applied in \cite{KaidV} to
estimate the charm production cross section in proton-antiproton
collisions. We refine this model by introducing the helicity
amplitudes and adjusting the two independent strong couplings to
the LCSR estimates.
In what follows, in Sect.~2 we present the LCSR results for strong couplings
of charmed and strange hadrons.
In Sect.~3 we demonstrate how the QGS model works for relatively simple
processes of meson pair production and trace the relation between
the model parameters and strong couplings in QCD.
In Sect.~4 we use the QGS model for $p\bar{p}$ binary reactions
with charmed and strange hadrons, employing the strong couplings from
LCSR and predict the charm production cross sections.
Sect.~5 contains the concluding discussion.
The two appendices contain: (A) the formulae for helicity amplitudes
and (B) the derivation of the absorption factor in $p\bar{p}$ binary processes.
\section{Strong couplings from QCD light-cone sum rules}
The strong couplings of the $\Lambda_c$-baryon with
the nucleon and $D$ or $D^*$ meson are formally defined as
the following hadronic matrix
elements:
\begin{eqnarray}
\langle {\Lambda_c}(P-q) |D(-q) N(P)\rangle &=& g_{{\Lambda_c} ND} \,
\bar{u}_{\Lambda_c}(P-q)\,i\gamma_5\,u_N(P), \nonumber
\\
\langle \Lambda_c(P-q) |D^*(-q) N(P) \rangle &=&
\bar{u}_{\Lambda_c}(P-q)\left(g^V_{ \Lambda_c
ND^*}\slashed{\epsilon}+i\,\frac{g^T_{ \Lambda_c ND^*}}{m_{
\Lambda_c}+m_N}\sigma_{\mu\nu}\epsilon^\mu q^\nu\right)u_N(P). \,
\hspace{0.5 cm} \label{eq:CC1}
\end{eqnarray}
Note that the above couplings are defined in \cite{LcDN}
as residues at the $D^{(*)}$ and $\Lambda_c$ poles in double
dispersion relations for the correlation functions with on-shell
nucleon state, hence all three hadrons are on their mass-shell.
The same definitions are valid for the $\Sigma_c$-baryon couplings
as well as for the corresponding strange hadrons with the
following replacements: ${\Lambda_c}(\Sigma_c)\to \Lambda (\Sigma) $
and $D^{(*)}\to K^{(*)}$ in the above.
The $\Lambda_c ND^{(*)}$ and $\Sigma_c ND^{(*)}$ strong couplings
were calculated from LCSR in \cite{LcDN}, where one can find the
detailed description of the sum rule derivation. The results for
the strong couplings which will be used in this paper are
collected in Table~\ref{tab:res}. Note that in \cite{LcDN} two
different interpolating currents for $\Lambda_c$ and $\Sigma_c$
baryons were used. With the procedure of eliminating the negative
parity baryons suggested in that paper, the results agree within
the uncertainties. In this paper we will only use the strong
couplings obtained for the pseudoscalar interpolating current for
$\Lambda_{(c)}$ and Ioffe current for $\Sigma_{(c)}$,
respectively, because the sum rules in these cases have a
comparatively lower background of higher states. In
Table~\ref{tab:res} also the ratios of strong couplings obtained
from LCSR are presented, generally they have smaller estimated
uncertainties, because of the common inputs used in the sum rules.
Turning to strange hadrons, we employ the same LCSR method as in
\cite{Braun:2000kw,Braun:2006hz} and, replacing $c$-quark with the
$s$-quark in the correlation function, calculate the $\Lambda
NK^{(*)}$ and $\Sigma NK^*$ couplings. The inputs used in LCSR
consist of universal nucleon DA's which are taken from
\cite{Lenz:2009ar} and explained in detail in \cite{LcDN}. In
particular we use for the virtual $c$ quark in the correlation
function the value $m_c(m_c)=1.28 \pm 0.03$ GeV. The
flavour-specific input parameters which we adopt here for the sum
rules involving strange hadrons are: the strange quark mass
$m_s(\mbox{2 GeV})= 98 \pm 16$ MeV and the renormalization scale
$\mu_{s} =1.0 \pm 0.2 ~{\rm GeV}$. Furthermore, one and the same
range $M^2=2.0\pm 0.5$ GeV$^2$ of the Borel parameter in the
$\Sigma$ and $\Lambda$ channels is adopted, whereas for the $K^*$
channel we use $\widetilde{M}^2=1.0\pm 0.5 $ GeV$^2$. The
threshold parameter in the LCSR for $\Lambda$($\Sigma$) strong
couplings is taken $s_0=2.55\pm 0.10$ GeV$^2$ ($s_0=2.75\pm 0.10$
GeV$^2$). The criteria of choosing the input parameters and the
quark-hadron duality ansatz in LCSR are the same as the ones used
and discussed in \cite{LcDN}. The two-point QCD sum rules for the
decay constants of $\Lambda$ and $\Sigma$ baryons with
pseudoscalar and Ioffe currents respectively, are taken from
\cite{Liu:2008yg}. Using the same definitions and notation as for
the decay constants of charmed baryons in \cite{LcDN}), we obtain:
\begin{eqnarray}
\lambda_{\Lambda}^{\mathcal{(P)}}=(0.87^{+0.23}_{-0.13}) \times
10^{-2} \,\, {\rm GeV^2}\,, ~~
\lambda_{\Sigma}^{\mathcal{(I)}}=(2.6^{+0.3}_{-0.2} ) \times
10^{-2} \,\, {\rm GeV^2} \,.
\end{eqnarray}
The resulting estimates of the strange baryon strong couplings
and their ratios obtained from LCSR are presented in
Table~\ref{tab:res}.
Note that ${\Lambda_c}$ and $\Sigma_c$ belong to different $SU(3)$ multiplets
(as opposed to $\Lambda$ and $\Sigma$), and this circumstance
explains a substantial difference between the ratios of tensor and vector
strong couplings for these baryons.
\begin{table}[t]
\begin{center}
\begin{tabular}{|c|c||c|c||c|c|}
\hline&&&&&\\[-3mm]
Strong & LCSR &Strong &LCSR&Ratio & LCSR \\
coupling&estimate&coupling&estimate&of couplings&estimate\\[1mm]
(charmed)&&(strange)&&$\big(\frac{\mbox{charmed}}{\mbox{strange}}\big)$&\\
\hline
&&&&&\\[-3mm]
$g_{\Lambda_c N D} $ & $10.7^{+5.3}_{-4.3}$ &$g_{\Lambda N K} $&
$7.3^{+2.6}_{-2.8}$&
$\frac{g_{\Lambda_c N D}}{g_{\Lambda N K}} $&$1.47^{+0.58}_{-0.44}$\\[3mm]
\hline &&&&&\\[-3mm]
$g^V_{\Lambda_c N D^{\ast}} $ &
$-5.8^{+2.1}_{-2.5}$&$g^V_{\Lambda N K^{\ast}} $&$-6.1^{+2.1}_{-2.0}$&$\frac{g^V_{\Lambda_c N D^{\ast}}}{g^V_{\Lambda N K^{\ast}}} $
&$0.95^{+0.35}_{-0.28}$\\[3mm]
$g^T_{\Lambda_c N D^{\ast}}$ & $3.6^{+2.9}_{-1.8}$& $g^T_{\Lambda N K^{\ast}} $& $12.8^{+5.8}_{-5.2}$&&\\[3mm]
$\frac{g^T_{\Lambda_c N D^{\ast}}}{g^V_{\Lambda_c N D^{\ast}}} $&
$-0.63^{+0.16}_{-0.28}$
&$\frac{g^T_{\Lambda N K^{\ast}}}{g^V_{\Lambda N K^{\ast}}}$& $-2.1^{+0.5}_{-0.6}$&&\\[3mm]
\hline
&&&&&\\[-3mm]
$g_{\Sigma_c N D} $ & $1.3^{+1.0}_{-0.9}$ & $g_{\Sigma N K} $
& $1.1^{+0.6}_{-0.5}$ && \\[3mm]
\hline &&&&&\\[-3mm]
$g^V_{\Sigma_c N D^{\ast}} $ &$1.0^{+1.3}_{-0.6}$ &$g^V_{\Sigma N K^{\ast}} $
&$1.7^{+0.9}_{-0.8}$&$\frac{g^V_{\Sigma_c N D^{\ast}}}{g^V_{\Sigma N K^{\ast}}} $&$0.56^{+0.42}_{-0.20}$\\[3mm]
$g^T_{\Sigma_c N D^{\ast}} $ & $2.1^{+1.9}_{-1.0}$&$g^T_{\Sigma
N K^{\ast}} $&$3.6^{+1.5}_{-1.2}$ &
&\\[3mm]
$\frac{g^T_{\Sigma_c N D^{\ast}}}{g^V_{\Sigma_c N D^{\ast}}} $ & $2.1\pm 0.5$ &$\frac{g^T_{\Sigma N K^{\ast}}}{g^V_{\Sigma N K^{\ast}}} $ & $2.1 ^{+0.6}_{-0.3}$&&\\[3mm]
\hline
\end{tabular}
\end{center}
\caption{\it Numerical results for the strong couplings of charmed \cite{LcDN}
and strange baryons and their ratios obtained from LCSR with nucleon
DA's.}
\label{tab:res}
\end{table}
In the literature (see e.g., \cite{TitovKampfer}), the strong
couplings of charmed hadrons are estimated assuming $SU(4)_{fl}$
symmetry and equating dimensionless couplings, for example,
assuming $g^{V(T)}_{\Lambda_c N D^{\ast}}\simeq
g^{V(T)}_{\Lambda N K^{\ast}} $. Such symmetry relations are
difficult to justify from the point of view of QCD. Indeed,
because of the large mass difference of $c$ and $s$ quarks, the
kinematical factors: masses and four-momenta entering the complete
hadronic matrix elements differ significantly. In our approach
we are not relying on any form of the $SU(4)_{fl}$ symmetry.
Still, it is interesting to compare the strong couplings for the
charmed and strange baryons obtained from LCSR and collected in
Table~\ref{tab:res}. We find that the values of the dimensionless
$g^V$ couplings are in the same ballpark, whereas there is a
significant difference between $g^T_{\Lambda N K^{\ast}} $ and
$g^T_{\Lambda_c N D^{\ast}}$. The strange baryon couplings were
also calculated with the Nijmegen potential model
\cite{Stoks:1999bz} of low-energy scattering, assuming $SU(3)_{fl}$-symmetry.
Expressed in terms of the dimensionless $g$-couplings defined in
(\ref{eq:CC1}) the results of \cite{Stoks:1999bz} with their sign conventions
are:
\begin{eqnarray}
g_{\Lambda N K} = 13.4 \div 17.5 \,,& g^V_{\Lambda N K^{\ast}} =
-(4.3 \div 6.1)\,,&
g^T_{\Lambda N K^{\ast}} =12.4 \div 16.3, \nonumber \\
g_{\Sigma N K} =-(4.1 \div 5.3)\,, & g^V_{\Sigma N K^{\ast}} = -(2.4 \div 3.5)\,,&
g^T_{\Sigma N K^{\ast}} = -(1.3 \div 4.6)\,.
\end{eqnarray}
Comparing with our predictions for the strange-baryon
couplings given in Table 1, we observe an agreement for vector-meson couplings
within uncertainties. Also the convention-independent
relative signs of $T$ and $V$ couplings agree.
Meanwhile, the LCSR predictions for
$g_{\Lambda N K}$ and $g_{\Sigma N K}$ are systematically
lower than the intervals for these couplings
obtained in the potential model.
\section{ The QGS model for meson pair production }
In the QGS model, the amplitudes of binary reactions, such as
$p\bar{p}\to \Lambda_c \bar{\Lambda}_c $ or
$p\bar{p}\to \bar{D}D$, are described by planar diagrams
depicted in Fig.~1. These diagrams have a dual interpretation.
From the $s$-channel point of view, annihilation of the slow
$u\bar{u}$ or $d\bar{d}$ pair from the initial proton and
antiproton is followed by a creation of the
$c\bar{c}$-pair. The spectator quarks and antiquarks from
the initial proton and antiproton coalesce with the created quark and
antiquark to form the final state
charmed hadrons.
\begin{figure}[t]
\centering \hspace{0 cm}
\includegraphics[scale=0.75]{fig1.eps}
\caption{\it The planar diagram of charmed baryon
(a) and meson (b) pair production in $p\bar{p}$ collisions.}
\label{fig:planar}
\end{figure}
The intermediate state in $s$-channel represents a sort of a
diquark-antidiquark (Fig.1~a) or quark-antiquark (Fig.1~b) string.
On the other hand, in the $t$-channel a virtual hadronic state
with the quantum numbers of a charmed meson or baryon is
exchanged. In the $s\gg |t|$ limit, this exchange is described by
the dominant Regge pole. For instance, the amplitude of
$p\bar{p}\to \Lambda_c \bar{\Lambda}_c $ is approximated by the
(degenerate) $D^{*},D^{**}$ Regge-trajectory
$\alpha_{D^*}(t)=\alpha_{D^*}(0)+\alpha'_{D^*}t$ (we use the
linear approximation). The QGS-model parameters are obtained
\cite{KaidV,Kaidalov:1981jw,Kaidalov:1981rw,Kaidalov:1999zb} using
the quark-parton description of the $s$-channel planar diagram.
Replacing the $c$-quark by $s$-quark in the planar diagrams of
Fig.~1 we reproduce the QGS model for the production of strange
baryons and mesons. The strange-hadron pair production cross
section in $p\bar{p}$ collisions calculated in this model
\cite{KaidV} agrees well with the experimental data. Importantly,
there is a strong flavour dependence of the binary reactions in
QGS model, encoded in the slopes and intercepts of the Regge
trajectories as well as in the scale factors $s_0$ entering the
Regge amplitudes. The relative suppression of the charmed hadron
production corresponds, in terms of the $s$-channel picture, to a
comparatively smaller probability to create a heavy
quark-antiquark pair within the intermediate string.
To discuss the QGS model in more detail, we first consider a
relatively simple binary reaction involving no spins or
helicities: $\pi^+\pi^-\to M \overline{M}$, with pseudoscalar
mesons ($M=\pi^0,K^+\!,\bar{D}^0$) of various flavours in the
final state and with isospin and/or flavour exchange in
$t$-channel. The planar diagram of this process is shown in Fig.
\ref{fig:pion}. At large $s$ and small $|t|\ll s $, the scattering
amplitude is written \cite{KaidV} in the following Regge-pole
form: \begin{equation} T^{(\pi^+\pi^-\to M \overline{M})}(s,t)= g^{(\pi M)}(t)
\frac{s}{\bar{s}}\left(\frac{s}{s_0^{\pi
M}}\right)^{\alpha_{R}(t)-1} , \label{eq:Reggeampl} \end{equation} where
$\bar{s}=1 ~\mbox{GeV}^2$ is a universal dimensional factor and the
energy dependence is determined by the Regge trajectory
$\alpha_R(t)$ with the corresponding quantum numbers
($R=\rho(a_2), K^{*(**)},D^{*(**)}$). In the above, $g^{\pi M}(t)$
is the residue function of the momentum transfer squared. In the
QGS model \cite{KaidV} the $\Gamma$-function dependence inspired
by Veneziano duality is adopted: \begin{equation}
g^{(\pi M)}(t)=C^{(\pi M)}g_0^2~\Gamma(1-\alpha_R(t))\,.
\label{eq:gt}
\end{equation}
The coefficient $C^{(\pi M)}$ is equal to the number of planar diagrams.
The amplitude (\ref{eq:Reggeampl}) for $\pi^+\pi^-\to
\pi^0\pi^0$ is determined by the $\rho(a_2)$-trajectory:
\begin{equation}
\alpha_R(t)=\alpha_\rho(t)=\alpha_\rho(0)+\alpha'_\rho t\,.
\label{eq:traject}
\end{equation}
In this case $C^{(\pi\pi)}=2$,
and the numerical values of the intercept $\alpha_\rho(0)$ and slope
$\alpha'_\rho$ taken from \cite{KaidV} are presented in
Table~\ref{tab:reggeM}.
\begin{figure}[t]
\centering
\includegraphics[scale=0.55]{fig2.eps}
\caption{\it The planar diagram of $\pi^+\pi^-\to
M \overline{M}$ ($q=u,s,c$ and
$M=\pi^0, K^+, \bar{D}^0$). An additional
diagram with $u\leftrightarrow d$ contributes to the
$\pi^0 \pi^0$-production. } \label{fig:pion}
\end{figure}
The universal parameter $g_0$ in QGS model
can be related to the $\rho\pi\pi$ strong coupling
defined as:
\begin{equation}
\langle \pi^-(p_1)\pi^0(p_2)| \rho^-(p_1+p_2)\rangle = g_{\rho\pi\pi}
~\epsilon^{(\rho)}_\mu (p_2-p_1)^\mu\,,
\label{eq:rhopipi}
\end{equation}
where $\epsilon^{(\rho)}$ is the polarization 4-vector of $\rho$-meson.
The numerical value $g_{\rho\pi\pi}\simeq 6.0$ with a negligible error
is then obtained from the measured \cite{PDG} width:
\begin{equation}
\Gamma(\rho\to \pi\pi)=\frac{g_{\rho\pi\pi}^2}{6\pi m_{\rho}^2}
(p^*_{\rho\pi\pi})^3\,,
\label{eq:Grhopipi}
\end{equation}
where $p^*_{\rho\pi\pi}=(m_\rho/2)\sqrt{1-4m_\pi^2/m_\rho^2}$ is the 3-momentum of
the pions in the rest frame of the $\rho$.
Combining the product of couplings defined in (\ref{eq:rhopipi})
with the $\rho$ propagator (neglecting the width) we calculate the
$\pi^+\pi^-\to \pi^0\pi^0$ scattering amplitude in a form of a
Feynman diagram with an ``elementary''
$\rho$-meson exchange in $t$-channel. The result is
\begin{equation}
T^{(\pi^+\pi^-\to \pi^0\pi^0)}_{diag}(s,t)= g_{\rho\pi\pi}^2
\frac{2s+t-4m_\pi^2}{m_\rho^2-t}\,.
\label{eq:amplFeynm}
\end{equation}
At $s\gg |t|,m_{\pi,\rho}^2$ the above amplitude correctly reproduces
the expected $s^J$ asymptotics, where $J=1$ is the spin of the vector meson
exchanged in $t$-channel. The Regge amplitude (\ref{eq:Reggeampl}),
being analytically continued in
the Mandelstam $\{s,t\}$ plane to $t\sim m_\rho^2$ has to
reproduce the Feynman diagram expression (\ref{eq:amplFeynm})
at $s\gg m_\pi^2,|t|$.
Substituting (\ref{eq:gt}) in (\ref{eq:Reggeampl}) and expanding
the $\Gamma$-function near $t=m_\rho^2$ where
$\alpha_\rho(m_\rho^2)=1$ we obtain a pole in the variable $t$:
\begin{equation}
\Gamma(1-\alpha_\rho(t))\simeq
\frac{1}{\alpha'_\rho(m_\rho^2-t)}\,,
\label{eq:gamma} \end{equation}
which
corresponds to the $\rho$-propagator pole in (\ref{eq:amplFeynm}).
Comparing the residues of the amplitudes (\ref{eq:Reggeampl}) and (\ref{eq:amplFeynm}) at large $s$, we obtain
\begin{equation}
\frac{C^{(\pi
\rho)}g_0^2}{\alpha'_\rho\bar{s}}=2g^2_{\rho\pi\pi}\,.
\label{eq:g2}
\end{equation}
Numerically, the above equation at the values
$\alpha'_\rho=0.9$ and $g^2_0/(4\pi)=2.7$ adopted in \cite{KaidV}
correctly reproduces the experimental value of $g_{\rho\pi\pi}$.
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|}
\hline process&Regge & intercept & slope
&scale param.&\\
&pole&$\alpha_R(0)$&$\alpha'_R(\mbox{GeV}^{-2})$&$s_0^{\pi M}(\mbox{GeV}^2)$&$C^{(\pi M)}$\\
\hline $\pi^+\pi^-\to \pi^0\pi^0$ &$\rho$&0.46 &0.9 &1.0&2\\ \hline
$\pi^+\pi^-\to K^+ K^-$ &$K^*$&0.32 &0.85&1.25&1\\ \hline
$\pi^+\pi^-\to D^0\bar{D}^0$&$D^*$&$-0.86$ &0.5&3.55&1\\ \hline
\end{tabular}
\end{center}
\caption{\it Parameters of Regge trajectories \cite{KaidV} involved in the
$\pi^+\pi^-\to M \overline{M}$.}
\label{tab:reggeM}
\end{table}
Repeating the same comparison for the binary reaction
$\pi^+\pi^-\to K^+K^-$ with the strangeness-exchange in the
$t$-channel, we replace the $\rho$ trajectory by the $K^*$
trajectory in the Regge-pole amplitude (\ref{eq:Reggeampl}). The
corresponding parameters of GGS model are presented in Table~\ref{tab:reggeM}. Note that the $SU(3)_{fl}$ violation in this model
(i.e. the effect of a heavier $s$-quark) is reflected in the
parameters of Regge trajectory, and also in the flavour-dependent
normalization scale $s_0^{\pi K}$, introduced in QGS
approach. We then compare the residue function near the pole
at $t=m_{K^*}^2$ with the diagram containing the $K^*$
propagator and the $K^*K\pi$ strong couplings. This diagram
yields an expression similar to
(\ref{eq:amplFeynm}):
\begin{equation}
T^{(\pi^+\pi^-\to
K^+ K^-)}_{diag}(s,t)= \frac{g_{K^*K\pi}^2}{m_{K^*}^2-t}
\left(2s+t-2(m_\pi^2+m_K^2)+\frac{(m_K^2-m_\pi^2)^2}{m_{K^*}^2}\right)\,,
\label{eq:amplFeynmD}
\end{equation}
with the same large $s$ asymptotic
behavior. The relation analogous to (\ref{eq:g2}) yields $
g_{K^*K\pi} =4.5 $ for the $K^{*0}K^+\pi^-$ strong coupling. This
is very close to the value extracted from the $K^*\to K\pi$ width
\cite{PDG}.
Turning to the charmed meson production in the two-pion
collisions, we consider the amplitude (\ref{eq:Reggeampl}) with
the $D^*$ Regge-trajectory: \begin{equation} T^{(\pi^+\pi^-\to D^0
\bar{D}^0)}(s,t)= g_0^2\,\Gamma(1-\alpha_{D^*}(t))
\frac{s}{\bar{s}}\left(\frac{s}{s_0^{\pi
D}}\right)^{\alpha_{D^*}(t)-1} . \label{eq:ReggepiD} \end{equation} In the
QGS approach, the flavour-dependence of the amplitude is reflected
by the substantial differences between the slope parameters of
$D^*$ and $\rho (K^*)$ trajectories on one hand, and between the
scale factors $s_0^{\pi D}$ and $s_0^{\pi \pi(\pi K)}$ on the
other hand, as can be seen from Table~\ref{tab:reggeM}. Hence as
we already mentioned, there is no $SU(4)_{fl}$ symmetry in this
model. Still there remains an important question if the universal
value of $g_0^2$ can be used also in the charm production
amplitude. The $D^*D \pi$ strong coupling is defined as
\footnote{ Note that $2g_{D^*D \pi}$ is equal to the $D^*D \pi$
coupling defined in \cite{BBKR}.} \begin{equation} \langle \pi^-(p_1)D^0(p_2)|
D^{*-}(p_1+p_2)\rangle = g_{D^*D \pi} ~\epsilon^{(D^*)}_\mu
(p_2-p_1)^\mu\,, \label{eq:DstDpi} \end{equation} and the ``elementary''
$D^*$-exchange diagram yields the same expression as in
(\ref{eq:amplFeynmD}), where $K\to D$ and $K^*\to D^*$ have to be
replaced. Continuing the Regge amplitude to $t=m_{D^*}^2$ and
comparing with the large $s$ limit of the $D^*$-exchange
(\ref{eq:amplFeynmD}), we obtain: \begin{equation}
\frac{g_0^2}{\alpha'_{D^*}\bar{s}}=2g_{D^*D\pi}^2 \,.
\label{eq:Dstdpi} \end{equation} Substituting the slope of the $D^*$
trajectory and using the universal value $g_0^2$ of the QGS model
\cite{KaidV} we obtain\footnote{Our estimate differs from the
smaller value quoted in \cite{KaidN} and based on the same model.
Note that in this earlier paper a larger value of the slope
$\alpha'_{D^*}=0.64$ was used.}:
$g_{D^*D\pi}=5.8\,$.
Interestingly, this value is close to the interval estimated from
QCD LCSR in \cite{BBKR} taking into account the gluon radiative
corrections \cite{KRWY}: $[g_{D^*D\pi}]_{\rm LCSR}= 5.0\pm 1.75$.
The only existing measurement of the total $D^*$ width combined
with the branching fraction yields a larger result: $
[g_{D^*D\pi}]_{\rm exp.}=8.95\pm 0.15 \pm 0.95$
\cite{Ahmed:2001xc}. We conclude that the $D^*$ trajectory slope
adopted in QGS model is consistent with the LCSR estimates of the
$D^{\ast}D\pi$ strong coupling.
\begin{figure}[h]
\centering
\includegraphics[scale=0.6]{fig3.eps}
\vspace{-0.3cm} \caption{ \it Dependence of the cross sections of
$\pi^+\pi^-\to K^+K^-$ (solid) and
$\pi^+\pi^-\to D^0\bar{D}^0$ (dashed)
on $p_{lab}$ in QGS model.}
\label{fig:compar}
\end{figure}
Note that one of the inputs used to determine the $D^*$ trajectory of
in QGS model is the Regge trajectory of $J/\psi$,
taken in \cite{KaidV} as:
\begin{equation}
\alpha_\psi(t)=-2.18 +0.33 t\,,
\label{eq:alphapsi}
\end{equation}
and supported by the QGS model analysis of inclusive charm production.
Here one can mention the estimate of the intercept $\alpha_\psi(0)$
obtained in \cite{AKO}, where a four-point correlation function of
heavy-quark-currents was first calculated using OPE in terms of
loop diagram and vacuum condensates and then, via optical theorem,
related to the photon-heavy meson scattering
cross section taken in the Regge form. The comparison of two
representations of the correlation function yields
$\alpha_\psi(0)=-(2 \sim 3)$,
consistent with (\ref{eq:alphapsi}). Note that the
perturbative loop approximation in \cite{AKO} yields $\alpha_\psi(0)=0$,
hence the estimated value of this parameter is entirely determined
by nonperturbative (gluon- and quark-condensate) effects.
To illustrate the QGS model for meson pair production, in
Fig.~\ref{fig:compar} we present the cross sections of the
processes discussed in this section. According to our definition
of the scattering amplitude, the differential cross section is:
\begin{equation}
\frac{d\sigma}{dt}(\pi^+\pi^-\to M \bar{M})= \frac{
|T^{(\pi^+\pi^-\to M \bar{M})}(s,t)|^2}{16\pi s(s-4m_\pi^2)}\,,
\end{equation}
where we substitute the Regge-pole amplitude (\ref{eq:Reggeampl})
for $M=K,D$ and integrate over $|t|$ from the kinematically
allowed minimal $|t_0|$ to $|t_{0}|+\Delta$. Here we choose for
definiteness $\Delta=0.6$ GeV$^2$, so that $|t|$ remains much
smaller than $s$ and hence the Regge-pole description is
applicable. Hereafter we refrain from predicting total cross
sections, because at large $|t| \sim s$ the behavior of the
scattering amplitudes is governed by mechanisms other than a
simple Regge-pole model. Still, differential cross sections
rapidly decrease with $|t|$, hence, our results for the integrated
cross sections provide an order-of-magnitude estimate also for the
total cross sections. As we can see from Fig.~\ref{fig:compar},
the charmed meson production cross section is strongly suppressed
with respect to the strange-meson one.
\section{ $p\bar{p}$- production of
charmed and strange hadrons }
The QGS model
of the charmed baryon-pair production
in $p\bar{p}$ collision is described by the planar diagram in Fig.~1a.
This amplitude, similar to $\pi\pi\to D\bar{D}$, is approximated
by the $D^*$ Regge-trajectory.
The amplitude of $p\bar{p}\to \Lambda_c\bar{\Lambda}_c$ presented in \cite{KaidV} has a helicity-averaged form
\begin{equation}
|T^{(p\bar{p}\to \Lambda_c\bar{\Lambda}_c)}(s,t)|=
g^{(p{\Lambda_c})}(t)
\frac{s}{\bar{s}}\left(\frac{s}{s_0^{p\Lambda_c}}\right)^{\alpha_{D^*}(t)-1}\,,
\label{eq:Reggeampllambdac}
\end{equation}
where the residue function $g^{(p{\Lambda_c})}(t)= C^{p{\Lambda_c}}g_0^2\Gamma(1-\alpha_{D^*}(t))$
contains the same universal coupling $g_0^2$. Importantly, the scale
factor $s_0^{p\Lambda_c}$
obtained following \cite{KaidV}
(see Table~\ref{tab:reggeinp}) is not equal to $s_0^{\pi D}$, reflecting
the difference between baryon and meson production in this model.
Here we consider a more elaborated version of the QGS model
for this process with the helicity
amplitudes (see App.~A).
The differential cross section has the following expression
\begin{eqnarray}
\frac{d\sigma}{d t}(p\bar{p}\to \Lambda_c\bar{\Lambda}_c)= \frac{1}{32 \pi
s(s-4m_p^2)}\bigg
[|H(++,++)|^2+2|H(+-,++)|^2
\nonumber\\
+2|H(++,-+)|^2 +|H(--,++)|^2
+|H(-+,-+)|^2
+|H(+-,-+)|^2 \bigg ] \,, \label{eq:diffLc}
\end{eqnarray}
where in the helicity amplitudes
$H(\lambda_1\lambda_2;\lambda_3\lambda_4)$
the notation $\lambda_{1,2},(\lambda_{3,4})$
denotes the helicities of the proton and antiproton (${\Lambda_c}$ and ${\bar{\Lambda}_c}$), respectively.
The $s,t$ dependence of the amplitudes is not shown for brevity.
At fixed $s$, the region of the momentum transfer
squared $t$ is given by $t_1<t<t_0$, where:
\begin{equation}
t_{0(1)}=m_p^2+m_{\Lambda_c}^2-\frac{s}{2} +(-) \frac12
\sqrt{(s-4m_p^2)(s-4m_{\Lambda_c}^2)}\,.
\end{equation}
We assume that each helicity amplitude has the Regge form
(\ref{eq:Reggeampllambdac}). The residue functions are
fixed by continuing these Regge
amplitudes to the point $t=m_{D^*}^2$ and matching them to the
helicity amplitudes of $p\bar{p}\to \Lambda_c\bar{\Lambda}_c$ with
an ``elementary'' $D^*$ exchange. The latter are given in App.~A.
(see eq.(\ref{eq:helicityDst})) and contain
two independent strong $\Lambda_c D^* N$
couplings defined in (\ref{eq:CC1}), so that the coupling
$g_{{\Lambda_c} ND^*}^V$ ($g_{{\Lambda_c} ND^*}^T$)
enters the helicity-nonflip (-flip) amplitudes.
We arrive at the following
expression for the cross section:
\begin{eqnarray}
\frac{d\sigma}{dt}(p\bar{p}\to {\Lambda_c}\bar{\Lambda}_c)&=&
\frac{C_{A}^{(p\bar{p}\to {\Lambda_c}\bar{\Lambda}_c)}\!(s,t)}{4\pi s(s-4m_p^2)}
|R_{D^*}(s,t)|^2 \nonumber\\
&\times& \Big(|g_{{\Lambda_c} ND^*}^V|^2
+\frac{|t|}{(m_{{\Lambda_c}}+m_N)^2}|g_{{\Lambda_c} ND^*}^T|^2 \Big)^2\,,
\label{eq:dsigdt} \end{eqnarray} where the function \begin{equation}
R_{D^*}(s,t)=\alpha'_{D^*}\Gamma(1-\alpha_{D^*}(t)) \, s \,
\left(\frac{s}{s_0}\right)^{\alpha_{D^*}(t)-1} \label{eq:dsigmaLc}
\end{equation} is determined by the Regge-pole parameters. As opposed to
$\pi\pi\to M \overline{M}$, where we used the original QGS model
with the universal normalization parameter $g_0^2$, there is now a
more subtle substructure of the Regge amplitudes with the strong
couplings determining the helicity-flip and helicity-nonflip
amplitudes. Furthermore, we modified the above cross section with
respect to (\ref{eq:diffLc}) by multiplying it with the so called
absorption factor $C_{A}^{(p\bar{p}\to {\Lambda_c}\bar{\Lambda}_c)}$ which
is included following \cite{KaidV}. This factor derived in App.~B
takes into account the initial- and final-state
rescattering of the baryons and antibaryons and suppresses the
cross sections.
The related processes
of charmed baryon production: $p\bar{p}\to \Sigma_c\bar{\Lambda}_c$ and
$p\bar{p}\to \Sigma_c\bar{\Sigma}_c$, have a similar
description in QGS model, in particular, they are also
dominated by the same $D^*$ Regge-pole exchange.
Their cross sections depend on the
combinations of couplings $g^{V,T}_{\Sigma_cND^*}$
and $g_{{\Lambda_c} N D^*}^{V,T}$. The corresponding
expressions in terms of helicity amplitudes have
minor differences with respect to (\ref{eq:diffLc})
which we will not discuss here for brevity.
The numerical analysis yields
a substantial suppression of the $\Sigma_c$ production
versus $\Lambda_c$ production, due to the difference in the
strong couplings inferred from LCSR.
This suppression will be discussed
below in more detail.
The charmed-meson production, $ p\bar{p} \to \bar{D} D$, is
described in QGS model by the planar diagram depicted in Fig.~1b.
In this case the $t$-channel exchange involves $\Lambda_c$ and
$\Sigma_c$ Regge-trajectories. Their parameters presented in
Table~\ref{tab:reggeinp} are assumed equal. However according to
our predictions, the strong couplings of these baryons to mesons
and nucleons quite differ from each other, hence there is a
significant difference between the cross sections of charged and
neutral charmed meson-pair production. Indeed, in the planar
diagram model, the process $ p\bar{p} \to D^- D^+$ can only be
mediated by the $\Sigma_c^{++}$ exchange in $t$-channel, whereas
in $ p\bar{p} \to \bar{D}^0 D^0$ both trajectories $\Lambda_c$ and
$\Sigma_c^+$ enter the amplitude. Note that this is a
characteristic feature of the planar diagram mechanism. Inelastic
scattering in the final state ($D^0 \bar{D}^0 \to D^{+} D^{-}$)
due to nonplaner diagrams can enhance $D^{+} D^{-}$ production
cross section. Moreover, in a model where these processes are
mediated by intermediate charmonium states in $s$-channel,
$p\bar{p}\to \{\bar{c}c\} \to D\bar{D}$, there is no correlation
between the flavours of initial and final hadrons, so that both
charged and neutral $D$ mesons are produced with equal rates. Such
a model may indeed work as an additional mechanism for charmed
meson-pair production slightly above the threshold, (see e.g.,
\cite{Kerbikov:1994xx}) but the resulting cross section is
much smaller than the one generated by $t$-channel exchanges. Let as
also mention that, according to the model \cite{Haidenbauer:2010nx}
based on the baryon-antibaryon potential, the initial state
inelastic interaction could significantly enhance the $D^{+} D^{-}$
production cross section in the near-threshold region.
Therefore, the charged charmed meson cross section
can serve as a useful check of different charm-production models.
The decomposition in the helicity amplitudes in $ p\bar{p} \to \bar{D}D$
is simpler than for the baryon-pair production because
only the helicities of the initial proton and antiproton are
involved. For the $\bar{D}^0D^0$
production we follow the same method of matching the Regge-pole
amplitude to the ``elementary'' $\Lambda_c$-exchange at $t=m_{\Lambda_c}^2$
and obtain the cross section:
\begin{eqnarray}
\frac{d\sigma}{dt}(p\bar{p}\to
\bar{D}^0 D^0)= \frac{C_{A}^{(p\bar{p}\to\bar{D}^0 D^0 )}\!(s,t)}{32 \pi s(s-4m_p^2)}
|R_{\Lambda_c}(s,t)|^2 (m_{\Lambda_c}^2 -t)|g_{{\Lambda_c} ND}|^4\,,
\end{eqnarray}
where
\begin{eqnarray}
R_{\Lambda_c}(s,t)=\alpha'_{\Lambda_c}\Gamma\big({1 \over 2}-\alpha_{\Lambda_c}(t) \big) \, \sqrt{s}
\, \left(\frac{s}{s^{pD}_0}\right)^{\alpha_{\Lambda_c}(t)-1/2},
\end{eqnarray}
and the $\Sigma_c$ exchange contribution is neglected due
to much smaller couplings.
The absorption factor $C_{A}^{(p\bar{p}\to\bar{D}^0 D^0 )}$ in the
above cross section has a form similar to the one in
(\ref{eq:dsigdt}).
\begin{table}
\begin{center}
\begin{tabular}{|l|c|c|c|c|}
\hline & & & & \\[-2mm]
process&Regge pole & $\alpha_R(0)$& $\alpha'_R({\rm GeV}^{-2}$)
&$s_0^{p H}({\rm GeV}^2$)\\
\hline
& & & & \\
$p \bar{p} \to\Lambda_c \bar{\Lambda}_c, \Sigma_c \bar{\Sigma}_c$ &$D^{\ast}$& $-0.86$ & $0.5$ & $5.76$\\
\hline
& & & & \\[-2mm]
$p \bar{p} \to \Lambda \bar{\Lambda},\Sigma \bar{\Sigma}$ &$K^{\ast}$& $0.32$ & $0.85$ & $2.43$\\
\hline
& & & & \\[-3mm]
$p \bar{p} \to D^{0} \bar{D}^{0}$ &$\Lambda_c,\Sigma_c$ &&&\\
& &$-1.82$ &$0.5$ &$3.30$\\
& & & & \\[-4mm]
$p \bar{p} \to D^{+} D^{-}$ &$\Sigma_c$
&&&\\
\hline
& & & & \\[-3mm]
$p \bar{p} \to K^{+} K^{-}$ &$\Lambda, \Sigma$
&&&\\[-2mm]
& &$-0.64$ &$0.85$ &$1.93$\\
& & & & \\[-4mm]
$p \bar{p} \to K^{0} K^{0}$ &$\Sigma$ &&&\\
\hline
\end{tabular}
\end{center}
\caption{\it Parameters of the Regge trajectories determining the $p
\bar{p}$ amplitudes of charmed and strange hadron-pair production
in QGS model \cite{KaidV}.} \label{tab:reggeinp}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[scale=0.6]{fig4a.eps}\\\vspace{-0.6cm}
\includegraphics[scale=0.6]{fig4b.eps}
\includegraphics[scale=0.6]{fig4c.eps}
\vspace{-0.3cm} \caption{\it Differential cross sections of $ p
\bar{p} \to \Lambda \bar{\Lambda}$ and $ p \bar{p} \to \Sigma
\bar{\Lambda}$ at $p_{lab}=6 $ GeV, and $ p \bar{p} \to K^{+}
K^{-}$ at $p_{lab}=4 $ GeV. The data points are from
\cite{Becker:1978kk,Brabson:1973qx}. The solid curves
are given by QGS model, with the ratio of tensor to vector strong couplings
from LCSR (dashed curves indicate the uncertainties) and
the vector strong couplings fitted to the
measured cross-section normalization.
}\label{fig:Lamdiff}
\end{figure}
Turning to the numerical analysis of the cross sections
we notice that LCSR predictions for strong couplings have a typical
error of $\sim 50 \%$, hence their fourth powers in
the cross sections introduce large uncertainties.
This mainly concerns the $g^V$-couplings. The ratios $g^T/g^V$
are predicted from LCSR with much smaller uncertainties and
moreover, the helicity-flip contributions to the cross sections
proportional to $g^T$-couplings are
kinematically suppressed at small $t$.
In order to decrease the uncertainty of the predicted
cross sections for charmed hadrons we consider also
the strange hadron pair-production in $p\bar{p}$ collisions.
Extending the (modified) QGS model to these processes,
allows us to test it, because in this
case some (albeit, quite old) experimental data are available.
Moreover,
we use the fact that the ratios of the strong couplings
of charmed and strange hadrons given in Table~\ref{tab:res}
have comparatively smaller uncertainties, than the individual
couplings, because
the same nucleon DA's are used in the sum rules in
both cases of charmed and strange hadrons.
Hence, we can constrain the couplings of strange hadrons
by fitting the model cross sections to experimental data
and then use the calculated ratios of the couplings
to reproduce the charm cross sections with smaller uncertainties.
Let us start with the
process $p\bar{p}\to \Lambda\bar{\Lambda}$. Its cross section
in the QGS model has the
same form as (\ref{eq:dsigdt}), with the strong couplings
$g_{\Lambda N K^*}^{V,T}$, the $K^*$ Regge-trajectory and
the absorption factor $C_{A}^{(p\bar{p}\to\Lambda\bar{\Lambda})}$.
We first calculate the differential cross section ${d\sigma \over dt}
(p\bar{p} \to \Lambda \bar{\Lambda})$, without taking into account
the absorption factor, and fit the slope of the $t$-dependence to
the exponential form $\exp(-L_{R}(s) |t|)$.
In this cross section we use the ratio of tensor and vector
strong couplings, $g^T_{\Lambda N K^{\ast}}/g^V_{\Lambda N
K^{\ast}}$, obtained from LCSR (see Table~\ref{tab:res}).
The slope $L_{R}$ is then used to
calculate the absorption factor $C_{A}^{(p\bar{p}\to\Lambda\bar{\Lambda})}$
as explained in App.~B.
Note that the overall normalization of the cross
section depending on the vector strong coupling $g^V_{\Lambda
N K^{\ast}}$ does not play role in this determination.
On the other hand, due to difference of the slopes $L_R$ for
the Regge amplitudes of strange and charmed hadron production
the resulting absorption factor turns out to be almost twice
larger for $p\bar{p}\to \Lambda\bar{\Lambda}$ than for
$p\bar{p}\to \Lambda_c\bar{\Lambda}_c$ at small $t$ (see App.~B),
in accordance with the estimates in \cite{KaidV}.
After inserting the
calculated absorption factor in the differential cross section
of $p\bar{p}\to \Lambda\bar{\Lambda}$, in Fig.\ref{fig:Lamdiff} we
compare the latter with the data points \cite{Becker:1978kk}
at $p_{lab}=6 $ GeV and at small $t$ where we expect the QGS model to work.
Note that not only the Regge amplitude itself
but also the absorption factor contribute to $t$-dependence,
making it steeper. As can be seen from this figure, the agreement
of the shape of the differential cross sections with the data is not very good,
which can possibly be traced back to slightly oversimplified model
of $t$ dependence for this particular (not yet sufficiently large)
energy in our model. Still in the integrated cross section
which is our main interest
here, we expect that the imperfection of the shape does not
play an important role.
As a next step, we fit the overall normalization of the cross
section to the data and obtain the interval of the vector strong
coupling \begin{equation} |g^V_{\Lambda N K^{\ast}}|=5.5_{-0.3}^{+0.2},
\label{eq:expgVL} \end{equation} which is within the broader interval of the
LCSR prediction given in Table~{\ref{tab:res}. After that, we
combine the above estimate with the calculated ratio
$g^V_{\Lambda_c N D^{\ast}}/ g^V_{\Lambda N K^{\ast}}$ (see
Table~{\ref{tab:res}) and estimate the vector coupling for the
charm case $|g^V_{\Lambda_c N D^{\ast}}|= 5.2_{-1.6}^{+1.9}$,
again in agreement with the interval of LCSR prediction. We use
the above ``rescaled'' interval for $g^V_{\Lambda_c N D^{\ast}}$
in obtaining the charmed baryon cross section (\ref{eq:dsigdt}),
thereby decreasing the resulting uncertainty. Note that the ratio
$g^T_{\Lambda_c N D^{\ast}}/g^V_{\Lambda_c N D^{\ast}}$ is again
taken from the LCSR prediction. The cross section of $\Lambda_c$
pair production we are interested in is then calculated in two
steps: first we fix the exponential slope $L_R$ in order to obtain
the absorption factor and second, include this factor in the
cross section. To estimate the cross sections of $ p \bar{p} \to
\Sigma_c \bar{\Lambda}_c,\bar{\Sigma}_c \Sigma_c$ and $p
\bar{p}\to \bar{D}^0 D^0$, we use a similar procedure employing
the available data on strange hadron pair production (see
Fig.~\ref{fig:Lamdiff}). In particular, fitting of the
corresponding strong couplings yields $|g^V_{\Sigma N
K^{\ast}}|=3.9^{+0.1}_{-0.2}$ and $|g_{\Lambda N K}|=
13.9^{+0.9}_{-0.7}$. The LCSR predictions for these couplings
given in Table~{\ref{tab:res} are only marginally consistent with
the above intervals. Note that the predictions of the potential
model \cite{Stoks:1999bz} with the scatteing potentials obeying a
(slightly broken) $SU(3)_{fl}$ symmetry, for the same couplings
are in a better agreement with the fitted values.
Differential cross sections of $ p \bar{p} \to \Lambda_c
\bar{\Lambda}_c ,\Sigma_c \bar{\Lambda}_c , \Sigma_c
\bar{\Sigma}_c $ and $ p \bar{p} \to D \bar{D}$ are displayed in
Fig. \ref{fig:Lcdiff2} as a function of $t$ at $p_{lab}=15\, {\rm
GeV}$. As expected, their slope is much smaller than in the case
of strange hadron production. The integrated cross section $\sigma(t_0,\Delta)$ of
charmed baryon or meson pair-production is defined as the integral
of the differential cross section over the region of small
momentum transfers: $\mbox{max} \{t_1,t_0-\Delta\}<t<t_0$, where
we adopt $\Delta=0.6 \, {\rm GeV^2}$. These cross sections plotted
as a function of $p_{lab}$ in the region accessible to
$\overline{P}ANDA$ are presented in Fig.~\ref{fig:Lctot}. The
summary of our results for the cross sections is also displayed in
Table~{\ref{tab:finres}}.
Let us emphasize that the uncertainties
of the predicted cross sections
are still quite large, even after we narrowed them using the strange hadron
production data. Note that we only quote the uncertainties
stemming from the LCSR estimates of the strong couplings. The QGS model itself
has ``systematical'' uncertainties, which is difficult
to assess quantitatively, as it is the case for any phenomenological
hadronic model not directly related to QCD.
The predictive
power of the model concerns mostly the ratios of cross sections,
where the ``intrinsic'' uncertainties of the method to a large extent
cancel. One important prediction concerns the suppression
of $\Sigma_c$- with respect to $\Lambda_c$-production cross section.
\begin{figure}[t]
\centering
\includegraphics[scale=0.6]{fig5a.eps}
\includegraphics[scale=0.6]{fig5b.eps}\\
\vspace{-0.3cm} \caption{ \it Differential cross sections of
$ p\bar{p} \to \Lambda_c \bar{\Lambda}_c $,
and $ p \bar{p} \to D\bar{D}$
at $p_{lab}=15\, {\rm GeV}$ calculated in QGS model. The dashed lines
indicate the uncertainties caused by LCSR estimates of strong couplings.}
\label{fig:Lcdiff2}
\end{figure}
This suppression is more significant than predicted in \cite{KaidV}
where simple relations based on the nonrelativistic quark-diquark model
for these reactions are used
\begin{eqnarray}
\frac{\sigma(p \bar{p} \to \Lambda_c \bar{\Lambda}_c)}{\sigma(p
\bar{p} \to \Sigma_c \bar{\Lambda}_c)}=
\frac{\sigma(p \bar{p} \to
\Sigma_c \bar{\Lambda}_c)}{\sigma(p \bar{p} \to \Sigma_c
\bar{\Sigma}_c)}= 3 \,.
\label{cross section:relation}
\end{eqnarray}
Our predictions for these ratios
at $p_{lab}= 15 $ GeV are:
\begin{eqnarray}
\frac{\sigma(p \bar{p} \to \Lambda_c \bar{\Lambda}_c)}{\sigma(p
\bar{p} \to \Sigma_c \bar{\Lambda}_c)}=5.1^{+1.0}_{-2.0}, \, &
\qquad & \frac{\sigma(p \bar{p} \to \Sigma_c
\bar{\Lambda}_c)}{\sigma(p \bar{p} \to \Sigma_c \bar{\Sigma}_c)} =
4.6^{+0.9}_{-1.8} \,.
\end{eqnarray}
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
\hline
& & \\
channel & $\frac{ d \sigma}{dt} \big|_{t=t_0} (nb~\mbox{GeV}^{-2})$ & $\sigma(t_0,\Delta) (nb)$ \\
& & \\
\hline
& & \\
$ p \bar{p} \to \Lambda_c \bar{\Lambda}_c$ & $ 130 \, (30 \div 470)$ & $60 \, (15 \div 210)$\\
& & \\
$ p \bar{p} \to \Sigma_c \bar{\Lambda}_c$ & $24 \, (5.0 \div 140)$ & $ 12 \, (2.0 \div 70)$\\
& & \\
$ p \bar{p} \to \Sigma_c \bar{\Sigma}_c$ & $ 5.0 \, (1.0 \div 45)$ & $ 3.0 \, (0.4 \div 24)$ \\
& & \\
$ p \bar{p} \to D^0 \bar{D}^0 $ & $52 \, (13 \div 200)$ & $20 \, (5.0 \div 75)$ \\
& & \\
$ p \bar{p} \to D^+ \bar{D}^- $ & $< 0.01$ & $<0.01$ \\
& & \\
\hline
\hline
\end{tabular}
\end{center}
\caption{\it Differential and integrated cross sections with
$\Delta=0.6~ {\rm GeV^2}$ for charmed hadron production at
$p_{lab} =15 \, {\rm GeV}$.} \label{tab:finres}
\end{table}
Due to the suppression of $\Sigma_c$ couplings
versus $\Lambda_c$ couplings, also the $D^0\bar{D}^0$
production cross section is expected to be significantly larger than
the $D^+\bar{D}^-$ one. It will be very interesting to test
all these predictions experimentally.
\begin{figure}[t]
\centering
\includegraphics[scale=0.6]{fig6a.eps}
\includegraphics[scale=0.6]{fig6b.eps}\\
\includegraphics[scale=0.6]{fig6c.eps}
\includegraphics[scale=0.6]{fig6d.eps}\\
\vspace{-0.3cm} \caption{\it The integrated
cross sections $\sigma(t_0,\Delta)$ of charmed baryon and meson pair
production in $ p\bar{p}$ collisions
in QGS model. The dashed lines
indicate the uncertainties introduced by the strong couplings
obtained from LCSR.}
\label{fig:Lctot}
\end{figure}
\section{Conclusion}
In this paper we bring together the strong couplings of charmed and strange
baryons, both predicted within one and the same QCD based method of LCSR.
We have demonstrated that it is possible to avoid
$SU(4)_{fl}$ approximation.
The relations between couplings are nontrivial
because they stem from the nonperturbative dynamics
which is quite different for heavy and light (also strange) quarks.
The LCSR predictions for strong couplings can be significantly improved in future
by calculating radiative gluon corrections and taking into account soft gluon
components of the nucleon DA's.
The main task of this paper is to estimate the charm production
cross sections in $p\bar{p}$ collisions.
For that purpose we have selected the most (in our opinion) ``QCD-friendly''
model
of hadronic reactions, that is, the Kaidalov's QGS model. This
approach has revealed itself as a very useful tool for
hadronic reactions with different flavours, also for inclusive
production of hadrons. In this paper we used a more detailed description
of binary processes with baryons in terms
of helicity amplitudes and employed the strong couplings
of initial protons and final charmed baryons (mesons) with the
intermediate charmed mesons (baryons)
calculated from LCSR.
Strictly speaking, the QGS model is applicable only at
sufficiently large energies, beyond the upper limit of the
$\bar{P}ANDA$ energy region. Hence the cross sections calculated
here can only be considered as an order of magnitude estimates,
also because the model is only valid at small momentum transfers
and the absorption factor is only taken in the first
approximation. Still the relations between cross sections are less
influenced by the uncertainties and are almost independent of the
absorption factors. In future, the model adopted in this paper can
be developed further, taking into account of the subleading Regge
trajectories and a more elaborated absorption ansatz.
Finally, turning to the comparison of our results with the
charm-production estimates in the literature, we observe that our
prediction for the dominant $\Lambda_c\bar{\Lambda}_c $ production
is (within estimated uncertainties) consistent with the one
obtained in the original QGS model \cite{KaidV}: $\sigma
(p\bar{p} \to \Lambda_c\bar{\Lambda}_c) \simeq 100$ nb,
at $p_{lab}=15~ \mbox{GeV}$, whereas the predictions for the ratios of
cross sections obtained here and in \cite{KaidV} differ. For example,
we do not exclude a larger charmed meson cross section
than $\sigma (p\bar{p} \to D^0\bar{ D^0}) \simeq 5$ nb predicted
in \cite{KaidV}.
A model of exclusive charm production cross sections
based on the QGS and Regge-poles can be found
in \cite{TitovKampfer}, where the $SU(4)_{fl}$ symmetry is used
and a
different form of the cross section is adopted,
adding a $t$-dependent dipole ``residual factor''.
Numerically, our predicted intervals
for the differential cross sections at $t_{0}$ turn out to be larger than
the ones in \cite{TitovKampfer}.
The other models in the literature are based on
radically different approaches. E.g., in
\cite{Haidenbauer:2009ad}} a hadronic baryon-antibaryon potential
derived from the coupled channel approach is used, predicting the
cross section of $\Lambda_c\bar{\Lambda}_c$ production up to a few $\mu b$
near the threshold, i.e., much larger than obtained here.
On the opposite side are the typically smaller cross sections obtained
from perturbative approaches, such as the inclusive charm production
estimate in the parton model \cite{Braaten}
and the approach \cite{Goritschnig:2009sq}
to $p\bar{p} \to \Lambda_c\bar{\Lambda}_c$
employing distribution amplitudes of initial and final baryons.
Concluding, this paper contains an attempt to apply QCD
predictions for hadronic strong couplings to the models of
exclusive hadronic reactions. Our estimates for charm production
cross sections contain rather large uncertainties. Still even the
lower limit of the cross sections predicted here allows one to
expect an appreciable number of charmed baryons and mesons
produced at $\bar{P}ANDA$. We look forward to other applications
of the strong couplings presented in this paper and their future
improvements.
\section* {Acknowledgments}
This work is supported by the German Ministry for Education and Research
(BMBF) under contract 06SI9192. T.M. also thanks A. Titov for a discussion
on the subject of the paper.
\section*{Appendix A: Helicity amplitudes}
The helicity amplitudes
of $\bar{p}p\to \bar{\Lambda}_c\Lambda_c$ scattering
via $D^*$ meson exchange are obtained from the
initial invariant scattering amplitude
\begin{eqnarray}
H(\lambda_1 \lambda_2; \lambda_3 \lambda_4) &=& {1 \over
m_{D^{\ast}}^2 - t} \bar{u}_{\Lambda_c} (p_3, \lambda_3) \bigg[ g^V_{
\Lambda_c ND^*}\slashed{\epsilon}+i\,\frac{g^T_{ \Lambda_c
ND^*}}{m_{ \Lambda_c}+m_N}\sigma_{\mu\nu}\epsilon^\mu q^\nu \bigg]
u_N(p_1,\lambda_1) \nonumber \\
&& \bar{v}_{\bar{N}}(p_2,\lambda_2) \bigg[ g^V_{ \Lambda_c
ND^*}\slashed{\epsilon}^{\ast}+i\,\frac{g^T_{ \Lambda_c ND^*}}{m_{
\Lambda_c}+m_N}\sigma_{\rho \tau}{\epsilon^{\ast}}^\rho q^\tau
\bigg] v_{\Lambda_c} (p_4, \lambda_4)\,,
\end{eqnarray}
where the baryon bisponors are distinguished by their indices
so that ($p_1, \lambda_1$), ($p_2, \lambda_2$), ($p_3, \lambda_3$)
and ($p_4, \lambda_4$) are the four-momenta and helicities of the
proton, antiproton, $\Lambda_c$ and $\bar{\Lambda}_c$
respectively; $\epsilon_{\mu}$ is the polarization vector of
the virtual $D^{\ast}$ meson. Generally, there are 16 different
helicity amplitudes for $\bar{p}p\to \bar{\Lambda}_c\Lambda_c$
process, only six of them are independent due to symmetries
\cite{Bourrely:1980mr}.
In the c.m. frame of proton-antiproton pair we choose the ${x,z}$ plane
for the process, and the 3-momentum
of proton in the $z$ direction, so that the 3-momentum of $\Lambda_c$
has the angular coordinates $(\theta, \varphi=0)$.
Then the kinematics is as follows:
\begin{eqnarray}
p_1 = {1 \over 2}( \sqrt{s} , 0 ,0 , \sqrt{s- 4 m_N^2} ) \,,
~~p_2 = {1 \over 2}( \sqrt{s} , 0 ,0 , - \sqrt{s- 4 m_N^2} )
\,,
\nonumber \\
p_3 = {1 \over 2} ( \sqrt{s} , \sqrt{s- 4 m_{\Lambda_c}^2} \sin
\theta, 0, \sqrt{s- 4 m_{\Lambda_c}^2} \cos \theta ) \,,
\nonumber \\
p_4 = {1 \over 2} ( \sqrt{s} , -\sqrt{s- 4 m_{\Lambda_c}^2}
\sin \theta,0, -\sqrt{s- 4 m_{\Lambda_c}^2} \cos \theta )
\end{eqnarray}
Summing over the polarization of $D^{\ast}$ meson:
\begin{eqnarray}
\sum_{\lambda=1, 2, 3, 4} \epsilon^{\mu}(q, \lambda) \epsilon^{\nu
\ast}(q, \lambda) =-g^{\mu \nu} + {q^{\mu} q^{\nu} \over m_{D^*}^2
}\,,
\end{eqnarray}
and substituting explicitly the bispinors with various helicities
in the chosen frame we obtain rather bulky expressions of helicity
amplitudes, which however greatly simplify in the limit of large
$s$ where we compare them with the Regge amplitudes. The six
helicity amplitudes are:
\begin{eqnarray}
H(++,++) &=& { s \over t- m_{D^{\ast}}^2} 2 (g^V_{ \Lambda_c
ND^*})^2 \,, \nonumber \\
H(+-,++) &=& { s \over t- m_{D^{\ast}}^2} {2 \sqrt{-t} g^V_{
\Lambda_c ND^*} g^T_{ \Lambda_cND^*} \over m_{ \Lambda_c}+m_N}
\nonumber \\
H(++,-+) &=& -{ s \over t- m_{D^{\ast}}^2} {2 \sqrt{-t} g^V_{
\Lambda_c ND^*} g^T_{ \Lambda_cND^*} \over m_{ \Lambda_c}+m_N}
\nonumber \\
H(--,++) &=& -{ s \over t- m_{D^{\ast}}^2} {2 t (g^T_{ \Lambda_c
ND^*})^2 \over (m_{\Lambda_c}+m_N )^2 }
\,, \nonumber \\
H(-+,-+) &=& { s \over t- m_{D^{\ast}}^2} 2 (g^V_{ \Lambda_c
ND^*})^2
\,, \nonumber \\
H(+-,-+) &=& { s \over t- m_{D^{\ast}}^2} {2 t (g^T_{ \Lambda_c
ND^*})^2 \over (m_{ \Lambda_c}+m_N )^2}\,.
\label{eq:helicityDst}
\end{eqnarray}
It is clear that in the large $s$ limit only three helicity
amplitudes are independent. The amplitudes can be related to
each other through the following relations:
\begin{eqnarray}
H(-\lambda_1 -\lambda_2; -\lambda_3 -\lambda_4) &=&
(-1)^{\lambda_1-\lambda_2-\lambda_3+\lambda_4} H(\lambda_1
\lambda_2; \lambda_3 \lambda_4)\,, \nonumber \\
H(\lambda_2 \lambda_1; \lambda_4 \lambda_3)&=&
(-1)^{\lambda_1-\lambda_2-\lambda_3+\lambda_4} H(\lambda_1
\lambda_2; \lambda_3 \lambda_4)\,,
\end{eqnarray}
following from the parity and charge-conjugation
invariance.
\section*{Appendix B: Absorption factor }
Here we derive the absorption factor $C_{A}(s,t)$, multiplying
the cross section.
The absorption is generated by the (quasi) elastic rescattering of
the initial proton and antiproton as well as of the final hadron pair,
both are approximated by the pomeron exchange \cite{KaidV}.
Here we make a simplifying assumption that the elastic rescattering
amplitudes dominate, they do not change the helicities
and are the same in the initial
and final states, independent of the flavour
content of the latter.
Consider a process $p\bar{p}\to B\bar{B}$ with generic $B$ hadrons
in the final-state. The amplitude in QGS model, having the form
(\ref{eq:Reggeampllambdac}) has predominantly
exponential behavior at small $t$, the main source of it
is the Regge-pole factor $(s/s_0)^{\alpha{(t)}}$.
Therefore, the $p\bar{p}\to B\bar{B}$
amplitude can be cast in the exponential form
\begin{equation}
T_R(s,t)=f_R(s) \exp\left(-\frac{L_R}{2}|t|\right)\,.
\label{eq:Texp}
\end{equation}
We then switch to the impact parameter representation,
where the 2-dimensional vector $\vec{b}$ is conjugate to the
transverse momentum transfer $\vec{q}_\perp$:
\begin{equation}
T_R(s,b)=\int
T_R(s,t) \, {\rm
exp}(i\vec{q}_\perp\cdot\vec{b})\frac{d\vec{q}_\perp}{2\pi}\,,
\label{eq:Tb}
\end{equation}
and at high energies $t\equiv q^2\simeq -|\vec{q}|^2$. The
angular integration yields:
\begin{equation}
T_R(s,b)=\frac12
\int\limits_0^\infty d|t|J_0\left
(\sqrt{|t|}b\right)T_R(s,-|t|)\,,
\end{equation}
where $J_0$ is the Bessel
function and $b=|\vec{b}|$. Substituting the exponential
representation (\ref{eq:Texp}) in the above integral, and
integrating over $t$ we obtain
\begin{equation}
T_R(s,b)=\frac{f_R(s)}{L_R}\exp\left(-\frac{b^2}{2L_R}\right)\,.
\label{eq:Trb}
\end{equation}
The rescattering in the initial and final state is dominated by the
pomeron amplitude $T_P(s,t)$ which is
predominantly imaginary and has an exponential form in the
momentum transfer $T_P(s,t)=T_P(s,0)\exp(-L_P|t|/2)$. The
forward-scattering amplitude is expressed via total $p\bar{p}$ cross section
using the optical theorem: $\mbox{Im}
T_P(s,0)=2p^*\sqrt{s}\sigma_{p\bar{p}}^{tot}(s)$, where $p^*$ is
the 3-momentum in the c.m. system of the $p\bar{p}$ collision.
Hence one obtains for the pomeron-mediated elastic rescattering
amplitude:
\begin{equation}
T_{P}(s,t)=2ip^*\sqrt{s}\,\sigma_{p\bar{p}}^{tot}(s)\exp(-L_P|t|/2).
\end{equation}
Employing the same Fourier-transformation to the impact
parameter space as in (\ref{eq:Tb}), it is easy to get the impact parameter
representation for this amplitude: \begin{equation}
T_{P}(s,b)=\frac{2ip^*\sqrt{s}\sigma_{p\bar{p}}^{tot}(s)}{L_P}
\exp\left (-\frac{b^2}{2 L_P}\right).
\end{equation}
The absorption contribution added to the initial Reggeon amplitude
in the $b$ space yields:
\begin{equation}
T(s,b)=T_R(s,b)\left[1+i\frac{T_P(s,b)}{8\pi p^*\sqrt{s}} \right]=
T_R(s,b)\bigg[1-\chi(s,b)\bigg]\,, \label{eq:absb}
\end{equation}
where
\begin{equation}
\chi(s,b)=\frac{\sigma_{p\bar{p}}^{tot}(s)}{4\pi L_P}\exp\left(-\frac{b^2}{2L_P}\right)\,,
\label{eq:chifact}
\end{equation}
and the normalization factor multiplying $T_P$
corresponds to the convention of impact parameter representation
adopted in \cite{KaidV}.
Substituting (\ref{eq:Trb}) in (\ref{eq:absb}) and performing the inverse
Fourier transformation to the $t$-dependent amplitude we finally obtain:
\begin{equation}
T(s,t)=T_R(s,t)\left[1-
\frac{\sigma_{p\bar{p}}^{tot}(s)}{4\pi(L_P+L_R)}
\exp\left(\frac{L_R^2|t|}{2(L_P+L_R)}\right) \right]\,.
\label{absorbtion:first}
\end{equation}
This expression takes into account absorption in the amplitude
in the first approximation.
To obtain $C_A(s,t)$ one simply has to square the expression in brackets
multiplying $T_R(s,t)$ in the above.
At high energies this effect should be resummed (exponentiated),
however at small $t$'s we are considering here
the resummation effects, as we checked numerically, are small,
so that the first approximation for $C_A$ is sufficient.
For the numerical evaluation of the absorption factor in
(\ref{absorbtion:first}), the data on $\sigma_{p\bar{p}}^{tot}(s)$
in a parameterized form are taken from \cite{PDG}, so that
$\sigma_{p \bar{p}}^{tot}(s)$ changes from 52.5~mb to 47.9~mb in
the interval $p_{lab}= 10$ GeV to 20 GeV. The slope of the pomeron
mediated elastic $\bar{p}p$ scattering is taken from
\cite{Okorokov:2008zf}, e.g., $L_P=12.1 \,\,{\rm GeV}^{-2}$ at
$p_{lab}= 15$ GeV. Finally, the slopes of Regge-pole amplitudes
fitted to the exponential form (\ref{eq:Texp}) are (in units
GeV$^{-2}$): $L_R=2.6$ ($p_{lab}= 6$ GeV) and $2.5$ ($p_{lab}= 4$
GeV) for $p\bar{p}\to \Lambda \bar{\Lambda}$,
$\Sigma\bar{\Lambda}$ and $K^+K^-$, respectively. For
$p\bar{p}\to\Lambda_c \bar{\Lambda}_c$, $\Sigma_c
\bar{\Lambda}_c$, $\Sigma_c \bar{\Sigma}_c$, $D \bar{D}$, the
corresponding slopes are $L_R=0.6$, $ 0.4$, $0.2$, $1.2$
($p_{lab}= 15$ GeV), respectively. For numerical illustration, we
quote the absorption factors calculated at the same energy
$p_{lab}=15 $ GeV and at $t=t_0$ for strange and charmed baryon
production: $C_A^{(p\bar{p}\to \Lambda \bar{\Lambda})}=0.09$ and
$C_A^{(p\bar{p}\to\Lambda_c
\bar{\Lambda}_c)}=0.04$. These values are in agreement with ($t$-averaged)
values presented in \cite{KaidV} and indicate a strong absorption effect
on one hand and a large difference between this effect for strange and
charmed baryons.
|
\section{Introduction\label{1}}
The well--known Be~star of spectral type B0IVe
$\gamma$~Cassiopei\ae\ (27~Cas, HR~264, HD~5394, HIP~4427, MWC~9, ADS~782A),
is one of the first two Be~stars ever discovered \citep[see][]{secchi66} and
a~member of a visual multiple system. It exhibits spectral and brightness
variations on several timescales. It underwent two major shell phases
in 1935--36 and 1939--40. Afterwards, it appeared briefly as a~normal B~star.
Emission strength of the Balmer lines and the brightness of the star
in the visual region had been rising slowly during
the rest of the 20$^{\rm th}$ century.
The observational history of the star has been summarized in detail
by~\citet{hec2002}.
In 1976, $\gamma$~Cas was identified as an X-ray source. This discovery started a long debate
over whether the source of X-rays is the star itself or whether
$\gamma$~Cas is an X-ray binary with a~mass accreting compact binary companion.
In an effort to prove the duplicity of $\gamma$~Cas, \citet{cowley1976}
measured radial velocities (RVs hereafter) on a rich collection
of~photographic spectra obtained in the years 1941--1967.
They were unable to find any RV changes exceeding 20~{km~s$^{-1}$}\ or to detect
any coherent periods between 2\fd5 and 4000\fd0.
\citet{jarad87} measured RVs on 81 medium-dispersion (30~{\ANG~mm$^{-1}$})
photographic spectra using
the cross-correlation technique. They concluded that the RVs vary with
a short period of 1\fd16885, a semi-amplitude of 27.7~{km~s$^{-1}$}, and a well-defined
phase curve. Combining their RVs with those measured
by \citet{cowley1976}, they found a period of 0\fd705163 with a semi-amplitude
of only 8.6 {km~s$^{-1}$} . They preferred the shorter period, which they
interpreted as either a rotational or pulsation period of the star.
\citet{robinson2000} published a~detailed study of the X-ray flux of $\gamma$~Cas.
They found that the X-ray flux varied with a~period $P$~=~1$^{\rm d}\!\!.$ 12277, which
they tentatively interpreted as the~rotational period of~$\gamma$~Cas. They
used this finding as one of the arguments against the~binary scenario
for the X-ray flux. More recently, \citet{smith2006} have reported
a coherent periodicity of 1\fd21581\p0\fd00004 from the 1998-2006
optical photometry, prewhitened for variations on longer time scales.
The latter authors pointed out that the initial period
estimate of 1\fd12277 was probably an alias of the
correct period near 1\fd21581.
\cite{hec2000} measured RVs of the steep H$\alpha$ emission wings
in a series of 295 Ond\v{r}ejov Reticon spectra spanning nearly 2500~days
from 1993 to 2000.
After removing the long-term RV changes, they discovered periodic RV
variations with a period $P$~=~203\fd59$\pm$ 0\fd29, semi-amplitude
$K_1$~=~4.68~{km~s$^{-1}$}, and eccentricity $e$~=~0.26, which they interpreted
as the binary motion around a common centre of gravity with
a low-mass companion. They demonstrate that the published RVs from
the photographic spectra can also be reconciled with the 203\fd59 period
and discuss the possible properties of the system.
Their result was confirmed by~\citet{mirosh2002}, who also measured
the RVs of the H$\alpha$ emission wings in a series of 130 electronic echelle
spectra, secured with the 1~m~reflector of~the~Ritter Observatory between
1993 and 2002. These two studies differ in the technique of RV measurements.
While Harmanec et al. (2000) measured the RVs manually, sliding the direct
and reversed continuum-normalized line profiles within a range of intensities
on the computer screen until the best match was obtained, Miroshnichenko et al. (2002)
also matched the original and reversed profiles, but used an automatic
procedure. They arrived at a period of
205\fd50$\pm$ 0\fd38, semi-amplitude of 3.80$\pm$ 0.12 {km~s$^{-1}$}, and a {\sl circular}
orbit. They discuss several possible reasons why their results
differ significantly from those of \citet{hec2000}.
\citet{mirosh2002} also document the cyclic long-term spectral variations
of $\gamma$~Cas over a time interval from 1973 to 2002.
To shed more light on the differences between these two RVs studies, to obtain
truly reliable orbital elements of $\gamma$~Cas, and to exclude possible
1~d aliases of the 204~d period, we combined our efforts and
analysed the two sets of spectra, complemented by more recent observations
from Ond\v{r}ejov and additional spectra from the Dominion Astrophysical
(DAO), Haute Provence (OHP) and Castanet-Tolosan Observatories.
The RVs in all these spectra were measured by both measuring
techniques -- alternatively used by \citet{hec2000} and
\citet{mirosh2002} -- and analysed
in several different ways. Here we present the results
of our investigation. We also studied the long-term and
rapid spectral variations of $\gamma$~Cas in our data.
\section{Spectral variations\label{2}}
We have collected and analysed series of electronic spectra from
five observatories. They cover the time interval from 1993 to 2010.
The journal of the observations is in~Table~\ref{spectra},
where the wavelength range, time interval covered, the number of spectra, and
the spectral lines are given. For more details on the individual
datasets, readers are referred to Appendix~\ref{apen}.
\begin{table}[h]
\caption{Journal of spectral observations.}
\begin{tabular}{ccclrllll}
\hline
\hline
Origin of &$\Delta\lambda$ &$\Delta T$ &Lines &$N$\\
spectra &(\accent'27A) &(RJD) & &\\
\hline
Ond &6300--6730 &49279--55398 &Ha, He, Si &439 \\
Rit &6528--6595 &49272--52671 &Ha &204 \\
OHP &4000--6800 &50414--52871 &Ha, He, Si &34 \\
DAO &6155--6755 &52439--54911 &Ha, He, Si &136 \\
OHP/Cst &various &53997--55422 &Ha &13 \\
\hline
\end{tabular}
\tablefoot{Ond = 2~m reflector of the Astronomical Institute AS CR Ond\v{r}ejov,
Rit = 1~m reflector of the Ritter Observatory of the~University of~Toledo,
DAO = 1.22~m reflector of the Dominion Astrophysical Observatory,
OHP = 1.52~m reflector of the Haute Provence Observatory,
Cast = Castanet-Tolosan; $\Delta\lambda$ = the wavelength region covered,
$\Delta T$ = the time interval spanned by each dataset, where
times are given in the {\sl reduced} Julian dates RJD = HJD-2400000,
Lines: Ha~=~H$\alpha$, He~=~\ion{He}{i}~6678~\ANG, Si~=~Si\,II~6347~\ANG , and~Si\,II~6371~\ANG .
}
\label{spectra}
\end{table}
We focused our study on the lines in the H$\alpha$ region, which are
available for all spectra, although several echelle spectra
cover almost the whole visible region of the electromagnetic spectrum.
In particular, we studied the following spectral lines: H$\alpha$, \ion{He}{i}~6678~\ANG, Si\,II~6347~\ANG,
and Si\,II~6371~\ANG. No dramatic changes were found in these line profiles.
The \ion{He}{i}~6678~\ANG ~ and \ion{Si}{II} lines exhibit double-peaked emissions
with the well-known $V/R$ variations (changes in the relative
strength of the shorter wavelength, ``violet'', to the longer wavelength,
``red'', peak) on the timescale of several years.
Over the whole time interval covered by our spectra, the H$\alpha$ line
was observed as a~strong, basically single--peaked emission, having a peak
intensity between 3.5 and 5.0 in the units of the continuum level.
Its $V/R$ variations manifest themselves as a relative shift in the
emission peak with respect to the centre of the emission profile.
Several shallow absorptions can be noted in the H$\alpha$ line in some of
the studied spectra, but most of them are the telluric water vapour lines.
Ocassionally, some weak shallow absorptions of probably stellar origin
were seen, but they disappeared in less than several
tens of days, and we found no regularity in their appearance and disappearance.
The \ion{He}{i}~6678~\ANG ~ line consists of a double--peaked emission
filling a large part of the rotationally broadened photospheric
(or pseudophotospheric) absorption. The whole line is very weak
and can only be measured reliably on the spectra with high $S/N$.
The emission peaks rise only a few percent above the continuum level.
Nevertheless, the time variations are seen most prominently
in this line. The red peak of the \ion{He}{i}~6678~\ANG ~ emission disappeared almost
completely at certain times. It is hard to say whether these variations
represent only real long-term changes or whether they are partly
caused by line blending.
The Si\,II~6347~\ANG ~ and Si\,II~6371~\ANG ~ double emission lines are even weaker than the \ion{He}{i}~6678~\ANG ~ line, and
their peak intensity never exceeds 5\% of the continuum level.
The more recent evolution of the H$\alpha$ line profile is shown in Fig.~\ref{haevol}.
All line profiles shown were obtained
after RJD~=~52225 and were not included in the study by \citet{hec2000}.
A similar sequences of the \ion{He}{i}~6678~\ANG, Si\,II~6347~\ANG , and Si\,II~6371~\ANG ~ line profiles are
shown in Fig.~\ref{heevol}. All displayed spectra are from Ond\v{r}ejov, to
compare the data with the same resolution.
There are, however, the huge differences in the flux scale between H$\alpha$, \ion{He}{i}~6678~\ANG,
Si\,II~6347~\ANG , and Si\,II~6371~\ANG ~ lines.
\begin{figure}[h]
\centering
\includegraphics[width=\hsize]{17922fg1.eps}
\caption{Recent evolution of the H$\alpha$ line profile.
Mid--exposure times of the displayed spectra are in RJD = HJD--2400000.}
\label{haevol}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=\hsize]{17922fg2.eps}
\caption{Recent evolution of the \ion{He}{i}~6678~\ANG ~ (left), Si\,II~6347~\ANG , and Si\,II~6371~\ANG ~ (right)
line profiles. Mid--exposure times of the displayed spectra are in
RJD = HJD--2400000. The vertical axis is in the units of the continuum flux.}
\label{heevol}
\end{figure}
We fitted the H$\alpha$ line profiles with a Gaussian profile to obtain
their peak intensity ($I_{\rm p}$ hereafter) and full width at half maximum
(FWHM hereafter). This procedure naturally returns a value of $I_{\rm p}$,
which is slightly less than the very maximum of the emission
profile, which is affected by both the $V/R$ variations and blending with
the neighbouring telluric lines. We do believe, however, that the fitted
Gaussian provides an objective measure of the gradual changes in the
emission-line strength. We omitted the saturated H$\alpha$ line profiles, of course.
One example of a Gaussian fit is in Fig.~\ref{fitexam} to show
where the FWHM and $I_{\rm p}$ were measured.
\begin{figure}[h]
\centering
\includegraphics[width=\hsize]{17922fg3.eps}
\caption{An example of the Gaussian fit to an H$\alpha$ line profile,
which also shows the derived quantities $I_{\rm p}$ and FWHM.
Dashed line: the observed H$\alpha$ profile; solid line: the Gaussian fit to it;
dotted line: the continuum level.}
\label{fitexam}
\end{figure}
Figure~\ref{fwhmic} shows the time variations of the FWHM and $I_{\rm p}$.
An interesting finding is that the secular variations of these two quantities
are anticorrelated with each other. The apparently
increased scatter of both dependencies between RJDs\,$\approx$\,50000 and
52000 is caused solely by the lower resolution in intensity of
the Ond\v{r}ejov Reticon spectra taken prior RJD\,=\,52000.
\footnote{Although the original Reticon detector and the currently used
CCD detector were attached to the same camera of the coud\'e
spectrograph and have the same {\sl spectral} resolution, the control
electronics of the Reticon detector allowed distinguishing only 4000 steps
in intensity, while the CCD detector recognizes 60000 intensity steps.
This leads to a systematic difference for strong emission lines.}
\begin{figure}[h]
\centering
\includegraphics[width=\hsize]{17922fg4.eps}
\caption{The secular time variations of $I_{\rm p}$ (upper panel)
and FWHM (bottom panel) of the H$\alpha$ line.
The time on abscissa is in RJD = HJD-2400000. The open symbols
denote measurements from the Ond\v{r}ejov Reticon detector, capable of
distinguishing only 4000 intensity steps, which is why these measurements
deviate systematically from the rest. See the text for details.}
\label{fwhmic}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=\hsize]{17922fg5.eps}
\caption{Comparison of the automatically measured H$\alpha$ emission-wing
RVs (ordinate = $y$) with those measured manually (abscissa = $x$). The dashed
line is a fitted linear function $y~=~0.968\,x$.}
\label{compall}
\end{figure}
\section{RV changes\label{3}}
\subsection{Long--term and periodic RV changes and the new
orbital solutions\label{3long}}
Similar to \citet{hec2000} and \citet{mirosh2002}, we measured
the steep wings of the H$\alpha$ emission in all unsaturated profiles.
The reasons the emission wings should provide a good estimate
of the true orbital motion of the Be primary around the common centre of
gravity with the secondary were recently summarized in detail
by \citet{rudzjak2009}. To them, we can add that \citet{poec1978} modelled
the H$\alpha$ emission of $\gamma$~Cas , and their model showed that the H$\alpha$ emission
wings originate in regions that are much closer to the star than the radiation
that is forming the upper part of the line.
\begin{figure}[h]
\centering
\includegraphics[width=\hsize]{17922fg6.eps}
\includegraphics[width=\hsize]{17922fg7.eps}
\caption{Plots of RVs vs. RJD = HJD-2400000:
{\sl Top panel:} Manually measured H$\alpha$ emission-wing RVs.
{\sl Bottom panel:} Automatically measured H$\alpha$ emission-wing RVs.
The solid lines in all panels represent the long--term RV change as
derived with the program HEC13. The squares in both panels denote
the local `systemic' velocities calculated with the program SPEL for
individual data subsets. The dashed line represents Hermite--polynomial
fit computed with program HEC36. See the text for details.}
\label{13hahe}
\end{figure}
Moreover, we also attempted to measure the RVs of other available
spectral lines (\ion{He}{i}~6678~\ANG, Si\,II~6347~\ANG ~ and Si\,II~6371~\ANG) to see if they undergo similar
time changes. We primarily measured also the outer emission wings of
these lines, but for the \ion{He}{i}~6678~\ANG ~ line it was possible to obtain
a relatively accurate RV of the central absorption core.
The RVs measured on the emission wings of the Si\,II~6347~\ANG ~ and Si\,II~6371~\ANG ~ lines were
averaged.
Two methods of RV measurement were used as follows.
\begin{enumerate}
\item {\sl Manual measurements} were carried out in the program SPEFO,
written by Dr.~J.~Horn and more recently improved by Dr.~P.~\v{S}koda and
Mr.~J.~Krpata \citep[see][]{horn1996,skoda1996}.
The user can slide the direct and reversed line profile on the computer
screen until a perfect match of the selected parts of the profiles
is achieved. The advantage of this, admittedly a bit tedious procedure, is
that the user actually sees all measured line profiles and can avoid any
flaws and blends with the telluric or weak stellar lines.
It is invaluable for measurements of weak spectral lines, where any
automatic method can be easily fooled by the noise.
The same measuring technique has also been used by \citet{hec2000}.
One of us, JN, also re-measured all Reticon spectra used in their study.
Plots of the old~vs.~new measurements are available as Fig.~\ref{cmphecjn}
in Appendix~\ref{apen}. Considering the good agreement of both
measurements, we used the mean value of the original and new RV
measurements for each studied feature.
All lines (whenever available) were measured with this technique.
\item {\sl Automatic measurements} were obtained with a program written
by AM, which also shifts the direct and reversed line-profile images
for a selected range of relative intensities in the continuum units
to find a minimum difference between them. The advantages of this method
are that it is fast and impersonal. A potential danger is that it can
be fooled by flaws and blends in some particular cases.
Only the H$\alpha$ emission wings were measured with this method.
\end{enumerate}
We denote the RVs measured by the first method as {\sl manual}
and those measured by the second method as {\sl automatic}
to distinguish them in following sections.
In Fig.~\ref{compall} the automatic H$\alpha$ emission RVs are plotted
vs. the manually measured ones to see whether there is any systematic
difference between the two methods. We fitted the data with a linear
relation and somewhat surprisingly the slope was found to be
$0.968$$\pm$$0.009$; i.e., the automatic method finds a slightly
narrower total range of RV variations than the manual one.
All individual RV measurements on the steep wings of the H$\alpha$ emission
are published in detail in Table~2 for the manual, and in Table~3
for the automatic measurements.\footnote{Tables 2 and 3 are published
only in the electronic form.}
The H$\alpha$ emission RVs measured manually and automatically are plotted
vs.~time in Fig.~\ref{13hahe}. Additional time plots for other
measured features can be found in Appendix~\ref{apen}.
Figure~\ref{13hahe} shows that $\gamma$~Cas exhibits
long-term RV variations over several years, which seem
to correlate with those of the peak intensity of the emission. To be able
to search for periodic RV changes on a shorter timescale, one has
first to~remove the long-term ones.
To check how robust the result is or how much it depends on the specific
way of secular-changes removal, we applied three different approaches
to this goal.
The~first was to use the program HEC13
written by PH, which is based on the smoothing technique developed
by~\citet{vondrak1969, vondrak1977} and which uses some subroutines
kindly provided by Dr.~Vondr\'ak\footnote{The program HEC13 with brief instructions how to use it is available
to interested users at http://astro.troja.mff.cuni.cz/ftp/hec/HEC13.}.
The level of smoothing is controlled by a smoothing parameter $\epsilon$
(the lower the value of $\epsilon$, the higher the smoothing),
and the smoothing routine can operate either through individual data points
or through suitably chosen normal points, which are the weighted mean
values of the observed quantity (RV in our case) over the chosen constant
time intervals. In both cases, the \hbox{$O\!-\!C$}\
residua are provided for all individual observations. In these particular
cases the following specifications for smoothing were used:
$\epsilon$~=~5$\times$10$^{{-16}}$ and 200~d normals for the H$\alpha$ emission-wing
RVs measured by both methods, and
$\epsilon$~=~1$\times$10$^{{-16}}$ and 200~d normals
for the RVs measured on the \ion{He}{i}~6678~\ANG ~ absorption core.
These particular choices of $\epsilon$ make the smoothing function
follow only the secular RV variations.
The solid lines in the three panels of Fig.~\ref{13hahe} show the estimated
long-term changes, which were then substracted from the original RVs.
These prewhitened RVs were than searched for periodicity
from 3000\D0 down to 0\D5 with the program HEC27 (also written by PH),
based on the PDM technique developed by \citet{stell78}.
The $\theta$-statistics periodograms for the manually and automatically
measured H$\alpha$ emission RVs are plotted in Figs.~\ref{thetajn} and \ref{thetamr},
respectively. The upper panels in both plots show the range
of periods from 3000\D0 down to 50\D0, while the periods from 2\D0 down
to 0\D5 are shown in the lower panels. The periodograms are flat,
with $\theta$ close to a value of 1 for all periods between 2 and 50 d,
which is why we do not show these parts of the periodograms in the figures.
One can see that the combination of RVs from several observatories, which are different
from each other in their local times, safely excluded the one-day aliases.
\begin{figure}[h]
\centering
\includegraphics[width=\hsize]{17922fg8.eps}
\caption{Stellingwerf's $\theta$ statistics for the manually measured RVs
of the H$\alpha$ emission wings plotted vs. frequency $f$.
{\sl Upper panel:} Periods from 3000\fd0 down to 50\fd0.
{\sl Bottom panel:} Periods from 2\fd0 downto 0\fd5.
The panels have a different scale on the ordinate.}
\label{thetajn}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=\hsize]{17922fg9.eps}
\caption{Stellingwerf's $\theta$ statistics for the automatically measured RVs
of the H$\alpha$ emission wings plotted vs. frequency $f$.
{\sl Upper panel:} Periods from 3000\fd0 down to 50\fd0.
{\sl Bottom panel:} Periods from 2\fd0 downto 0\fd5.
The panels have a different scale on the ordinate.}
\label{thetamr}
\end{figure}
The deepest minimum in both periodograms at a frequency
$f$~$\approx$~0.004910 d$^{\rm -1}$ corresponds to a period of $P$\,$\approx$\,203\fd0.
The two shallower peaks at lower frequencies in Figs.~\ref{thetajn}
and~\ref{thetamr} correspond to the integer multiples of the 203~d period.
The 203~d period was also detected in the measured RVs of other spectral lines,
though with a larger scatter in the RV curves. These additional results are
presented in the online Appendix.
Since no obvious signs of the secondary companion are seen in the spectrum,
we adopted $\gamma$~Cas as a single-line spectroscopic binary.
We derived a number of orbital solutions for the H$\alpha$ emission
RVs, prewhitened for the long-term changes with HEC13.
We used the program SPEL (written by Dr.~J.~Horn and never published)
for this purpose. The program has already been used in several previous
studies, e.g., \citet{hec1983, hec1984}, \citet{koubsky1985},
\citet{stefl1990}, and \citet{horn1992, horn1994}.
Our first goal was to decide whether the orbital eccentricity found
by \citet{hec2000} is real or whether the orbit is actually circular
as concluded by \citet{mirosh2002}.
In Table~\ref{elementhec} the eccentric-orbit solutions for the manually and
automatically measured H$\alpha$ emission-wing RVs are compared.
\citet{ls71} have pointed
out that observational uncertainties may cause the estimated eccentricity
to be biased when the eccentricity is low. The probability
that the true eccentricity is zero can be calculated, and this is given
in the column ''L-S test''. If this probability is greater than 0.05 we accept the hypothesis that the
true eccentricity is zero at a 5\% confidence level.
To shed more light on the problem, we split both manually and automatically
measured RVs into a number of data subsets, each of them covering
a time interval not longer than three consecutive orbital periods
and containing enough observations to define the orbital RV curve.
We used the original, not the prewhitened RVs.
The phase diagrams for a period $P$~=~203\D52 for all selected data subsets
are shown in Fig.~\ref{subsets}. We derived the elliptical-orbit solutions
for them, again testing the reality of the non-zero orbital eccentricity
after \citet{ls71}. To always find the solution with the
smallest rms error, we started the trial solutions for each subset with
initial values of $\omega$~ for four possible orientations
of the orbit, namely 45$^{\circ}$, 135$^{\circ}$, 225$^{\circ}$, and
315$^{\circ}$.
The corresponding orbital elements for the selected subsets, together
with the Lucy-Sweeney test, are summarized in Table~\ref{subtabman} for
manually measured RVs and in Table~\ref{subtabaut} for automatically measured RVs.
The results show very convincingly that the true orbit must be circular
(or has a very low eccentricity, which is beyond the accuracy limit of
our data). Although the L--S test detected a definite eccentricity for
several subsets, the individual values of the longitude of periastron
are basically accidental, and that is probably the strongest argument for
the eccentric-orbit solutions not being trusted.
\begin{table}[h]
\setcounter{table}{3}
\caption{Eccentric-orbit solutions based on the H$\alpha$ emission-wing RVs
measured manually and automatically and prewhitened for the long-term changes
with the program HEC13.}
\begin{tabular}{lll}
\hline
\hline
Solution No.: &1) &2) \\
Element &Manual &Automatic \\
\hline
$P$ (d) &203.36\p0.10 &203.08\p0.11 \\
$T_{\rm RVmin}$ (RJD) &52083\p17 &52080\p13 \\
$\omega$ ($^{\circ}$) &19\p30 &147\p22 \\
$e$ &0.048\p0.027 &0.072\p0.030 \\
$K_{\rm 1}$ ({km~s$^{-1}$}) &3.88\p0.10 &3.93\p0.12 \\
$\gamma$ ({km~s$^{-1}$}) &0.164\p0.071 &0.180\p0.081 \\
rms ({km~s$^{-1}$}) &1.772 &1.948 \\
L--S test &0.183 &0.064 \\
No. of RVs &757 &700 \\
\hline
\end{tabular}
\label{elementhec}
\end{table}
\begin{table}[h]
\caption{The eccentric-orbit solutions for the subsets of manually measured
H$\alpha$ emission-wing RVs, shown as phase plots in Fig.\ref{subsets}.}
\begin{tabular}{llllll}
\hline
\hline
$N_{\rm s}$&\multicolumn{1}{c}{$e$}&\multicolumn{1}{c}{$\omega$}&\multicolumn{1}{c}{$K_{\rm 1}$}&\multicolumn{1}{c}{L--S}&\multicolumn{1}{c}{$N$}\\
& &\multicolumn{1}{c}{($^{\circ}$)}&\multicolumn{1}{c}{({km~s$^{-1}$})} &\multicolumn{1}{c}{test}& \\
\hline
2 &0.052\p0.073&141\p38 &4.30\p0.24 &0.734&25 \\
3 &0.313\p0.082&165\p12 &5.43\p0.44 &0.001&42 \\
4 &0.416\p0.083&108.9\p8.3&6.46\p0.58 &0.000&42 \\
5 &0.14\p0.14 &160\p41 &4.37\p0.60 &0.672&25 \\
7 &0.10\p0.11 &322\p47 &6.88\p0.88 &0.674&50 \\
9 &0.109\p0.067&201\p36 &4.10\p0.30 &0.299&42 \\
10 &0.120\p0.040&263\p16 &5.20\p0.20 &0.026&61 \\
12 &0.086\p0.055&149\p28 &4.64\p0.26 &0.353&46 \\
13 &0.18 \p0.10 &308\p29 &3.56\p0.36 &0.270&70 \\
14 &0.102\p0.040&147\p22 &4.31\p0.21 &0.038&59 \\
16 &0.116\p0.067&105\p31 &3.87\p0.22 &0.287&30 \\
\hline
\end{tabular}
\tablefoot{$N_{\rm s}$~=~ a number of RV subset (the same numbering being
also used in Fig.~\ref{subsets}),
L--S test~=~probability that the eccentricity found is false,
$N$~=~number of RVs in the subset.}
\label{subtabman}
\end{table}
\begin{table}[h]
\caption{The eccentric-orbit solutions for the subsets of automatically measured
H$\alpha$ emission-wing RVs.}
\begin{tabular}{llllll}
\hline
\hline
$N_{\rm s}$&\multicolumn{1}{c}{$e$}&\multicolumn{1}{c}{$\omega$}&\multicolumn{1}{c}{$K_{\rm 1}$}&\multicolumn{1}{c}{L--S}&\multicolumn{1}{c}{$N$}\\
& &\multicolumn{1}{c}{($^{\circ}$)}&\multicolumn{1}{c}{({km~s$^{-1}$})} &\multicolumn{1}{c}{test}& \\
\hline
2 &0.30\p0.14 &322\p24 &4.57\p0.81 &0.174&24 \\
4 &0.42\p0.11 &285\p17 &5.82\p0.78 &0.004&44 \\
5 &0.34\p0.10 &6\p20 &4.74\p0.62 &0.014&25 \\
7 &0.16\p0.12 &87\p49 &6.0\p1.1 &0.311&52 \\
9 &0.15\p0.19 &92\p53 &3.38\p0.50 &0.583&40 \\
10 &0.098\p0.050&48\p23 &4.12\p0.19 &0.147&60 \\
12 &0.40\p0.21 &336.0\p9.5&5.9\p1.2 &0.054&36 \\
13 &0.21\p0.14 &82\p44 &3.94\p0.53 &0.157&59 \\
14 &0.167\p0.068&111\p20 &4.57\p0.33 &0.026&46 \\
16 &0.26\p0.13 &227\p24 &4.61\p0.56 &0.201&27 \\
\hline
\end{tabular}
\tablefoot{$N_{\rm s}$~=~a number of a RV subset (the same numbering being
also used in Fig.~\ref{subsets}),
L--S test~=~probability that the eccentricity found is false,
$N$~=~number of RVs in the subset.}
\label{subtabaut}
\end{table}
\begin{figure}[h]
\centering
\includegraphics[width=\hsize]{17922fg10.eps}
\includegraphics[width=\hsize]{17922fg11.eps}
\includegraphics[width=\hsize]{17922fg12.eps}
\caption{Phase diagrams for subsets of~H$\alpha$ emission-wing RVs measured manually.
The ordinates of all plots are RVs in {km~s$^{-1}$} and the individual
subsets vertical axis is in RV and the different ranges reflect the fact that
the original, not prewhitened RVs are used.
Trial orbital solutions for these subsets are in
Tabs.~\ref{subtabman} and \ref{subtabaut}.}
\label{subsets}
\end{figure}
Adopting the circular orbit from now on, we attempted to remove the long-term
RV variations with a different approach. Using the manually measured
H$\alpha$ emission-wing RVs, we treated each data subset spanning no more than three orbital
periods (last four subsets) and two orbital periods (remaining subsets) as
data coming from separate spectrographs, allowing SPEL to derive individual `systemic velocities'
for each subset. This way, the orbital solution was again free of
the long-term changes but they were removed in discrete velocity steps.
The corresponding orbital solution is in Table~\ref{elementga}, and individual
velocity levels ($\gamma$'s) are shown in the top
panel of Fig.~\ref{13hahe} and listed in Table~\ref{gammas} in
Appendix~\ref{apen}.
At first sight, this way of removing the long-term changes might seem
less accurate, but it leads to smaller rms error for the solution than
the removal via HEC13. There are two reasons for that:
\begin{enumerate}
\item The HEC13 program computes normal points from RV subsets covering
constant time intervals, while for $\gamma$~velocities the length of subsets
was chosen more suitably to our data distribution in time.
\item The HEC13 program computes the normal points as weighted means, but
the $\gamma$~velocities computed with the SPEL program are elements of the
Keplerian orbital model.
\end{enumerate}
\begin{table}[h]
\caption{Orbital elements obtained using RVs measured manually (3)
and automatically (4) on the H$\alpha$ line, with the removal of the long--term
changes using different $\gamma$~velocities for individual data subsets.}
\begin{tabular}{lll}
\hline
\hline
Solution No.: &3) &4) \\
Element &Manual &Automatic \\
\hline
$P$(d) &203.65\p0.13 &203.47\p0.14 \\
$T_{\rm RVmin}$(RJD) &52081.42\p0.81 &52081.99\p0.95 \\
$K_1$({km~s$^{-1}$}) &4.084\p0.10 &4.14\p0.13 \\
rms({km~s$^{-1}$}) &1.657 &1.908 \\
No. of RVs &757 &700 \\
\hline
\end{tabular}
\label{elementga}
\end{table}
For completeness, a solution for an eccentric orbit was
also derived but the L-S test gave the probability of 0.18, reassuring us
that the eccentricity is spurious.
Also for this method of the long-term removal the rms error of the
orbital solution is better for the manually than for the automatically
measured RVs.
We tested yet another method of removing the long--term
variations. It can only be used when one has a RV curve
uniformly enough covered by observations. The RVs are averaged over chosen
time intervals, and the Hermite polynomials are fitted through these averaged
(normal) points. We used the program HEC23 to compute the normal points and
program HEC36 to fit them\footnote{Both programs, written by PH, and
the instructions how to use them, are available at
{\sl http://astro.troja.mff.cuni.cz/ftp/hec/HEC36\,.}}.
We tentatively averaged the RVs over a 300~d and a 400~d interval.
New orbital solutions were derived using RVs prewhitened this way.
The rms error of the resulting solution was approximately the same as
the rms error of the solution for RVs prewhitened HEC13.
We decided to use this approach in another way. We used the systemic
velocities derived with SPEL as normal points and fitted them with
the Hermite polynomials using HEC36. The RJDs of RVs in a subset were averaged and
the mean RJD was used as the epoch of the $\gamma$~velocity. The same approach
to computing epochs of normal points is also used in the program HEC13. This way we effectively removed one of
the disadvantages of the previous method since HEC36 connects the normal
points with a smooth curve, thus removing the discontinuous shifts
introduced with the second method. The Hermite--polynomial fit is shown in
the first two panels of Fig.~\ref{13hahe}.
The \hbox{$O\!-\!C$}\ residua were again used to derive circular-orbit solutions with SPEL.
These are presented in Table~\ref{elem36}. As expected, the improvement
in the resulting rms errors with the second method is relatively small.
\begin{table}[h]
\caption{Orbital elements obtained using the H$\alpha$ RVs measured manually (5) and
automatically (6) and removing the long--term RV~variations via
a Hermite--polynomial fit through the locally derived $\gamma$~velocities.}
\begin{tabular}{lll}
\hline
\hline
Solution No.: &5) &6) \\
Element &Manual &Automatic \\
\hline
$P$(d) &203.523\p0.076 &203.371\p0.089 \\
$T_{\rm RVmin}$(RJD) &52081.89\p0.62 &52082.07\p0.76 \\
$K_1$({km~s$^{-1}$}) &4.297\p0.090 &4.26\p0.11 \\
$\gamma$({km~s$^{-1}$}) &0.018\p0.064 &-0.018\p0.075 \\
rms ({km~s$^{-1}$}) &1.592 &1.825 \\
No. of RVs &757 &700 \\
\hline
\end{tabular}
\label{elem36}
\end{table}
The net orbital RV curves and the corresponding \hbox{$O\!-\!C$}\ residua from
the orbital solutions are shown in Fig.~\ref{fazeham} for the manual
and in Fig.~\ref{fazeha} for the automatically measured H$\alpha$ RVs.
The rms errors per one observation of the Keplerian fit no.~5 based on
manually measured RVs (no.~6 in the case of automatic measurements)
are shown as short absciss\ae\ in the upper right corners of both figures.
The RVs prewhitened for the long-term changes, on which these best solutions
5 and 6 are based, are also presented in Tables~2 and 3.
\begin{figure}[h]
\centering
\includegraphics[width=\hsize]{17922fg13.eps}
\caption{{\sl Top panel:} The net orbital RV curve corresponding to
the circular-orbit solution (5) of Table~\ref{elem36}.
{\sl Bottom panel:} The \hbox{$O\!-\!C$}\ residua from the orbital solution.
Both plots have the same velocity scale, and the epoch of RV$_{\rm min}$
is used as a reference epoch.
The short abcissa in the right upper corner
of the top panel denotes the rms of 1 observation for the Keplerian
fit no.~5.}
\label{fazeham}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=\hsize]{17922fg14.eps}
\caption{{\sl Top panel:} The net orbital RV curve corresponding to
the circular-orbit solution (6) of Table~\ref{elem36}.
{\sl Bottom panel:} The \hbox{$O\!-\!C$}\ residua from the orbital solution.
Both plots have the same velocity scale, and the epoch of RV$_{\rm min}$
is used as a reference epoch.
The short abcissa in the right upper corner
of the top panel denotes the rms of 1 observation for the Keplerian
fit no.~6.}
\label{fazeha}
\end{figure}
We did several preliminary tests of the possible rapid variations of $\gamma$~Cas.
Although the results clearly demonstrated their presence,
the quantitative results were inconclusive so we decided to postpone
analysis of rapid spectral changes for a future study,
based on dedicated series of whole-night spectral observations.
\section{Interpretation of results\label{4}}
\subsection{Spectral variations}
We have confirmed the continuing presence of long--term variations in
the H$\alpha$ emission profile using the Gaussian fit and RV measurements.
The former procedure gives more objective results than a direct measurement
of the peak height and the full width of the line at half maximum
(FWHM), because the observed height of the highest peak of a very strong
emission line partly reflects the long-term $V/R$ changes, while
the measured FWHM can be affected by the presence of several telluric
water vapour-lines of variable strength. Figure~\ref{fwhmic} shows that
the $I_{\rm p}$ and FWHM of the H$\alpha$ emission are {\sl anti--correlated}
over the interval covered by our observations and seem to also show some
variability on shorter time scales. We carried out period searches
for both these quantities prewhitened for the long-term changes,
but no significant periodicities were detected, and trial plots
of prewhitened data did not show any evidence of variations connected
with the orbital period.
It is probable that the observed long-term changes reflect some
variations in the density and/or the extent of the circumstellar disk around
the Be primary. Using optical interferometry, \citet{quirrenbach1997} and
\citet{tycner2006} were able to resolve the envelope around the primary
component of $\gamma$~Cas and estimate that its inclination must be higher
than 46$^\circ$ and 55$^\circ$, respectively. Adopting a reasonable assumption
that the rotational axis of the disk is roughly
perpendicular to the binary orbit, we can conclude that the binary
system is also seen under an orbital inclination higher than some $45^\circ$.
We then suggest the following possible interpretation of the observed changes.
It is very probable that the envelope around the Be primary is
rotationally supported and that its linear rotational velocity
decreases with the axial distance from the star. Regardless of the process
causing secular variations of the disk, one can assume that the increase
in the height of the emission peak reflects the presence of more emission
power at lower rotational velocities, thus implying either an increase
in the density of the outer parts of the disk and/or an increase
in the geometrical extent of the disk. This must
naturally decrease the observed FWHM of the emission as observed.
This qualitative scenario seems to be supported by the line-profile
calculations published by \citet{silaj2010}.
\subsection{Orbital motion}
As mentioned in Sect.~\ref{1}, one of the principal motivations of this
study was to resolve the differences in the orbital solutions
obtained by~\citet{hec2000} and by~\citet{mirosh2002} and to arrive
at a more definitive set of the orbital elements. We were able
to combine both independent datasets and complement them with more recent
spectra. In Sect.~\ref{3} we carried out a number of various tests,
analysing separately the RVs measured by a manual and an automatic technique.
We also tested the effect of the different ways of data prewhitening on
the resulting elements. Special attention was payed to the test of whether
the orbit is actually circular or has a significant eccentricity.
The principal results are the following.
\begin{enumerate}
\item The binary orbit of $\gamma$~Cas is circular, at least within the limits of
the accuracy of our data, as concluded by \citet{mirosh2002}.
\item The resulting value of the orbital period is now well constrained
by the data at hand, and it is robust with respect to different ways of analysis.
From all experiments, we were finding values between 203\fd0 and 203\fd6,
close to the value already found by \citet{hec2000}.
An inspection of all trial solutions shows that the solutions based
on manually measured RVs have rms errors that are systematically lower for
$\approx$~15\% in comparison to the solutions for the automatically
measured H$\alpha$ emission-wing RVs. We therefore conclude that
solution~5 of Table~\ref{elem36} is the best we can offer and suggest
the following linear ephemeris for the epoch of RV minimum
to be used in the future studies of this binary:
\begin{eqnarray}
T_{\rm min.RV} &=& {\rm HJD}\,(2452081.89\pm0.62)\nonumber\\
&+& (203\fd523\pm0\fd076) \times E.\label{efe}
\end{eqnarray}
\item The semiamplitude of the orbital motion is close to 4~{km~s$^{-1}$}\ for
all solutions. The recommended value from solution~5
is $K_1 = 4.297$$\pm$$0.090$~{km~s$^{-1}$}, implying the mass function
$f(M) = 0.00168$~M$_{\odot}$. For comparison, \citet{hec2000} and \citet{mirosh2002}
obtained semiamplitudes of $4.68$$\pm$$0.25$~{km~s$^{-1}$} and $3.80$$\pm$$0.12$~{km~s$^{-1}$},
respectively. If we adopt the inclination value i\,=\,45$^{\circ}$ and
the primary star mass M$_{1}$\,=\,13~M$_{\odot}$ suggested by \citet{hec2000},
we can estimate the secondary star mass M$_{2}$\,=\,0.98~M$_{\odot}$.
If the system is at a post mass-transfer phase, then the secondary might be
a hot helium star that could be directly detectable in the UV region of
the electromagnetic spectrum.
\end{enumerate}
{\sl Note:} After our paper was accepted for publication, we had the
privilege to read a preliminary version of another study of $\gamma$~Cas
kindly communicated to us by Dr.~Myron~A.~Smith and his coauthors.
They analysed in particular a smaller and partly independent
set of RVs to obtain their own orbital solution.
Their results are compatible with our final solution within
the respective error bars.
\begin{acknowledgements}
We gratefully acknowledge the use of spectrograms of $\gamma$~Cas from the public
archives of the Elodie spectrograph of the Haute Provence Observatory.
Our thanks are also due to Drs. P.\,Chadima, M.\,Dov\v{c}iak, P.\,Hadrava,
J.\,Kub\'at, P.\,Mayer, P.\,\v{S}koda, S.\,\v{S}tefl, M.\,Wolf, and P.\,Zasche,
who obtained some of the spectra used in this study.
Constructive criticism and a careful proofreading of the original
version by an anonymous referee helped to improve the paper and are
gratefully acknowledged.
We thank Dr. M.A.~Smith and his collaborators for
allowing us to see a preliminary version of their new complex study
of $\gamma$~Cas before it was submitted for publication and for useful comments.
This research was supported by the grants 205/06/0304,
205/08/H005, and P209/10/0715 of the Czech Science Foundation,
from the Research Programme MSM0021620860 {\sl Physical study of objects and
processes in the solar system and in astrophysics} of the Ministry
of Education of the Czech Republic, and from the research project
AV0Z10030501 of the Academy of Sciences of the Czech Republic.
The research of JN was supported from the grant SVV-263301
of the Charles University of Prague, while PK was supported from
the ESA PECS grant 98058.
We acknowledge the use of the electronic database from the CDS Strasbourg, and
the electronic bibliography maintained by the NASA/ADS system.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
The non-flaring Sun has been studied at radio wavelengths for several decades.
(For a recent review, see {\it e.g.} \opencite{Shibasaki2011}).
Most of the quiescent solar microwave emission is due to free--free bremsstrahlung (FF) emission from the lower corona and upper chromosphere.
At microwave frequencies, the quiet--Sun emission can be modeled as an optically thin $\approx$1 MK corona on top of an optically thick $\approx$10\,000 K chromosphere ({\it e.g.} \opencite{Zirin1991}; but see \opencite{LandiChiuderiDrago2003} and \citeyear{LandiChiuderiDrago2008} for a nuance).
In active regions (AR), an additional broadband gyroresonance component, typically peaking around 2800 GHz (10.7 cm), is present when the coronal magnetic field is strong enough \cite{Zlotnik1968a,Zlotnik1968b,WhiteKundu1997}.
During flares, additional thermal bremsstrahlung and gyrosynchroton components occur ({\it e.g.} \opencite{Dulk1985}).
Recent efforts to understand microwave emission from the Sun's upper atmosphere have concentrated on either the spectral aspect, with multi-frequency measurements at Sun center \cite{Zirin1991,Borovik1992,LandiChiuderiDrago2003,LandiChiuderiDrago2008},
or by studying its spatial profile at a few frequencies ({\it e.g.} \opencite{Selhorst2003}, \opencite{Selhorst2005}, \opencite{Selhorst2010}, \opencite{Krissinel2005} for recent work on the topic).
However the two approaches have not yet been reconciled into producing a density model that satisfactorily explains the equatorial limb profile in the whole microwave band, let alone the polar one.
The 10.7 cm spatially integrated solar radio flux is known to be well correlated with the sunspot number ({\it e.g.} \opencite{Tapping1994}).
Qualitatively, the reason is simple: active regions tend to have both enhanced densities (leading to increased free--free emission), and an additional gyroresonance (GR) component.
But, to our knowledge, a systematic observational study of the fraction of the full-Sun 10.7 cm flux that is due to GR and that is due to FF emission remains to be done.
Section~\ref{sect:ATA} of this article introduces the Allen Telescope Array in the context of solar observations.
Section~\ref{sect:obs} presents the three sets of ATA solar observations obtained so far, including raw brightness temperature $[T_{\textrm{B}}]$ maps and equatorial/polar profiles.
For comparison, cross-calibration with other instruments or semi-empirical models is also presented.
In Section~\ref{sect:budget}, a multi-frequency flux budget for each observation is presented, and the fraction of the solar 10.7 cm flux that is due to gyroresonance emission is determined.
In Section~\ref{sect:CTBC}, we derive ``Coronal Thermal Bremsstrahlung Contribution'' (CTBC) maps and profiles from observations, a format which facilitates cross-frequency comparisons,
and allows for the quick discrimination between regions on the Sun where only optically thin free--free emission is relevant, and regions where additional physics ({\it e.g.} optical thickness, gyroresonance) is present.
The results are summarized in Section~\ref{sect:ccl}.
\section{Solar observations with the Allen Telescope Array}\label{sect:ATA}
The Allen Telescope Array \cite{Welch2009} is a radio interferometer near Hat Creek, California, USA.
It consists of 42 6.1-m antennas located in an optimized ``pseudo-random'' configuration within a 300-m diameter.
Two digital correlators can be tuned to analyze two 100-MHz wide bands (1024 channels each), independently chosen between 0.5 GHz and 11.2 GHz. Each of these observing frequencies can be changed (under computer control) in a few seconds.
Through careful design, the interferometer system is phase-stable and is fully calibrated (in amplitude and in phase) with respect to standard cosmic sources.
The gains and signal levels of the receivers and digital back-end can be reconfigured under computer control to handle expected solar signal levels.
Temporal resolution as high as one second is currently available from the correlators.
The correlator output can be transformed into images through standard techniques using a well-developed
data analysis package ({\sf Miriad}:\url{http://bima.astro.umd.edu/miriad/} \opencite{Sault1995}) for semi-automated calibration and mapping with pipeline software for flagging, calibration, and imaging specific to the ATA \cite{Keating2009}.
The ATA is well-suited for imaging/spectroscopy of large--scale, slowly varying solar phenomena.
It is the only instrument currently available that can provide instantaneous two-dimensional images of the full Sun at frequencies near 2800 MHz.
The 6.1-m antennas provide a field of view that scales with frequency ($HPBW \approx \frac{2.8^{\circ}}{\nu (Ghz)}$; {\it viz.} full-Sun field of view when $\nu \lesssim $6 GHz).
The interferometer's angular resolution also scales with frequency ($HPBW \approx \frac{4.5'}{\nu (Ghz)}$).
Even though this is insufficient to resolve the internal structure of active regions or burst sources, it is well-suited to separating the various spatial contributions from across the disk. Furthermore, the compact array includes (projected) baselines as short as $\approx$6 m, corresponding to fringe spacings as large as $\approx$one degree.
Therefore in addition to imaging, the system can provide information on large--scale structures ({\it e.g.} coronal holes) with size scales of 10--20 arcminutes as well as an independent measurement of the total solar flux.
Although the ATA has dual linearly polarized feeds, for this application the correlator combines these to measure circular polarization, thereby providing an additional diagnostic of solar sources.
Investigations of compact radio sources \cite{Law2010} indicate that the accuracy of Stokes $V$ maps is $\approx$1\%.
\section{Summary of Observations}\label{sect:obs}
\begin{figure}[p]
\centering
\includegraphics[width=10cm]{fig_2009Oct01_Tb.ps}
\caption{
1-Oct-2009 observation (two-minute integration): brightness temperature $T_{\textrm{B}}$ maps and profiles.
Top left: $T_{\textrm{B}}$ maps, with photospheric limb and FWHM beam sizes in the lower--right corner.
Top right: Corresponding SOHO/EIT EUV image.
Bottom left: equatorial $T_{\textrm{B}}$ profile.
Bottom right: polar $T_{\textrm{B}}$ profile.
All maps in this work have solar North oriented ``up''.
}
\label{fig:img:Oct01:4a}
\end{figure}
\begin{figure}[p]
\centering
\includegraphics[width=12cm]{fig_2010Jan29_Tb_NoRP.ps}
\caption{
29-Jan-2010. Brightness temperature maps (with photospheric limb contour and FWHM beam sizes in the lower--right corners) and profiles.
}
\label{fig:img:Jan29:4a}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=12cm]{fig_2010Mar21_Tb.ps}
\caption{
21-Mar-2010: Brightness temperature $T_{\textrm{B}}$ maps at 1430, 2100, 2750, 3500, 4500, and 6000 MHz,
along with HPBW beamsizes in the lower--right corner.
Bottom row: $T_{\textrm{B}}$ equatorial and polar profiles (through Sun center), and the corresponding SOHO/EIT image.
The profiles (polar in particular) display an important contribution by the active region.
}
\label{fig:img:Mar21:4a}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=12cm]{fig_2010Mar21_Stokes_V.ps}
\caption{
Stokes $V$ maps for the 21-Mar-2010 observations.
The non-negligible level of polarization (varying with frequency, peaking at 11\% at 4.5 GHz) in the AR is a good indicator that magnetobremsstrahlung emission is also present.
As there is some spatial smearing happening (ATA HPBW at 6 GHz is $\approx$45'', while the typical magnetic size scale in an AR is $\approx$10'', \protect\opencite{Lee1993a} ), the degree of polarization may be underestimated.
}
\label{fig:img:Mar21:4a:V}
\end{figure}
Three sets of observations were taken:
\begin{enumerate}
\item {\bf 1-October-2009} (Figure \ref{fig:img:Oct01:4a}):
In this first test, the Sun was observed for two minutes around 21:56:36 UT, at 2750 MHz only.
No flare activity was detected during that time (the GOES X-ray digitization was bottoming out at the lowest observable A0.3 level).
Two small active regions were present, both near the western limb of the Sun.
\item {\bf 29-January-2010} (Figure \ref{fig:img:Jan29:4a}):
The Sun was observed with 31 antennas at 1430, 2750, and 3500 MHz, for $\approx$15 minutes, from 21:44 to 22:01 UT.
The GOES X-ray level was low and stable between the A2 and A3 level (no flares).
Radio Solar Telescope Network (RSTN) lightcurves were flat.
The only feature is a single on-disk active region.
\item {\bf 21-March-2010} (Figure~\ref{fig:img:Mar21:4a}):
The Sun was observed at six frequencies centered at 1430, 2100, 2750, 3500, 4500, and 6000 MHz,
for eight hours beginning at 16:30 UT,
in 1-minute intervals, alternating between frequency pairs.
There were several active regions: a main on-disk AR (``A'' in Figure~\ref{fig:img:Mar21:4a}), and several smaller ones near both equatorial limbs (``B'' to ``D'' in Figure~\ref{fig:img:Mar21:4a}).
The Sun was slightly active: the GOES X-ray baseline was A6--A8, and four small B-class flares (B1.0 to B2.2) occurred during the time the data were accumulated.
These have negligible influence on our results.
We have also plotted the Stokes $V$ maps for this observations (Figure~\ref{fig:img:Mar21:4a:V}), which will be discussed in Section~\ref{sect:budget}.
\end{enumerate}
Each ATA map was CLEANed \cite{Hogbom1974} around bright active region features, and the rest of the image ($\approx$quiet Sun) was then treated with a Maximum Entropy Method (MEM).
Image dynamic range (ratio of brightest pixel to rms noise of image) for the 21-Mar-2010 observations is about 700 at 1430 MHz, and 250 at 6000 MHz.
We discuss the accuracy of the flux calibration in the next sub-sections, by comparing total map fluxes to that given by other (spatially integrated) instruments (Section~\ref{sect:totmapflux}),
and by comparing the brightness temperature at Sun center with other published results (Section~\ref{sect:tbsuncenter}).
We then say a few words about the brightness temperature profiles and Stokes $V/I$ maps.
\subsection{Total Map Fluxes:} \label{sect:totmapflux}
\begin{figure}[h]
\centering
\includegraphics[height=11cm]{full_sun_fluxes_comparisons.eps}
\caption{
Full-map fluxes from ATA and comparisons with other observatories.
Black: ATA
Red: Penticton, adjusted noon value.
Blue: Radio Solar Telescope Network (RSTN) spectrum, averaged between the Palehua and Sagamore Hill observatories.
Purple: Nobeyama Radio-polarimeter (NoRP) values.
}
\label{fig:sp:comparisons}
\end{figure}
Figure~\ref{fig:sp:comparisons} displays the total fluxes in each map at each frequency for all three periods of observations,
with comparisons with other instruments: the 10.7 cm flux from Penticton (British Columbia), the Nobeyama Radiopolarimeter (NoRP; Japan), and two observatories that are part of the Radio Solar Telescope Network (RSTN; Palehua, Hawaii and Sagamore Hill, Massachussets), all adjusted to 1 AU.
There is generally good agreement, typically within the 10--15\% level.
\subsection{Center Pixel $T_{\textrm{B}}$:}\label{sect:tbsuncenter}
\begin{figure}[h]
\centering
\includegraphics[width=12cm]{full_sun_centerTb.ps}
\caption{
{\bf Center pixel $T_{\textrm{B}}$.}
21-Mar-2010 observations (solid, blue) center $T_{\textrm{B}}$ are higher than expected at low frequencies because of contribution from a nearby AR.
The dashed, blue line was taken instead for pixels about 500'' further south, corresponding to $\approx$2 HPBW (at 1430 MHz) away from main AR.
The value of $T_{\textrm{B}}$ around Sun center is highly oscillatory (spatially) at high frequency, probably due to the lack of small baselines at these frequencies.
Taking the modes of the maps lead to values in between the extremes shown (between the dashed and the solid blue lines).
}
\label{fig:sp:centerpixelTb}
\end{figure}
We compared the brightness temperature $T_{\textrm{B}}$ of the pixel nearest Sun center to the results published by \inlinecite{Zirin1991} (see Figure~\ref{fig:sp:centerpixelTb}).
There is good agreement, except at 6 GHz, where we observe a high amount of on-disk spatial oscillations, and a low on-disk average brightness temperature,
which we attribute to a lack of short baselines at these frequencies.
Hence, the values at high frequencies, particularly at 6 GHz, must be used with caution.
\subsection{$T_{\textrm{B}}$ profiles:}\label{sect:tbprofiles}
We point out the equatorial limb brightening observed at all frequencies and for all three observations, and the absence of such at the poles.
This has already been observed before ({\it e.g.} \opencite{Christiansen1955}), at higher spatial accuracy.
Despite this limitation, we note that ATA's multi-frequency capability allows us to study limb brightening in a new light (see introduction for references on the topic).
We shall make a few comparisons with simple atmospheric models later in this work, but a thorough investigation is beyond the scope of the present article.
Suffice it to say for now that the equatorial limb is thought to be brighter than the polar limb because the solar Equator comprises many closed loops, which tend to retain plasma,
while polar regions, predominently composed of ``open field lines'' where the plasma can easily escape with the solar wind, are less dense.
\subsection{Stokes $V/I$ Maps:}\label{sect:stokesV}
Figure~\ref{fig:img:Mar21:4a:V} shows a clear and consistent circular polarization signature accross all frequencies for the disk AR (source ``A''),
and are consistent with SOHO/MDI magnetic maps of the underlying photosphere.
The maximum Stokes $V/I$ occurs at 4.5 GHz, where it can reach 11\%.
This value is probably a lower limit, as there is certainly some spatial smearing in effect:
the typical magnetic size scale in a sunspot is about 10'' \cite{Lee1993a},but the instrumental resolution is about 1' at 4.5 GHz.
Furthermore, solar rotation moves the source horizontally by $\approx$1' over the course of data accumulation.
Free--free thermal bremsstrahlung emission in active regions is not expected to be able to reach such high levels of circular polarizations ({\it e.g.} Gelfreikh, \citeyear{GaryKeller2004}),
and, in the absence of meaningful gyrosynchroton emission from radio bursts, we therefore conclude that it is mostly due to gyroresonance emission.
\section{Flux Budget:} \label{sect:budget}
Tables~\ref{tab:budget:Oct01}, \ref{tab:budget:Jan29}, and \ref{tab:budget:Mar21} contain a flux budget for the most prominent sources in each of the three observations.
The ``full-Sun'' flux was computed by adding together all pixel values within 1.3 $R_S$ of Sun center.
Additionally, for the 21-Mar-2010 observations, source spectra have been plotted (Figure~\ref{fig:sp:Mar21}).
As most of the features were spatially unresolved (at least at the lowest frequencies), we have plotted the fluxes, $S_{\nu}$, of each feature, instead of the more usual brightness temperatures.
\begin{table}[h]
\small
\centering
\caption{Flux budget [SFU] for {\bf 1-Oct-2009} observations (sources are background-subtracted, relative errors are below the 10\% level):
}
\begin{tabular}{llll}
\hline
Frequency & Full-Sun & Limb source in & Limb source \\
& (within 1.3 $R_S$) & upper right quadrant & in lower--right quadrant\\
\hline \hline
2750 MHz & 71.4 (100\%) & 1.1 (1.6\%) & 2.5 (3.5\%) \\
\hline
\end{tabular}
\label{tab:budget:Oct01}
\end{table}
\begin{table}[h]
\caption{ Flux budget [SFU] for {\bf 29-Jan-2010} observations (sources are background-subtracted):
}
\begin{tabular}{lll}
\hline
Frequency & Full-Sun flux & Active Region flux\\
& (within 1.3 $R_S$) & \\
\hline \hline
1430 MHz & 57.5 (100\%) & 3.8 (6.6\%) \\
2750 MHz & 71.6 (100\%) & 3.6 (5.0\%) \\
3500 MHz & 89.5 (100\%) & 3.6 (4.5\%) \\
\hline
\end{tabular}
\label{tab:budget:Jan29}
\end{table}
\begin{table}[h]
\caption{Flux budget [SFU] for {\bf 21-Mar-2010} observations (sources are background-subtracted, see top left plot of Figure~\ref{fig:img:Mar21:4a} for letter labeling):
}
\centering
\small
\begin{tabular}{llllll}
\hline
Frequency & Full-Sun & Main AR & Left limb & Right limb, & Right limb, \\
& & ``A'' & ``B'' & ``C'' & ``D'' \\
\hline \hline
1430 MHz & 61.3 (100\%) & 5 (8.1\%) & 2 (3.3\%) & 1 (1.6\%) & 0.5 (0.8\%) \\
2100 MHz & 86.3 (100\%) & 7.8 (9.0\%) & 2.5 (2.9\%) & 1.5 (1.7\%) & 0.4 (0.5\%) \\
2750 MHz & 76.3 (100\%) & 6.0 (7.9\%) & 1.7 (2.2\%) & 1.1 (1.4\%) & 0.35 (0.5\%) \\
3500 MHz & 83.0 (100\%) & 5.8 (7.0\%) & 1.6 (1.9\%) & 1.1 (1.3\%) & 0.27 (0.3\%) \\
4499 MHz & 114.7 (100\%) & 5.2 (4.5\%) & 1.7 (1.5\%) & 1.6 (1.4\%) & 0.8 (0.7\%) \\
6000 MHz & 130.0 (100\%) & 4.6 (3.5\%) & 1.8 (1.4\%) & 1.5 (1.2\%) & 0.9 (0.7\%) \\
\hline
\end{tabular}
\label{tab:budget:Mar21}
\end{table}
\begin{figure}[ht!]
\centering
\includegraphics[width=7.5cm]{fig_Mar21_sp_RSTN_log.ps}
\caption{
Spectra for the 21-Mar-2010 observations.
All values were normalized using interpolated values of the solar fluxes observed by NoRP.
From top to bottom:
full-Sun value, divided by ten (black),
background-subtracted main active region labelled ``A'' in Figure~\ref{fig:img:Mar21:4a} (red),
background-subtracted region ``B'' (orange),
and background-subtracted region ``C'' (gray).
Source ``D'', being even weaker, is not displayed.
Free--free emission is expected to have a flat flux spectrum (while it is optically thin), consistent with all sources besides ``A''.
The ``bump'' at mid-frequencies for source ``A'' is a strong indicator of gyroresonance emission.
}
\label{fig:sp:Mar21}
\end{figure}
Our observations show that for these particular cases, discrete background-subtracted ARs on the Sun typically contain a few percent of the total solar flux at various frequencies.
The weaker ARs' generally flat spectra lead us to conclude that they were producing mostly optically thin thermal bremsstrahlung emission.
However, the main AR in the 21-Mar-2010 observations (``A'' in Figure~\ref{fig:img:Mar21:4a}) possesses an additional component that exhibits an increase centered around 2.8 GHz, typical of GR emission.
In the next section, we will merge these multi-frequency observations into a single quantity, that we dubbed ``Coronal Thermal Bremsstrahlung Contribution'' or ``$\eta$'' function,
and use it to determine, amongst other things, the fraction of emission that is due to GR.
\section{Coronal Thermal Bremsstrahlung Contribution ``CTBC''} \label{sect:CTBC}
\subsection{Theory:}
In the microwave range, the brightness temperature at Sun center can be modeled as the sum of an optically thick chromosphere ($\approx$10\,000 K)
and an optically thin bremsstrahlung emission from an exponential hydrostatic corona ({\it e.g.} \opencite{Zirin1991}):
\begin{equation}
T_{\textrm{B}} = T_{\textrm{chromo}} + \tau \, T_{\textrm{corona}}
\end{equation}
with the free--free optical depth $\tau$ given by:
\begin{equation}
\tau = \zeta \, \nu^{-2} \, \frac{EM}{T^{3/2}}
\end{equation}
where $EM$ is the line-of-sight (LOS) emission measure, $\nu$ the frequency, $T$ the temperature of the coronal plasma,
and $\zeta=\zeta(T/\nu)$ a slowly varying function of $T/\nu$, approximately given by \cite{Dulk1985}:
\begin{equation}
\zeta(T/\nu) = 9.786 \times 10^{-3} \ln \left( 4.7 \times 10^{10} \times \frac{T}{\nu} \right)
\end{equation}
We have assumed the refractive index to be unity, appropriate for the microwave regime outside of active regions ({\it e.g.} \opencite{LandiChiuderiDrago2008} and \opencite{Shibasaki2011}).
We define, for every pixel, the $CTBC$ ``$\eta$'' function as follows:
\begin{equation}
\eta(\nu) \equiv \frac{ T_{\textrm{B}}(\nu)-T_{\textrm{chromo}} }{\zeta \, \nu^{-2}}
\end{equation}
where $T_{\textrm{B}}(\nu)$ is the observed brightness temperature at frequency $\nu$, $T_{\textrm{chromo}}$ is set to 10\,880 K, the value fitted by \inlinecite{Zirin1991}, and, to compute $\zeta$, we assume a temperature of 1 MK.
As stated earlier, $\zeta$ depends only weakly on the exact value of $T$.
For each pixel where optically thin free--free thermal emission is the only emission mechanism, one expects:
\begin{equation}\label{eq:eta}
\eta = \frac{EM}{\sqrt{T}}
\end{equation}
{\bf which is independent of frequency}.
(More generally, $\eta$=$\Sigma \,\frac{EM_i}{\sqrt{T_i}}$ or \\ $\eta$ = $\int \frac{\left(\frac{\mathrm{d}EM}{\mathrm{d}T}\right)}{\sqrt{T}} \mathrm{d}T$).
{\bf Provided the physical assumptions are valid, this does not depend on the spatial distribution of the coronal features.}
In the absence of flares, deviations indicate other physical phenomena are at work: optical thickness (at low frequencies),
or gyroresonance emission \cite{Zlotnik1968a,Zlotnik1968b,WhiteKundu1997}.
Furthermore, if $\eta(\nu)$ for a feature is indeed the same at all frequencies,
then the physics of the emission mechanism and the actual coronal density and temperature structure can be separated,
and the latter independently studied and modeled.
\subsection{Observations:}
\begin{figure}[p]
\centering
\includegraphics[width=10cm]{fig_2009Oct01_EM_sqrtT_NoRP.ps}
\caption{
1-Oct-2009: $\eta$ maps (left) and profiles (right, top and bottom).
Overplotted are expected $\eta$-profiles from a 1-MK exponential corona with base electron density
$n_0=4.4\times 10^8$ cm$^{-3}$ (dotted line) and $n_0=5.8\times 10^8$ cm$^{-3}$ (dashed line), convolved with ATA beamsize.
}
\label{fig:img:Oct01:4c}
\end{figure}
\begin{figure}[p]
\centering
\includegraphics[width=12cm]{fig_2010Jan29_EM_sqrtT_NoRP.ps}
\caption{
29-Jan-2010 $\eta$ maps and equatorial (lower left) and (lower middle) polar profiles.
All maps (and hence their derived profiles also) were convolved to the largest beamsize (i.e. the 1430 MHz beam in this case).
Overplotted are expected $\eta$-profiles from a 1-MK exponential corona with base electron density
$n_0=4.4\times 10^8$ cm$^{-3}$ (dotted line) and $n_0=5.8\times 10^8$ cm$^{-3}$ (dashed line), convolved with ATA beamsize.
}
\label{fig:img:Jan29:4c}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=12cm]{fig_2010Mar21_EM_sqrtT_NoRP.ps}
\caption{
$\eta$ maps and profiles.
All images were convolved to the same elliptical gaussian beam, 1.05 times the largest (1430 MHz) one.
The bottom right image is the displays the standard deviation of the other six.
Overplotted are expected $\eta$-profiles from a 1-MK exponential corona with base electron density
$n_0=4.4\times 10^8$ cm$^{-3}$ (dotted line) and $n_0=5.8\times 10^8$ cm$^{-3}$ (dashed line), convolved with ATA beamsize.
Notice difference between equatorial and polar solid black profiles, due to beam anisotropy.
}
\label{fig:img:Mar21:4c}
\end{figure}
$\eta$-maps and profiles are displayed in Figures~\ref{fig:img:Oct01:4c}, \ref{fig:img:Jan29:4c}, and \ref{fig:img:Mar21:4c}, along with synthetic $\eta$-profiles for simple exponential coronae.
For proper comparison, all maps were convolved to the worst point-spread-function (psf) of the set, i.e. to the psf of the lowest frequency.
A first glance reveals that:
\begin{itemize}
\item There is rough agreement of the $\eta$-profiles across frequencies, consistent with Equation (\ref{eq:eta}).
The largest discrepancies appear in the vicinity of active regions (on-disk and at the limb) in the case of the 21-Mar-2010 observation.
\item The $\eta$-value of the quiet Sun at disk center is about 8$\times$10$^{23}$ K$^{-1/2}$ cm$^{-5}$.
The value of the limb spike is more complicated to estimate, as the profile is convolved with the instrument's psf, but can be estimated to be $\approx$10$^{25}$ K$^{-1/2}$ cm$^{-5}$.
\item The equatorial $\eta$-profiles can be constrained by hydrostatic exponential 1 MK coronae with base densities $\approx$4.4 and $\approx$5.8$\times$10$^8$ cm$^{-3}$,
but neither faithfully reflect the actual profile.
\item Equatorial limb brightening is observed, although some of it is due to the presence of active regions (21-Mar-2010 observations).
At high frequencies, particularly 6 GHz, a ring appears.
As discussed earlier, the lack of short baselines at these frequencies makes the center of the Sun (large spatial structure) less bright than it should be,
increasing the contrast with the limbs, which are themselves better-imaged small scale structures.
\end{itemize}
The $\eta$-profiles for 29-Jan-2010 appear to show little GR contribution (as they superpose well in all three frequencies, even in the AR -- Figure~\ref{fig:prof:Jan29Mar21:mainAR}).
From the AR's $\eta$-flux and its gaussian size $r_{\mathrm{source}}$ estimated from the observed size $r_{\mathrm{obs}}$ and the beamwidth $r_{\mathrm{beam}}$ using $r_{\mathrm{source}}=\sqrt{r_{\mathrm{obs}}^2-r_{\mathrm{beam}}^2}$,
the peak $\eta$ can be estimated, and amounts to $\approx$10$^{25}$ K$^{-1/2}$ cm$^{-5}$, which compares well with other observations (see Table~\ref{tab:losem}).
Assuming a single temperature of $\approx$3 MK, this means an average line-of-sight emission measure (LOS EM) of $\approx$1.6$\times$10$^{28}$ cm$^{-5}$.
\begin{table}[h]
\caption{Approximate temperatures, emission measures, and $\eta$ values for various observations or models.
}
\centering
\small
\begin{tabular}{llccc}
\hline
Source & Detail & T & LOS EM & $\eta$ \\
& & [MK] & [cm$^{-5}$] & [K$^{-1/2}$ cm$^{-5}$] \\
\hline \hline
EUV observations, & AR core & 5 & 5$\times$10$^{28}$ & 2.2$\times$10$^{25}$ \\
Aschwanden and Boerner 2011 & AR average & 1.6 & 10$^{27}$ & 7.9$\times$10$^{23}$ \\
\hline
Exponential 1 MK & Sun center & 1 & 6.3$\times$10$^{26}$ & 6.3$\times$10$^{23}$ \\
corona with base density& limb spike & 1 & 8.4$\times$10$^{27}$ & 8.4$\times$10$^{24}$ \\
5$\times$10$^{8}$ cm$^{-3}$& & & & \\
\hline
29-Jan-2010 & Main AR & & & $\approx$10$^{25}$ \\
ATA observations & & & & \\
\hline
\end{tabular}
\label{tab:losem}
\end{table}
The spectrum of the 21-Mar-2010 disk AR (Figure~\ref{fig:sp:Mar21}, source ``A'': {\it red line}) indicates some gyroresonance contribution.
The frequency-dependence of the $\eta$ maps over the disk AR (Figures~\ref{fig:img:Mar21:4c} and \ref{fig:prof:Jan29Mar21:mainAR}) and the presence of circular polarization (Figure~\ref{fig:img:Mar21:4a:V}) further support this.
The $\eta$ value of the main AR is smallest at 6 GHz:
this value can be be taken as upper boundary to the amount of FF emission in the disk AR.
It is found that {\it at least} 63\% (representing 6.4\% of the total solar flux) of the 10.7 cm flux from the main AR is due to GR emission, the rest mostly from FF emission.
If one assumes that there is {\it only} FF emission at 17 GHz, then an $\eta$-map derived from the Nobeyama Radioheliograph (NoRH, 17 GHz)
at around 01:00 UT on 2010-Mar-22 gives us the correct FF flux, and hence the correct fraction of the emission that is due to GR: 76\% (or 7.8\% of total solar flux).
\begin{figure}[h!]
\centering
\includegraphics[width=12cm]{fig_2010Jan29_2010Mar21_EM_sqrtT_mainAR_NoRP.ps}
\caption{
$T_{\textrm{B}}$ {\it (top row)} and $\eta$ {\it (bottom row)} radial profiles accross main AR for the 29-Jan-2010 {\it left column} and 21-Mar-2010 {\it right column} observations.
Otherwise, as in Figure~\ref{fig:img:Jan29:4c}.
}
\label{fig:prof:Jan29Mar21:mainAR}
\end{figure}
\section{Summary and Conclusions}\label{sect:ccl}
The Allen Telescope Array has been used to obtain full--disk solar maps at several frequencies simultaneously.
We have established a multi-frequency flux budget, finding that a single active region amounted for up to $\approx$8\% of the total solar flux at 2750 MHz, the actual fraction varying with frequency and active region.
Using the multi-frequency capability, a simple method to find regions of the Sun where optically thin free--free emission dominates has been devised.
Regions where the $\eta$-value varies with frequency are likely regions with a gyroresonance component.
The emission from an active region observed on 29-Jan-2010 appeared to be purely free--free thermal bremsstrahlung in origin.
However, the emission from an active region observed on 21-Mar-2010 appeared to be $\approx$75\% due to gyroresonance, and $\approx$25\% due to free--free emission at 2750 MHz.
Although its current operational status is uncertain, the ATA's combination of high-quality mapping and frequency coverage put it in a unique position to address global coronal studies.
Moreover, its high spectral resolution below 1 GHz enable it to explore a region that has been poorly investigated so far.
\begin{acks}
This work was supported by NASA Contract No. NAS 5-98033.
We thank the anonymous referee for comments that have helped to improve this article.
\end{acks}
\bibliographystyle{spr-mp-sola}
|
\section*{APPENDIX}
In this Appendix, we calculate the loop diagrams for the M1 radiative decays. We present the calculation of $J/\psi \to \eta_c \gamma$, the calculations of the decays $\psi^\prime \to \eta_c \gamma$
($\psi^\prime \to \eta^\prime_c \gamma$) are obtained by replacing $m_{J/\psi}$ with $m_{\psi^\prime}$ and $g_2^2$ with $g_2 g_2^\prime$ ($g_2^{\prime\, 2}$).
The triangle loop diagrams in Fig.~\ref{triangle} have a similar form as the triangle loop diagrams in the hadronic decays, therefore
the notation used here will be the almost the same as that of Ref.~\cite{Guo:2010ak}. We refer the reader to that paper for explicit expressions for the integrals.
The amplitude from Fig.~\ref{triangle}(a) is
\begin{eqnarray}
\nonumber
i\mathcal{M}_{1a}&=&-2ig_{2}^2 \lambda_{1(3)}
\int\frac{d^{4}l}{(2\pi)^4}\frac{\epsilon_{ijk}q_{i}\epsilon^{\gamma}_{j}(2l-q)_{k}\vec{\epsilon}^{\,J/\psi}\cdot\vec{l}}{8(l_{0}-\frac{\vec{l}^2}{2m_D}+i\epsilon)(l_{0}+\frac{\vec{l}^2}{2m_D}+b_{DD}-i\epsilon)
(l_{0}-q_{0}-\frac{(\vec{l}-\vec{q})^2}{2m_{D^{*}}}-\Delta+i\epsilon)}\\
&=&4g_2^2\lambda_{1(3)}\epsilon_{ijk}q_{i}\epsilon^{\gamma}_{j}\epsilon^{J/\psi}_{k}|\vec{q}\,|^2I^{(2)}_1(q,m_D,m_D,m_{D^*})\, ,
\end{eqnarray}
where $b_{DD} = 2 m_D - m_{J/\psi}$, $\lambda_1=\frac{2}{3}(e\beta+\frac{e}{m_c})$ is relevant for loops with neutral $D$ mesons and
$\lambda_3=-\frac{1}{3}(e\beta-\frac{2e}{m_c})$ is relevant for loops with charged and strange $D$ mesons.
Here $I^{(2)}_1(q,m_1,m_2,m_3)$ only differs from the function defined in Ref.~\cite{Guo:2010ak} by omitting a factor of $m_1m_2m_3$ from the denominator.
Fig.~\ref{triangle}(b) contributes
\begin{eqnarray}
i\mathcal{M}_{1b} &=& 2g_2^2\lambda_{1(3)}\epsilon_{ijk} q_{i}\epsilon^{\gamma}_{j}\epsilon^{J/\psi}_{k}|\vec{q}\,|^2 \\
&\times& (2I^{(2)}_0(q,m_D,m_{D^*},m_{D^*})+4I^{(2)}_1(q,m_D,m_{D^*},m_{D^*})-I^{(1)}(q,m_D,m_{D^*},m_{D^*}))\, , \nonumber
\end{eqnarray}
where the functions $I^{(2)}_0(q,m_1,m_2,m_3)$ and $I_1(q,m_1,m_2,m_3)$ are again the same as functions in Ref.~\cite{Guo:2010ak} up to a factor of $m_1m_2m_3$.
Fig.~\ref{triangle}(c) contributes
\begin{eqnarray}
i\mathcal{M}_{1c} &=& 2g_2^2\lambda_{1(3)}\epsilon_{ijk} q_{i}\epsilon^{\gamma}_{j}\epsilon^{J/\psi}_{k}|\vec{q}\,|^2 \\
&\times&(2I^{(2)}_0(q,m_{D^*},m_{D^*},m_D)+6I^{(2)}_1(q,m_{D^*},m_{D^*},m_D)-I^{(1)}(q,m_{D^*},m_{D^*},m_D))\, . \nonumber
\end{eqnarray}
Fig.~\ref{triangle}(d) contributes
\begin{eqnarray}
i\mathcal{M}_{1d} &=& 2g_2^2\lambda_{2(4)}\epsilon_{ijk} q_{i}\epsilon^{\gamma}_{j}\epsilon^{J/\psi}_{k}|\vec{q}\,|^2 \\
&\times&(2I^{(2)}_0(q,m_{D^*},m_D,m_{D^*})+4I^{(2)}_1(q,m_{D^*},m_D,m_{D^*})-I^{(1)}(q,m_{D^*},m_D,m_{D^*})) \, , \nonumber
\end{eqnarray}
where $\lambda_2=-\frac{2}{3}(e\beta-\frac{e}{m_c})$ is relevant for loops with neutral $D$ mesons and
$\lambda_4=\frac{1}{3}(e\beta+\frac{2e}{m_c})$ is relevant for loops with charged and strange $D$ mesons.
Finally, Fig.~\ref{triangle}(e) gives
\begin{eqnarray}
i\mathcal{M}_{1e} &=& 2g_2^2\lambda_{2(4)}\epsilon_{ijk} q_{i}\epsilon^{\gamma}_{j}\epsilon^{J/\psi}_{k}|\vec{q}\,|^2 \\
&\times& (2I^{(2)}_0(q,m_{D^*},m_{D^*},m_{D^*})+8I^{(2)}_1(q,m_{D^*},m_{D^*},m_{D^*})-I^{(1)}(q,m_{D^*},m_{D^*},m_{D^*})) \, . \nonumber
\end{eqnarray}
In addition to the triangle diagrams with the couplings of $D$ and $D^*$ mesons to the magnetic field from Eq.~(\ref{mag}), there are also two triangle diagrams with the coupling
of the photon to charged $D$ and $D^*$ mesons that arises due to gauging their kinetic terms. These are shown in Fig.~\ref{gauge}. The sum of these two diagrams yields
\begin{eqnarray}
i\mathcal{M}_{2}=4g_2^2e\epsilon_{ijk}q_{i}\epsilon^{\gamma}_{j}\epsilon^{J/\psi}_{k}|\vec{q}\,|^2\left(\frac{1}{m_D}I^{(2)}_1(m_D,m_{D^*},m_D)-\frac{1}{m_{D^*}}I^{(2)}_1(m_{D^*},m_D,m_{D^*})\right).
\end{eqnarray}
So far we have only included the interactions coupling the photon to $D$ and $D^*$ mesons. There are additional diagrams where the photons couple to $\bar{D}$ and $\bar{D}^*$ mesons that give an equal contribution.
Fig.~\ref{contact} shows the loop diagrams with the contact interaction that arises from gauging the coupling $g_2$.
Fig.~\ref{contact}(a) yields
\begin{eqnarray}
\nonumber
i\mathcal{M}_{3a}&=&i2g_2^2e
\int\frac{d^{4}l}{(2\pi)^4}\frac{\epsilon_{ijk}(2l+q)_{i}\epsilon^{\gamma}_{j}\epsilon^{J/\psi}_{k}}{4(l_{0}-\frac{\vec{l}^2}{2m_D}+i\epsilon)
(l_{0}+q_{0}+\frac{(\vec{l}+\vec{q})^2}{2m_{D^{*}}}+b_{DD^{*}}-i\epsilon)}\\
&=&-g_2^2e\epsilon_{ijk} q_{i}\epsilon^{\gamma}_{j}\epsilon^{J/\psi}_{k}I'(m_D,m_{D^*})\, ,
\end{eqnarray}
where
\begin{eqnarray}
I'(m_D,m_{D^*})=\frac{\mu_{DD^*}}{4\pi}\left(\frac{m_{D^*}-m_D}{{m_{D^*}+m_D}}\right)
\sqrt{\frac{\mu_{DD^*}}{m_D+m_{D^*}}|\vec{q}|^2+2\mu_{DD^*}(b_{DD^*}+q_0)} \, ,
\end{eqnarray}
for $\mu_{DD^*}=m_Dm_{D^*}/(m_D+m_{D^*})$ and $b_{DD^*}=m_D+m_{D^*}-m_{J/\psi}$.
Fig.~\ref{contact}(b) is related to Fig.~\ref{contact}(a) by charge conjugation so
\begin{eqnarray}
i\mathcal{M}_{3b}=i\mathcal{M}_{3a}=-g_2^2e\epsilon_{ijk} q_{i}\epsilon^{\gamma}_{j}\epsilon^{J/\psi}_{k}I'(m_D,m_{D^*})\, .
\end{eqnarray}
The graphs in Fig.~\ref{contact}(c) and Fig.~\ref{contact}(d) both vanish, so the total contribution to the amplitude from loops with contact interactions is $2i\mathcal{M}_{3a}$.
Only diagrams with charged and strange $D$ mesons in the loop will contribute.
The total amplitude from loop diagrams in Fig.~\ref{triangle} with neutral $D$ mesons is given by
\begin{eqnarray}
\nonumber
i\mathcal{M}^n_{1}&=&\epsilon_{ijk}\vec{q}_{i}\epsilon^{\gamma}_{j}\epsilon^{J/\psi}_{k}\{g_2^2|\vec{q}|^2[4\lambda_1I^{(2)}_1(q,m_D,m_D,m_{D^*})+2\lambda_1(2I^{(2)}_0(q,m_D,m_{D^*},m_{D^*})\\
&&\nonumber+4I^{(2)}_1(q,m_D,m_{D^*},m_{D^*})-I^{(1)}(q,m_D,m_{D^*},m_{D^*}))+2\lambda_1(2I^{(2)}_0(q,m_{D^*},m_{D^*},m_D)\\
&&\nonumber+6I^{(2)}_1(q,m_{D^*},m_{D^*},m_D)-I^{(1)}(q,m_{D^*},m_{D^*},m_D))+2\lambda_2(2I^{(2)}_0(q,m_{D^*},m_D,m_{D^*})\\
&&\nonumber+4I^{(2)}_1(q,m_{D^*},m_D,m_{D^*})-I^{(1)}(q,m_{D^*},m_D,m_{D^*}))+2\lambda_2(2I^{(2)}_0(q,m_{D^*},m_{D^*},m_{D^*})\\
&&+8I^{(2)}_1(q,m_{D^*},m_{D^*},m_{D^*})-I^{(1)}(q,m_{D^*},m_{D^*},m_{D^*}))]\} \, .
\end{eqnarray}
Adding the five diagrams with the photon coupling to a $\bar{D}$ or $\bar{D}^*$ doubles this contribution. The contribution from diagrams of Fig.~\ref{triangle} with
charged and strange charmed mesons in the loops is obtained by substituting $\lambda_1$ with $\lambda_3$ and $\lambda_2$ with $\lambda_4$.
In addition, the contributions from the diagrams in Fig.~\ref{gauge} and Fig.~\ref{contact} with charged and strange charmed mesons in the loops
need to be included. The full decay amplitude is
\begin{eqnarray}
i\mathcal{M}_{full}=i\mathcal{M}_0+2(i\mathcal{M}^n_{1}+i\mathcal{M}^c_{1}+i\mathcal{M}^s_{1}+i\mathcal{M}^c_{2}+i\mathcal{M}^s_{2}+i\mathcal{M}^c_{3a}+i\mathcal{M}^s_{3a}) \, ,
\end{eqnarray}
where the superscript $c (s)$ indicates a contribution from loops with charged (strange) $D$ mesons.
The decay rate for $J/\psi\rightarrow\gamma\eta_c$ is given by
\begin{eqnarray}
\Gamma[J/\psi\rightarrow\gamma\eta_c]=\frac{1}{8\pi}(\sqrt{m_{J/\psi}m_{\eta_c}}M_{full})^2\frac{|\vec{q}\,|}{m_{J/\psi}^2}.
\end{eqnarray}
\input{charmloops.bbl}
\end{document}
|
\section{Introduction}
OJ287 is one of the brightest AGN in the sky. Since it is also highly variable, it has become one of the favorite objects for both professional and amateur astronomers to follow. In addition, it lies close to the ecliptic, which means that its image has been recorded by chance in hundreds of photographic plates since 1891.
In 1982 one of the authors (A.S.) put together the historical light curve of OJ287 based on published measurements. These were partly photometric measurements since the discovery of OJ287 as an extragalactic object in 1968, partly studies of photographic plate archives from years prior to 1968 that are kept in various observatories, in particular at Harvard and at Sonneberg. There appeared to be a 12 year outburst cycle (see Figure 1), and moreover, it was obvious that the next cyclic outburst was due very shortly. This prediction was distributed to colleages world wide, and indeed, OJ287 did not disappoint us but produced the expected event in the following January [1,2]. Observations showed a sharp decline in the percentage polarization during the outburst maximum, indicating that the outburst was produced essentially by unpolarized light [3]. This is different from ordinary outbursts in OJ287 which are characterized by an increase in the percentage polarization at the maximum light. In radio wavelengths the outbursts were found to follow the optical outbursts with a time delay of between 2 months and a year, depending on the observing frequency [Ref. 4].
\begin{figure}
\includegraphics[width=5cm,angle=270]{Historical.eps}
\caption{The optical light curve of OJ287 from 1891 to 2010. The observations are taken from Ref. 22, complemented by unpublished data from R.Hudec and M.Basta}
\end{figure}
Sillanp\"a\"a et al. [Ref. 5] suggested that OJ287 is a binary black hole system where a smaller companion periodically perturbs the accretion disk of a massive primary black hole. They stated that the best way to verify this hypothesis was to study future outbursts and to show that the major axis of the binary system precesses as expected in General Relativity. The next expected outburst was in 1994; it came as scheduled [Ref. 6, 7]. At this point it became obvious to us that we are indeed dealing with a relativistic binary system, and it became necessary to develop the model in greater detail. In the binary model there should be two disk crossings per 12 yr orbital period. Thus the 1994 outburst should have an equal pair whose timing was calculated to be at the beginning of 1995 October [Ref. 8 - 10]. This prediction was also verified [Ref. 11].
Figure 1 shows the optical light curve from 1891 to 2010. The main gap in the light curve is between years 1893 and 1897; observations from three consecutive years are missing there. There are no major gaps since 1897. (Ref.12 makes a peculiar statement that ``there are about 10 year long gaps in the data''; no such gaps are seen even in their own historical light curve.)
Alternative explanations have also been put forward. Quasiperiodic oscillations in an accretion disk were suggested [Ref. 13], and several binary toy models without relativistic precession have also been proposed [Ref. 14 - 16]. The latter models all predicted the next main outburst of OJ287 in the autumn of 2006, while the precessing binary model gave a prediction one year earlier, at the beginning of 2005 October [Ref. 17, 18]. The second major outburst was expected in late 2007 in all binary models, while in the single black hole model there was no reason to expect a second major outburst. In the precessing binary model the date was given with high accuracy, with the last prediction prior to the actual event being 2007 September 13 [Ref. 19, 20]. In the single accretion disk model and in the non-precessing binary models the nature of the radiation at these outbursts should have been polarized synchrotron radiation, while the precessing binary model predicted unpolarized bremsstrahlung radiation [Ref. 9]. In addition, the precessing binary model predicted a series of further outbursts for the interval 2007 - 2010, but they were expected to show up as an increased level of synchrotron radiation [Ref. 17]. Also, the companion black hole should affect the disk of the primary in a predictable way, leading to the wobble of the jet [Ref. 21]. In contrast, the non-precessing models predicted simultaneous brightening both in radio and in optical, at least for the second outburst [Ref. 16] since in these models disk impacts play no role or a minor role, and flux enhancements are purely jet phenomena.
With these predictions in mind, a multiwavelength campaign of observing OJ287 during 2005 - 2010 was set up, with one of the authors (A.S.) among the leaders.
\section{Five ``smoking gun'' results}
In the following, we describe five ``smoking gun'' observations which produced expected results from the point of view of the precessing binary black hole model, but which were surprising and unexpected in other theories.
\subsection{Timing the 2005 outburst}
The 2005 outburst was well covered by observations. The points in Figure 2 are daily averages, 92 in all, formed from altogether 2329 observations in V-band and in R-band. The latter are transformed to V-band by adding 0.4 magnitudes to the R-band value. Finally the flux values are calculated in a standard way (see e.g. Ref. 22).
\begin{figure}
\includegraphics[width=5cm,angle=270]{Figure05.eps}
\caption{The optical light curve of OJ287 during the 2005 outburst. The data points are based on Refs. 23, 24 and 25. The dashed line is the theoretical fit based on Ref. 9.}
\end{figure}
According to Ref. 10, the impact causing the 2005 outburst was expected 22.3 years after the impact of the 1983 outburst. In addition, in Ref. 9 it is estimated that the 2005 outburst should be delayed after the impact. The 1983 outburst is also delayed but not as much, the difference being 0.44 yr. The timing uncertainty was estimated to be $\pm 0.1$ yr. The rapid flux rise started in the latter outburst at 1983.00; thus the corresponding rapid flux rise of the 2005 outburst was expected at 1983.00 + 22.30 + 0.44 = 2005.74. Actually, the outburst was one week late and did not begin until 2005.76 (Ref. 23), but anyway the timing was well within the stated error limits. The comment in Ref. 12 that ``No one expected a major burst at this point'' is rather strange, and it fails to understand that any prediction has its associated error limits. Only a few polarization measurements were carried out at that time, and unfortunately, even those happened during secondary flares. Thus the polarization state of the primary outburst remains unknown from observations. (In contrast, Ref. 12 states that ``the whole burst was rather strongly polarized'', based on two measurements among more than 2000 light curve points, an extraordinary extrapolation!)
Sillanp\"a\"a et al. [Ref. 5] stated that the binary system should show forward precession and thus the disk crossings should follow each other at shorter intervals than the orbital period. The required amount of precession is easily calculated, and it turns out to be $39.1^{\circ}$ per period. It is so much higher than e.g. in binary pulsars (by a factor of $10^4$) that we immediately realise the importance of OJ287 in testing General Relativity.
\begin{figure}
\includegraphics[width=5cm,angle=270]{ValtonenFigure4.eps}
\caption{The optical - UV spectrum of OJ287 during the 2005 outburst. Data points are based on Ref. 25. The solid line is the bremsstrahlung fit, as predicted in Ref. 9. The observational points are corrected for the internal extinction in OJ287, taken from Ref. 35, and the assumed Galactic extinction law (Ref. 37). The standard extinction in our Galaxy is also taken into account.}
\end{figure}
We may also calculate the mass of the primary. Its value, $1.84\times10^{10}$ solar mass, seemed rather high when it was first calculated, but subsequent work on black hole mass functions now places it among the fairly common upper mass range black holes (common in the same sense as O-type stars are common among main sequence stars, see Refs. 26 - 29). This mass value places OJ287 right on the mean correlation of the black hole mass - host galaxy K-magnitude relation, with $M_K\sim-28.9$ (Refs. 30, 31).
In Ref. 12 it is claimed that OJ287 is significantly, slightly more than one standard deviation, off the mean correlation. Based on this, the authors state that the measurement in Ref. 30 ``is most likely spurious''. The reason for the one standard deviation offset in Ref. 12 may be traced to an incorrect way of transforming optical magnitudes to K-magnitudes. The process of transforming the R-band measurement of the host galaxy magnitude (Refs. 32 -34) to the K-band is composed of several steps, each containing its associated uncertainties. One has to have a theory of the stellar composition of the host galaxy (not trivial for a merger) and of its (passive) cosmological evolution. One has to measure the neutral hydrogen column density of the host galaxy, and then to transform it to the extinction in R-band. For the hydrogen column density there exists a measurement in Ref. 35, albeit with large error bars, while for the extinction curves large variations from galaxy to galaxy have been found (Ref. 36). As a result of these large uncertainties, one can safely say that the R-band measurements of the magnitude of the host galaxy are consistent with the direct measurement in the K-band (which does not require the above mentioned transformations), and that the black hole mass - host galaxy K-magnitude correlation holds in OJ287 within measurement errors. In any case, a displacement by one standard deviation from the mean correlation cannot be used as an argument for the correctness or otherwise of a single point in a correlation diagram.
\begin{figure}
\includegraphics[width=7cm,angle=0]{notkuva-1.ps}
\caption{The optical light curve of OJ287 during the 2007 September outburst. The upper panel shows the measured magnitudes, while the lower panel shows the percentage polarization. Data points are published in Ref. 22.}
\end{figure}
\subsection{Nature of radiation at the 2005 outburst}
Impact outbursts are expected to consist of bremsstrahlung radiation, and thus the optical polarization of OJ287 should go down during them. As mentioned above, polarization information for the basic 2005 outburst is not available. However, bremsstrahlung may also be recognized by its spectrum, and this is the part of the campaign that was successfully carried out.
\begin{figure}
\includegraphics[width=5cm,angle=270]{Fig_lowpol.ps}
\caption{The optical light curve of OJ287 during 2006-2008. Only low polarization (less than $10\%$) data are shown; they are based on ref. 24.}
\end{figure}
We had XMM-Newton observations both before the 2005 outburst (2005 April), and during the outburst (2005 November 3-4). Fortunately, the November observation happened at the time when the source was at its basic outburst level, in between two secondary bursts. Thus we would expect to see an additional pure bremsstrahlung spectrum above the underlying synchrotron power-law. A preliminary report of these observations has appeared in Ref. 25, and a more detailed report is under preparation.
\begin{figure}
\includegraphics[width=5cm,angle=270]{AFig2007.eps}
\caption{The optical light curve of OJ287 during the 2007 September outburst. Only low polarization (less than $10\%$) data points are shown. The data points are based on Ref. 22. The dashed line is the theoretical fit based on Ref 9. The arrow points to September 13.0, the predicted time of origin of the rapid flux rise.}
\end{figure}
In Figure 3 we show the difference between the 2005 November flux and the 2005 April flux. The values have been corrected for the Galactic extinction and for the extinction in the OJ287 host galaxy. For the latter, we use the measuments in Ref. 35 and the standard Galactic extinction curve (Ref. 37). The solid line shows the bremsstrahlung spectrum at the expected temperature of $3\times 10^{5~\circ}$K. Note that a raised synchrotron spectrum, as one might have expected in some other theories, would have a downward slope toward higher frequencies, and it is entirely inconsistent with observations. Incidentally, the effect of the extinction in the host galaxy of OJ287 is such that it causes an apparent spectral break in the normal (non-outburst) spectrum in the optical region, while it really happens at the AGN source somewhere in the UV (Ref. 35).
\subsection{Timing and nature of the 2007 September 13 outburst}
The 2007 September 13 outburst was an observational challenge, as the source was visible only for a short period of time in the morning sky just before the sunrise. Therefore a coordinated effort was made starting with observations in Japan, then moving to China, and finally to central and western Europe. A crucial role was played by the NOT telescope in the Canary Islands and by the Calar Alto telescope in mainland Spain, which were able to make polarization observations. The observed points with estimated error bars are given in Figure 4, where the contributions by participating observatories are shown by colour codes.
A comparison of the two panels shows immediately that there were two kinds of outbursts in 2007 September. Three outbursts were highly polarised, with the degree of polarization above $15\%$, while the biggest outburst had polarization below $10\%$. Thus it is not difficult to decide which was the expected bremsstrahlung event. Later in the year there were more highly polarized outbursts, but if we look at the light curve composed of low polarization states only (Figure 5), the September 13 outburst clearly stands out.
In Figure 6 we look at the low polarization light curve in more detail around the September 13 event. A theoretical light curve is also drawn, and an arrow points to the expected moment of the beginning of the sharp flux rise. We see that the observed flux rise coincides within 6 hours with the expected time. The accuracy is about the same as we were able to predict the return of Halley's comet with in 1986!
\subsection{2007 - 2010 outbursts}
Ref. 18 gave a detailed prediction of the whole light curve of OJ287 during the campaign period (Figure 7). In addition to the two impact outbursts, it was expected that the tidal forcing mechanism of Sillanp\"a\"a et al. [Ref. 5] would raise the general level of activity of OJ287, starting from the spring of 2007 and continuing until the spring of 2009. The detailed structure of minor bursts in Figure 7 is immaterial, since it is due to Poisson noise in a simulation with a finite number of disk particles. This prediction is best compared with the low-polarization light curve of Figure 5. In general outline OJ287 behaved just as expected, except that the optical flux declined fast in the spring of 2008, sooner than we would have thought.
\begin{figure}
\includegraphics[width=5cm,angle=270]{Figure_sundelius.ps}
\caption{The predicted optical light curve of OJ287 during 2000-2012. The data is based on Ref. 18.}
\end{figure}
At this point we may remind the reader that Sillanp\"a\"a et al. [Ref. 5] also interpreted some sharp flux decreases as ``eclipses''. At these times the secondary may move across our line of sight, between us and the AGN optical continuum source. One such event was predicted in 2008 [Ref. 9], but it is not included in the light curve prediction of Figure 7. However, if we take the differential of observed minus predicted flux (Figure 8), the eclipse-like feature becomes quite obvious. The two previous ``eclipses'' in the same sequence occurred in 1989 [Ref. 38] and in 1998 [Ref. 39]. The astrophysical reason for the eclipses could be gravitational deflection of the jet stream by the secondary, extinction in gas clouds circling the secondary, or something else.
Figure 7 shows also a prominent outburst at the end of 2009. It also came as expected [Ref. 40].
In our model the accretion disk is optically thick but geometrically thin, it possesses a strong magnetic field [Ref. 41] and the disk is connected to the jet by magnetic field lines [Ref. 42]. Ref. 12 presents an entirely different model which they then strongly criticise, and finally try to make a case for quasi-periodic oscillations in an accretion disk of a single black hole. It is shown in Ref. 40 that the probability that such a model would explain the good match between the theory and observations is less than one in part in $10^8$, not to mention that the other ``smoking gun'' observations also remain unexplained in such a model. Actually, there is no evidence presented in favour of a single black hole model in Ref. 12, while the criticism of a binary model is misdirected and consists of a number of incorrect statements.
\begin{figure}
\includegraphics[width=5cm,angle=270]{2004figure11.eps}
\caption{The observed optical flux of OJ287 during 2004-2010 minus the prediction in Ref. 18. The scatter is 1.4 mJy, and the only significant deviation from the prediction occurs in the spring of 2008 which suggests an eclipse.}
\end{figure}
\subsection{Turning jet}
The accretion disk as a whole is also affected by the companion in its 12 year orbit. On the other hand, in our model the jet and the disk are connected. Thus the jet direction should be strongly influenced by the companion.
There are three periodicities that could be expected to show up: the 12 yr orbital cycle, the 120 yr precession cycle (or half of it due to symmetry) and the Kozai cycle [Ref. 43], which also happens to be 120 yr. The 12 yr orbital cycle produces the tidal enhancements in accretion flow, as postulated by Sillanp\"a\"a et al. [Ref. 5], but in addition this enhancement can be stronger or weaker depending on where we are in the precession cycle. These two tidal effects pretty much explain the overall appearance of the light curve [Ref. 18]. In addition, there is a modulation in the long term base emission level (unexplained by the tidal enhancement) which is in tune with the Kozai cycle. This cycle may also appear in polarization data [44]. The jet orientation is delayed relative to the disk wobble. Theoretically the delay should be of the order of ten years; fitting with the optical data gives the best fit with a 13 yr delay.
The jet wobble shows up in observations in several ways. First, the mean angle of optical polarization varies. The binary model predicts, among other things, a quick change in the optical polarization angle by nearly $90^{\circ}$ around 1995, which was observed [Ref. 12]. In radio, we should see a similar rapid change in the position angle of the parsec scale jet. Depending on the actual value of the delay in radio jet orientation, the change could already be under way (Figure 9), or it may be delayed by another 12 cycle (Figure 10, Ref. 45). There are recent observations which suggest the first alternative [Ref. 46], but the interpretation of these observations is not yet clear-cut.
There are also longer periods that are expected in the binary model: the period of the black hole spin (about 1300 yr, Refs. 47,48) which may show up in the structure of the megaparsec scale jet [Ref. 49]. Also the time scale of the binary settling in the nucleus of OJ287 after a merger of two galaxies, about $10^8$yr [Ref. 50], may be connected with the overall curvature of the magaparsec jet. In the shorter time scales, the half-period of the last stable orbit around the Kerr black hole of $\sim50$ days may also show up in observations [Ref.51].
\begin{figure}
\includegraphics[width=5cm,angle=270]{Fig_PositionAngle_15.eps}
\caption{The observed position angle of the radio jet of OJ287 (points) compared with the binary model, with a 3 year response time of the jet orientation changes.}
\end{figure}
\begin{figure}
\includegraphics[width=5cm,angle=270]{Fig_PositionAngle_16.ps}
\caption{The observed position angle of the radio jet of OJ287 (points) compared with the binary model, with a 14 year response time of the jet orientation changes.}
\end{figure}
\section{Testing General Relativity}
Using the OJ287 binary, we may test the idea that the central body is actually a black hole. One of the most important characteristics of a black hole is that it must satisfy the so called no-hair theorem or theorems (Refs. 52-57). A practical test was suggested in Refs. 58, 59. In this test the quadrupole moment $Q$ of the spinning body is measured. If the spin of the body is $S$ and its mass is $M$, we determine the value of $q$ in
\begin{eqnarray}
Q = -q \frac{S^2}{Mc^2}.
\end{eqnarray}
For black holes $q=1$, for neutron stars and other possible bosonic structures $q > 2$ (Refs. 60, 61)
We calculate the two-body orbit using the third Post-Newtonian (3PN) order orbital dynamics, which includes the leading order
general relativistic, classical spin-orbit and radiation reaction effects (Refs. 62-64).
The 3PN-accurate equations of motion can be written schematically as
\begin{eqnarray}
\ddot { {\vek x}} \equiv
\frac{d^2 {\vek x}} { dt^2} &=&
\ddot { {\vek x}}_{0} + \ddot { {\vek x}}_{1PN} \nonumber
+ \ddot { {\vek x}}_{SO}+\ddot { {\vek x }}_{Q}\\
&& + \ddot { {\vek x}}_{2PN} + \ddot { {\vek x}}_{2.5PN}
+ \ddot { {\vek x}}_{3PN} \,,
\end{eqnarray}
where ${\vek x} = {\vek x}_1 - {\vek x}_2 $ stands for the
center-of-mass
relative separation vector between the black holes with masses $m_1$ and $m_2$ and
$ \ddot { {\vek x}}_{0} $ represents the Newtonian acceleration given by
$ \ddot { {\vek x}}_{0} = -\frac{ G\, m}{ r^3 } \, {\vek x} $; $m= m_1 + m_2$ and $ r = | {\vek x} |$.
The PN contributions occurring at the conservative 1PN, 2PN, 3PN and the reactive 2.5PN orders, denoted
by $\ddot { {\vek x}}_{1PN}$, $\ddot { {\vek x}}_{2PN}$, $\ddot { {\vek x}}_{3PN}$ and
$\ddot { {\vek x}}_{2.5PN}$ respectively, are non-spin by nature, while $ \ddot { {\vek x}}_{SO}$ is the spin-orbit term of the order 1.5PN.
The quadrupole-monopole interaction term $\ddot { {\vek x}}_Q $, entering at the 2PN order, reads
\begin{eqnarray}
\ddot { {\vek x}}_Q & =- q \, \chi^2\,
\frac{3\, G^3\, m_1^2 m}{2\, c^4\, r^4}
\, \biggl \{
\biggl [ 5(\vek n\cdot \vek s_1)^2
-1 \biggr ] {\vek n}
-2(\vek n\cdot \vek s_1) {\vek s_1} \biggr \},
\end{eqnarray}
where parameter $q$, whose value is $1$ in general relativity, is introduced to test the black hole `no-hair'
theorem. The Kerr parameter $\chi$ and the unit vector
${\vek s}_1$ define the spin of the primary black hole by the relation
${\vek S}_1 = G\, m_1^2 \, \chi\, {\vek s}_1/c$
and $\chi$ is allowed to take values between $0$ and $1$ in general relativity. The unit vector {\vek n} is along the direction of {\vek x}.
In Figure 11 we show the distribution of $q$-values allowed by ``good'' orbits. By ``good'' we mean an orbit which gives the correct timing of 9 outbursts within the range of measurement accuracy. Obviously the range of timing at each of the 9 outbursts means that a set of solution orbits are possible. Here we have used a representative set of such orbits.
We note that the distribution peaks at $q=1$, thus confirming the no-hair theorem. It is also the first test of general relativity that has been performed at higher than the 1.5 Post-Newtonian order. Thus it forms a milestone in our study of the correct theory of gravitation.
\begin{figure}
\includegraphics[width=5cm,angle=270]{FigureFit38000.eps}
\caption{The distribution of the test parameter $q$ among 598 solutions of the orbit. The result is consistent with General Relativity ($q=1$), and excludes the cases of no relativistic spin-orbit coupling ($q=0$) and of a material body ($i.e.$ not a black hole, $q$ greater than 2).}
\end{figure}
\section{Conclusions}
Prior to the 2005 - 2010 multiwavelength campaign there were several ideas about the nature of OJ287. Fortunately, these models made completely different predictions about the behaviour of OJ287 during these years. One of the key differences was the timing of the first outburst: the precessing binary model, initially proposed by Sillanp\"a\"a et al. [Ref. 5] with subsequent refinements [Refs. 9, 10, 18] predicted the outburst in October 2005, the other models in October 2006. The result of a scientific enquiry is seldom as clear-cut as this: the outburst came within one week of the time expected in the precessing binary model and its spectrum agreed with the bremsstrahlung spectrum at the predetermined temperature. The prediction for the second outburst turned out to be accurate within 6 hours, and the lack of polarization again suggested strongly the bremsstrahlung origin. All flux values predicted for this period turned out to be accurate with the standard scatter of 1.4 mJy, which is only ten percent of the variability range in optical. The only exception occurred in 2008; however, this was the time when lower flux values were expected due to an ``eclipse''. We have ``eclipse'' in quotation marks, as the reason for the sudden fade at the time when the secondary passes through our line of sight is not known.
The optical variability data specifies the binary model except for the exact direction of the jet relative to our line of sight. However, the resolved parsec scale radio jet allows us to get a handle on this parameter, too. The remaining unknown is the delay between the wobble of the accretion disk, due to the effect of the secondary, and the reorientation of the jet in the sky. A major reorientation may already have started, or it may come after one orbital period, depending on the details of the jet/disk connection [Ref. 65].
The success of the binary model has encouraged us in using it to test theories of gravitation. Any theory which can be presented as Newton's law plus Post-Newtonian terms may be studied, as different laws of gravitation produce different impact times on the accretion disk. We have used the Post-Newtonian terms of general relativity up to third order, and have found that the orbit solutions agree with the theory. Our test parameter $q$ should have the value of exactly 1, and indeed the possible solutions cluster around this value. The parameter values $q=0$ or $q=2$ can be rejected at the 4 standard deviation level at present. This is the first time that it has been possible to study general relativity at higher than the 1.5PN order [Ref. 66].
\section{Acknowledgement}
We would like to thank all the participants in this campaign for the extraordinary amout of work dedicated to solving the riddle of OJ287. In particular, we thank Kari Nilsson and Stefano Ciprini, who have put together and harmonized huge amounts of data. Rene Hudec has made an invaluable contribution in collecting historical data, Tapio Pursimo and Leo Takalo were responsible for the key polarization observations at NOT, and Seppo Mikkola, Harry Lehto and Achamveedu Gopakumar have been key persons in turning the data into a viable binary black hole model.
|
\section{Introduction}
Many fundamental questions about output feedback pole assignment for general linear systems
have been answered by appealing to algebraic geometry, and more specifically
to the geometry of Grassmann manifolds.
This body of work has led to important contributions in systems theory:
Hermann and Martin gave necessary and sufficient conditions for complex static output
feedback control~\cite{HM1, HM2, HM3}, Brockett and Byrnes used Schubert calculus to count
the number of pole-assigning feedback laws~\cite{BB81}, and then Rosenthal~\cite{Ro94} and
Ravi, Rosenthal, and Wang~\cite{RRW1,RRW2} solved these problems for complex dynamic
compensators using quantum Schubert calculus.
For a description of the earlier literature, we recommend~\cite{By89}.
This line of work on complex feedback has led to a solution of the problem of pole-assignment in the
real case, some of which is found in~\cite{EG,RSW,W92}.
Likewise, it has influenced work in algebraic geometry~\cite{HV,So00b,SS}, some of
which is surveyed in~\cite{So01}.
The \demph{Lagrangian Grassmannian} and \demph{orthogonal Grassmannian} are subsets of the
usual Grassmannian, and in principle they should also appear in systems theory.
This was realized by Helmke, Rosenthal, and
Wang~\cite{HRW06}, who studied the control of linear systems with symmetric and Hamiltonian
state-space realizations by static symmetric output feedback.
They gave necessary and sufficient conditions for pole placement by static symmetric complex
feedback, linking this problem to the Schubert calculus on Lagrangian Grassmannians, and then used
this link to count the number of complex feedback laws
when there are finitely many of them.
We continue this line of research.
We first identify spaces of linear systems
with a natural symmetry as certain
subvarieties of the space of rational curves in a Grassmannian.
More specifically, we consider linear systems with McMillan degree $n$
whose transfer function $G(s)$ (a square matrix of rational functions which is
defined at $s=\infty$) has one of the following four symmetries:
\begin{enumerate}
\item $G(s)^T=G(s)$ \quad symmetric,
\item $G(s)^T=G(-s)$ \quad Hamiltonian ($n$ must be even),
\item $G(s)^T=-G(-s)$ \quad skew-Hamiltonian, and
\item $G(s)^T=-G(s)$ \quad skew-symmetric ($n$ must be even).
\end{enumerate}
Symmetries (1)-(2) were studied in~\cite{HRW06} and occur naturally in
systems theory~\cite{BD,F83}.
Stabilization of symmetric systems (1) by real symmetric output feedback was
considered in~\cite{MaHe}, where it was shown that there may be no real feedback
laws placing $n$ real poles when $n\geq m$.
The symmetries (3)-(4) are natural to consider from the point of view of algebraic
geometry.
Theorem~\ref{Th:three} gives an example of a real $m$-input $m$-output skew-symmetric
linear system of McMillan degree $2\binom{m}{2}$ such that every feedback law is real when
placing real poles, demonstrating that it is feasible to place real
poles with real skew-symmetric feedback.
Let $A$ be a nondegenerate bilinear form on $\C^{2m}$.
The annihilator \DeCo{$H^{\perp_A}$} of a plane $H$ in $\C^{2m}$ is
the set of $v\in\C^{2m}$ such that $A(v,w)=0$ for all $w\in H$.
The annihilator of an $m$-plane in $\C^{2m}$ is
also an $m$-plane, and so the association $H\mapsto H^{\perp_A}$ defines an involution \DeCo{$\iota_A$} on the Grassmannian
\demph{$G(m,2m)$}.
If $A$ is skew-symmetric, then the set of fixed points of $\iota_A$ is the
\demph{Lagrangian Grassmannian, $LG(m)$}, which is a manifold of dimension
$\binom{m{+}1}{2}$.
If $A$ is symmetric, then the set of fixed points of $\iota_A$ has two isomorphic
components, either of which forms the $\binom{m}{2}$-dimensional
\demph{orthogonal Grassmannian, $OG(m)$}, also called the \demph{spinor variety}~\cite{FuPr}.
It is classical (e.g., proved by Gaussian elimination \cite{BW66}) that any two invertible complex symmetric matrices $A,B$ are (transpose) \textit{congruent}: there exists an invertible matrix $X$ such that $X^{\top}AX = B$.
Similarly, any two invertible complex skew-symmetric matrices are congruent.
Thus, we will always assume that our forms are $\langle x,y\rangle= x^{\top}Ay$, where $A$ is either
\begin{equation}\label{Eq:forms}
\DeCo{O_{2m}}\ :=\ \left[\begin{matrix}0&I_m\\ I_m&0\end{matrix}\right]
\qquad\mbox{or}\qquad
\DeCo{J_{2m}}\ :=\ \left[\begin{matrix}0&I_m\\-I_m&0\end{matrix}\right]\,,
\end{equation}
where $I_m$ is the $m\times m$ identity matrix.
We omit the subscripts $m$ and $2m$ when the dimensions are clear from context.
A general linear subspace $H\in G(m,2m)$ is the row space of a matrix of the form
$[I_m \co F]$, where $F$ is an $m\times m$ matrix.
A calculation shows that $\iota_J(H)$ is spanned by $[I_m\co F^T]$ and
$\iota_O(H)$ is spanned by $[I_m\co -F^T]$, so that $H\in LG(m)$
if and only if $F$ is symmetric and $H\in OG(m)$ if and only if $F$
is skew-symmetric.
We reach the same conclusion for $H$ of the form $[F\co I_m]$.
Such isotropic planes form dense open subsets of $LG(m)$ and $OG(m)$.
If we associate an $m$-input $m$-output proper transfer function $G(s)$ of McMillan
degree $n$ to the row space of the matrix
\[
[I_m\ :\ G(s)]\,,
\]
we obtain a map $\gamma\colon \P^1\to G(m,2m)$ of degree $n$, where $\P^1$ is the
complex projective line.
The image is the \demph{Hermann-Martin curve}~\cite{HM3} of $G(s)$, which we will identify
with $G(s)$.
The set of all such proper transfer functions forms a dense open subset in the space
of rational curves of degree $n$ in the Grassmannian $G(m,2m)$~\cite{Ro94}.
Our first main result identifies sets of transfer functions with symmetries as natural
subvarieties of the space of rational curves in the Grassmannian.
\begin{theorem}\label{Th:one}
The following hold over the ground field $\C$.
$(1)$ The set of symmetric linear systems with $m$ inputs and $m$ outputs of McMillan degree
$n$ is an irreducible quasiprojective manifold of dimension $(m{+}1)n + \tbinom{m{+}1}{2}$.
It is a dense open
subset of the space of rational curves of degree $n$ in $LG(m)$.
$(2)$ The set of Hamiltonian linear systems with $m$ inputs and $m$ outputs of even McMillan
degree $n$ is an irreducible quasiprojective manifold of dimension $mn + \tbinom{m{+}1}{2}$.
It is a subset
of the space of rational curves $\gamma$ in the Grassmannian $G(m,2m)$ that satisfy:
\begin{equation}\label{Eq:Hamil_curves}
\gamma(-s)\ =\ \iota_J(\gamma(s))\,.
\end{equation}
$(3)$ The set of skew-Hamiltonian linear systems with $m$ inputs and $m$ outputs of McMillan
degree $n$ is an irreducible quasiprojective manifold of dimension $mn + \tbinom{m}{2}$.
It is a subset
of the space of rational curves $\gamma$ in the Grassmannian $G(m,2m)$ that satisfy:
\begin{equation}\label{Eq:Sk_Hamil_curves}
\gamma(-s)\ =\ \iota_O(\gamma(s))\,.
\end{equation}
$(4)$ The set of skew-symmetric linear systems with $m$ inputs and $m$ outputs of even
McMillan degree $n=2\ell$ is an irreducible quasiprojective manifold of dimension $(m{-}1)n + \tbinom{m}{2}$.
It is naturally a dense open
subset of the space of rational curves of degree $\ell$ in
$OG(m)$.
\end{theorem}
The proof of Theorem~\ref{Th:one} is straightforward and given in Section~\ref{S:one},
following a proof of a version of the Kalman Realization
Theorem~\cite[Theorem~6.2-4]{Ka80} for symmetric transfer functions.
We do not know if Hamiltonian systems form dense open subsets of the space of curves
satisfying~\eqref{Eq:Hamil_curves}, or if skew-Hamiltonian systems form dense open subsets
of the space of curves satisfying~\eqref{Eq:Sk_Hamil_curves}, for these spaces of curves have
yet to be studied.
As sets of linear systems with symmetry are identified with irreducible quasiprojective algebraic
varieties, the notion of genericity for complex systems makes sense.
That is, a property is \demph{generic} if it holds on a nonempty Zariski open subset
(which is therefore dense) of the corresponding space.
Because symmetric and skew-symmetric transfer functions are open subsets of the moduli
spaces of rational curves in the Lagrangian Grassmannian and orthogonal Grassmannian,
respectively, output feedback control by either static or dynamic symmetric and
skew-symmetric linear systems is related to Schubert calculus on these Grassmannians,
both classical (for static feedback laws) and quantum (for dynamic feedback).
The main result of~\cite{HRW06} concerned static symmetric feedback.
We establish the analogous result for static skew-symmetric feedback.
The symmetry of skew-Hamiltonian and skew-symmetric linear systems is preserved by
static skew-symmetric output feedback, so it is natural to place poles with
static skew-symmetric controllers.
The poles of a skew-Hamiltonian transfer function are invariant under multiplication by
$-1$, so there are essentially only $\lfloor n/2\rfloor$ poles to place.
Here, $\lfloor x \rfloor$ is the greatest integer not exceeding the real number $x$.
Similarly, poles of a skew-symmetric linear system occur with even
multiplicity, and therefore a skew-symmetric linear system of even McMillan degree $n$
has only $n/2$ poles to place.
Our second main theorem gives necessary and sufficient conditions for pole placement with
static skew-symmetric feedback.
\begin{theorem}\label{Th:two}
A generic strictly proper skew-symmetric (respectively skew-Hamiltonian) transfer
function $G(s)$ with $m$ inputs, $m$ outputs, and McMillan degree $n$ is pole-assignable
with complex static skew-symmetric feedback compensators if and only if
$\lfloor n/2\rfloor \leq \binom{m}{2}$.
\end{theorem}
In particular, when $m=4$, we see that a skew-symmetric system of McMillan degree $12$ or
less is pole-assignable with complex static skew-symmetric feedback compensators, and
skew-Hamiltonian systems with $m=4$ and McMillan degree $13$ or less are pole-assignable.
We remark that our proof does not determine the dense open subset of pole-assignable
symmetric linear systems.
Our third main result counts the number of feedback laws for a generic skew-symmetric system
of McMillan degree $2\binom{m}{2}$.
It also shows that there exist systems with real feedback
laws, in a strong way---for these systems, every feedback law placing real poles is real.
\begin{theorem}\label{Th:three}
A generic skew-symmetric linear system of McMillan degree
$2\binom{m}{2}$ has exactly
\[
\DeCo{d_m}\ :=\
\binom{m}{2}! \frac{1!\dotsb(m-2)!}{1! 3!\dotsb(2m-3)!}
\]
static complex skew-symmetric
controllers that place a given general set of $\binom{m}{2}$ poles.
Moreover, for every $m$, there exists a real skew-symmetric linear system with $m$ inputs and outputs and
McMillan degree $2\binom{m}{2}$ such that for every choice of $\binom{m}{2}$ real poles,
there are are $d_m$ feedback laws, and every one is real.
\end{theorem}
We prove these theorems in Section~\ref{S:two}.
The argument for Theorem~\ref{Th:two} is influenced by the proof in~\cite{HRW06},
but it is a considerable simplification.
The skew-symmetric system with $d_m$ real feedback laws
comes from the Wronski map in the Schubert calculus, and the result on reality is a
restatement of a theorem of Purbhoo~\cite{Purbhoo}.
We do not address questions about dynamic feedback.
If we use a dynamic compensator of McMillan degree $q$ to place the poles of a
linear system of McMillan degree $n$ with one of these symmetries, then a calculation shows
that the resulting system (of McMillan degree $n+q$) has the same symmetry as the original
system if and only if the compensator had that same symmetry.
Thus it is natural to consider dynamic control when both the system and compensator have the
same symmetry.
A dimension count gives the necessary condition that $n+q$ be at most
\[
(m{+}1)q+\binom{m+1}{2}\,,\quad
mq+\binom{m+1}{2}\,,\quad
mq+\binom{m}{2}\,,\quad\mbox{and}\quad
(m{-}1)q+\binom{m}{2}\,,
\]
for generic pole placement of symmetric, Hamiltonian, skew-Hamiltonian, and skew-symmetric
linear systems by dynamic controllers of the same symmetry.
We do not know if these conditions are sufficient---this requires the generic surjectivity of
the corresponding pole-placement map.
If this dimension condition is necessary and sufficient, then the quantum Schubert calculus
for $LG(m)$ and $OG(m)$~\cite{KT03,KT04} may be used to count the number of dynamic
compensators for symmetric or skew-symmetric systems (also~\cite{Ruffo} for symmetric
compensators).
Counting dynamic compensators for
Hamiltonian and skew-Hamiltonian systems requires a deeper
study of the corresponding spaces of curves, for it will involve orbifold quantum cohomology~\cite{CR02}.
Our results and analysis also apply to discrete-time linear systems with
these symmetries, in the same way that continuous-time transfer functions are related
to discrete-time transfer functions when there are no symmetries.
We thank Joachim Rosenthal who explained his work to us and encouraged us to extend his
results to skew-symmetric transfer functions. We also thank Uwe Helmke for discussions on the
systems theory background.
\section{Geometry of state-space realization with symmetries}\label{S:one}
We study the state-space realizations of transfer functions with symmetry and
identify the spaces of such transfer functions as
certain subvarieties of the moduli spaces of rational curves in Grassmannians.
Portions of this material are classical or can be found in~\cite{HRW06}, but we include some
proofs for completeness. We work entirely over the complex numbers.
Write \Blue{$X^{\top}$} for the transpose of a matrix $X$ and \Blue{$X^{-\top}$} for
$(X^{-1})^{\top}=(X^{\top})^{-1}$.
Recall that a square matrix $X$ is \demph{symmetric} if $X^{\top}=X$ and
\demph{skew-symmetric} if $X^{\top}=-X$.
Let $J$ be the $2\ell\times 2\ell$ matrix,
\[
\DeCo{J}\ :=\ \left[\begin{matrix}0&I\\-I&0\end{matrix}\right],
\]
where $I$ is the $\ell\times\ell$ identity matrix.
Note that $J^{\top}=-J=J^{-1}$.
A $2\ell\times 2\ell$ matrix $X$ is \demph{Hamiltonian} if $XJ$ is symmetric and
\demph{skew-Hamiltonian} if $XJ$ is skew-symmetric.
Let $m$ and $n$ be positive integers, which we assume are fixed throughout.
We write $\ell$ for $\lfloor n/2\rfloor$.
Suppose that we have a time-dependent complex linear system with inputs $u\in\C^m$,
outputs $y\in\C^m$, and McMillan degree $n$.
This has a minimal state-space realization:
\begin{equation}\label{Eq:State-Space}
\begin{array}{rcl}
\dot{x}&=& Ax + Bu\\
y&=& Cx + Du\,,
\end{array}
\end{equation}
where $A\in\C^{n\times n}$, $B\in\C^{n\times m}$, $C\in\C^{m\times n}$, and
$D\in\C^{m\times m}$, and corresponding (proper) transfer function,
\[
\DeCo{G(s)}\ :=\ C(sI-A)^{-1}B + D\,.
\]
The \demph{poles} of the transfer function $G(s)$ are the eigenvalues of $A$.
The transfer function $G(s)$ is \demph{strictly proper} if $D=0$.
Given a strictly proper linear system, a static linear feedback law is given by an
$m\times m$ matrix $F$, where we set $u=Fy+v$.
The resulting linear system is
\begin{equation}\label{Eq:controlled_system}
\dot{x}\ =\ (A+BFC)x + Bv\qquad y\ =\ Cx\,,
\end{equation}
and its transfer function has poles at the roots of
\[
\varphi(s)\ :=\ \det(sI-(A+BFC))\,.
\]
A fundamental problem is: When is it possible to choose $F$ to obtain a given choice of
monic polynomial $\varphi(s)$?
A system is \demph{pole-assignable} if the map $F\mapsto \varphi(s)$ is dominant (its image
is a dense subset of the set of polynomials $\varphi(s)$).
Our main result concerns the pole-assignability of generic linear systems with
skew-Hamiltonian and skew-symmetric symmetry.
We first study the spaces of linear systems with symmetry.\smallskip
The complex general linear group $GL(n)$ of invertible $n\times n$ matrices acts
on the space of realizations~\eqref{Eq:State-Space} via
\begin{equation*}\label{GLn_action}
X.(A,B,C,D)\ \longmapsto\ (X^{-1}AX,\, X^{-1}B,\, CX, D)\,,
\end{equation*}
where $X\in GL(n)$, and it preserves the transfer function.
If we restrict this action to the dense open set of
minimal state space realizations,
then the Kalman Realization Theorem~\cite[Theorem~6.2-4]{Ka80} identifies the orbits with
transfer functions and shows that $GL(n)$ acts without fixed points.
In particular, if $(A,B,C,D)$ and $(\alpha,\beta,\gamma,\delta)$ are both minimal state
space realizations of the same transfer function, then there is a unique $X\in GL(n)$ such
that
\[
X.(A,B,C,D)\ =\ (\alpha,\beta,\gamma,\delta)\,.
\]
We first extend these classical facts to transfer
functions with symmetries.
\begin{definition}
A transfer function $G(s)$ is
\demph{symmetric},
\demph{Hamiltonian},
\demph{skew-Hamiltonian}, or
\demph{skew-symmetric} if for all $s\in \C$ we have,
\[
G(s)^{\top}\ =\ G(s)\,,\quad
G(s)^{\top}\ =\ G(-s)\,,\quad
G(s)^{\top}\ =\ -G(-s)\,,\quad\mbox{or}\quad
G(s)^{\top}\ =\ -G(s)\,,
\]
respectively.
\end{definition}
State-space realizations may also have symmetries.
\begin{definition}\label{Def:state-space}
A realization~\eqref{Eq:State-Space} is \demph{symmetric} if
$A$ is symmetric, $B=C^{\top}$, and $D$ is symmetric.
Symmetric realizations have symmetric transfer functions:
\[
G(s)^{\top}\ =\ B^{\top}(sI-A^{\top})^{-1}C^{\top} + D^{\top}
\ =\ C(sI-A)^{-1}B + D\ =\ G(s)\,.
\]
A realization~\eqref{Eq:State-Space} with $n$ even is \demph{Hamiltonian}
if $A$ is Hamiltonian, $JB=C^{\top}$, and $D$ is symmetric.
Note that $A^{\top}=JAJ$ and $B^{\top}=CJ$.
Hamiltonian realizations have Hamiltonian transfer functions:
\begin{eqnarray*}
G(-s)^{\top}\ =\
B^{\top}(-sI-A^{\top})^{-1}C^{\top}+D^{\top} &=& CJ(sJIJ-JAJ)^{-1}JB + D^{\top}\\
&=& C(sI-A)^{-1}B + D^{\top}\ \ =\ G(s)\,.
\end{eqnarray*}
A realization~\eqref{Eq:State-Space} is \demph{skew-Hamiltonian} if
$A$ is skew-symmetric, $B=C^{\top}$, and $D$ is skew-symmetric.
Skew-Hamiltonian realizations have skew-Hamiltonian transfer functions:
\begin{eqnarray*}
-G(-s)^{\top}&=& -B^{\top}(-sI-A^{\top})^{-1}C^{\top} - D^{\top}\\
&=&-C(-sI+A)^{-1}B + D\ =\ G(s)\,.
\end{eqnarray*}
Finally, a realization~\eqref{Eq:State-Space} with $n$ even is
\demph{skew-symmetric} if $A$ is skew-Hamiltonian, $JB=C^{\top}$, and $D$ is skew-symmetric.
In this case, $A^{\top}=-JAJ$ and $B^{\top}=CJ$.
Skew-symmetric realizations have skew-symmetric transfer functions:
\begin{eqnarray*}
-G(s)^{\top}&=&-B^{\top}(sI-A^{\top})^{-1}C^{\top} - D^{\top}\\
&=&-CJ(-sJIJ+JAJ)^{-1}(-JB) +D\ =\ G(s)\,.
\end{eqnarray*}
\end{definition}
\begin{remark}
In symmetric and Hamiltonian realizations, the matrix $A$ has the same type
(symmetric or Hamiltonian, respectively), while for skew-Hamiltonian and skew-symmetric
realizations, the matrix $A$ has the opposite type; it is skew-symmetric or
skew-Hamiltonian, respectively.
\end{remark}
While there is no {\it a priori} reason that a transfer function
with one of these symmetries
would have a minimal state-space realization with the
same symmetry,
that is indeed the case.
The \demph{orthogonal group $O(n)$} is the subgroup of $GL(n)$ consisting of complex
matrices $X$ with $X^{\top}X=I$, and when $n$ is even, the \demph{symplectic group $Sp(n)$}
is the subgroup of $GL(n)$ consisting of complex matrices $X$ with $X^{\top}JX=J$.
We establish the analog of the Kalman Realization
Theorem for transfer functions with symmetry.
\begin{proposition}\label{prop:realizations}
A transfer function has one of the symmetry types---symmetric, Hamiltonian,
skew-Hamiltonian, or skew-symmetric---if and only if it has a complex minimal state-space
realization having the corresponding symmetry type.
Furthermore, if $(A,B,C,D)$ and $(\alpha,\beta,\gamma,\delta)$ are two such minimal
state-space
realizations of the same transfer function, then there is a unique matrix
$X\in O(n)$ (respectively $X\in Sp(n)$) such that
$X.(A,B,C,D)=(\alpha,\beta,\gamma,\delta)$ for symmetric and
skew-Hamiltonian transfer
functions, (respectively for Hamiltonian and skew-symmetric transfer functions).
\end{proposition}
Following the proof for symmetric transfer functions in~\cite{HRW06} (see also~\cite{FH95}),
we give the proof in the cases of skew-Hamiltonian and skew-symmetric transfer functions.
The case of Hamiltonian transfer functions is similar, and is due to Brockett and
Rahimi~\cite{BR72}. Also, the first half, concerning symmetric realizations, is due to Brockett~\cite{Br78}.
\begin{proof}
We prove the forward implication in the first statement as we have already shown that a
state-space realization having one of these symmetries gives a transfer function with the
same symmetry.
Suppose that $G(s)=-G(-s)^{\top}$ is a skew-Hamiltonian transfer function with minimal
state-space realization $(A,B,C,D)$.
Since
\[
-G(-s)^{\top}\ =\ -B^{\top}(-sI-A^{\top})^{-1} C^{\top}-D^{\top}\ =\ B^{\top}(sI+A^{\top})^{-1}C^{\top}-D^{\top}\,,
\]
$(-A^{\top}, C^{\top}, B^{\top}, -D^{\top})$ is also a minimal realization.
By the Kalman Realization Theorem, there is a unique invertible
matrix $X$ such that
\begin{eqnarray*}
(A,\ B,\ C,\ D)&=& X . (-A^{\top},\ C^{\top},\ B^{\top},\ -D^{\top}) \\
&=& (-X^{-1}A^{\top}X,\ X^{-1}C^{\top},\ B^{\top}X,\ -D^{\top})\\
&=& (-X^{-1}(-X^{-1}A^{\top}X)^{\top}X,\ X^{-1}(B^{\top}X)^{\top},\ (X^{-1}C^{\top})^{\top}X,\ -D^{\top})\\
&=& (X^{-1}X^{\top}AX^{-\top}X,\ X^{-1}X^{\top}B,\ CX^{-\top}X,\ -D^{\top}). \\
&=&(X^{-1}X^{\top}). (A,\ B,\ C,\ D).
\end{eqnarray*}
Since $(A,\ B,\ C,\ D) = I. (A,\ B,\ C,\ D)$, another application of the Kalman
Realization Theorem gives us that $X^{-1}X^{\top}=I$; thus, $X$ is symmetric.
An invertible complex symmetric matrix $X$ admits a Tagaki factorization $X=Y^{\top}Y$,
with $Y$ invertible \cite[Corollary 4.4.4]{HJ1}.
Then the realization $(YAY^{-1}, YB, CY^{-1})$ of the
transfer function $G(s)$ is skew-Hamiltonian.
Indeed,
\begin{eqnarray*}
(YAY^{-1})^{\top}& =& Y^{-\top}A^{\top}Y^{\top}\ =\ -Y^{-\top}X^{\top}AX^{-\top}Y^{\top}\\
& =& -Y^{-\top}Y^{\top}YAY^{-1}Y^{-\top}Y^{\top}\ =\ -YAY^{-1}\,.
\end{eqnarray*}
Similarly, $(YB)^{\top}=CY^{-1}$.
Indeed, $C=B^{\top}X$ so $B^{\top}=CX^{-1}$, and thus
\[
(YB)^{\top}\ =\ B^{\top}Y^{\top}\ =\ CX^{-1} Y^{\top}\ =\ C(Y^{-1}Y^{-\top})Y^{\top}
\ =\ CY^{-1}\,.
\]
Now suppose that $(A,B,C,D)$ and $(\alpha,\beta,\gamma,\delta)$ are minimal
skew-Hamiltonian realizations of the same transfer function.
Let $X\in GL(n)$ be the unique matrix such that
\[
X.(A,B,C,D)\ =\ (X^{-1}AX,X^{-1}B,CX,D)\ =\
(\alpha,\beta,\gamma,\delta)\,.
\]
Then we have
\[
\alpha\ =\ -\alpha^{\top}\ =\ -X^{\top}A^{\top}X^{-\top}\ =\ X^{\top}AX^{-\top}\,,
\]
and similarly $\beta=X^{\top}B$ and $\gamma=CX^{-\top}$ so that
$X^{-\top}.(A,B,C,D)=(\alpha,\beta,\gamma,\delta)$.
It follows that $X^{-\top}=X$, by the uniqueness of $X$.
But then $X^{\top}X=I$ and so $X\in O(n)$ is orthogonal.\medskip
Consider next the case that $G(s)=-G(s)^{\top}$ is a skew-symmetric transfer function with
minimal state-space realization $(A,B,C,D)$.
Since $J^{-1}=J^{\top}$, we see that $-G(s)^{\top}$ equals
\begin{eqnarray*}
-B^{\top}(sI-A^{\top})^{-1} C^{\top} -D^{\top} &=& B^{\top}J^{\top}J(-sI+A^{\top})^{-1}JJ^{\top}C^{\top}-D^{\top} \\
&=& (JB)^{\top}(-sJ^{-1}IJ^{-1}+J^{\top}A^{\top}J^{\top})^{-1}(CJ)^{\top}-D^{\top} \\
&=& (JB)^{\top}(sI-(-JAJ)^{\top})^{-1}(CJ)^{\top}-D^{\top} \,,
\end{eqnarray*}
and so $(-(JAJ)^{\top}, (CJ)^{\top}, (JB)^{\top}, -D^{\top})$ is also a minimal realization of $G(s)$.
Then there is a unique $X\in GL(n)$ such that
\begin{eqnarray*}
(A,\ B,\ C,\ D)&=& X.(-(JAJ)^{\top},\ (CJ)^{\top},\ (JB)^{\top},\ -D^{\top})\\
&=& (-X^{-1}(JAJ)^{\top}X,\ X^{-1}(CJ)^{\top},\ (JB)^{\top}X,\ -D^{\top})\,.
\end{eqnarray*}
Substituting the equality into itself and simplifying, we see that
\begin{equation} \label{Eq:SIB}
(A,B,C, D) = (X^{-1}J^{\top}X^{\top}JAJX^{-\top}J^{\top}X,\ X^{-1}J^{\top}X^{\top}JB,\ CJ X^{-\top}J^{\top}X,\ D).
\end{equation}
The right-hand side of~\eqref{Eq:SIB} is not immediately seen to have the form
$R.(A,B,C,D)$,
but if we set $R:=JX^{-\top}J^{\top}X$ and use that $-J^{\top}=J^{-\top}$ and $J^{-1}=-J$,
we obtain
\[
R^{-1}\ =\ X^{-1}J^{-\top}X^{\top}J^{-1}\ =\ X^{-1}J^{\top}X^{\top}J\,,
\]
which shows that~\eqref{Eq:SIB} is $(R^{-1}AR, R^{-1}B, CR, D)$.
We conclude that $R=I$, so that $-I=JX^{-\top}J^{\top}X=J^{-\top}X^{-\top}JX$, and so
$(JX)^{\top}=-JX$ is skew-symmetric.
Paralleling the argument for skew-Hamiltonian transfer functions, we will use this to
obtain a skew-symmetric realization.
The key ingredient is a factorization of the matrix $JX$.
An invertible complex skew-symmetric matrix $Z$ admits a Tagaki-like factorization $Z=Y^{\top}JY$.
Indeed, $Z = USU^{\top}$ for a unitary matrix $U$ and a block diagonal $S$ which is a direct sum of
$2\times 2$ blocks of the form
$\left[\begin{smallmatrix}0&a\;\\-a&0\end{smallmatrix}\right]$ with
$a \in \mathbb C \setminus\{0\}$ (e.g., see \cite[Problem 26 in Chapter 4.4]{HJ1}).
Thus, after block scaling with blocks
$\left[\begin{smallmatrix}\sqrt{a}&0\\0&\sqrt{a}\end{smallmatrix}\right]$ and
applying a permutation similarity, we arrive at the claimed factorization.
In the factorization $JX=Y^{\top}JY$ of the invertible skew-symmetric matrix $JX$,
the matrix $Y$ is also invertible.
Then the realization $(YAY^{-1}, YB, CY^{-1}, D)$ of the
transfer function $G(s)$ is skew-symmetric.
Indeed, as $D=-D^{\top}$ and we have $J=-J^{\top}=J^{-\top}$, $X^{\top}J^{\top}=-JX=-Y^{\top}JY$, and
$J^{-\top}X^{-\top}=Y^{-1}JY^{-\top}$, we obtain
\begin{eqnarray*}
(YAY^{-1}J)^{\top}& =& J^{\top}Y^{-\top}A^{\top}Y^{\top}\ =\
-J^{\top}Y^{-\top} (X^{-1}(JAJ)^{\top}X)^{\top} Y^{\top}\\
& =& -J^{\top}Y^{-\top}X^{\top}JAJX^{-\top} Y^{\top} \ =\
J^{\top}Y^{-\top}X^{\top}J^{\top}AJ^{-\top}X^{-\top} Y^{\top}\\
& =& -J^{\top}Y^{-\top}Y^{\top}JY AY^{-1}JY^{-\top} Y^{\top}\ =\ -Y AY^{-1}J\,.
\end{eqnarray*}
Similarly, $CY^{-1}=(JB)^{\top}XY^{-1}=(J YB)^{\top}$,
so the realization $(YAY^{-1}, YB, CY^{-1}, D)$ of $G(s)$ is skew-symmetric.
\smallskip
Finally, suppose that $(A,B,C,D)$ and $(\alpha,\beta,\gamma,\delta)$ are minimal
skew-symmetric realizations of the same transfer function.
Let $X\in GL(n)$ be the unique matrix such that
\[
X.(A,B,C,D)\ =\ (X^{-1}AX,X^{-1}B,CX,D)\ =\
(\alpha,\beta,\gamma,\delta)\,.
\]
Recall that $-A^{\top}=JAJ$ and $\alpha =-J\alpha^{\top}J$, and so
\[
\alpha \ =\ -J\alpha^{\top}J\ =\ -JX^{\top}A^{\top}X^{-\top}J
\ =\ JX^{\top}JAJX^{-\top}J \ =\ J^{-1}X^{\top}JAJ^{-1}X^{-\top}J\,.
\]
Similarly recall that $C^{\top}=JB$ and $\beta=J\gamma^{\top}$, so that
\[
\beta\ =\ J^{\top}\gamma^{\top}\ =\ J^{\top}(CX)^{\top}\ =\ J^{\top}X^{\top}C^{\top}\ =\ -JX^{\top}JB\ =\ J^{-1}X^{\top}JB\,.
\]
Since we also have $\gamma=CJ^{-1}X^{-\top}J$, we see that
$(\alpha,\beta,\gamma,\delta)=R.(A,B,C,D)$, where $R=J^{-1}X^{-\top}J$.
By the uniqueness of $X$, we have $X= J^{-1}X^{-\top}J$ so that
$X^{\top} JX=J$ and so $X\in Sp(n)$ is symplectic.
\end{proof}
We use this proposition to compute the dimensions of the corresponding spaces of transfer
functions/rational curves, the first part of the proof of Theorem~\ref{Th:one}.
\begin{cor}\label{C:dim}
The set of transfer functions with a fixed symmetry is an irreducible
quasiprojective complex algebraic manifold.
For symmetric, Hamiltonian, skew-Hamilton\-ian, and skew-symmetric transfer functions, these have respective
dimensions
\[
(m{+}1)n\ +\ \tbinom{m{+}1}{2}\,,\quad
mn\ +\ \tbinom{m{+}1}{2}\,,\quad
mn\ +\ \tbinom{m}{2}\,,\quad\mbox{and}\quad
(m{-}1)n\ +\ \tbinom{m}{2}\,.
\]
\end{cor}
\begin{proof}
The space of minimal complex symmetric realizations is a Zariski-open subset of an affine space.
By Proposition~\ref{prop:realizations}, the set of
transfer functions with a fixed symmetry type is identified with the set of orbits of
an algebraic group ($O(n)$ or $Sp(n)$) acting freely on the space of minimal symmetric
realizations, which is an open subset of a vector space.
The first statement of the corollary follows from this as the set of such orbits has a
natural structure as an irreducible smooth complex algebraic variety~\cite[Th.~9.16]{Lee}.
For the second, we note that the dimension of the orbit space is the
difference of the dimensions of the space of symmetric realizations and of
the group.
The orthogonal group $O(n)$ has dimension $\binom{n}{2}$ and the
symplectic group $Sp(n)$ has dimension $\binom{n{+}1}{2}$~\cite{FuH}.
Spaces of symmetric and Hamiltonian realizations both have the same dimension
\[
\binom{n{+}1}{2}\ +\ nm\ +\ \binom{m{+}1}{2}\,,
\]
while the spaces of skew-Hamiltonian and skew-symmetric realizations both have dimension
\[
\binom{n}{2}\ +\ nm\ +\ \binom{m}{2}\,.
\]
The corollary now follows.
\end{proof}
We now complete the proof of Theorem~\ref{Th:one} by identifying the Hermann-Martin
curves of symmetric and skew-symmetric transfer functions with dense open subsets of the
appropriate spaces of rational curves in the Lagrangian and Orthogonal Grassmannians.
First note that if $G(s)$ is symmetric (respectively skew-symmetric) then for $s\in\P^1$, the row space $K(s)$ of the matrix
$[I_m\co G(s)]$ lies in $LG(m)$ (respectively in $OG(m)$).
Thus the Hermann-Martin curve is a curve in $LG(m)$ (respectively in $OG(m)$).
To finish, we show that that these Hermann-Martin curves are
dense in the corresponding spaces of rational curves, which is a consequence of their
having the same dimension.
The space of rational curves in $LG(m)$ of degree $d$ has dimension
$d(m{+}1)+\binom{m{+}1}{2}$ and the space of rational curves in $OG(m)$ of degree $d$
has dimension $2d(m{-}1)+\binom{m}{2}$~\cite{KT03,KT04}.
Thus Theorem~\ref{Th:one} follows if we knew that a curve in $LG(m)$ of McMillan degree
$n$ has degree $n$ in $LG(m)$ and a curve in $OG(m)$ of McMillan degree $n=2\ell$ has
degree $\ell$ in $OG(m)$.
These facts are well-known.
The McMillan degree of a curve is it degree in the classical
Grassmannian $G(m,2m)$ in its Pl\"ucker embedding.
The inclusion $LG(m)\hookrightarrow G(m,2m)$ arises from a linear map on the standard
projective embedding of $LG(m)$, so rational curves of degree $n$ in $LG(m)$ have degree
$n$ in $G(m,2m)$, and hence McMillan degree $n$.
On the other hand, the inclusion $OG(m)\hookrightarrow G(m,2m)$ arises from
the second Veronese map on the natural projective embedding of $OG(m)$.
Thus a rational curve of degree $\ell$ in $OG(m)$ will have degree $2\ell$ in $G(m,2m)$,
and hence McMillan degree $2\ell$.
This completes the proof of Theorem~\ref{Th:one}.
\section{Static skew-symmetric state feedback control}\label{S:two}
Suppose that we have a strictly proper ($D = 0$) skew-Hamiltonian or skew-symmetric linear system
with a minimal state-space realization
\begin{equation}\label{Eq:system}
\dot{x}\ =\ Ax+Bu\qquad\quad y\ =\ Cx\,.
\end{equation}
If we introduce a static linear state-space feedback law $u=Fy+v$, then the new system
\begin{equation}\label{Eq:feedback}
\dot{x}\ =\ (A+BFC)x+Bv\qquad\quad y\ =\ Cx\,
\end{equation}
has the same symmetry as the original system when $F$ is skew-symmetric.
(The elementary calculation is given below.)
We investigate the control of such linear systems with complex skew-symmetric static
state-space feedback.
We first establish the necessary and sufficient conditions for generic pole placement of
Theorem~\ref{Th:two}, relate skew-symmetric feedback control to the Schubert calculus
on the orthogonal Grassmannian, and then prove Theorem~\ref{Th:three}, counting the number
of feedback laws that place a generic set of poles of a generic skew-symmetric transfer
function with McMillan degree $2\binom{m}{2}$.
We do not yet know how to count the controllers of a skew-Hamiltonian transfer function of
McMillan degree $n$ when $\binom{m}{2}=\lfloor n/2\rfloor$.
\begin{proof}[Proof of Theorem~$\ref{Th:two}$]
We give the proof for generic pole-assignability of skew-symmetric transfer
functions and indicate how the argument changes for skew-Hamiltonian transfer functions.
This follows and simplifies the arguments in~\cite{HRW06}.
We identify skew-symmetric $N\times N$ matrices with the vector space $\wedge^2\C^N$,
where the elementary decomposable tensor $e_i\wedge e_j$ ($i \neq j$) corresponds to the matrix having
1 in position $(i,j)$, $-1$ in position $(j,i)$, and 0 in other positions.
Suppose that~\eqref{Eq:system} is skew-symmetric, so that $A$ is skew-Hamiltonian,
$(AJ)^{\top}=-AJ$, and $C=B^{\top}J$.
If we have a feedback law $u=Fy+v$ where $F$ is skew-symmetric, then
the new system~\eqref{Eq:feedback} remains skew-symmetric as
$A+BFB^{\top}J$ is skew-Hamiltonian.
Thus, the characteristic polynomial
\begin{equation}\label{Eq:char_ss}
\varphi(s)\ :=\ \det(sI - (A + BFB^{\top}J))\ =\ \det(sJ - (AJ - BFB^{\top})),
\end{equation}
is the determinant of a skew-symmetric matrix and is therefore a square (its determinant is the square of its Pfaffian).
Thus it is natural to ask for skew-symmetric feedback laws $F$ which place these
$\ell=n/2$ roots (which are poles of the transfer function).
The pole placement map sends a
skew-symmetric matrix $F\in\wedge^2\C^m$ to the degree $2\ell$ polynomial
$\varphi(s)$.
Since this monic polynomial is a square, its last $\ell$ coefficients (those of
$s^{2\ell-1},\dotsc,s^{\ell}$) determine its first $\ell$ coefficients.
These coefficients are, up to a sign, the elementary symmetric functions of the
eigenvalues of $A+BFB^{\top}J$.
By the Newton identities, these coefficients determine, and are determined by, the Newton
power sums which are the traces of
$(A+BFB^{\top}J)^k$ for $k=1,\dotsc,\ell$.
To show generic pole-assignability, we only need to exhibit one choice of matrices $A,B$ for
which the map
\[
\DeCo{\Psi}\ \colon\
\wedge^2 \C^m\ni F\ \longmapsto\
(\tr((A+BFB^{\top}J)^k)\mid k=1,\dotsc,\ell)\ \in\ \C^\ell
\]
is dominant.
We do this by showing that the differential $d\Psi_0$ at $0\in\wedge^2\C^m$ is
surjective.
Let $\alpha_1,\dotsc,\alpha_\ell$ be distinct numbers and $\beta_1,\dotsc,\beta_m$ be
numbers such that the $\binom{m}{2}$ products $\beta_i\beta_j$ for $i<j$ are distinct, and
such that $\beta_i^\ell\neq\beta_j^\ell$, for every $i\neq j$.
Let $D=\diag(\alpha_1,\dotsc,\alpha_\ell)$ be the diagonal matrix with entries
$\alpha_1,\dotsc,\alpha_\ell$, and let $A$ be the block diagonal matrix
$\left[\begin{smallmatrix}D&0\\0&D\end{smallmatrix}\right]$.
Finally, let $B$ be the matrix with entries $\beta_j^{i-1}$ for $i=1,\dotsc,2\ell$ and
$j=1,\dotsc,m$.
For this choice of $A,B$, the differential $d\Psi$ is surjective at
$0\in\wedge^2\C^m$, for this implies that the image contains an open set in the
classical topology, and thus a Zariski-open subset.
To see surjectivity, note that the differential at $0$ is the linear map
\begin{equation}\label{Eq:dPsi0}
d\Psi_0\ \colon F\ \longmapsto\
(k\cdot \tr(A^{k-1}\cdot BFB^{\top}J)\mid k=1,\dotsc,\ell)\,.
\end{equation}
Consider this map on the basis element $e_i\wedge e_j$ of $\wedge^2\C^m$.
A direct calculation shows that $B(e_i\wedge e_j)B^{\top}$ is the vector $b_i\wedge b_j$,
where $b_1,\dotsc,b_m$ are the columns of $B$.
For our choice of $B$, the $(p,q)$-entry of $b_i\wedge b_j$ is
\[
\det\left[\begin{matrix}\beta_i^{p-1}&\beta_j^{p-1}\\
\beta_i^{q-1}&\beta_j^{q-1}\end{matrix}\right]
\ =\ (\beta_i\beta_j)^{p-1}(\beta_j^{q-p}-\beta_i^{q-p})\,.
\]
It follows that the map $d\Psi_0$ sends the vector $e_i\wedge e_j$ to the vector
\[
\Bigl(k\sum_{p=1}^{2\ell} (A^{k-1}(b_i\wedge b_j) J)_{p,p}\;\Big\vert\;
k=1,\dotsc,\ell\Bigr)\,.
\]
Since
$A^{k-1}=\diag(\alpha_1^{k-1},\dotsc,\alpha_\ell^{k-1},\alpha_1^{k-1},\dotsc,\alpha_\ell^{k-1})$,
\[
\bigl((b_i\wedge b_j)J\bigr)_{p,p}\ =\ \left\{
\begin{array}{rcl} -(b_i\wedge b_j)_{p,p+\ell}&\ &\mbox{if }p\leq\ell\\
(b_i\wedge b_j)_{p,p-\ell}&\ &\mbox{if }p>\ell
\end{array}\right.\ ,
\]
and $(b_i\wedge b_j)$ is skew-symmetric, this vector is
\[
\Bigl(2k\sum_{p=1}^{\ell}
\alpha_p^{k-1}(\beta_i\beta_j)^{p-1}(\beta_i^{\ell}-\beta_j^{\ell})\;\Big\vert\;
k=1,\dotsc,\ell\Bigr)\,.
\]
Thus, $d\Psi_0$ is represented by the $\ell\times\binom{m}{2}$ matrix which is the
product of the matrices
\[
\diag(2,4,\dotsc,2\ell)\cdot
\bigl(\alpha_p^{k-1}\bigr)_{p=1,\dotsc,\ell}^{k=1,\dotsc,\ell} \cdot
\bigl( (\beta_i\beta_j)^{p-1}\bigr)_{1\leq i<j\leq m}^{p=1,\dotsc,\ell}
\cdot \diag( \beta_i^\ell-\beta_j^\ell | 1\leq i<j\leq m)\,,
\]
and so its rank is the minimum of $\ell$ and $\binom{m}{2}$, which proves Theorem~\ref{Th:two}
for skew-symmetric linear systems.
\medskip
Suppose now that~\eqref{Eq:system} is skew-Hamiltonian, so that $A$ is skew-symmetric
and $C=B^{\top}$.
Under a skew-symmetric feedback law $u=Fy+v$,
the new system~\eqref{Eq:feedback} remains skew-Hamiltonian
as $A+BFB^{\top}$ is skew-symmetric.
The characteristic polynomial
\begin{equation}\label{Eq:char_sH}
\varphi(s)\ :=\ \det(sI - (A + BFB^{\top}))
\end{equation}
satisfies $\varphi(s)=(-1)^n\varphi(-s)$, so its nonzero roots $\lambda$ occur in pairs
$\pm\lambda$.
Thus it is natural to ask for skew-symmetric feedback laws $F$ which place these
$\ell=\lfloor n/2\rfloor$ pairs of roots (which are poles of the transfer function).
The pole placement map sends a
skew-symmetric matrix $F\in\wedge^2\C^m$ to the degree $n$ polynomial
$\varphi(s)$.
As before, we investigate the surjectivity of the pole-placement map by considering the
map associating Newton power sums, which is
\[
\Psi\ \colon\ \wedge^2\C^m\ni F\ \longmapsto
\ (\tr(A+BFB^{\top})^k\mid k=1,\dotsc,n)\ \in \C^\ell\,.
\]
Since $A+BFB^{\top}J$ is skew-symmetric, the trace is zero if $k$ is odd, which is why the
codomain of this map is $\C^\ell$.
Thus, the differential of $\Psi$ at $0$ is
\[
d\Psi_0\ \colon\ F\ \longmapsto \ (2k\cdot\tr(A^{2k-1}BFB^{\top})\mid k=1,\dotsc,\ell)\,.
\]
Let $D$ be the same diagonal matrix as before.
If $n$ is even, let $A$ be the block matrix
$\left[\begin{smallmatrix}0&D\\-D&0\end{smallmatrix}\right]$ and $B$ be the same as
before.
If $n$ is odd, then add a new first row and column of $0$s to $A$ and extend $B$ with the row
$(\beta_i^{2\ell} \, | \, i=1,\dotsc,m)$.
Then nearly the same calculation as before shows that $d\Psi_0$ is surjective when
$\ell\leq\binom{m}{2}$, which completes the proof of Theorem~\ref{Th:two}.
\end{proof}
Before proving Theorem~\ref{Th:three}, we first recall some standard matrix manipulations which
transform the problem of finding matrices $F$ which place the poles of the transfer
function~\eqref{Eq:feedback} into a geometric problem on a Grassmannian. We then make some
definitions.
These poles are the roots of the characteristic polynomial $\varphi(s)=\det(sI_n-(A+BFC))$
of the matrix in~\eqref{Eq:feedback}.
The rational function $\varphi(s)/\det(sI_n-A)$ equals the determinant of the product
\[
\left[\begin{matrix}
(sI_n-A)^{-1}&0&0\\-C(sI_n-A)^{-1}&I_m&0\\0&0&I_m\end{matrix}\right]
\left[\begin{matrix}sI_n-A-BFC&BF&-B\\0&I_m&0\\0&0&I_m\end{matrix}\right]
\left[\begin{matrix}I_n&0&0\\C&I_m&0\\0&F&I_m\end{matrix}\right]
\]
which is
\[
\left[\begin{matrix}
(sI_n-A)^{-1}&0&0\\-C(sI_n-A)^{-1}&I_m&0\\0&0&I_m\end{matrix}\right]
\left[\begin{matrix}sI_n-A&0&-B\\C&I_m&0\\0&F&I_m\end{matrix}\right]
\ =\
\left[\begin{matrix}I_n&0&-(sI_n-A)^{-1}B\\0&I_m&C(sI_n-A)^{-1}B\\0&F&I_m\end{matrix}\right].
\]
The transfer function $G(s)=C(sI_n-A)^{-1}B$ admits a left coprime factorization into
matrices of polynomials $D(s)^{-1}N(s)$, where $\det D(s)=\det(sI_n-A)$,
and so we have
\begin{equation}\label{Eq:geometry}
\varphi(s)\ =\ \det(sI_n-A)\det\left[\begin{matrix}I_m&G(s)\\F&I_m\end{matrix}\right]
\ =\ \det \left[\begin{matrix} D(s)&N(s)\\F&I_m\end{matrix}\right].
\end{equation}
Thus, $s$ is a pole of the transfer function of~\eqref{Eq:feedback}
if and only if the matrix on the right of~\eqref{Eq:geometry} does not have full rank.
Geometrically, if $K(s)$ is the row space of $[D(s)\co N(s)]$ and $H$ is the row space
of $[F\co I_m]$, which are both $m$-planes in $\C^{2m}$, then $K(s)\cap H\neq\{0\}$.
It follows that the feedback laws $F$ which place a given set of poles $s_1,\dotsc,s_n$
correspond to those $H\in OG(m)$ such that
\begin{equation}\label{Eq:SchubertVariety}
K(s_i)\cap H\ \neq\ \{0\}\,, \ \ \text{for each $i=1,\dotsc,n$.}
\end{equation}
When $G(s)$ is skew-symmetric, then $n=2\ell$ and $\varphi(s)$ has only $\ell$ distinct
roots, say $s_1,\dotsc,s_\ell$.
In this case, we also have that $K(s_i)$ is isotropic and the set of $H\in OG(m)$
satisfying~\eqref{Eq:SchubertVariety} defines a \demph{Schubert subvariety} of $OG(m)$, which
represents the first Chern class $c_1$ of the tautological bundle.
A consequence of the surjectivity of the map $d\Psi_0$~\eqref{Eq:dPsi0} for generic $A,B$ is
that these Schubert varieties meet generically transversally (in the open set
consisting of $H$ of the form $[F\co I_m]$).
Thus, when $\ell=\binom{m}{2}$ there are finitely many $H$
satisfying~\eqref{Eq:SchubertVariety} for $i=1,\dotsc,\ell$, and their count is bounded
above by the intersection number $\deg(c_1^\ell)$, which may be computed
using the Schubert calculus on $OG(m)$~\cite{FuPr}. It is equal to
\[
\DeCo{d_m}\ :=\
\binom{m}{2}! \frac{1!\dotsb(m-2)!}{1! 3!\dotsb(2m-3)!}\,,
\]
which is also the degree of $OG(m)$ in its natural embedding as the spinor variety.
We complete the proof of Theorem~\ref{Th:three} by exhibiting a specific skew-symmetric
transfer function $G(s)$ of McMillan degree $2\binom{m}{2}$ such that there are
exactly $d_m$ skew-symmetric feedback laws placing any given $\binom{m}{2}$
real poles, and all the feedback laws are real.
This also shows that generic systems have $d_m$ feedback laws.
The argument uses a result of Purbhoo~\cite{Purbhoo} concerning the reality of the Wronski
map, which we will transfer into the language of systems theory.
Suppose that $\C^{2m}$ has ordered basis $\be_1,\dotsc,\be_{2m}$.
Let \demph{$\gamma(s)$} be the vector-valued function $\gamma\colon\C\to\C^{2m}$ with
components
\[
\bigl(\,1\,,\, s\,,\,
\frac{s^2}{2!}\,,\, \dotsc\,,\, \frac{s^{m-1}}{(m{-}1)!}\frac{1}{\sqrt{2}}
\ ,\ \frac{(-s)^{2m-2}}{(2m-2)!}\,,\,\dotsc,
\frac{(-s)^m}{m!}\,,\, \frac{(-s)^{m-1}}{(m{-}1)!}\frac{1}{\sqrt{2}}\,\bigr)\,.
\]
If we set $\DeCo{v_i(s)}:=\left(\frac{d}{ds}\right)^{i-1} \gamma(s)$ for $i=1,\dotsc,m{-}1$
and $v_m(s):=\left(\frac{d}{ds}\right)^{m-1} \gamma(s)+
\frac{1}{\sqrt{2}}(e_m+(-1)^{m-1}e_{2m})$,
then the row span of $v_1(s),\dotsc,v_m(s)$ is isotropic.
While this defines a curve in $OG(m)$, it does not come from a skew-symmetric linear
system, as it does not correspond to a strictly proper transfer function.
However, the row span \demph{$K(s)$} of the vectors
$s^{2m-2}v_1(s^{-1}),\dotsc,s^{m-1}v_m(s^{-1})$, is still isotropic and it comes from a
strictly proper skew-symmetric transfer function.
We display this for $m=5$, giving a $5\times 10$ matrix with rows
$s^{2m-2}v_1(s^{-1}),\dotsc,s^{m-1}v_m(s^{-1})$:
\[
\left[\begin{matrix}
s^8&s^7&\frac{s^6}{2}&\frac{s^5}{3!}&\frac{1}{\sqrt{2}}\frac{s^4}{4!}&
\rule{3pt}{0pt}\frac{1}{8!}&
-\frac{s}{7!}&\frac{s^2}{6!}&-\frac{s^3}{5!}&\frac{1}{\sqrt{2}}\frac{s^4}{4!}\\
0&s^7&s^6&\frac{s^5}{2}&\frac{1}{\sqrt{2}}\frac{s^4}{3!}&
\rule{0pt}{17pt}\rule{3pt}{0pt} \frac{1}{7!}&
-\frac{s}{6!}&\frac{s^2}{5!}&-\frac{s^3}{4!}&\frac{1}{\sqrt{2}}\frac{s^4}{3!}\\
0&0&s^6&s^5&\frac{1}{\sqrt{2}}\frac{s^4}{2}&
\rule{0pt}{17pt}\rule{3pt}{0pt} \frac{1}{6!}&
-\frac{s}{5!}&\frac{s^2}{4!}&-\frac{s^3}{3!}&\frac{1}{\sqrt{2}}\frac{s^4}{2}\\
0&0&0&s^5&\frac{1}{\sqrt{2}}s^4&
\rule{0pt}{17pt}\rule{3pt}{0pt} \frac{1}{5!}&
-\frac{s}{4!}&\frac{s^2}{3!}&-\frac{s^3}{2!}&\frac{1}{\sqrt{2}}s^4\\
0&0&0&0&\sqrt{2}s^4&\rule{0pt}{17pt}\rule{3pt}{0pt} \frac{1}{4!}&
-\frac{s}{3!}&\frac{s^2}{2!}&-s^3&0
\end{matrix}\right].
\]
If we write $K(s)=[D(s)\co N(s)]$, then $D(s)$ is an upper triangular matrix with diagonal
\mbox{$(s^{2m-2},\dotsc,s^m,\sqrt{2}s^{m-1})$,} and hence is
invertible for all $s\neq 0$. Set $\DeCo{G(s)}:=D(s)^{-1}N(s)$, which is strictly proper and real.
Here is $G(s)$ when $m=5$:
\[
\left[\begin{matrix}
0& \frac{5}{2}\frac{1}{7!s^7}&-\frac{5}{2}\frac{1}{6!s^6}&
\frac{3}{2}\frac{1}{5!s^5}&-\frac{1}{\sqrt{2}}\frac{1}{4!s^4}\\
-\frac{5}{2}\frac{1}{7!s^7}&0\rule{0pt}{17pt}& \frac{1}{5!s^5}&
-\frac{1}{4!s^4}& \frac{1}{\sqrt{2}}\frac{1}{3!s^3}\\
\frac{5}{2}\frac{1}{6!s^6}& -\frac{1}{5!s^5}&0\rule{0pt}{17pt}&
\frac{1}{2}\frac{1}{3!s^3}&-\frac{1}{\sqrt{2}}\frac{1}{2!s^2}\\
-\frac{3}{2}\frac{1}{5!s^5}&\frac{1}{4!s^4}&
-\frac{1}{2}\frac{1}{3!s^3}&0\rule{0pt}{17pt}&\frac{1}{\sqrt{2}}\frac{1}{s}\\
\frac{1}{\sqrt{2}}\frac{1}{4!s^4}& -\frac{1}{\sqrt{2}}\frac{1}{3!s^3}&
\frac{1}{\sqrt{2}}\frac{1}{2!s^2}& -\frac{1}{\sqrt{2}}\frac{1}{s}&0\rule{0pt}{17pt}
\end{matrix}\right].
\]
Theorem~\ref{Th:three} follows from the following facts about the transfer function $G(s)$.
\begin{proposition}
The transfer function $G(s)$ is skew-symmetric with McMillan degree $2\binom{m}{2}$.
Any set of $\binom{m}{2}$ distinct real poles is placed by
exactly $d_m$ skew-symmetric feedback laws, with each one real.
\end{proposition}
\begin{proof}
Since the isotropic $m$-plane $K(s)$ is the the row space of $[I_m\co G(s)]$, we conclude
that $G(s)$ is skew-symmetric.
Let $V\simeq \C^{2m-1}\subset\C^{2m}$ be the subspace with ordered basis
\[
(\be_1,\dotsc,\be_{m-1},
(\be_m+(-1)^{m-1}\be_{2m})/\sqrt{2},\be_{m+1},\dotsc,\be_{2m-1})\,.
\]
The nondegenerate symmetric bilinear form on $\C^{2m}$
restricts to a nondegenerate symmetric bilinear form
on $V$, and the map $H\mapsto W:=H\cap V$ sends a maximal isotropic subspace $H$ of $\C^{2m}$
to a maximal isotropic subspace of $V$, inducing an isomorphism between $OG(m)$ and
the space $BOG(m{-}1)$ of maximal isotropic subspaces of $V\simeq\C^{2m-1}$.
The reason for this is that for each $W\in BOG(m{-}1)$ there are two maximal isotropic subspaces
$H$ of $\C^{2m}$ containing $W$, exactly one of which lies in $OG(m)$.
When $W$ is real, both isotropic subspaces $H$ containing $W$ are also real.
Also, $\gamma(s)$ is a rational normal curve in $V$, as it involves the monomials
$1,\dotsc,s^{2m-1}$.
Furthermore, $\DeCo{L(s)}:=K(s)\cap V$ is the $(m{-}1)$-plane osculating
$\gamma(s^{-1})$, and $L(s)$ is isotropic.
The problem of which isotropic subspaces $W$ of $V$ that meet $r = \binom{m}{2}$
osculating planes $L(s_1),\dotsc,L(s_{r})$ was studied by Purbhoo~\cite{Purbhoo} in the
context of the Wronski map from $BOG(m{-}1)\simeq OG(m)$, which extends the pole placement map
from $[F\co I_m]\mapsto\varphi(s)$ for the transfer function $G(s)$.
This map is surjective onto the space of polynomials of degree $2r$ which are
squares of polynomials, and it has finite fibers of algebraic degree $d_m$.
This implies that there are at most $d_m$ feedback laws placing a given set of
$r$ poles.
It also implies that any given isotropic plane $H$ meets at most $r$
subspaces of the form $L(s)$, including $L(\infty)=[0\co I_{m-1}]$.
Purbhoo~\cite[Theorem 3]{Purbhoo} showed that if
$s_1,\dotsc,s_{r}$ were real, then there are
exactly $d_m$ real isotropic planes $W$ in $BOG(m-1)$ such that
$L(s_i)\cap W\neq\{0\}$, for each $i=1,\dotsc,r$.
For each such $W$, let $H$ be the unique isotropic plane in $OG(m)$ containing $W$, which
is necessarily real.
Then $H\cap K(s_i)\neq 0$ for each $i$, and so $H$ corresponds to a real feedback law
if $H$ has the form $[I_m\co F]$.
But this is guaranteed for otherwise $H\cap K(\infty)\neq 0$, which would imply
$W\cap L(\infty)\neq\{0\}$, an impossibility as $H$ already meets the maximum number of
subspaces of the form $L(s)$.
Lastly, the transfer function has McMillan degree $2r$ since the image of the pole
placement map (a linear projection) meets the set of polynomials of this degree.
\end{proof}
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
|
\section{Motivation}
For the past 40 years, N--body simulations have allowed to numerically
study the evolution of the distribution of matter in the expanding
Universe \citep[cf.][]{Peebles:1971a, Bertschinger:1998,
Springel:2005a}. A significant number of simulation codes have been
developed for this purpose \citep[e.g.][to name just a
few]{Efstathiou:1985, Couchman:1991, Bryan:1997, Stadel:2001,
Springel:2001b, Teyssier:2002, Wadsley:2004}. All such approaches to
structure formation model the collisionless fluid of dark matter by a
set of massive particles (typically of equal mass) and differ in how
the gravitational forces are calculated at the positions of the
particles. The forces are applied to update the velocities which in
turn are used to update the positions. The system is then evolved
forward in time. From such simulations much has been learned about the
formation and evolution of cosmological structures and they have
become a standard tool in physical cosmology. While three dimensional
calculations have difficulty in sampling the six dimensional
phase-space well \citep[see e.g.][]{Buchert:1991} they have found a
very large range of applications and have driven much of the progress
that has been made in the past decades of understanding structure
formation. In quite a range of these applications the space density of
the dark matter fluid is required and in many others the phase-space
density is of great importance.
Some open questions that require detailed information about the dark
matter density and its velocity distribution are related to dark
matter detection. For the indirect detection techniques, the
predictions of the dark matter annihilation luminosity depend
sensitively on the density of the dark matter streams and the
distribution and the relative velocities of the particles. Assuming
well mixed phase-space and assuming a shape of the velocity
distribution function, this annihilation rate would scale with the
square of the space density, $\rho$. In current work, these estimates
are typically carried out by fitting spherical profiles to the main
dark matter halo and its subhaloes and then assign annihilation
luminosities by scaling the square of the smoothed halo and subhalo
profiles appropriately. This smoothes the dark matter fluid
sufficiently to avoid noisy estimates of the annihilation signal
\citep{2008Natur.454..735D, 2008Natur.456...73S}. However, in general
the annihilation rates depend on the relative velocity of the dark
matter particles interacting. Here one can distinguish between dark
matter annihilation within individual streams of dark matter as well
the contribution from stream-stream interactions which differ strongly
in the relative velocities \citep[see e.g.][]{Hogan:2001,
Afshordi:2009}. This fine grained dark matter phase-space structure
is equally important for considerations of dark matter direct
detection experiments where one applies velocity cuts in the
experimental analysis in order to reject certain backgrounds. The
annular modulation of the relative velocity and the main dark matter
streams in the solar neighborhood provided by the earths motion around
the sun allows one to potentially map the fine grained phase-space
structure should the experiments be able to detect dark matter.
The best current approaches to probe phase-space structures were
surveyed by \cite{2009MNRAS.393..703M}. These typically start with some
tessellation of phase-space such as a Delaunay triangulation or a
Voronoi tesselation \citep{1996MNRAS.279..693B}, or cartesian trees
\citep{2006MNRAS.373.1293S, 2010CoPhC.181.1438A}. The mass of the
particles found within the cells give the local densities. When only
the space density is required, adaptive kernel smoothing is often
employed. In fact, most images of dark matter simulations shown are
projections of kernel smoothed particle distributions. Configuration
space density estimators also play a particularly important role when
studying the topology and character of the cosmic web
\citep[e.g][]{Schaap:2000,Pelupessy:2003, Aragon-Calvo:2007,
Colberg:2008,Neyrinck:2008}. However, in all cases, control volumes
are defined such that they contain sufficient numbers of particles to
reduce sampling noise. Unfortunately, this will average over large
regions of configuration and phase-space and consequently effectively
degrade the spatial resolution of the calculation.
In any case, there is ample motivation to study the distribution,
evolution and current state of dark matter in the Universe further by
observations as well as simulations. In this contribution, we introduce
a novel way to analyze $N$-body simulations. Our approach naturally
arises from considering how the collisionless fluid of dark matter is
expected to evolve in phase-space.
As is well known, \citep[e.g.][]{Shandarin:1989}, at early enough times,
i.e. before shell crossing, the motion of the dark matter fluid is well
described by the Zel'dovich approximation \citep{ZelDovich:1970} as a
potential flow
\begin{eqnarray}
\mathbf{x}_t & = & \mathbf{q} + g_t\,\boldsymbol{\nabla}\phi(\mathbf{q}),\\
\mathbf{v}_t & = &\dot{g}_t\,\boldsymbol{\nabla}\phi(\mathbf{q}).
\end{eqnarray}
Here $\phi$ is a potential field that is proportional to the initial gravitational
potential field of perturbations, $\mathbf{q}$ are the initial particle positions and $g(t)$ is
the growth factor of linear perturbations. At early times, i.e. for $t\to 0$,
also $g\to 0$. These particles occupy a three dimensional submanifold $S$
of the entire six dimensional phase-space with the time-dependent mapping:
\begin{equation}
\mathbf{q} \mapsto \left(\mathbf{q}+ g_t\,\boldsymbol{\nabla}\phi(\mathbf{q}), \dot{g}_t\,\boldsymbol{\nabla}\phi(\mathbf{q})\right),
\end{equation}
where for $t\to0$
\begin{equation}
\mathbf{q} \mapsto \left(\mathbf{q}, \dot{g}_0\,\boldsymbol{\nabla}\phi(\mathbf{q})\right),
\end{equation}
the three-dimensional structure ($\mathbf{q}\in\mathbb{R}^3$) can be easily seen.
The map between $\mathbf{x}_t$ and $\mathbf{q}$ is
bijective until shell crossing occurs, i.e when more than one stream
of dark matter exists at one spatial location. We will refer to this
three dimensional submanifold as the dark matter sheet (even when
discussing it in one or two spatial dimensions). The volume of the
sheet continues to grow as structure forms and
evolves \citep{Shandarin:1989,Vogelsberger:2008}.
At the same time, current $N$--body simulations of structure formation
do already follow individual dark matter particles through phase-space
\citep[e.g.][]{Bertschinger:1998}. The $N$-body technique thus
corresponds to sampling the sheet at a finite number of points
$\mathbf{q}$ with the entire mass concentrated at their positions.
\cite{Vogelsberger:2008}, \cite{White:2009} and \cite{Vogelsberger:2011} developed a
powerful approach to augment cosmological simulations to record more
knowledge about the evolution of this dark matter sheet. They derive
an equation of motion for the distortion tensor around every particle
to linear order and then evolve it with every particle during the
simulation. They refer to this as the geodesic deviation equation
(GDE). This gives access to information about the evolution
of the stream density along every particle trajectory and allows to
track the number of caustics an infinitesimal fluid element surrounding
a dark matter particle will experience. This technique goes a long way
in obtaining more information about the fine grained phase-space
structure in dark matter haloes \citep{Vogelsberger:2011}.
More restricted calculations in this context have been carried out in
fixed potentials \citep{Stiff:2001} or one dimensions
for Newtonian gravity \citep{Alard:2005} and General Relativity
\citep{Rasio:1989}. Also in the context of stellar dynamics there is a
large body of literature which explores details of the
phase-space structure of stellar system. From a numerical point of
view the work of \cite{Cuperman:1971a} is particularly
remarkable. Some 40 years ago these authors realized that in one
dimension one can follow the phase-space boundary of a collisionless
fluid and they give a beautiful implementation and calculations
treating the phase-space fluid as a continuum. Studying the connection
of their formalism to the N--body technique is revealing and what
follows here is in some ways the extension to three dimensions with
the exception of our approach to velocity dimensions and the way the
Poisson equation is solved.
We suggest that the three-dimensional manifold can be decomposed by a
space-filling grid that connects a finite number of vertices
$\mathbf{q}$. The simplest version is to decompose the volume into
three dimensional simplices, i.e. tetrahedra, which have the nice
topological property of being either convex or degenerate. For any
choice of a regular lattice of vertices $\mathbf{q}$, such a
tetrahedral decomposition can be achieved by a Delaunay triangulation.
In contrast to the particle discretization, we can now think of the
dark matter mass being spread out over the corresponding volume
elements.
This mesh traces the dark matter sheet as it subsequently evolves in
phase-space. The motion of the mesh vertices are evolved using the
Vlasov-Poisson equation of motion leading to complex foldings of the
submanifold \citep[e.g.][]{Arnold:1982, Tremaine:1999}. Any such
folding is coincidental with a volume inversion of a simplex. This
volume inversion occurs when the simplex topologically evolves through
a degenerate state (where the tetrahedron is planar because one vertex
moves through the plane defined by the remaining three) which is
equivalent to the emergence of a caustic. Note how this corresponds
exactly to the sign changes of the distortion tensors of
\cite{Vogelsberger:2008} (their eq. 24).
The motion of the vertices does not change the connectivity of the
mesh so that at all times the simplex structure can be constructed
from knowledge of the $\mathbf{q}$. For cosmological $N$-body
simulations, there exists a unique mapping between a particle ID and
$\mathbf{q}$, so that the phase-space structure of the dark matter
sheet can be reconstructed at all times. Projecting the sheet onto
configuration space gives then a volume filling density field of the
dark matter fluid that we propose to use as the density field that
should be used to solve Poisson's equation in future solvers for
collisionless fluids. Current $N$-body solvers do not evolve the
vertices consistently with a density field construed in the proposed
way.
As a first step towards this goal, we analyze the results of standard
cosmological $N$-body simulations using this new definition of the
dark matter sheet. The plan of the paper is as follows. First we will
explain one and two dimensional analogues to introduce the relevant
concepts. We then describe the details of our implementation before we
apply the method to analyze cosmological large-scale structure as well as
the phase-space properties of a single galaxy cluster halo.
\section{Evolution of three-dimensional sheets in phase-space}
The distribution function $f(\mathbf{x}, \mathbf{p}; t)$ describes the density
of a fluid in phase-space. It evolves via
\begin{equation}
\frac{\partial f}{\partial t} = - \frac{\mathbf{p}}{m} \cdot \boldsymbol{\nabla}_x f -
\boldsymbol{\nabla}_x \phi \cdot \boldsymbol{\nabla}_p f,
\end{equation}
where $\phi$ is the gravitational potential and $m$ is the dark matter
particle mass. Fluid elements get stretched or compressed in coordinate space by
advection $\frac{\mathbf{p}}{m} \cdot \nabla_x f$, and in the momentum
coordinates by the gravitational forces $\frac{\mathbf{p}}{m} \cdot
\boldsymbol{\nabla}_x f$. Note that in a Lagrangian frame, the first term on the
right hand side is zero. Furthermore, the second term describes how
the fluid is stretched in momentum space and does not affect the space
density of the fluid parcel. This just states Liouville's theorem
\citep{gibbs1902elementary} that the volume in phase-space is
conserved. Hence, any fluid volume $\triangle \mathbf{x} \triangle
\mathbf{v}$ will remain constant. We are interested here in the space
density of the fluid, the projection of $f$ into coordinate
space. i.e. the integral $\rho(\mathbf{x}) = \int f(\mathbf{x}, \mathbf{v})
{\rm d}^{3}v$. The contribution to the space density of any stream of dark
matter is only affected by the volume it occupies in the space
coordinates, i.e. $\triangle \mathbf{x}$. Consequently, all that is necessary
to follow the evolution of the dark matter density is to follow the
Lagrangian evolution of fluid elements. The mass inside a volume
element is conserved and its contribution to the space density of dark
matter is described by the volume it occupies. Conversely, for a given
WIMP model one knows the initial velocity dispersion at any point in
space \citep[e.g.][]{Hogan:2001, Vogelsberger:2008}. Therefore, if one
knows the spatial part of the phase-space density one has information
about the density in velocity space. For a given shape of the initial
distribution function in the velocity directions (e.g. a Maxwellian)
one has a reliable measure of the intrinsic velocity density at all
times.
Using the Vlasov equation to describe DM is justified for most
particle physics inspired models of dark matter. For a WIMP scenario
with a $100\,$GeV particle e.g. there would be $10^{67}$ such
particles in the Milky Way alone. So an element in phase space that
contains a million such WIMPS, giving a well defined phase space
density, would have a spatial extent at mean density equivalent to a
cube 500 meters on a side. Such a volume would only be a few meters on
a side at the DM density expected in the solar neighbourhood. Clearly
we are interested in scales much larger than this and the
approximation of using a density in phase space is justified to a high
degree of confidence.
It is instructive to first describe a straightforward and well known
one--dimensional example of the evolution of a collisionless fluid
from which a number of lessons can be learned which apply equally well
in higher dimensions.
\subsection{The Zel'dovich pancake}
\begin{figure}
\centerline{\includegraphics[width=0.47\textwidth]{Figures/ZeldovichPedagogy}}
\caption{The one--dimensional plane wave collapse of Zel'dovich
\citep{ZelDovich:1970, Binney:2004}. The top panel gives the phase-space diagram showing the velocities of the particles at their
locations. The bottom panel gives the density of the dark matter
inside the stream, one computed with a seven point stencil (red
squares), and the other computed from the volume between two neighboring
points (solid line). Knowing the spatial volume between particles
along one stream is sufficient to obtain accurate density estimates
at and between the points. }\label{fig:zeldo}
\end{figure}
The phase-space diagram and the evolved density in a Zel'dovich plane
wave collapse is shown in Figure~\ref{fig:zeldo}. The initial sheet at
very early times would be coincident with the $x$-axis as the initial
velocity perturbation is small and the initial state models a nearly
homogeneous Universe. Sampling this initial state with particles of
equal mass results in a grid of uniformly placed particles. Their
configuration space volume is now simply related to their distances in the
$x$--direction. Figure~\ref{fig:zeldo} shows the results of computing
their local stream density from two approaches. In one, labelled
``neighbour'', we take the $V_i=x_{i+1}-x_i$ as the volume between
particle $i$ and $i+1$. One full particle mass is distributed in this
volume and the density at $(x_{i}+x_{i+1})/2$ is given by
$\rho_{neigh} = m_p/|V|$. The values shown as ``squares'' in the same
figure are computed including information from points further along
the stream, $\rho_{7pt}=6\,m_p/|x_{i+3}-x_{i-3} |$. It is defined at
the particle position $x_i$. A number of observations can be
made. Volumes defined in this way may be positive or negative
depending on whether particles have the same or opposite ordering that
they had initially. Volume elements may also become zero. The density
involving more points along the stream gives rise to some smoothing
and density extrema are clipped. The central high configuration space
densities are reached for two reasons. The primordial stream densities
along the sheet become larger and many streams overlap adding their
densities. The number of streams in space is always an odd number at
any location in space. Only at the caustics may one measure even
numbers.
The particle locations trace the sheet in phase-space. Any
unstructured space-filling grid that connects adjacent fluid elements
may be used to trace the dark matter sheet as it evolves in
phase-space. In fact, there is significant ambiguity here as
illustrated in Figure~\ref{fig:choices}. The two--dimensional analog
shown there has is based on triangles (the 2D simplex). The smallest possible elements
one may thus choose to follow would be the Delaunay triangulation of the
points. However, these would give two resolution elements per square
initial cell (case c). It would seem unreasonable that the mass of the fluid
would be conserved exactly in each element, given that we only have
information at the vertices. This could only be true if the mesh
points were not distorted very much and the gradients of the flow,
both in configuration and momentum space had length scales much larger
than the sides of the triangle. However, cases a) depicted in the
figure would likely be a better choice as a fundamental resolution
element as its eight nodes on the surface would be able to more
accurately describe the deformations caused by the flow pattern. The
mass inside that boundary would be conserved to a better degree than
choices for a fundamental resolution element with smaller area. Note,
however, that we still can use triangles to calculate the volume (area
in 2D) of the sheet.
\begin{figure}
\centerline{\includegraphics[width=0.37\textwidth]{Figures/MeshChoices}}
\caption{There are a number of possible cell elements (and that can be
further decomposed into simplices) one may consider as the
unit cell. These unit cells can be used to tessellate the three dimensional dark
matter sheet which connects the particles of an N--body simulation.
}\label{fig:choices}
\end{figure}
After considering some preliminaries in one and two dimensions, let us
proceed to the three--dimensional case.
\subsection{An Unstructured Grid to trace the Dark Matter Sheet in
phase-space}
The simplex in three dimensions is the tetrahedron. We will use it to
calculate volumes and tesselate our chosen fundamental volume
element. This is exactly equivalent to choosing line elements in one
and triangles in two dimensions as the elements which are summed in
volume calculations. Figure~\ref{fig:tets} shows one of the choices we
have employed to tesselate a cubical fundamental cell. The figure also
gives a numbering of vertices of the six tetrahedra which make up the
cell. The connectivity of vertices is chosen such that in a regular
uniform grid all tetrahedron volumes are positive. For much of the
calculation we keep the sign, because, as we will see, it can be a
useful diagnostic of the flow. If we shift a tetrahedron such that one
vertex coincides with the origin, the volume is simply given by the
determinant or, equivalently, from a scalar and a cross product
involving its other three vertices, $V_{tet}=\det\left|
\mathbf{a},\mathbf{b},\mathbf{c}\right|/6=\mathbf{a}\cdot(\mathbf{b}
\times \mathbf{c})/6$. This implies that the volume can be negative
if the tetrahedron has been turned inside out. This sign inversion
occurs when one vertex moves through the plane defined by the other
three vertices of the tetrahedron and is an efficient way to find
caustics. Within an N-body calculation, tracking the number of volume
inversions could be used to trace the
number of caustics crossed by a fluid element. This volume can now
be used straightforwardly to estimate the local stream density of
the fluid element described by the tetrahedron. In our case, one tetrahedron
contains one sixth of the mass of an N-body particle spread over
its volume so that
\begin{eqnarray}
\rho_{s} = \frac{m_P/6}{V_{tet}} = \frac{m_P}{|\mathbf{a}\cdot(\mathbf{b} \times \mathbf{c})|}.\label{equ:rhostream}
\end{eqnarray}
Alternatively one may choose a cubical region rather than a single tetrahedron
as the fundamental volume element to consider. This can be achieved e.g. by using the 24
($2\times 6+6\times2$, see Figure~\ref{fig:tets})
tetrahedra around each point which abut to a given vertex. Their
volumes are then thought of containing four times the mass of one
particle. This effectively averages the density field on a kernel of
the same size as four cubical volumes.
\begin{figure}
\centerline{\includegraphics[width=0.37\textwidth]{Figures/TetrahedraInCubicalCell}}
\caption{One of the possible decompositions of a cubical cell into
tetrahedra. This choice results in 6 tetrahedra all of equal volume
fraction. This particular case has two of the cube corners as vertices for
all tetrahedra and while the other corners are connected to two
tetrahedra each. }\label{fig:tets}
\end{figure}
In a simple implementation, the vertices of the tesselation correspond
to the particles of a standard $N$-body simulation. This implies that
the vertices are moving in a Lagrangian way so that the spatial
sampling is degrading over time in low-density regions and improving
in high-density regions. One implication of the Lagrangian motion of
the mesh vertices is that unresolved volume elements may not evolve
according to the properties of the underlying flow. One particular
such example is given in Figure~\ref{fig:saddlepoints} which
illustrates that volume elements covering a divergent flow can evolve
as if they were in a convergent flow if only the vertices are convergent.
This behaviour implies that single volume elements, i.e. tetrahedra, of
our space discretization are not to be trusted as perfect bags
of fluid. In the example from Figure~\ref{fig:saddlepoints}, the
single tetrahedron suggests that shell-crossing occurred across a
divergent flow which is unphysical -- e.g. implying spurious links
between haloes across unresolved void regions in cosmological context.
There are two possibilities to achieve a more robust density estimate:
(1) The volume estimates of several neighbouring tetrahedra are combined, so that
density estimates are based on a larger region of the flow, or (2)
local refinement of the volume elements is performed whenever e.g.
axis ratios or curvature in phase-space suggest that the volume
element might miss the flow properties. The second option is certainly
the more exciting possibility to achieve a density estimation method
that is well consistent with flow properties. It requires however that
the discretized dark matter sheet is refined while the $N$-body
simulation is run, so that the newly inserted vertices are evolved
with the flow. For this reason, we focus on an evaluation of the
averaging approach in this paper and consider refinement strategies in
a future publication.
\begin{figure}
\centerline{\includegraphics[width=0.47\textwidth]{Figures/degenerate_tets}}
\caption{The need to resolve critical points of the flow: (left) a
halo-void-halo configuration that, when probed with a tetrahedral
structure leads to a final state of the structure as given in the
middle panel. The correct motion of the fluid element could have
been more correctly been described by the right panel. The loss of
spatial resolution can be circumvented by smoothing or by adaptive
refinement of the volume elements. }\label{fig:saddlepoints}
\end{figure}
\subsection{Implementation: Computing the stream properties}
The choice of tetrahedra used for most of the plots in this paper were
given with a slightly different tesselation than the one shown in
Figure~\ref{fig:tets}. The connectivity list for the six tetrahedra
specifically is [4, 0, 3, 1], [7, 4, 3, 1], [ 7, 5, 4, 1], [7, 2, 5,
1], [7, 3, 2, 1], [7, 6, 5, 2] where the unit cube vertices are
labeled as in Figure~\ref{fig:tets}. A natural way to obtain the dark
matter stream density at the location of an N-body particle is to
consider the volume surrounding any vertex $\mathbf{x}$ as the union
of all the tetrahedra that share this vertex. The volume of this union
is computed using the modulus of the volumes, i.e. inverted tetrahedra
are not subtracted. Such a volume carries the mass of four particles
which allows us to estimate the local density at a particle,
$\rho_{s,p} = 4\, m_p/\sum_{i=1}^{24}V_{tet}^i$. We call this local average over
adjacent tetrahedra the ``primordial'' stream density at a
particle's position. It is a well defined quantity even after
shell-crossing. However, it is only defined for the vertices
$\mathbf{x}$ of the tessellation. The simpler definition of
equation~(\ref{equ:rhostream}), however, works very well and, from
visual inspections of dark matter renderings, appears perfectly justified.
The next step in calculating a configuration space density estimate,
as well as in evaluating other stream properties is an integration
through velocity space. This is achieved by finding all intersections
of tetrahedra with the point $\mathbf{y}$ at which the density or
other properties are to be determined. We refer to all these
intersections, which are not part of the primordial (or fundamental)
stream, as the ``secondary'' streams. Because we start from a complete
tessellation of the sheet, the number of tetrahedra enclosing any
spatial point is the number of streams contributing at this point.
At the heart of a fast algorithm is thus a way to speed up this search
for point-tetrahedron-collisions. To find the phase space properties
e.g. at the locations of all the simulation particles one at first
sight expects to require $6*N^2$ searches. I.e. for each of the
$N$--points check all $6N$ tetrahedra whether they overlap that
location. This indeed would be very numerically inefficient.
Fortunately, this can be radically improved using a chaining mesh that
organizes tetrahedra contained in mesh cells or use tree structures
which group tetrahedra into regions of space they occupy. One could
use six trees, e.g., to organize bounding boxes containing tetrahedra
which is advantageous as they are generally smaller than the
circumsphere of tetrahedra which would require less storage. We,
however, implemented two other versions. One with a chaining mesh and
another that employs three bounding-box oct-trees offset from each
other containing disjoint sets of tetrahedra so that tetrahedra
cutting along tree node boundaries only do so typically for one of the
trees. We then parallelized the search with OpenMP and arrived at a
very practical tool to carry out the tessellation and measure the
quantities we were interested in.
For the most straightforward case where the stream density is given by
equation~(\ref{equ:rhostream}) the total density at a given locations is
now just the sum of these $\rho_s$ for all the intersecting
tetrahedra. This simple approach is what we used for the renderings in
the following section.
Since $\mathbf{y}$ will typically lie inside a tetrahedron, we can
also interpolate the primordial stream density to $\mathbf{y}$ if we
defined them by averaging over a cubical volume at the vertices of
this tetrahedron as described above. A one-over-distance-squared
weighting (Shepard's method) works well. Also in this case, the total
configuration space density is simply obtained by finding all
tetrahedra that contain the point under consideration and summing over
all their stream densities interpolated to this location. When
evaluating velocities, we interpolate the velocity field inside of a
tetrahedron from the velocities of the four particles that constitute
its vertices. Again, this is achieved by Shepard's method and no further
averaging is needed.
To be more explicit, let us emphasize here that all of what follows below is
obtained solely by post-processing existing simulations. No line
of code has to be changed in the readers' favourite cosmology code. As
long as the code writes out the particle IDs at every snapshot of the
simulation, and the connectivity of the initial particle distribution is known
or can be constructed, one can post-process simulations that already
exist and measure, visualise and analyse it in new ways.
With these definitions and implementation details in hand, we now
proceed to apply the method to a number of N--body simulations of
large scale structure formation.
\section{Applications}
For our first applications, we chose to test the method on simulations
of the same physical volume, differing only in numerical resolution.
\subsection{$N$-body Simulations} We have carried out cosmological
$N$-body simulations of a volume of $40\,h^{-1}{\rm Mpc}$ length, run
with the tree-PM code {\sc Gadget-2} \citep{Springel:2005}. The
initial conditions for these single-mass-resolution simulations were
generated with the {\sc Music} code \citep{Hahn:2011} keeping
large-scale phases identical with changing mass and spatial
resolution. We assume a concordance $\Lambda$CDM cosmological model
with density parameters $\Omega_{\rm m}=0.276$,
$\Omega_{\Lambda}=0.724$, power spectrum normalization
$\sigma_8=0.811$, Hubble constant $H_0=100\,h\,{\rm km}{\rm
s}^{-1}{\rm Mpc}^{-1}$ with $h=0.703$ and a spectral index
$n_s=0.961$. The particle numbers, masses and force softenings of
these simulations are summarized in Table \ref{tab:sims}. Very
clearly, this box is much too small for a careful statistical analysis of
the cosmic web. However, we chose it here to give us the opportunity
to study how our method converges at different mass resolution both in
the collapsed objects as well as in the cosmic web. A number of the
quantities we measure should hence not be understood as final
numbers/answers. Similarly, the highest resolution simulation we
discuss here takes very little computational resources. However, as we
shall see, these simulations will suffice to demonstrate the
advantages of the new approach.
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
number particles & $m_p / h^{-1}{\rm M}_\odot$ & $\epsilon / h^{-1}{\rm kpc}$ \\
\hline
\hline
$32^3$ & $1.50\times10^{11}$ & 100 \\
$64^3$ & $1.87\times10^{10}$ & 50 \\
$128^3$ & $2.34\times10^{9}$ & 25 \\
$256^3$ & $2.92\times10^{8}$ & 10 \\
\hline
\end{tabular}
\end{center}
\caption{The specifics of the suite of $N$-body simulations used in this
paper. All simulations are of a $40\,h^{-1}{\rm Mpc}$ cosmological volume,
$m_p$ is the particle mass, $\epsilon$ the force softening.}\label{tab:sims}
\end{table}
\subsection{Large Scale Structure \& Streams}
Our method allows to separate physically distinct structures. The
number of streams at a given location can only ever be an odd number
as any fold will add two more streams to an existing one
\citep[e.g.][]{Arnold:1982, Shandarin:1989}. The notable exception is
at the location of caustics where points may sit such as to only
measure an odd number of additional streams. For the number of streams
defined at the particle positions, we can use this fact to select the
structures that are constituting the first caustic.
\begin{figure*}
\centerline{\includegraphics[width=0.9\textwidth]{Figures/particles_in_structures}
}
\caption{Particles with different numbers of streams in a slice
$0.5h^{-1}{\rm Mpc}$ thick. Top left: particles whose
primordial stream does not overlap with other parts of the
sheet. Top right: particles which are on their first caustic,
i.e. they measure a number of streams of two at their
location. Bottom left: particles for which the number of streams is
greater to or equal to three. Bottom right gives the average number of
streams on an infinitesimally thin slice. Distinct physical
components become clearly visible and separated.
}\label{fig:lss}
\end{figure*}
We illustrate the meaning of the local number of streams in
Figure~\ref{fig:lss}. The particles that record that they are part of
only their primordial stream clearly define the voids at the mass
scale that is resolved in the calculations. Particles that count two
streams surround the sheets formed between voids. When they undergo
their first caustic, they have already crossed through the sheet and
are turning around on the side opposite to from where they fell in
from. The particles which measure three or more streams are also
shown and they trace the location of the collapsed objects well. We
still consider all particles which count two streams or more as part
of collapsed objects. This is the same decomposition that can be made
in the GDE approach of \cite{Vogelsberger:2008} as shown for the
environment of Aquarius haloes in \cite{2011MNRAS.416.1377V} (their
Figure~B1) and \cite{Vogelsberger:2011} (their Figure~4).
\begin{figure}
\begin{center}
\includegraphics[width=0.47\textwidth]{Figures/massfraction_streamnum}
\end{center}
\caption{The mass fraction distribution in streams. Many more odd
numbers of streams are found than even ones. We show them as
separate lines with the odd ones displaced down by a factor of ten
for clarity. They differ by a factor approximately 4 and that
ratio does not depend much on the resolution of the underlying dark
matter simulation. An asymptotic slope of $\sim N_{stream}^{-1}$ develops for
the higher resolutions for an intermediate number of streams. At low
resolution steeper relations are inferred. For our most resolved
simulations about 90 per cent of the mass is in collapsed structures. Some
50 per cent of the mass is in locations with 20 streams or more. }\label{fig:streams}
\end{figure}
We can see the distributions of the mass fractions as a function of
the number of streams in Figure~\ref{fig:streams}. We plot them for
particles recording odd and even counts separately. One may have
expected that the fraction of particles that are on caustics vs
particles that have odd numbers of streams to decrease, as the caustics are
better resolved for the high resolution runs. Instead, the offset
between odd and even numbered mass fractions is approximately
constant. This is just a feature of cold dark matter simulations that
with increasing resolution also more smaller objects can be resolved.
This is illustrated directly in the bottom panel of that Figure. The
cumulative distribution of mass above a given number of streams
clearly does not converge. This just reflects that there are many
small scale density fluctuations that collapse even earlier when the
simulations can resolve them. It is quite plausible that the total
fraction of collapsed mass will ultimately approach the very high
value of 99 per cent that has been estimated analytically by
\cite{Shen:2006} from the ellipsoidal collapse model.
For the volume averaged fraction as a function of streams and the
corresponding cumulative distribution in Figure~\ref{fig:streams_volfrac},
we observe a similar lack of convergence. Smaller and smaller pancakes are
resolved as the resolution increases, making more and more volume
have had shell crossing in the past.
It is clear from these results that questions about the shell-crossed
mass and volume fractions in cold dark matter simulations are
intimately tied to a scale. Only when introducing such a scale,
e.g. through filtering of density perturbations or a constant force
softening across resolutions, we could hope to obtain a meaningful
measure of these quantities.
This is compatible with previous results on mass and volume fractions
in the various parts of the cosmic web e.g. by
\cite{HahnPorciani:2007} using a fixed scale, or by
\cite{Aragon-Calvo:2010a} using adaptive filtering. In both cases the
filtering scales are related to the non-linear scales today which is
the relevant scale for much of galaxy formation.
\begin{figure}
\begin{center}
\includegraphics[width=0.38\textwidth]{Figures/volfraction_streamnum}
\end{center}
\caption{The volume fraction distribution in streams; a resolution
study. For our most resolved simulations about 85\% of the volume is
in voids, around 7 (14) per cent is in collapsed structures
($N_{streams}$ larger or equal to three) for the $128^3$ ($256^3$)
run. As expected in CDM, the fraction of volume occupied by
collapsed objects does not converge. }\label{fig:streams_volfrac}
\end{figure}
\begin{figure*}
\centerline{\includegraphics[width=0.97\textwidth]{Figures/256_test}}
\caption{A rendering of the projected dark matter density in the
$256^3$ run using our density estimator and our custom GPU based
renderer.
}\label{fig:pretty}
\end{figure*}
\begin{figure*}
\centerline{\includegraphics[width=0.47\textwidth]{Figures/point-256-zoom}
\includegraphics[width=0.47\textwidth]{Figures/256-zoom}}
\caption{Comparison of the visual appearance of renderings of the dark
matter density in the $256^3$ run using our new density estimator with a
simpler density estimate based on the log of the number of dark matter particles
falling within given image pixels. While many of the well sampled
regions are clearly apparent in both, the detailed structure of
filaments, sheets and how they connect to voids becomes only
apparent in our new approach shown on the right.
}\label{fig:pretty-box}
\end{figure*}
\subsection{Visualization}
The method presented here is also an ideal tool to visualize the
data from current N--body simulations and thus to further help extracting
physical insights from the calculations. Figure~\ref{fig:pretty} gives
an example that we have obtained with a custom-written OpenGL based
renderer of the tetrahedral mesh. The primordial stream densities
are averaged over abutting tetrahedra as described in the
implementation section above. Then tetrahedra are projected, taking
advantage of the OpenGL primitives designed specifically for polygonal
meshes. Details of this algorithm and variations as well as efficient
implementations of current graphics hardware will be
given in a forthcoming publication (Kaehler, Hahn and Abel, in prep.).
We compare this to the visual impression one obtains by plotting
individual points of a calculation vs. our new density definition in
Figure~\ref{fig:pretty-box}. One can clearly see how filaments and
sheets in and surrounding voids can be distinguished easily now. The
visual impression is commensurate with the statistics we present next.
\begin{figure}
\centerline{\includegraphics[width=0.47\textwidth]{Figures/primary_vs_rhotot}}
\caption{Two dimensional mass-weighted histogram comparing the density of the
primordial dark matter sheet to the total sheet density in the
$128^3$ simulation. Much of the mass is contained at configuration
space densities about ten times the mean. The primordial stream
density also scatters very much but its median density is close to
the average density of the Universe. }\label{fig:primVSrhotot}
\end{figure}
\begin{figure}
\centerline{\includegraphics[width=0.37\textwidth]{Figures/hist_rho_massw}}
\caption{Mass-weighted density distributions. The top left panel
shows the histogram for four different densities defined at the
particle locations for the $256^3$ run. The density estimated from a
zeroth order estimate of the Voronoi tessellation, the total sheet
dark matter density, the primordial stream density and the secondary
stream density. The results of the resolution study is given in the
other three panels. The total space density is given in the top
right panel. The bottom left is the mass-weighted density pdf of
sheet density in the primordial stream. The bottom right panel gives
the density contributed by material that is not in the primordial
stream. }\label{fig:hist-rho-massw}
\end{figure}
\begin{figure}
\includegraphics[width=0.4\textwidth]{Figures/hist_rho_volw}
\caption{Volume-weighted density distributions. The top panel shows
the histograms for the $256^3$ run, the lower those for the $32^3$
run. The zeroth-order density estimated from a Voronoi tessellation is
shown with a dashed line, the total sheet dark matter density with a
solid line. At both resolutions, both the Voronoi and the stream
density approach a $\rho^{-1}$ power-law at high densities. Also, the
two methods produce different estimates at intermediate densities of
$\rho/\bar{\rho}\sim10$. The bottom panel also shows in grey the
density histograms from our method for all simulations to aid the
comparison.}\label{fig:hist-rho-volw}
\end{figure}
\subsection{Dark Matter Densities}
Figure~\ref{fig:primVSrhotot} gives the relation of the density in the
primordial stream to the total density evaluated at the locations of
all the particles, i.e. a mass-weighted histogram. At low densities
these are identical as this material is traced by the original sheet
and no folding has occurred. There is an enormous scatter at higher
densities which we can quantify
further. Figure~\ref{fig:hist-rho-massw} gives the mass-weighted
density distribution for all the simulations we have analyzed. The
top panel summarizes the individual contributions for the $256^3$
simulation. The primordial stream density distribution peaks slightly
below mean density while the total mass weighted density distribution
is at much higher densities. The reason is that all the streams not
inside the primordial stream contributing to the density at the
location of the particle contribute much more to the total density at
high densities. That distribution is given by the green line in the
top panel labelled as ``secondary''. There is an apparent power law
part in the primordial stream densities visible. We will discuss this
further when considering volume-weighted distributions. These are
given in Figure~\ref{fig:hist-rho-volw} where we show the volume-weighted dark
matter density. The total densities we estimate with our method are
labelled as ``Sheet'' . We also indicate the resolution of the dark matter
simulation used to compute it. The median of the stream density is
$1.2$ but its average is 26 times the mean. We also do not expect
these numbers to converge as one continues to increase the
resolution.
\begin{figure}
\centerline{\includegraphics[width=0.47\textwidth]{Figures/voronoi_vs_rhotot}}
\caption{Two dimensional histogram comparing the zeroth--order Voronoi
density estimate vs. the total sheet density. The correspondence is
quite good. The largest difference is observed for values between a
third and thirty times the mean density of dark matter. The
zeroth--order Voronoi density estimators overestimates the volumes
in regions around filaments and sheets. }\label{fig:vorVSrhotot}
\end{figure}
We also compared our new density estimates with the corresponding
results from another density estimator, which finds the unique Voronoi
cells around each particle. The density in that volume is then simply
defined as the mass of the enclosed particle divided by that volume
element. Following \cite{van-de-Weygaert:2009}, we refer to this as the
zeroth--order Voronoi density. Albeit, this is clearly more noisy than
the DTFE density estimator developed by \cite{Schaap:2000} and
\cite{Pelupessy:2003} since the density is defined for the smallest
region. DTFE is more advanced and employs averaging over nearby
tetrahedral cells. These authors give comparisons of that estimator
to the smooth kernel estimates obtained with the otherwise popular
Smoothed-Particle Hydrodynamics estimator. Like DTFE, the simple
zeroth--order Voronoi estimator we use tessellates the entire volume
and has no parameters and as such is a well suited benchmark for
comparison.
Quite strikingly, at low and high densities our density estimate is
very similar to the zeroth--order Voronoi volume based density
estimator (Figure~\ref{fig:vorVSrhotot}). Both methods do not
converge at the very low density tail when varying resolution.
This is physical in that the simulations model smaller voids when
the resolution is increased. These can achieve lower densities than
their larger counterparts in lower resolution simulations. So, at higher
resolutions, it is not surprising to see a tail growing at those lowest
densities. Also, the peak distribution shifts continuously to lower
densities as more particles are employed. Significant differences in
the volume fraction at a given density of our method with the
zeroth--order Voronoi method are seen at intermediate densities around
the mean density. This is understandable as the volumes that the
Voronoi tessellation provides, tend to connect particles in the voids
to the particles in the sheets and filaments. At that point it spreads
particles in filaments into volumes that are larger and consequently
estimates lower densities. In fact, these estimate are very
significantly different. So much so that, if we integrate the density
from our method over the volumes computed from the zeroth--order
Voronoi estimator, the total mass in the box is overestimated by more
than a factor of ten. It certainly would be interesting to compare our
density estimate not only to this zeroth-order Voronoi density but also
to other density estimators such as DTFE, adaptive Kernel softening,
etc.. This is, however, beyond the scope of this first exposition of
our approach.
Next, we will now apply our method to measuring a number of well studied
quantities in dark mater haloes.
\subsection{Radial profiles of haloes}
\cite{Navarro:1996} have discovered a universal radial profile of the
dark matter density in virialized haloes. This is one of the key
findings of cosmological N--body simulations and a large body of
literature has largely confirmed the finding. We will give these
profiles next. To get the best possible estimate, we chose 100,000
test positions and bin these in 50 radii, spaced
logarithmically in radius. Figure~\ref{fig:NFW} summarizes our
findings. The density profiles computed from the dark matter sheet are
somewhat shallower and have about 50\% larger central densities at all
resolutions for the single halo we have analyzed. Physically, it is
conceivable that volume elements formed by particles on radial orbits
oscillate such that the bounding regions have a higher probability to
be found at large radii while still contributing to the density
interior to small radii. Similarly, one can picture particles of the
sheet orbiting the center at larger radii such that the volume element
they span can contribute to the center. Our method has a well defined
density at all radii and it is bound to be a constant at the lowest of
radii where one averages always over the same tetrahedra. At large
radii, the new density estimate and the Voronoi estimates all agree
extraordinarily well. This is true even in the infall region. The
Voronoi estimates at all radii are perfectly consistent with a simpler
estimate based on the particle mass binned in shells divided by the
shell volume (not shown).
The masses included within a radius do converge also quite well with
our estimate being consistently slightly higher at all radii.
This is understandable as mass from particles outside a given
radius can contribute if they span a volume element that has nodes
inside the radius.
\begin{figure}
\begin{center}
\includegraphics[width=0.45\textwidth]{Figures/halo_profiles}
\end{center}
\caption{Dark Matter density profile in the most massive halo of
$2\times 10^{14}M_\odot$ with $R_{vir}\approx 1{\rm Mpc}$ at
redshift zero. The overdensity in the top, the enclosed dark matter
(middle) as calculated from the density in the top panel, and the
number of streams (bottom panel) are shown for all the different
resolution simulations studied in this paper. The density profile
estimated from our method starts to differ from the conventional
estimate at scales as large as one third of the virial radius. As
expected in CDM, the number of streams does not converge also in
radial profiles as the resolution is increased. }\label{fig:NFW}
\end{figure}
The number of streams that contribute to the profile are given in the
lower panel of Figure~\ref{fig:NFW}. Not surprisingly, these increase
with increasing resolution. If one scales them by factors of eight
between increasing resolutions, some closer convergence is
observed. Between the $128^3$ and $256^3$ simulations, there remains a
difference of about 30\% which is likely just due to streams
contributing from larger radii to the regions inside the particles
spanning the volume element.
It certainly would be interesting that the dark matter density profile
in the central parts of haloes could be different than one estimates
by measuring dark matter particles inside a given radius. It is
also suggestive how well our density profiles converge from $128^3$ to
$256^3$ particles.
However, if the mass profile were indeed different, also the forces
contributing to the particle motions would be changed. So even if our
density estimator were more accurate, one could not prove that the
result shown in Figure~\ref{fig:NFW} is the correct physical one until
one has evolved the dark matter sheet consistently, i.e. using
accelerations created by the density distribution of the actual sheet
elements. We discuss some potential approaches to carry out such
simulations in the discussion section.
\subsection{Velocity dispersions and the dark matter ``entropy''}
While our method gives access to the full fine-grained phase-space
structure, we chose to only show moments to serve as an example of what
the method is good for, and to be able to compare to work done on
this previously. While the collisionless fluid does not experience
microphysical collisions, the scattering provided by the time-varying
gravitational potential leads to mixing in phase-space.
The velocity dispersion of the particles is a measure of the effective
pressure of the dark matter, which is of relevance for understanding
the dynamical structure of orbits, i.e. e.g. the expectations of how observable
stars move in the DM potential.
Figure~\ref{fig:veldisp} summarizes the radial profiles of the
velocity dispersion for the same halo we have analyzed above for
density profiles. We again draw particles at random positions in
spherical shells for which we measure the stream-density-weighted
bulk velocity, subtract it from the stream-density-weighted local
velocity dispersion, before we finally average it to obtain the velocity
dispersion in radial shells.
These velocity dispersions differ quite significantly from the
similarly termed quantities presented previously \citep[e.g.][and
references therein]{Navarro:2010} . While this may seem surprising, one
has to keep in mind that we measure the dispersion at a single
point, i.e. we do not carry out any averaging over volume.
Dispersions quoted in the past measured a combination of a bulk and
local dispersion. This will not only have large sampling error but
also confuse turbulent bulk motions with actual microphysical
velocity distributions. Indeed, we can see that our measured velocity
dispersion does not converge at scales of about one half the virial
radius, as only about one thousand streams contribute there for our
highest resolution results. The distribution functions at this location will be
quite anisotropic and a single temperature will be a bad fit (see
below). The halo is remarkably cold in the center -- having less than half the
velocity dispersion expected from the virial velocity. In the same
figure, we also show the pseudo phase-space density of
\cite{Taylor:2001} which has been found to be a perfect power-law
entirely independent of resolution \citep{Navarro:2010}. When we
measure the average of only the fine-grained quantities, as shown in
Figure~\ref{fig:veldisp}, this perfect power-law disappears. This may
suggest that much of the measure is dominated by large-scale bulk
flows. It is worthwhile to explore this further with higher resolution
simulations, where one can more confidently separate thermal motions
from the bulk velocity dispersion.
\begin{figure}
\begin{center}
\includegraphics[width=0.45\textwidth]{Figures/halo_sigma_profiles}
\end{center}
\caption{Radial spherically averaged profiles of the velocity
dispersion (top), the dark matter ``entropy''
$\sigma^2/(\rho/\bar{\rho})^{2/3}$ (middle) and the pseudo
phase-space density (bottom) for the same halo as in
Figure~\ref{fig:NFW}. The velocity dispersion is remarkably flat
inside about one tenth of the virial radius. The dark matter
``entropy'' profile also shows signs of already converging at the
modest resolutions employed here. Using the microscopic velocity
dispersion of our approach which removes the bulk flows does not
give the typical powerlaw behavior in the pseudo-phase-space
density found when using the total dispersion of particles at those
radii.
}\label{fig:veldisp}
\end{figure}
It is though just as interesting to check the actual distribution
function of dark matter velocities at a given point. The seminal work
of \cite{Lynden-Bell:1967b} discussed this in the context of stellar
systems. The global distribution has been measured from simulations
many times \cite[see e.g.][]{Hoeft:2004, Wise:2007a, Navarro:2010,
Vogelsberger:2009} but, to the best of our knowledge, this was never
done at individual points in the simulations. Figure~\ref{fig:violent}
summarizes the distributions found at three different points in the
most massive halo. We show it at the center where a relatively hot
component overlays a colder one. At the center, the distributions of
the individual velocity components have peaks that almost coincide and
widths which are quite similar as well. They are not too far from an
isotropic Maxwell-Boltzmann distribution in their cores. As we step
out in radius, the situation changes rapidly and the microphysical flow
structure clearly shows more and more anisotropy. Interestingly, a
quite hot component is seen along the $x$ direction. At the same the
velocity distribution in the $y$ direction is the coldest at all
radii. These distributions are consistent with the visual impression
obtained from the velocity dispersion slice in
Figures~\ref{fig:clusterpanels} and \ref{fig:veldisp} which also shows
that some of the hottest DM fluid elements are found just inside the
virial radius.
It is remarkable how much physics can be learned from even these low
resolution simulations analyzed here. For the halo we just
discussed, there are not even 600,000 dark matter particles inside the
virial radius of $1.4h^{-1}{\rm Mpc}$. At the same time, there are
already enough streams to compute meaningful measures of the structure
of phase-space. We are certainly looking forward to carrying out a
more detailed analysis on higher resolution simulations. This point is
born out by the visual impression given by infinitesimal slices as shown in
\ref{fig:clusterpanels}, which we will describe next.
\begin{figure}
\begin{center}
\includegraphics[width=0.47\textwidth]{Figures/velo_distribution}
\end{center}
\caption{The velocity distribution function binned in velocity space
at three points at varying distance from the center of the halo
(left panels) and the individual distributions of the velocity
components (right). We used 200, 80 and 30 bins for the points at
the center, 0.4 $h^{-1}$Mpc and $1 h^{-1}$Mpc from the center along
one axis. These had 150,000, 13,000 and 400 individual streams that
contributed. }\label{fig:violent}
\end{figure}
\subsection{Slices of Density and Dark Matter ``Entropy''}
To aid in the interpretation of the profiles we have just presented,
we also give two dimensional slices through the dark matter density
and ``entropy'', which we define analogously to the adiabats used when
studying e.g. galaxy clusters hydrodynamically simply as $S_{DM} =
\sigma^2/(\rho/\bar{\rho})^{2/3}$, which then has units of the
square of a velocity. Also, the average number of streams contributing
to every point on the slice is given. Material from the voids falls in
perfectly cold. We can think of the velocity dispersion as a measure
of the temperature of the fluid. It is ill defined in the single
stream regions falling from the voids. However, these carry very
little mass. Then we see a region that extends to about two Mpc from
the center which hosts multi-stream material of the order of about ten
streams. The virial radius, which is approximately at one Mpc, shows a
marked increase in the velocity dispersion and a much smoother density
structure. Even on this scale, we can see the cold central isothermal
part of the object, both in the velocity dispersion and in the
entropy. Substructure is easily seen as cold low entropy material
embedded in the hot halo. Many of the structures seen here are very
reminiscent of adiabatic hydrodynamical simulations of galaxy clusters
\citep[e.g.][]{Frenk:1999} and even first star formation \citep{Abel:2002c}
where gas enters haloes predominantly through filaments and shock
heats, resulting in a halo with rising entropy profiles with radius.
\begin{figure*}
\begin{center}
\includegraphics[width=0.97\textwidth]{Figures/cluster_panels}
\end{center}
\caption{Infinitesimally thin slice through the $256^3$ simulations
for the most massive halo. We show the density in units of the mean
density (top left), the stream-density weighted velocity dispersion
in kilometers per second (top right), the dark matter entropy
[(km/s)$^2$], computed from the density and stream weighted velocity
dispersion $\sigma^2/(\rho/\bar{\rho})^{2/3}$, and the average number
of streams (not stream density weighted) at the bottom
right. Clearly, our approach gives information so far thought
inaccessible from current simulations. }\label{fig:clusterpanels}
\end{figure*}
\subsection{General interpolation to any point in space}
There are large advantages to have well-defined grids which allows one
to interpolate to any point in space. This is a very obvious
observation of course, it is, however, a large step forward in
understanding dark matter simulations. This has led to a number of
approaches being devised that allow such interpolation, such as the
methods discussed in the introduction. Here we discuss but a few
approaches on how to use the tessellated dark matter sheet for
interpolation.
When probing the sheet at the particle locations, we find the
primordial stream densities, total space densities, number of streams,
velocities etc.. Hence, we have sampled the full volume and have a
non-uniform distribution of the fields we aim to interpolate. One may
choose to achieve further interpolation by using a distance-weighted
estimate from the nearest particle locations. An efficient way to
find two dimensional slices is to take all tetrahedron edges and
compute their intersections with the plane to be interpolated
to. Along every edge one can now linearly interpolate the values of
the vertices to the plane. The resulting scattered data on the plane
is then triangulated again and interpolated with linear interpolation
between nodes. As an example the slice of the total sheet density is
shown in Figure~\ref{fig:DensityVsCIC} which also gives a visual clue
to how cloud in cell interpolation would sample the density field.
\begin{figure}
\begin{center}
\includegraphics[width=0.47\textwidth]{Figures/DensityTetEdgesCIC}
\end{center}
\caption{The logarithm of the density in an infinitesimally thin slice
in units of the mean density for the $128^3$ simulation. The white
dots show the location of the particles which would contribute to a
cloud in cell interpolation on a grid with cells as large as the
mean particle spacing. The squares at the top left show the area to
which these individual particles would contribute to at their
locations.
}\label{fig:DensityVsCIC}
\end{figure}
Similarly, this allows us to extract one dimensional skewers at
arbitrary resolution from N--body simulations. As an example,
Figure~\ref{fig:velSkewer} compares the velocity along a random line
through the volume for different resolutions. The large scale modes are
all consistent by design. It is interesting to see that convergence is
quite slow and suggests to extend this analysis rigorously to much
higher resolutions.
\begin{figure}
\begin{center}
\includegraphics[width=0.47\textwidth]{Figures/VelocitySkewer}
\end{center}
\caption{The velocity field along a one dimensional line extracted
using tetrahedron edges to interpolate to a slice plane. The
differences in resolution are understandably quite large given the
large range of mass resolution of the simulations. The large scale
features remain recognizable even at the lowest of resolutions.
The ``N-shaped'' infall regions are seen for many structures.
}\label{fig:velSkewer}
\end{figure}
\section{Discussion}
\cite{Vogelsberger:2008}, \cite{White:2009} and \cite{Vogelsberger:2011} developed the
GDE formalism to allow a calculation of the primordial stream
density. Their approach modifies the simulation code to integrate an
evolution equation of the tidal tensor along with every particle
trajectory. In principle, this can be much more accurate since the
local stretching of the dark matter sheet is calculated at every time
step of the calculation. It will be of great interest to compare our
approach to theirs in detail. This will require to carry out the
calculations with both methods on an identical simulation to
facilitate a particle by particle comparison. This should be
particularly interesting given that our method can, in addition to the
primordial stream density, also provide the total space density and
number of streams at every location. Since our approach also gives
full fine-grained phase-space information, it seems plausible one
could combine both approaches to a hybrid which inherits the advantages of
both.
More generally, both the GDE and our approach suggest a number of
possible approaches to improve the accuracy of N--body calculations.
Almost all current cosmological N--body solvers employ the particle
mesh method at least for the largest scales in the calculation. The
cloud in cell approximation is used to interpolate the dark matter
particles to a grid on which the gravitational potential will be
evaluated before differencing it to obtain the gravitational forces on
the particles. Since one integrates twice to get the potential from
the density field and only differentiates once, this method gives
reasonably smooth gravitational forces. However, it inherently models
a very noisy inaccurate density distribution obtained from CIC which
will have the largest relative errors in poorly sampled regions such
as voids, pancakes and filaments (see Figure~\ref{fig:DensityVsCIC}).
We have shown that our density estimator would have significantly more
fidelity and reliability for these large regions. It is in principle
quite straightforward to modify an existing particle mesh code to make
use of our density estimator and then derive more accurate potentials and
forces from it. It only involves the interpolation step to the
grid. When interpolating the contribution back to the particle positions
one could make use of the known analytical solutions to the Newtonian
potential of homogeneous polyhedra \citep{Waldvogel:1979}. Similarly,
these analytic formulae could be applied in direct summation and
tree-based codes. A priori it may seem difficult to imagine how to
construct trees efficiently when considering that the tetrahedra may
become exceedingly distorted and elongated and would cover many nodes
of the tree. However, any new code would most likely ever only be
employed using local mesh refinement given that the tessellation we
suggest gives many opportunities to discover the regions of the flow
which may be prone to errors. The local curvature of the flow
compared to the tetrahedra edges is one measure but also the axis
ratio of individual tetrahedra provides an estimate where the flow
would benefit from refinement. The key to any such new method will
have to be to fully consider the dark matter as a fluid so that spurious
particle-particle interactions may be avoided and multi-mass
resolution becomes feasible. Given a locally refined mesh, tree
structures will remain useful in rapidly finding neighbours and retain
$n\,\log n$ scaling.
There are a number of improvements possible that will help to develop
our GPU assisted volume rendering further. Using vertex values and
some form of Shepard's method to carry out distance dependent
weighting should still likely be very fast on current GPUs even when
drawing billions of tetrahedra.
Higher order interpolation in fact could be another avenue to improve
on the method suggested here. We have only implemented the very
simplest of ideas. Namely that volume elements in phase-space are
uniformly filled with the dark matter fluid. This is similar in spirit
to donor cell methods used decades ago for hydrodynamical flows. We
believe that it will be possible to improve on our approach
significantly. Up-winded schemes with linear reconstruction were a
large gain in accuracy in numerical fluid dynamics and similar
improvements are certainly possible here.
Given that one now has a natural grid that can be used to interpolate
any state variables as well as the full fine-grained phase-space
structure, one can also define differentials on it. Consequently, it
becomes possible to study vorticity, divergences as well as carry out
the Cauchy-Stokes decomposition of the dark matter velocity
fields. This way one can separate bulk, shear and rotational
components of the velocity fields which undoubtedly will
make it possible to track down the physical origin of dark matter
density profiles as well as to better understand the internal structure
of haloes and the cosmic web.
There is a remarkably large number of applications where we think our
method can aid to gain new insights. Whether it is gravitational
lensing to find more accurate lensing potentials to studying the origin
of the angular momentum profiles\citep[see e.g. Fig. (12)
in][]{Bullock:2001a}. Obviously the connection between dark matter and the
baryons they host can be explored much more fully now as well.
As we were preparing this manuscript, \cite{Shandarin:2011} posted a
paper on the electronic preprint server which explores the same basic
idea as the one we present here. The concept of tessellating
phase-space and tracking the dark matter sheet is identical to
ours. Details of the implementation and what to think of as
fundamental parts of the approach are not the same. Their choice of
tessellation is quite different.
They pick the minimal combination of tetrahedra of the unit cube
possible which has five elements where one of them is twice the volume
of the other four and itself does not tesselate the space
uniformly. Consequently they alternately rotate adjacent cubes such
that the edges of tetrahedra never cross. This is effective albeit
likely more cumbersome for a practical implementation. The powerlaw
\cite{Shandarin:2011} gives for the volume fraction as a function of
the number of streams ($f_V(N_{stream}) = 0.93\,N_{stream}^{-2.82\pm
0.05}$) is to be compared with our
Figure~\ref{fig:streams_volfrac}. Their power-law fits approximately
our $32^3$ run and likely just reflects the fact that the single
simulation they study had approximately two times worse mass
resolution than our $32^3$ run with an effective gravitational
softening length about five times larger than ours. So both approaches
do agree. We at this time would not attach a special meaning to this
power-law as it clearly is strongly resolution dependent with our
highest resolution run giving something close to $N_{stream}^{-2}$.
Our description also discusses the fine-grained structure in the
velocity directions of phase space, discusses halo properties and
profiles and gives visualizations of the dark matter density not given
by \cite{Shandarin:2011}.
\section{Summary}
We presented a novel approach to better understand the dynamics of cold
collisionless fluids. We apply it by post-processing cosmological N--body
simulations and document the significant improvement it represents
over previous attempts to quantify the macroscopic and microscopic
aspects of the dark matter fluid flow. In particular, we show new
results for density estimates, a dark matter ``entropy'', bulk
velocities, velocity distribution functions -- many of which are
computed here for the first time. We are confident that our approach
to tracing the dark matter sheet in phase-space gives important
physical insights which were inaccessible with previous approaches.
\section*{Acknowledgements}
T.A. is grateful for numerous conversations with Greg
Bryan in the past ten years on how one might solve for dark matter dynamics
directly in phase-space and acknowledges support by the National
Science foundation through award number AST-0808398 and the LDRD
program at the SLAC National accelerator laboratory as well as the
Terman fellowship at Stanford University. He also acknowledges help
by Patrick Abel in constructing tetrahedra from paper which helped
considerably in understanding the many possible options of
tessellations of the dark matter sheet.
|
\section{Introduction}
Studies in astroparticle physics link astrophysics, cosmology,
particle and nuclear physics and involve hundreds of scientific
groups linked by regional networks (like ASPERA/ApPEC \cite{aspera})
and national centers. The exciting progress in these studies will
have impact on the fundamental knowledge on the structure of
microworld and Universe and on the basic, still unknown, physical
laws of Nature (see e.g. \cite{book,newbook} for review).
In the proposal \cite{Khlopov:2008vd} it was suggested to organize a
Virtual Institute of Astroparticle Physics (VIA), which can play the
role of an unifying and coordinating structure for astroparticle
physics. Starting from the January of 2008 the activity of the
Institute takes place on its website \cite{VIA} in a form of regular
weekly videoconferences with VIA lectures, covering all the
theoretical and experimental activities in astroparticle physics and
related topics. The library of records of these lectures, talks and
their presentations is now accomplished by multi-lingual Forum. In
2008 VIA complex was effectively used for the first time for
participation at distance in XI Bled Workshop \cite{archiVIA}. Since
then VIA videoconferences became a natural part of Bled Workshops'
programs, opening the virtual room of discussions to the world-wide
audience. Its progress was presented in \cite{BledVIA9,BledVIA10}. Here the
current state-of-art of VIA complex, integrated since the end of
2009 in the structure of APC Laboratory, is presented in order to
clarify the way in which VIA discussion of open questions beyond the
standard model took place in the framework of XIV Bled Workshop.
\section{The current structure of VIA complex}
\subsection{The forms of VIA activity}
The structure of
VIA complex is illustrated on Fig. \ref{homevia}.
\begin{figure}
\begin{center}
\includegraphics[scale=0.3]{home_page3.eps}
\caption{The home page of VIA site}
\label{homevia}
\end{center}
\end{figure}
The home page, presented on this figure, contains the information on
VIA activity and menu, linking to directories (along the upper line
from left to right): with general information on VIA (About VIA),
entrance to VIA virtual lecture hall and meeting rooms (Rooms), the
library of records and presentations (Previous) of VIA Lectures (Previous $\rightarrow$ Lectures), records of
online transmissions of Conferences(Previous $\rightarrow$ Conferences), APC Seminars (Previous $\rightarrow$ APC Seminars) and APC Colloquiums (Previous $\rightarrow$ APC Colloquiums)
and courses, Calender of the past and future VIA events
(All events) and VIA Forum (Forum). In the upper right angle there are links to Google search
engine (Search in site) and to contact information (Contacts). The
announcement of the next VIA lecture and VIA online transmission of
APC Colloquium occupy the main part of the homepage with the record
of the most recent VIA events below. In the announced time of the event
(VIA lecture or transmitted APC Colloquium) it is sufficient to
click on "to participate" on the announcement and to Enter as Guest
in the corresponding Virtual room. The Calender links to the program
of future VIA lectures and events. The right column on the VIA
homepage lists the announcements of the regularly up-dated hot news of Astroparticle
physics.
In 2010 special COSMOVIA tours were undertaken in Switzerland
(Geneva), Belgium (Brussels, Liege) and Italy (Turin, Pisa, Bari,
Lecce) in order to test stability of VIA online transmissions from
different parts of Europe. Positive results of these tests have
proved the stability of VIA system and stimulated this practice at
XIII Bled Workshop. These tours assumed special equipment, including, in particular, the use of the sensitive audio system KONFTEL 300W \cite{konftel}. The records of the videoconferences at the previous XIII Bled Workshop are available on VIA site \cite{VIAbled10}.
In 2011 VIA facility was effectively used for the tasks of the Paris Center of Cosmological Physics (chaired by G. Smoot), for the public programme "The two infinities" (conveyed by J.L.Robert) for Post-graduate programme assumed by the agreement between the University Paris Diderot and the University of Geneva. It has effectively supported participation at distance at meetings of the Double Chooz collaboration: the experimentalists, being at shift, took part in the collaboration meeting in such a virtual way.
It is assumed that the VIA Forum can continue and extend the
discussion of questions that were put in the interactive VIA events.
The Forum is intended to cover the topics: beyond the standard
model, astroparticle physics, cosmology, gravitational wave
experiments, astrophysics, neutrinos. Presently activated in
English, French and Russian with trivial extension to other
languages, the Forum represents a first step on the way to
multi-lingual character of VIA complex and its activity.
One of the interesting forms of Forum activity is the educational
work. For the last four years M.Khlopov's course "Introduction to cosmoparticle physics" is given in the form of VIA videoconferences and the records of these lectures and their ppt presentations are put in the corresponding directory of the Forum \cite{VIAforum}. Having attended the VIA course of lectures in order to be
admitted to exam students should put on Forum a post with their
small thesis. Professor's comments and proposed corrections are put
in a Post reply so that students should continuously present on
Forum improved versions of work until it is accepted as
satisfactory. Then they are admitted to pass their exam. The record
of videoconference with their oral exam is also put in the
corresponding directory of Forum. Such procedure provides completely
transparent way of estimation of students' knowledge.
\subsection{VIA lectures, online transmissions and virtual meetings} First tests of VIA
system, described in \cite{Khlopov:2008vd,archiVIA,BledVIA9,BledVIA10},
involved various systems of videoconferencing. They included skype,
VRVS, EVO, WEBEX, marratech and adobe Connect. In the result of
these tests the adobe Connect system was chosen and properly
acquired. Its advantages are: relatively easy use for participants,
a possibility to make presentation in a video contact between
presenter and audience, a possibility to make high quality records
and edit them, removing from records occasional and rather rare
disturbances of sound or connection, to use a whiteboard facility
for discussions, the option to open desktop and to work online with
texts in any format. The regular form of VIA meetings assumes that
their time and Virtual room are announced in advance. Since the
access to the Virtual room is strictly controlled by administration,
the invited participants should enter the Room as Guests, typing
their names, and their entrance and successive ability to use video
and audio system is authorized by the Host of the meeting. The
format of VIA lectures and discussions is shown on Fig. \ref{ellis},
illustrating the talk "New physics and its experimental probes" given by John Ellis from CERN in the framework
of XIV Workshop. The complete record of this talk and other VIA
discussions are available on VIA website \cite{VIAbled11}.
\begin{figure}
\begin{center}
\includegraphics[scale=0.3]{john_bled_2.eps}
\caption{Videoconference Bled-Marburg-Liege-Geneva-Moscow-Paris with lecture by John Ellis,
which he gave from his office in CERN,
Switzerland, became a part of the program of
XIV Bled Workshop.}
\label{ellis}
\end{center}
\end{figure}
The ppt or pdf file of presentation is uploaded in the system in
advance and then demonstrated in the central window. Video images of
presenter and participants appear in the right window, while in the
upper left window the list of all the attendees is given. To protect
the quality of sound and record, the participants are required to
switch out their microphones during presentation and to use lower
left Chat window for immediate comments and urgent questions. The
Chat window can be also used by participants, having no microphone,
for questions and comments during Discussion. The interactive form of VIA lectures provides oral discussion, comments and questions during the lecture. Participant should use in this case a "raise hand" option, so that presenter gets signal to switch our his microphone and let the participant to speak. In the end of presentation
the central window can be used for a whiteboard utility
as well as the whole structure of windows can be changed,
e.g. by making full screen the window with the images of participants of discussion.
Regular activity of VIA as a part of APC includes online transmissions of all the APC Colloquiums and of some topical APC Seminars, which may be of interest for a wide audience. Online transmissions are arranged in the manner, most convenient for presenters, prepared to give their talk in the conference room in a normal way, projecting slides from their laptop on the screen. Having uploaded in advance these slides in the VIA system, VIA operator, sitting in the conference room, changes them following presenter, directing simultaneously webcam on the presenter and the audience.
\section{\label{Bled} VIA Sessions at Bled Workshop}
\subsection{The program of discussions}
In the course of XIV Bled Workshop meeting the list of open questions
was stipulated, which was proposed for wide discussion with the use
of VIA facility.
The list of these questions was put on VIA Forum (see \cite{VIAforum11}) and all the
participants of VIA sessions were invited to address them during VIA
discussions. Some of them were covered in the VIA lecture "New physics and its experimental probes" given by John Ellis (see the records in \cite{VIAbled11}). During the XIV Bled Workshop the test of minimal necessary equipment was undertaken. VIA Sessions were supported by personal laptop with WiFi Internet connection only.
It proved the possibility to provide effective interactive online VIA videoconferences even in the absence of any special equipment. Only laptop with microphone and webcam together with WiFi Internet connection was shown to be sufficient not only for attendance, but also for VIA presentations and discussions.
Another application at Bled Workshop was related with VIA records of closed meetings. The presentation was given in the regime of VIA online transmission and recorded, but the admission to the virtual room was restricted by a very short list of distant participants and the link to the record was available to a restricted list of users. Such use of VIA facility may be of interest for closed collaboration meetings.
\subsection{VIA discussions}
VIA sessions of XIV Bled Workshop have developed from
the first experience at XI Bled Workshop \cite{Bregar:2008zz} and
their more regular practice at XII and XIII Bled Workshops \cite{BledVIA9,BledVIA10}.
They became a regular part of the Bled Workshop's programme.
In the framework of the program of XIV Bled Workshop, John Ellis, staying in his office in CERN,
gave his talk "New physics and its experimental probes" and took part in the
discussion, which provided a brilliant demonstration of the interactivity of VIA in the way
most natural for the non-formal atmosphere of Bled Workshops. The
advantage of the VIA facility has provided distant
participants to share this atmosphere and contribute the discussion. VIA sessions were finished by the discussion of puzzles
of dark matter searches (see \cite{VIAbled11}). N.S. Manko\v c Bor\v stnik and G. Bregar presented possible dark matter candidates
that follow from the approach, unifying spins and charges, and Maxim Khlopov presented composite
dark matter scenario, mentioning that it can offer the solution for the puzzles of direct
dark matter searches as well as that it can find physical basis in the above approach. H.B.Nielsen informed about his macroscopic candidate for dark matter.
The comments by Rafael Lang from his office in USA were very important
for clarifying the current status of experimental constraints on the possible properties
of dark matter candidates (Fig. \ref{dm})
VIA sessions provided participation at distance in Bled discussions
for John Ellis and A.Romaniouk (CERN, Switzerland), K.Belotsky, N.Chasnikov,
A.Mayorov and E. Soldatov (MEPhI, Moscow), J.-R. Cudell (Liege,
Belgium), R.Weiner (Marburg, Germany) H.Ziaeepour (UK), R.Lang (USA) and many
others.
\begin{figure}
\begin{center}
\includegraphics[scale=0.3]{bled.eps}
\caption{Bled Conference Discussion Bled-Moscow-CERN-UK-Marburg-Liege-USA}
\label{dm}
\end{center}
\end{figure}
\section{Conclusions}
Current VIA activity is integrated in the structure of APC
laboratory and includes regular weekly videoconferences with VIA
lectures, online transmissions of APC Colloquiums and Seminars, a
solid library of their records and presentations, together with the work of
multi-lingual VIA Internet forum.
The Scientific-Educational complex of Virtual Institute of
Astroparticle physics can provide regular communications between
different groups and scientists, working in different scientific
fields and parts of the world, get the first-hand information on the
newest scientific results, as well as to support various educational
programs at distance. This activity would easily allow finding
mutual interest and organizing task forces for different scientific
topics of astroparticle physics and related topics. It can help in
the elaboration of strategy of experimental particle, nuclear,
astrophysical and cosmological studies as well as in proper analysis
of experimental data. It can provide young talented people from all
over the world to get the highest level education, come in direct
interactive contact with the world known scientists and to find
their place in the fundamental research.
VIA applications can go far beyond the particular tasks of astroparticle physics and give rise to an interactive system of mass media communications.
VIA sessions became a natural part of a program of Bled Workshops,
opening the room of discussions of physics beyond the Standard Model
for distant participants from all the world. The experience of VIA applications at Bled Workshops plays important role in the development of VIA facility as an effective tool of $e-science$ and $e-learning$.
\section*{Acknowledgements}
The initial step of creation of VIA was
supported by ASPERA. I am grateful to P.Binetruy, J.Ellis and S.Katsanevas for
permanent stimulating support, to J.C. Hamilton for support in VIA
integration in the structure of APC laboratory,
to K.Belotsky, A.Kirillov and K.Shibaev for assistance in
educational VIA program, to A.Mayorov, A.Romaniouk and E.Soldatov for fruitful
collaboration, to M.Pohl, C. Kouvaris, J.-R.Cudell,
C. Giunti, G. Cella, G. Fogli and F. DePaolis for cooperation in the tests of VIA online
transmissions in Switzerland, Belgium and Italy and to D.Rouable for help in
technical realization and support of VIA complex.
I express my gratitude to N.S. Manko\v c Bor\v stnik, G.Bregar, D. Lukman and all
Organizers of Bled Workshop for cooperation in the organization of VIA Sessions at XIV Bled Workshop.
|
\section{Introduction}
Let $\mathcal{M}(r,n)$ be the moduli space of \textit{framed sheaves} on $\mathbb{CP}^2$, that is, the moduli space of pairs $(E,\alpha)$ modulo isomorphism, where $E$ is a torsion-free sheaf on $\mathbb{CP}^2$ of rank $r$ with $c_2(E)=n$, locally trivial in a neighborhood of a fixed line $l_{\infty}$, and $\alpha\colon E\vert_{l_{\infty}}\stackrel{\sim}{\rightarrow} \mathcal{O}_{l_{\infty}}^{\oplus r}$ is the \textit{framing at infinity}. $\mathcal{M}(r,n)$ is a nonsingular quasi-projective variety of dimension $2rn.$ This moduli space also admits a description in terms of linear data, the so-called \textit{ADHM} data (see, for example, Chapter 2 of Nakajima's book \cite{book:nakajima1999}).
As described in \cite[Chapter 3]{book:nakajima1999}, by using the ADHM data description, the moduli space $\mathcal{M}(r,n)$ can be realized as a hyper-K\"ahler quotient. By fixing a complex structure within the hyper-K\"ahler family of complex structures on $\mathcal{M}(r,n)$, one can define a holomorphic symplectic form on $\mathcal{M}(r,n).$
In addition to the case of $\mathbb{CP}^2$, the only other relevant result in the literature about symplectic structures on moduli spaces of framed sheaves is due to Bottacin \cite{art:bottacin2000}.
Let $X$ be a complex nonsingular projective surface, $D$ an effective divisor such that $D=\sum_{i=1}^n C_i$, where $C_i$ is an integral curve for $i=1,\ldots, n$, and $F_D$ a locally free $\mathcal{O}_D$-module. We call $(D,F_D)$-\emph{framed vector bundle} on $X$ a pair $(E,\alpha)$, where $E$ is a locally free sheaf on $X$ and $\alpha\colon E\vert_D\stackrel{\sim}{\rightarrow} F_D$ is an isomorphism. Let us fix a Hilbert polynomial $P.$ Bottacin constructs Poisson structures on the moduli space $\mathcal{M}_{lf}^*(X;F_D,P)$ of framed vector bundles on $X$ with Hilbert polynomial $P$, that are induced by global sections of the line bundle $\omega_X^{-1}(-2\,D).$ In particular, when $X$ is the complex projective plane, $D=l_\infty$ is a line and $F_D$ the trivial vector bundle of rank $r$ on $l_\infty$, this yields a symplectic structure on the moduli space $\mathcal{M}_{lf}(r,n)$ of framed vector bundles on $\mathbb{CP}^2$, induced by the standard holomorphic symplectic structure of $\mathbb{C}^2=\mathbb{CP}^2\setminus l_\infty.$ It is not known if this symplectic structure is equivalent to the one given by the ADHM construction.
Bottacin's result can be seen as a generalization to the framed case of the construction of Poisson brackets and holomorphic symplectic two-forms on the moduli spaces of Gieseker-stable torsion-free sheaves on $X.$ We recall briefly the main results for torsion-free sheaves. In \cite{art:mukai1984}, Mukai proved that any moduli space of simple sheaves on a $K3$ or abelian surface has a non-degenerate holomorphic two-form. Its closedness was later proved by Mukai in \cite{art:mukai1987}. Mukai's result was generalizated to moduli spaces of simple vector bundles on symplectic K\"ahler manifolds by Ran \cite{art:ran1996} and to moduli spaces of Gieseker-stable vector bundles over surfaces of general type and over Poisson surfaces by Tyurin \cite{art:tyurin1988}; a more thorough study of the Poisson case was made by Bottacin in \cite{art:bottacin1995}. In \cite{art:ogrady1992}, by using these results O' Grady defined closed two-forms on algebraic varieties parametrizing flat families of coherent sheaves. In all these cases, the symplectic two-form is defined in terms of the Atiyah class.
In the present paper we define a modified Atiyah class of a family of framed sheaves, which allows us to describe a framed version of the Kodaira-Spencer map and to construct closed two-forms on the moduli spaces of framed sheaves, that under some conditions are symplectic.
More precisely, let $X$ be a nonsingular projective surface over an algebraically closed field $k$ of characteristic zero, $D\subset X$ a divisor, $F_D$ a locally free $\mathcal{O}_D$-module and $S$ a Noetherian $k$-scheme of finite type. A flat family of $(D,F_D)$-framed sheaves parametrized by $S$ is a pair $\mathcal{E}:=(E,\alpha)$ on $S\times X$, such that $E$ is flat over $S$ and all the restrictions to the fibres $\{s\}\times X$ are $(D,F_D)$-framed sheaves on $X.$
Let $\mathcal{E}=(E,\alpha)$ be a $S$-flat family of $(D,F_D)$-framed sheaves. We introduce the \emph{framed sheaf of first jets} $\jfr(\mathcal{E})$ as the subsheaf of the sheaf of first jets $\jf(E)$ (introduced by Atiyah in \cite{art:atiyah1957}) consisting of those sections whose $p_S^*(\Omega_S^1)$-part vanishes along $S\times D.$ We define the \emph{framed Atiyah class} $at(\mathcal{E})$ of $\mathcal{E}$ as an extension class of $\jfr(\mathcal{E})$ in
\begin{equation*}
\mathrm{Ext}^1(E,\left(p_S^*(\Omega^1_S)(-S\times D)\oplus p_X^*(\Omega_X^1)\right)\otimes E),
\end{equation*}
where $p_S^*(\Omega^1_S)(-S\times D)=p_S^*(\Omega^1_S)\otimes\mathcal{O}_{S\times X}(-S\times D)=p_S^*(\Omega^1_S)\otimes p_X^*(\mathcal{O}_X(-D)).$
Starting from the framed Atiyah class $at(\mathcal{E})$, one can define a section $\mathcal{A}t_S(\mathcal{E})$ in
\begin{equation*}
\mathrm{H}^0(S,\mathcal{E}xt_{p_S}^1(E, p_S^*(\Omega^1_S)\otimes p_X^*(\mathcal{O}_X(-D))\otimes E)),
\end{equation*}
where $p_S\colon S\times X \rightarrow S$ is the projection.
In the same way as in the nonframed case (cf. \cite[Section 10.1.8]{book:huybrechtslehn2010}), by using $\mathcal{A}t_S(\mathcal{E})$ one can define the \emph{framed} Kodaira-Spencer map associated to $\mathcal{E}$:
\begin{equation*}
KS_{fr}\colon (\Omega_S^1)^\vee\longrightarrow \mathcal{E}xt_{p_S}^1(E,p_X^*(\mathcal{O}_X(-D))\otimes E).
\end{equation*}
$(D,F_D)$-framed sheaves are a particular case of \emph{framed modules}, whose theory was developed by Huybrechts and Lehn \cite{art:huybrechtslehn1995-I, art:huybrechtslehn1995-II}. In order to construct moduli spaces parametrizing these objects, they defined a notion of semistability depending on a polarization and a rational polynomial $\delta$ of degree less than the dimension of the ambient variety and positive leading coefficient. They proved that there exists a coarse moduli space parametrizing semistable framed modules and a fine moduli space parametrizing stables ones. Moduli spaces of stable $(D,F_D)$-framed sheaves turn out to be open subschemes of the fine moduli spaces of stable framed modules (cf. \cite{art:bruzzomarkushevich2011}).
Let $X$ be a nonsingular projective surface over $k$ equipped with an ample line bundle $\mathcal{O}_X(1).$ We consider as above pairs $(D,F_D).$ Let $\delta\in \mathbb{Q}[n]$ be a stability polynomial and $P$ a numerical polynomial of degree two.
Let us denote by $\mathcal{M}^*_\delta(X;F_D,P)$ the moduli space of $(D,F_D)$-framed sheaves on $X$ with Hilbert polynomial $P$ that are stable with respect to $\mathcal{O}_X(1)$ and $\delta.$ Let $\mathcal{M}^*_\delta(X;F_D,P)^{sm}$ be the smooth locus of $\mathcal{M}^*_\delta(X;F_D,P)$ and $\tilde{\mathcal{E}}=(\tilde{E},\tilde{\alpha})$ a \emph{universal object} over $\mathcal{M}^*_\delta(X;F_D,P)^{sm}.$
The first result we obtain by using the framed Atiyah class is the following:
\begin{theorem}
The framed Kodaira-Spencer map defined by $\tilde{\mathcal{E}}$ induces a canonical isomorphism
\begin{equation*}
KS_{fr}\colon T\mathcal{M}^*_\delta(X;F_D,P)^{sm}\stackrel{\sim}{\longrightarrow} \mathcal{E}xt_p^1(\tilde{E}, \tilde{E}\otimes p_X^*(\mathcal{O}_X(-D))),
\end{equation*}
where $p$ is the projection from $\mathcal{M}^*_\delta(X;F_D,P)^{sm}\times X$ to $\mathcal{M}^*_\delta(X;F_D,P)^{sm}.$
\end{theorem}
Thus we get a generalization to the framed case of the corresponding statement for the moduli space of Gieseker-stable torsion-free sheaves on $X$ (cf. \cite[Theorem 10.2.1]{book:huybrechtslehn2010}).
From this theorem it follows that for any point $[(E,\alpha)]$ of $\mathcal{M}^*_\delta(X;F_D,P)^{sm}$, the vector space $\mathrm{Ext^1}(E,E(-D))$ is naturally identified with the tangent space $T_{[(E,\alpha)]}\mathcal{M}^*_\delta(X;F_D,P).$ For any $\omega\in \mathrm{H}^0(X,\omega_X(2\,D))$, we can define a skew-symmetric bilinear form
\begin{eqnarray*}
&&\mathrm{Ext}^1(E,E(-D))\times \mathrm{Ext}^1(E,E(-D))\stackrel{\circ}{\longrightarrow} \mathrm{Ext}^2(E,E(-2\,D))\\
&&\stackrel{tr}{\longrightarrow} \mathrm{H}^2(X, \mathcal{O}_X(-2\,D))\stackrel{\cdot\, \omega}{\longrightarrow} \mathrm{H}^2(X,\omega_X)\cong k.
\end{eqnarray*}
By varying $[(E,\alpha)]$, these forms define an exterior two-form $\tau(\omega)$ on $\mathcal{M}^*_\delta(X;F_D,P)^{sm}.$
We prove that $\tau(\omega)$ is a closed form (cf. Theorem \ref{thm:closed}) and provide a criterion of its non-degeneracy (cf. Proposition \ref{prop:non-dege}). In particular, if the line bundle $\omega_X(2\,D)$ is trivial, the two-form $\tau(1)$ induced by $1\in \mathrm{H}^0(X,\omega_X(2\,D))\cong k$ defines a symplectic structure on $\mathcal{M}^*_\delta(X;F_D,P)^{sm}.$
As an application, we construct holomorphic symplectic structures on moduli spaces of $(D,F_D)$-framed sheaves on some birationally ruled surfaces. In particular, for $\mathbb{CP}^1\times \mathbb{CP}^1$, $\mathbb{C}\times \mathbb{CP}^1$, with $C$ elliptic curve, and the second Hirzebruch surface $\mathbb{F}_2$, under a suitable choice of $D$ and for $F_D=\mathcal{O}_D^{\oplus r}$, we get a generalization to the non-locally free case of Bottacin's construction of symplectic structures induced by non-degenerate Poisson structures (cf. \cite{art:bottacin2000}).
For $X=\mathbb{CP}^2$ or the blowup of $\mathbb{CP}^2$ at a point, under a suitable choice of $D$ and $F_D$, we obtain new examples of non-compact holomorphic symplectic varieties, not covered by Bottacin's construction.
This paper is structured as follows. In Section \ref{sec:preliminaries}, we briefly introduce the theory of framed modules and framed sheaves and state the main theorems about the existence of moduli spaces parametrizing these objects. In Section \ref{sec:atiyah} we recall the definition of the Atiyah class for a flat family of coherent sheaves. In Section \ref{sec:framedatiyah} we give the definitions of the framed version of the Atiyah class and of the Kodaira-Spencer map. In Section \ref{sec:framedkodaira} we prove that the framed Kodaira-Spencer map is an isomorphism for the moduli space of stable $(D,F_D)$-framed sheaves and, in Section \ref{sec:framedtwoforms}, we construct closed two-forms on it. Finally, in Section \ref{sec:example} we apply our result to some particular birationally ruled surfaces, and define a symplectic structure on the moduli spaces of (stable) $(D,F_D)$-framed sheaves on them.
\subsection*{Conventions}
All schemes are Noetherian of finite type over an algebraically closed field $k$ of characteristic zero. A \emph{variety} is a reduced separated scheme. A \emph{polarized variety of dimension} $d$ is a pair $(X,\mathcal{O}_X(1))$, where $X$ is a smooth connected projective variety of dimension $d$, defined over $k$, and $\mathcal{O}_X(1)$ a very ample line bundle. The \emph{canonical line bundle} of $X$ is denoted by $\omega_X$ and its associated divisor class by $K_X.$
Let $(X,\mathcal{O}_X(1))$ be a polarized variety and $S$ a scheme. We denote by $\mathcal{X}$ the cartesian product $S\times X$, and by $p_S$, $p_X$ the projections from $\mathcal{X}$ to $S$ and $X$ respectively. Let $D$ be an effective divisor on $X$: we denote by $\mathcal{D}$ the cartesian product $S\times D$ and for any coherent sheaf $F$ on $\mathcal{X}$, $F(-\mathcal{D})$ is the tensor product $F\otimes p_X^*(\mathcal{O}_X(-D)).$
We denote by $NS(X)$ the Néron-Severi group of a smooth connected projective variety $X$, that is, the image of the map $c_1\colon \mathrm{Pic}(X)\rightarrow \mathrm{H}^2(X,\mathbb{Z}).$
As usual, a polarized variety of dimension two is called a \emph{polarized surface}.
Finally, we denote by $\HE^i(E^\bullet, G^\bullet)$ the $i$-th hyper-Ext group of two finite complexes of locally free sheaves $E^\bullet$ and $G^\bullet$ on a scheme $Y$, that is, the $i$-th hyper-cohomology group of the total complex associated to the double complex $C^\bullet(\mathcal{H}om^\bullet(E^\bullet, G^\bullet)).$ As it is explained in \cite[Section 10.1]{book:huybrechtslehn2010}, for any coherent sheaf $N$ on $Y$, one can define the trace map $\mathrm{tr}\colon \HE^i(E^\bullet,E^\bullet\otimes N)\rightarrow \mathrm{H}^i(Y,N).$
Moreover, if $E$ is a coherent sheaf on $Y$ that admits a finite locally free resolution $E^\bullet\rightarrow E$, we have $\HE^i(E^\bullet, E^\bullet\otimes N)\cong \mathrm{Ext}^i(E,E\otimes N)$ for any locally free sheaf $N.$ In this case, we denote by $\mathrm{Ext}^i(E,E\otimes N)_0$ the kernel
\begin{equation*}
\ker[\mathrm{tr}\colon \mathrm{Ext}^i(E,E\otimes N)\rightarrow \mathrm{H}^i(Y,N)].
\end{equation*}
\subsubsection*{Acknowledgements}
This paper is largely based upon the author’s PhD thesis \cite{phd:sala2011}. The author thanks his supervisors Ugo Bruzzo and Dimitri Markushevich for suggesting this problem and for their constant support. He thanks the referees for useful remarks and suggestions which helped to improve the paper. He also thanks Francesco Bottacin, Laurent Manivel and Christoph Sorger for interesting discussions. This paper was mostly written while the author was staying at SISSA, Université Lille 1 and IHÉS. He thanks those institutions for hospitality and support. He was partially supported by PRIN ``Geometria delle varietà algebriche e dei loro spazi di moduli'', co-funded by MIUR (cofin 2008), by the grant of the French Agence Nationale de Recherche VHSMOD-2009 Nr. ANR-09-BLAN-0104, by the European Research Network ``GREFI-GRIFGA'' and by the ERASMUS ``Student Mobility for Placements'' Programme.
\section{Preliminaries on framed sheaves}\label{sec:preliminaries}
In this section we introduce the notions of $(D,F_D)$-framed sheaves and framed modules. Moreover, we recall the main results about the construction of moduli spaces of these objects.
Let $(X,\mathcal{O}_X(1))$ be a polarized surface.
\begin{definition}\label{def:locfree}
Let $D$ be an effective divisor of $X$ and $F_D$ a coherent sheaf on $X$, which is a locally free $\mathcal{O}_{D}$-module. We say that a coherent sheaf $E$ on $X$ is $(D,F_D)$\textit{-framable} if $E$ is torsion-free, locally free in a neighborhood of $D$, and there is an isomorphism $E\vert_D\stackrel{\sim}{\rightarrow} F_D.$ An isomorphism $\alpha\colon E\vert_D\stackrel{\sim}{\rightarrow} F_D$ will be called a $(D,F_D)$-\textit{framing} of $E.$ A $(D,F_D)$-\textit{framed sheaf} is a pair $\mathcal{E}:=(E,\alpha)$ consisting of a $(D,F_D)$-framable sheaf $E$ and a $(D,F_D)$-framing $\alpha.$ Two $(D,F_D)$-framed sheaves $(E,\alpha)$ and $(E',\alpha')$ are isomorphic if there is an isomorphism $f\colon E\rightarrow E'$ such that $\alpha'\circ f\vert_D=\alpha.$
\end{definition}
\begin{remark}
Our notion of framing is the same as the one used in M. Lehn's Ph.D. thesis \cite{phd:lehn1993} and in T. A. Nevins' papers \cite{art:nevins2002-I, art:nevins2002-II}. \triend
\end{remark}
We shall call $(D,F_D)$\emph{-framed vector bundles} on $X$ those $(D,F_D)$-framed sheaves whose underlying coherent sheaf is locally free.
Note that $(D,F_D)$-framed sheaves are a particular type of \emph{framed modules}.
\begin{definition}[\cite{art:huybrechtslehn1995-I, art:huybrechtslehn1995-II}]
Let $F$ be a coherent sheaf on $X.$ A $F$-\textit{framed module} on $X$ is a pair $(E,\alpha)$, where $E$ is a coherent sheaf on $X$ and $\alpha\colon E\rightarrow F$ a morphism of coherent sheaves.
\end{definition}
In \cite{art:huybrechtslehn1995-I, art:huybrechtslehn1995-II}, Huybrechts and Lehn developed the theory of framed modules. In the case of polarized surfaces, they introduced a notion of (semi)stability depending on the polarization $\mathcal{O}_X(1)$ and a rational polynomial $\delta(n)=\delta_1 n +\delta_0$, with $\delta_1>0.$ We shall call $\delta$ a \emph{stability polynomial}. When the framing $\alpha$ is zero, this reduces to Gieseker's (semi)stability condition for torsion-free sheaves.
Fix a coherent sheaf $F$ and a numerical polynomial $P$ of degree two. Let us denote by $\underline{\mathcal{M}}^{ss}_\delta(X;F,P)$ (resp. $\underline{\mathcal{M}}^{s}_\delta(X;F,P)$) the contravariant functor from the category of schemes to the category of sets, that associates with every scheme $S$ the set of isomorphism classes of flat families $(G,\beta\colon G\rightarrow p_X^*(F))$ of semistable (resp. stable) $F$-framed modules with Hilbert polynomial $P$ parametrized by $S.$ The main result in their papers is the following:
\begin{theorem}[\cite{art:huybrechtslehn1995-I, art:huybrechtslehn1995-II}]
There exists a projective scheme $\mathcal{M}^{ss}_\delta(X;F,P)$ which corepresents the moduli functor $\underline{\mathcal{M}}^{ss}_\delta(X;F,P)$ of semistable $F$-framed modules. Moreover, there is an open subscheme $\mathcal{M}^{s}_\delta(X;F,P)$ of $\mathcal{M}^{ss}_\delta(X;F,P)$ that represents the moduli functor $\underline{\mathcal{M}}^{s}_\delta(X;F,P)$ of stable $F$-framed modules, i.e., $\mathcal{M}^{s}_\delta(X;F,P)$ is a fine moduli space for stable $F$-framed modules.
\end{theorem}
``Fine'' means the existence of a \emph{universal} $F$-framed module over $\mathcal{M}^{s}_\delta(X;F,P)$, that is, a pair $(\tilde{E},\tilde{\alpha}\colon \tilde{E}\rightarrow p_X^{*}(F))$, where $\tilde{E}$ is a coherent sheaf on $\mathcal{M}^{s}_\delta(X;F,P)\times X$, flat over $\mathcal{M}^{s}_\delta(X;F,P)$, such that for any scheme $S$ and any family of stable $F$-framed modules $(G,\beta)\in \underline{\mathcal{M}}^{s}_\delta(X;F,P)(S)$ parametrized by $S$, there exists a unique morphism $g\colon S\rightarrow \mathcal{M}^{s}_\delta(X;F,P)$ such that the pull back of $(\tilde{E},\tilde{\alpha})$ is isomorphic to $(G,\beta).$
By using Huybrechts and Lehn's result, Bruzzo and Markushevich constructed a moduli space parametrizing \emph{all} isomorphism classes of $(D,F_D)$-framed sheaves on $X$, under some mild assumptions on the divisor $D$ and on $F_D.$ More precisely, they proved the following:
\begin{theorem}[\cite{art:bruzzomarkushevich2011}]\label{thm:bruzzomarkushevich}
Let $X$ be a surface, $D\subset X$ a big and nef curve and $F_D$ a Gieseker-semistable locally free $\mathcal{O}_D$-module. Then for any numerical polynomial $P$ of degree two, there exists a polarization $\mathcal{O}_X(1)$ and a stability polynomial $\delta(n)=\delta_1 n+\delta_0$ for which there is an open subset $\mathcal{M}^*(X;F_D,P)\subset \mathcal{M}^{s}_\delta(X;F_D,P)$ of the moduli space of stable $F_D$-framed modules, which is a fine moduli space for $(D,F_D)$-framed sheaves on $X$ with Hilbert polynomial $P.$ In particular, $\mathcal{M}^*(X;F_D,P)$ is a quasi-projective separated scheme.
\end{theorem}
Now we give a characterization of the smoothness of the moduli space $\mathcal{M}^*(X;F_D,P).$
\begin{theorem}[\cite{art:bruzzomarkushevich2011}]
Let $[(E,\alpha)]$ be a point in $\mathcal{M}^*(X;F_D,P).$ The following statements hold:
\begin{itemize}
\item[(i)] The Zariski tangent space of $\mathcal{M}^*(X;F_D,P)$ at $[(E,\alpha)]$ is naturally isomorphic to the Ext-group $\mathrm{Ext}^1(E,E(-D)).$
\item[(ii)] If $\mathrm{Ext}^2(E,E(-D))_0$ vanishes, then $\mathcal{M}^*(X;F_D,P)$ is smooth at $[(E,\alpha)].$
\end{itemize}
\end{theorem}
\begin{corollary}\label{cor:smoothness}
Let $X$ be a rational surface over $\mathbb{C}$, $D\subset X$ a smooth connected big and nef curve of genus zero and $F_D\cong \mathcal{O}_D^{\oplus r}.$ Let $n\in \mathbb{Z}$ and $c\in NS(X)$ such that $\int_{D} c=0.$ Then the moduli space $\mathcal{M}^*(X;F_D,r,c,n)$ parametrizing isomorphism classes of $(D,F_D)$-framed sheaves on $X$ of rank $r$, first Chern class $c$ and second Chern class $n$ is a smooth quasi-projective variety of dimension
\begin{equation*}
\dim_{\mathbb{C}} \mathcal{M}^*(X;F_D,r,c,n)=2rn-(r-1)\int_X c^2.
\end{equation*}
\end{corollary}
\proof
By \cite[Proposition 2.1]{art:gasparimliu2010}, for any point $[(E,\alpha)]\in \mathcal{M}^*(X;F_D,P)$ we have
\begin{equation*}
\mathrm{Ext}^0(E,E(-D))=\mathrm{Ext}^2(E,E(-D))=0.
\end{equation*}
Hence the moduli space $\mathcal{M}^*(X;F_D,P)$ is smooth. The formula for the dimension follows from the Hirzebruch-Riemann-Roch theorem.
\endproof
\section{The Atiyah class}\label{sec:atiyah}
In this section we recall the notion of the \emph{Atiyah class} for flat families of coherent sheaves. The Atiyah class was introduced by Atiyah in \cite{art:atiyah1957} for the case of coherent sheaves and by Illusie in \cite{book:illusie1971, book:illusie1972} for any complex of coherent sheaves (see \cite[Section 10.1.5]{book:huybrechtslehn2010} for a description of the Atiyah class in terms of \v{C}ech cocycles). Atiyah's approach involves the notion of the \emph{sheaf of first jets} of a fixed coherent sheaf associated to the sheaf of one-forms (for a generalization of the sheaf of first jets to quotients of the sheaf of one-forms, see Maakestad's paper \cite{art:maakestad2010-I}).
Let $Y$ be a scheme.
\begin{definition}
Let $E$ be a coherent sheaf on $Y.$ We call \textit{sheaf of the first jets} $\jf(E)$ of $E$ the coherent sheaf of $\mathcal{O}_Y$-modules defined as follows:
\begin{itemize}
\item as a sheaf of $k$-modules, we set $\jf(E):=(\Omega_Y^1\otimes E)\oplus E$,
\item for any $y\in Y$, $a\in \mathcal{O}_{Y,y}$ and $(z\otimes e, f)\in \jf(E)_y$, we define
\begin{equation*}
a(z\otimes e, f):=(az\otimes e +d(a)\otimes f, af),
\end{equation*}
where $d$ is the exterior differential of $Y.$
\end{itemize}
\end{definition}
The sheaf $\jf(E)$ fits into an exact sequence of coherent sheaves
\begin{equation}\label{eq:atiyahclass}
0\longrightarrow \Omega^1_Y\otimes E \longrightarrow \jf(E)\longrightarrow E\longrightarrow 0.
\end{equation}
\begin{definition}
Let $E$ be a coherent sheaf on $Y.$ We call \textit{Atiyah class} of $E$ the class $at(E)$ in $\mathrm{Ext}^1(E,\Omega_Y^1\otimes E)$ associated to the extension \eqref{eq:atiyahclass}.
\end{definition}
It is a well known fact that the Atiyah class $at(E)$ is the obstruction for the existence of an algebraic connection on $E.$ This means the following:
\begin{proposition}{\normalfont (\cite[Proposition 3.4]{art:maakestad2010-I}).}
Let $E$ be a coherent sheaf on $Y.$ The Atiyah class $at(E)$ is zero if and only if there exists a connection on $E.$
\end{proposition}
\section{The Atiyah class for framed sheaves}\label{sec:framedatiyah}
In this section we turn to the framed case. First, we introduce the \emph{framed sheaf of first jets} $\jfr(\mathcal{E})$ of a $S$-flat family $\mathcal{E}=(E,\alpha)$ of $(D,F_D)$-framed sheaves as the subsheaf of the sheaf of first jets $\jf(E)$ consisting of those sections whose $p_S^*(\Omega_S^1)$-part vanishes along $S\times D.$ Then we define a \emph{framed} version of the Atiyah class and of the Newton polynomials\footnote{In the nonframed case, the Newton polynomials are introduced, for example, in \cite[Section 10.1.6]{book:huybrechtslehn2010}.} for $S$-flat families. We use this \emph{relative} approach, because later on we shall want to consider the framed Atiyah class of a \emph{universal framed sheaf}.
From now on we fix the pair $(D,F_D)$ and we just say \emph{framed sheaf} for a $(D,F_D)$-framed sheaf.
\begin{definition}
Let $S$ be a scheme. A \textit{flat family of framed sheaves parametrized by} $S$ is a pair $\mathcal{E}=(E,\alpha)$ where $E$ is a coherent sheaf on $\mathcal{X}$, flat over $S$, and $\alpha\colon E\rightarrow p_X^*(F_D)$ is a morphism such that for any $s\in S$ the pair $(E\vert_{\{s\}\times X},\alpha\vert_{\{s\}\times X})$ is a $\left(\{s\}\times D,p_X^*(F_D)\vert_{\{s\}\times D}\right)$-framed sheaf on $\{s\}\times X.$
\end{definition}
Let $\mathcal{E}=(E,\alpha)$ be a flat family of framed sheaves parametrized by a scheme $S.$ We define a subsheaf $\jfr(\mathcal{E})$ of $\jf(E)$, that we shall call \emph{framed sheaf of first jets} of $\mathcal{E}.$
For a point $x\in \mathcal{X}$ such that $x\notin \mathcal{D}$, we set $\jfr(\mathcal{E})_x:=\jf(E)_x.$
Fix $x\in \mathcal{D}.$ By definition of a flat family of framed sheaves, $\left(E\vert_{\{s\}\times X}\right)_x$ is a free module for $s\in S$ such that $p_S(x)=s$, hence $E_x$ is a free module (cf. \cite[Lemma 2.1.7]{book:huybrechtslehn2010}). Therefore there exists an open neighborhood $V\subset \mathcal{X}$ of $x$ such that $E\vert_V$ is a locally free $\mathcal{O}_V$-module.
We denote by $E_V$ the restriction of $E$ to $V.$ Let $\mathcal{D}':=V\cap \mathcal{D}$ and $\mathcal{U}=\{U_i\}_{i\in I}$ a cover of $\mathcal{D}'$ over which $p_X^{*}(F_D)\vert_{\mathcal{D}'}$ trivializes, and choose on any $U_i$ a set $\{e_i^0\}$ of basis sections of $\Gamma(p_X^{*}(F_D)\vert_{\mathcal{D}'},U_i).$ Let $g^0_{ij}$ be transition functions of $p_X^{*}(F_D)\vert_{\mathcal{D}'}$ with respect to chosen local basis sections (i.e., $e^0_i=g^0_{ij}e^0_j$), constant along $S.$ Let us fix a cover $\mathcal{W}=\{W_i\}_{i\in I}$ of $V$ over which $E_V$ trivializes with sets $\{e_i\}$ of basis sections such that $W_i\cap \mathcal{D}'=U_i$ for any $i\in I$ and
\begin{eqnarray*}
e_i\vert_{\mathcal{D}'}&=& e_i^0, \\
g_{ij}\vert_{\mathcal{D}'}&=& g_{ij}^0.
\end{eqnarray*}
Let $z^1_i, \ldots, z^s_i$ and $z^{s+1}_i, \ldots, z^t_i$ be the local coordinates of $S$ and $X$ on $W_i$, respectively, and $f_i=0$ the local equation of $\mathcal{D}$ on $W_i.$ Define $\jfr(\mathcal{E})_x\subset \jf(E)_x=(\Omega_{\mathcal{X},x}^1\otimes E_x)\oplus E_x$ as the $\mathcal{O}_{\mathcal{X},x}$-module spanned by the basis obtained by tensoring all the elements of the set $\{f_i d z^1_i, \ldots, f_i d z^s_i, d z^{s+1}_i,\ldots, d z^t_i\}$, where $\{dz^1_i, \ldots, dz^t_i\}$ is a basis of $\Omega_{\mathcal{X},x}^1$, by the elements of the basis $\{e_i\}:=\{e^1_i, \ldots, e^r_i\}$ of $E_x$ and then adding the elements of $\{e_i\}.$ Thus an arbitrary element of $\jfr(\mathcal{E})_x$ is of the form
\begin{equation*}
h_i + f_i \sum_{n=1}^r \sum_{k=1}^s \chi_{n,k}\, e_i^n\otimes d z_i^k+\sum_{m=1}^r \sum_{l=s+1}^t \psi_{m,l}\, e_i^m\otimes d z_i^l,
\end{equation*}
where $h_i\in E_x$ and $\chi_{n,k}, \psi_{m,l}\in \mathcal{O}_{X,x}$, for $m,n=1, \ldots r$, $k=1,\ldots, s$ and $l=s+1,\ldots, t.$
If $x$ is also a point in the open subset $W_j$ of $\mathcal{W}$, let us denote by $l_{ij}\in \mathcal{O}_V^*(W_i\cap W_j)$ the transition function on $W_i\cap W_j$ of the line bundle associated to the divisor $\mathcal{D}_V$ and by $J_{ij}$ the Jacobian matrix of change of coordinates. Let us define the following matrices:
\begin{equation*}
L_{ij}:=\left(\begin{array}{cc}
l_{ij}I_s & 0_{s,t-s} \\
0_{t-s,s} & I_{t-s}
\end{array}\right)
\end{equation*}
and
\begin{equation*}
F_i:=\left(\begin{array}{cc}
f_i I_s & 0_{s,t-s} \\
0_{t-s,s} & I_{t-s}
\end{array}\right)
\end{equation*}
where $I_k$ is the identity matrix of order $k$ and $0_{k,l}$ is the $k$-by-$l$ zero matrix.
The matrix which expresses a change of basis in $\jfr(\mathcal{E})_x$ under a change of basis in $E_x$ is:
\begin{equation*}
\left(\begin{array}{cc}
L_{ij}\otimes g_{ij} & (F_i^{-1}\otimes \mathrm{id})\cdot dg_{ij} \\
0 & g_{ij}
\end{array}\right),
\end{equation*}
where the block at the position (1,2) is a regular matrix function, because $g_{ij}$ is constant along $\mathcal{D}_V.$ The matrix corresponding to a change of local coordinates is:
\begin{equation*}
\left(\begin{array}{cc}
L_{ij}\cdot J_{ij}\otimes \mathrm{id} & 0 \\
0 & \mathrm{id}
\end{array}\right).
\end{equation*}
The framed sheaf of first jets $\jfr(\mathcal{E})$ of $\mathcal{E}$ fits into an exact sequence of coherent sheaves of $\mathcal{O}_{\mathcal{X}}$-modules:
\begin{equation}\label{eq:framedatiyahclass}
0\longrightarrow \left(p_S^*(\Omega^1_S)(-\mathcal{D})\oplus p_X^*(\Omega_X^1)\right)\otimes E \longrightarrow \jfr(\mathcal{E}) \longrightarrow E\longrightarrow 0.
\end{equation}
\begin{definition}
Let $\mathcal{E}=(E,\alpha)$ be a flat family of framed sheaves parametrized by a scheme $S.$ We call \textit{framed Atiyah class} of the family $\mathcal{E}$ the class $at(\mathcal{E})$ in
\begin{equation*}
\mathrm{Ext}^1(E,\left(p_S^*(\Omega^1_S)(-\mathcal{D})\oplus p_X^*(\Omega_X^1)\right)\otimes E)
\end{equation*}
associated to the extension \eqref{eq:framedatiyahclass}.
\end{definition}
Let us consider the short exact sequence
\begin{equation*}
0\longrightarrow p_S^{*}(\Omega_S^1)(-\mathcal{D})\oplus p_X^*(\Omega_X^1)\stackrel{i}{\longrightarrow} \Omega^1_{\mathcal{X}}\stackrel{q}{\longrightarrow} p_S^{*}(\Omega_S^1)\vert_{\mathcal{D}}\longrightarrow 0.
\end{equation*}
After tensoring by $E$ and applying the functor $\mathrm{Hom}(E,\cdot)$, we get the long exact sequence
\begin{eqnarray*}
&&\cdots\rightarrow\mathrm{Ext}^1(E, \left(p_S^{*}(\Omega_S^1)(-\mathcal{D})\oplus p_X^*(\Omega_X^1)\right)\otimes E)\stackrel{i_*}{\longrightarrow}\\
&&\mathrm{Ext}^1(E,\Omega^1_{\mathcal{X}}\otimes E)\stackrel{q_*}{\longrightarrow}\mathrm{Ext}^1(E, p_S^{*}(\Omega_S^1)\vert_{\mathcal{D}}\otimes E)\rightarrow\cdots.
\end{eqnarray*}
By construction, the image of $at(\mathcal{E})$ under $i_*$ is $at(E)$, which is equivalent to saying that we have the commutative diagram
\begin{equation*}
\begin{tikzpicture}[xscale=3.2,yscale=-1.2]
\node (A0_0) at (0, 0) {$0$};
\node (A0_1) at (1, 0) {$\left(p_S^*(\Omega^1_S)(-\mathcal{D})\oplus p_X^*(\Omega_X^1)\right)\otimes E$};
\node (A0_2) at (2.5, 0) {$\jfr(\mathcal{E})$};
\node (A0_3) at (3.5, 0) {$E$};
\node (A0_4) at (4, 0) {$0$};
\node (A1_0) at (0, 1) {$0$};
\node (A1_1) at (1, 1) {$\Omega_{\mathcal{X}}^1\otimes E$};
\node (A1_2) at (2.5, 1) {$\jf(E)$};
\node (A1_3) at (3.5, 1) {$E$};
\node (A1_4) at (4, 1) {$0$};
\path (A0_1) edge [->]node [auto] {$\scriptstyle{}$} (A1_1);
\path (A0_0) edge [->]node [auto] {$\scriptstyle{}$} (A0_1);
\path (A0_1) edge [->]node [auto] {$\scriptstyle{}$} (A0_2);
\path (A1_0) edge [->]node [auto] {$\scriptstyle{}$} (A1_1);
\path (A0_3) edge [thin, double distance=1.5pt]node [auto] {$\scriptstyle{}$} (A1_3);
\path (A1_1) edge [->]node [auto] {$\scriptstyle{}$} (A1_2);
\path (A0_3) edge [->]node [auto] {$\scriptstyle{}$} (A0_4);
\path (A0_2) edge [->]node [auto] {$\scriptstyle{}$} (A1_2);
\path (A1_2) edge [->]node [auto] {$\scriptstyle{}$} (A1_3);
\path (A0_2) edge [->]node [auto] {$\scriptstyle{}$} (A0_3);
\path (A1_3) edge [->]node [auto] {$\scriptstyle{}$} (A1_4);
\end{tikzpicture}
\end{equation*}
Moreover, $q_*(at(E))=q_*(i_*(at(\mathcal{E})))=0$, hence we get the commutative diagram
\begin{equation*}
\begin{tikzpicture}[xscale=3.2,yscale=-1.2]
\node (A0_0) at (0, 0) {$0$};
\node (A0_1) at (0.7, 0) {$\Omega_{\mathcal{X}}^1\otimes E$};
\node (A0_2) at (2, 0) {$\jf(E)$};
\node (A0_3) at (3.5, 0) {$E$};
\node (A0_4) at (4, 0) {$0$};
\node (A1_0) at (0, 1) {$0$};
\node (A1_1) at (0.7, 1) {$p_S^{*}(\Omega_S^1)\vert_{\mathcal{D}}\otimes E$};
\node (A1_2) at (2, 1) {$\left(p_S^{*}(\Omega_S^1)\vert_{\mathcal{D}}\otimes E\right)\oplus E$};
\node (A1_3) at (3.5, 1) {$E$};
\node (A1_4) at (4, 1) {$0$};
\path (A0_1) edge [->]node [auto] {$\scriptstyle{}$} (A1_1);
\path (A0_0) edge [->]node [auto] {$\scriptstyle{}$} (A0_1);
\path (A0_1) edge [->]node [auto] {$\scriptstyle{}$} (A0_2);
\path (A1_0) edge [->]node [auto] {$\scriptstyle{}$} (A1_1);
\path (A0_3) edge [thin, double distance=1.5pt]node [auto] {$\scriptstyle{}$} (A1_3);
\path (A1_1) edge [->]node [auto] {$\scriptstyle{}$} (A1_2);
\path (A0_3) edge [->]node [auto] {$\scriptstyle{}$} (A0_4);
\path (A0_2) edge [->]node [auto] {$\scriptstyle{}$} (A1_2);
\path (A1_2) edge [->]node [auto] {$\scriptstyle{}$} (A1_3);
\path (A0_2) edge [->]node [auto] {$\scriptstyle{}$} (A0_3);
\path (A1_3) edge [->]node [auto] {$\scriptstyle{}$} (A1_4);
\end{tikzpicture}
\end{equation*}
\begin{example}\label{ex:linebundle}
Let $F$ be a line bundle on $D.$ Let $\mathcal{L}=(L, \alpha)$ be a flat family of framed sheaves parametrized by $S$ with $L$ line bundle on $\mathcal{X}.$ In this case $V$ is the whole $\mathcal{X}$ and we choose transition functions $g_{ij}^0$ and $g_{ij}$ for $p_X^*(F_D)$ and $L$, respectively, such that
\begin{equation*}
g_{ij}\vert_{\mathcal{D}}= g_{ij}^0.
\end{equation*}
Recall that $dg_{ij} g_{ij}^{-1}$ is a cocycle representing $at(L)$ (cf. \cite[Proposition 12]{art:atiyah1957}). By the choice of $g_{ij}^0$, we get that $d_S(g_{ij})$ vanishes along $\mathcal{D}$, where $d_S$ is the exterior differential of $S.$ Hence $dg_{ij} g_{ij}^{-1}$ can be also interpreted as a cocycle representing $at(\mathcal{L}).$ Moreover, it vanishes under the restriction of the de Rham differential $\tilde{d}:=d\vert_{p_S^*(\Omega^1_S)(-\mathcal{D})\oplus p_X^*(\Omega_X^1)}.$
\triend
\end{example}
Now we shall provide another way to describe the framed Atiyah class of a flat family of framed sheaves $\mathcal{E}=(E,\alpha)$ by using finite locally free resolutions of $E$, but in this case the costruction is \emph{local over the base}, as we will explain in the following. First, we recall a result due to B{\u{a}}nic{\u{a}}, Putinar and Schumacher that will be very useful later on.
\begin{theorem}{\normalfont (\cite[Satz 3]{art:banicaputinarschumacher1980}).}\label{thm:banica}
Let $p\colon R\rightarrow T$ be a flat proper morphism of schemes of finite type over $k$, $T$ smooth, $E$ and $G$ coherent $\mathcal{O}_R$-modules, flat over $T.$ If the function $y\mapsto \dim \mathrm{Ext}^l(E_y,G_y)$ is constant for $l$ fixed, then the sheaf $\mathcal{E}xt_p^l(E,G)$ is locally free on $T$ and for any $y\in T$ we have
\begin{equation*}
\mathcal{E}xt_p^i(E,G)_y\otimes_{\mathcal{O}_{T,y}}\left(\sfrac{\mathcal{O}_{T,y}}{m_y}\right)\cong \mathrm{Ext}^i(E_y,G_y)\, \mbox{ for }\, i=l-1, l.
\end{equation*}
Moreover, the same statement is true for complexes.
\end{theorem}
Let $\mathcal{E}=(E,\alpha)$ be a flat family of framed sheaves parametrized by a smooth scheme $S.$ Since the projection morphism $p_S\colon \mathcal{X}\longrightarrow S$ is smooth and projective, there exists a finite locally free resolution $E^\bullet \rightarrow E$ of $E$ (see, e.g., \cite[Proposition 2.1.10]{book:huybrechtslehn2010}).
Let us fix a point $s_0\in S.$ By the flatness property, the complex $(E^\bullet)\vert_{\{s_0\}\times D}$ is a finite resolution of locally free sheaves of $E\vert_{\{s_0\}\times D}\cong F_D.$ Let us denote by $F^\bullet$ the complex $(E^\bullet)\vert_{\{s_0\}\times D}.$ Define $\mathcal{F}^\bullet:=F^\bullet \boxtimes \mathcal{O}_S.$ The complex $\mathcal{F}^\bullet$ is $S$-flat since $(E^\bullet)\vert_{\{s_0\}\times D}$ is a complex of locally free $\mathcal{O}_D$-modules and the sheaf $\mathcal{O}_{\mathcal{D}}$ is a $S$-flat $\mathcal{O}_{\mathcal{X}}$-module. Moreover, for any $s\in S$, the complex $(\mathcal{F}^\bullet)\vert_{\{s\}\times X}$ is quasi-isomorphic to $F$, hence we get
\begin{equation*}
\mathrm{Hom}((E^\bullet)\vert_{\{s\}\times X}, (\mathcal{F}^\bullet)\vert_{\{s\}\times X})=\mathrm{Hom}(E\vert_{\{s\}\times X}, F_D)\cong\mathrm{End}(F_D).
\end{equation*}
By applying Theorem \ref{thm:banica}, we get that the natural morphism of complexes between $E^\bullet$ and $\mathcal{F}^\bullet$ on $\{s_0\}\times X$ extends to a morphism of complexes
\begin{equation*}
\alpha_\bullet \colon E^\bullet\longrightarrow\mathcal{F}^\bullet.
\end{equation*}
Let $U\subset S$ be a neighborhood of $s_0$ such that the following condition holds
\begin{equation*}
(\alpha_\bullet)\vert_{\{s\}\times D}\, \mbox{ \emph{is an isomorphism for any} } s\in U.
\end{equation*}
Let $\mathcal{X}_U=U\times X$ and $\mathcal{D}_U=U\times D.$ For any $i$, the pair $\mathcal{E}_U^i:=(E^i\vert_{\mathcal{X}_U},\alpha_i\vert_{\mathcal{X}_U}\colon E^i\vert_{\mathcal{X}_U} \rightarrow\mathcal{F}^i\vert_{\mathcal{X}_U})$ is a flat family $\mathcal{E}_U^i$ of $(D,E^i\vert_{\{s_0\}\times D})$-framed sheaves parametrized by $U.$
Thus we proved the following:
\begin{proposition}\label{prop:locallyfree}
Let $\mathcal{E}=(E,\alpha)$ be a flat family of framed sheaves parametrized by a smooth scheme $S$ and $E^\bullet \rightarrow E$ a finite locally free resolution of $E.$ Let $s_0$ be a point in $S.$ Then there exists a complex $\mathcal{F}^\bullet$, a morphism of complexes $\alpha_\bullet\colon E^\bullet \rightarrow \mathcal{F}^\bullet$ and an open neighborhood $U\subset S$ of $s_0$ with the following property: for any $i$ the sheaf $\mathcal{F}^i\vert_{\{s_0\}\times D}$ is a locally free $\mathcal{O}_D$-module and the pair $\mathcal{E}_U^i:=(E^i\vert_{\mathcal{X}_U}, \alpha_i\vert_{\mathcal{X}_U})$ is a flat family of $(D,\mathcal{F}^i\vert_{\{s_0\}\times D})$-framed sheaves parametrized by $U.$
\end{proposition}
If for any $i$, we consider the short exact sequence associated to $\jfr(\mathcal{E}_U^i)$
\begin{equation*}
0\longrightarrow \left(p_U^*(\Omega^1_U)(-\mathcal{D}_U)\oplus p_X^*(\Omega_X^1)\right)\otimes E^i\vert_{\mathcal{X}_U} \longrightarrow \jfr(\mathcal{E}_U^i) \longrightarrow E^i\vert_{\mathcal{X}_U}\longrightarrow 0,
\end{equation*}
we get a class $at_{U}(\mathcal{E})$ in
\begin{eqnarray*}
&&\HE^1\left(E^\bullet\vert_{\mathcal{X}_U},\left(p_U^*(\Omega^1_U)(-\mathcal{D}_U)\oplus p_X^*(\Omega_X^1)\right)\otimes E^\bullet\vert_{\mathcal{X}_U}\right)\cong\\
&&\cong \mathrm{Ext}^1\left(E\vert_{\mathcal{X}_U},\left(p_U^*(\Omega^1_U)(-\mathcal{D}_U)\oplus p_X^*(\Omega_X^1)\right)\otimes E\vert_{\mathcal{X}_U}\right).
\end{eqnarray*}
By construction, $at_{U}(\mathcal{E})$ is independent of the resolution and is the image of $at(\mathcal{E})$ with respect to the map on Ext-groups induced by the natural morphism $i^*\colon \Omega_S^1\rightarrow \Omega_U^1$, where $i\colon U\hookrightarrow S$ is the inclusion morphism.
\subsection{Framed Newton polynomials}
Let $\mathcal{E}=(E,\alpha)$ be a flat family of framed sheaves parametrized by a smooth scheme $S.$ Let $at(\mathcal{E})^i$ denote the image in $\mathrm{Ext}^i\left(E,\tilde{\Omega}_{\mathcal{X}}^i\otimes E\right)$ of the $i$-th product
\begin{equation*}
at(\mathcal{E})\circ \cdots \circ at(\mathcal{E})\in \mathrm{Ext}^i\left(E,(\tilde{\Omega}_{\mathcal{X}}^1)^{\otimes i}\otimes E\right)
\end{equation*}
under the morphism induced by $(\tilde{\Omega}_{\mathcal{X}}^1)^{\otimes i}\rightarrow \tilde{\Omega}_{\mathcal{X}}^i$, where $\tilde{\Omega}_{\mathcal{X}}^1:=p_S^*(\Omega^1_S)(-\mathcal{D})\oplus p_X^*(\Omega_X^1)$ and $\tilde{\Omega}_{\mathcal{X}}^i:=\Lambda^i(\tilde{\Omega}_{\mathcal{X}}^1)$ is the $i$-th exterior power of $\tilde{\Omega}_{\mathcal{X}}^1.$
\begin{definition}
The $i$-th \emph{framed Newton polynomial} of $\mathcal{E}$ is
\begin{equation*}
\gamma^i(\mathcal{E}):=\mathrm{tr}(at(\mathcal{E})^i)\in \mathrm{H}^i(\mathcal{X},\tilde{\Omega}_{\mathcal{X}}^i).
\end{equation*}
\end{definition}
Fix a finite locally free resolution $E^\bullet\rightarrow E$ of $E$ and a point $s_0\in S.$ Let $U\subset S$ be a neighborhood of $s_0$ as in Proposition \ref{prop:locallyfree}. The restriction of $\gamma^i(\mathcal{E})$ to $\mathcal{X}_U$ coincides with the class $\gamma_{U}^i(\mathcal{E})$ defined as
\begin{equation*}
\gamma_{U}^i(\mathcal{E}):=\mathrm{tr}(at_{U}(\mathcal{E})^i)\in \mathrm{H}^i(\mathcal{X}_U,\tilde{\Omega}_{\mathcal{X}_U}^i).
\end{equation*}
The restricted de Rham differential $\tilde{d}$ introduced in Example \ref{ex:linebundle} induces $k$-linear maps
\begin{equation*}
\tilde{d}\colon \mathrm{H}^i(\mathcal{X},\tilde{\Omega}_{\mathcal{X}}^{i})\longrightarrow \mathrm{H}^{i}(\mathcal{X},\tilde{\Omega}_{\mathcal{X}}^{i+1}(\mathcal{D})).
\end{equation*}
\begin{proposition}\label{prop:tildedclosed}
The $i$-th \emph{framed Newton polynomial} of $\mathcal{E}$ is $\tilde{d}$-closed.
\end{proposition}
\proof
Let $E^\bullet\rightarrow E$ be a finite locally free resolution of $E$ and $s_0\in S.$ Let $U\subset S$ be a neighborhood of $s_0$ as in Proposition \ref{prop:locallyfree}. The cohomology class $\gamma_{U}^i(\mathcal{E})$ is $\tilde{d}\vert_U$-closed by the same arguments as in the nonframed case (cf. \cite[Section 10.1.6]{book:huybrechtslehn2010}), in particular the fact that we can reduce to the case of line bundles by using the splitting principle, and by Example \ref{ex:linebundle}. Since the restriction of $\gamma^i(\mathcal{E})$ to $\mathcal{X}_U$ is $\gamma_{U}^i(\mathcal{E})$, we get that $\gamma^i(\mathcal{E})$ is closed with respect to $\tilde{d}.$
\endproof
\subsection{The Kodaira-Spencer map for framed sheaves}
Let $\mathcal{E}=(E,\alpha)$ be a flat family of framed sheaves parametrized by a scheme $S.$ Consider the framed Atiyah class $at(\mathcal{E})$ in $\mathrm{Ext}^1\left(E,\left(p_S^*(\Omega^1_S)(-\mathcal{D})\oplus p_X^*(\Omega_X^1)\right)\otimes E\right)$ and the induced section $\mathcal{A}t(\mathcal{E})$ under the relative local-to-global map
\begin{equation*}
\mathrm{Ext}^1\left(E,\left(p_S^*(\Omega^1_S)(-\mathcal{D})\oplus p_X^*(\Omega_X^1)\right)\otimes E\right)\longrightarrow \mathrm{H}^0(S, \mathcal{E}xt_{p_S}^1(E,\left(p_S^*(\Omega^1_S)(-\mathcal{D})\oplus p_X^*(\Omega_X^1)\right)\otimes E)),
\end{equation*}
coming from the relative local-to-global spectral sequence
\begin{equation*}
\mathrm{H}^i(S,\mathcal{E}xt_{p_S}^j(E,\left(p_S^*(\Omega^1_S)(-\mathcal{D})\oplus p_X^*(\Omega_X^1)\right)\otimes E))\Rightarrow \mathrm{Ext}^{i+j}(E,\left(p_S^*(\Omega^1_S)(-\mathcal{D})\oplus p_X^*(\Omega_X^1)\right)\otimes E).
\end{equation*}
By considering the $S$-part $\mathcal{A}t_S(\mathcal{E})$ of $\mathcal{A}t(\mathcal{E})$ in $\mathrm{H}^0(S,\mathcal{E}xt_{p_S}^1(E,p_S^*(\Omega^1_S)(-\mathcal{D})\otimes E))$, we define the framed version of the Kodaira-Spencer map.
\begin{definition}
The \textit{framed Kodaira-Spencer map} associated to the family $\mathcal{E}$ is the composition
\begin{eqnarray*}
KS_{fr}\colon (\Omega_S^1)^\vee&\stackrel{\mathrm{id}\otimes \mathcal{A}t_S(\mathcal{E})}{\longrightarrow}& (\Omega_S^1)^\vee\otimes \mathcal{E}xt_{p_S}^1(E,p_S^*(\Omega^1_S)(-\mathcal{D})\otimes E)\rightarrow\\
&\longrightarrow& \mathcal{E}xt_{p_S}^1(E,p_S^*((\Omega_S^1)^\vee\otimes\Omega^1_S)\otimes p_X^*(\mathcal{O}_X(-D))\otimes E)\rightarrow\\
&\longrightarrow& \mathcal{E}xt_{p_S}^1(E,p_X^*(\mathcal{O}_X(-D))\otimes E).
\end{eqnarray*}
\end{definition}
\subsection{Closed two-forms via the framed Atiyah class}
Let $S$ be a smooth affine scheme and $\mathcal{E}=(E,\alpha)$ a flat family of framed sheaves parametrized by $S.$ Let $\gamma^{0,2}$ denote the component of $\gamma^2(\mathcal{E})$ in $\mathrm{H}^0(S, \Omega^2_S)\otimes \mathrm{H}^{2}(X,\mathcal{O}_X(-2\,D)).$
\begin{definition}
Let $\tau_S$ be the homomorphism given by
\begin{equation*}
\tau_S\colon \mathrm{H}^0(X,\omega_X(2\,D))\cong \mathrm{H}^2(X,\mathcal{O}_X(-2\,D))^\vee\stackrel{\cdot \,\gamma^{0,2}}{\longrightarrow} \mathrm{H}^0(S, \Omega^2_S),
\end{equation*}
where $\cong$ denotes Serre's duality.
\end{definition}
\begin{proposition}\label{prop:closedness}
For any $\omega\in \mathrm{H}^0(X,\omega_X(2\,D))$, the associated two-form $\tau_S(\omega)$ on $S$ is closed.
\end{proposition}
\proof
We can write
\begin{equation*}
\gamma^{0,2}=\sum_l \mu_l\otimes \nu_l,
\end{equation*}
for elements $\mu_l\in \mathrm{H}^0(S, \Omega^2_S)$ and $\nu_l\in \mathrm{H}^2(X,\mathcal{O}_X(-2\,D)).$ Since $\tilde{d}(\gamma^2(\mathcal{E}))=0$ (cf. Proposition \ref{prop:tildedclosed}), the component of $\tilde{d}(\gamma^{0,2})$ in $\mathrm{H}^0(S, \Omega^3_S)\otimes \mathrm{H}^{2}(X,\mathcal{O}_X(-2\,D))$ is zero, which means
\begin{equation*}
\sum_l d_{S}(\mu_l)\otimes \nu_l=0.
\end{equation*}
Therefore
\begin{equation*}
d_S(\tau_S(\omega))=d_S\left(\sum_l \mu_l \cdot \omega(\nu_l)\right)=\sum_l d_{S}(\mu_l)\cdot\omega(\nu_l)=0.
\end{equation*}
\endproof
Fix $\omega\in \mathrm{H}^0(X,\omega_X(2\,D)).$ Since $S$ is smooth, it follows from the definitions of the framed Kodaira-Spencer map and the framed Newton polynomial that the two-form $\tau_S(\omega)$ at a point $s_0\in S$ coincides with the following composition of maps:
\begin{eqnarray*}
&&T_{s_0} S\times T_{s_0} S\stackrel{KS \times KS}{\longrightarrow}\mathrm{Ext}^1(E\vert_{\{s_0\}\times X},E\vert_{\{s_0\}\times X}(-D))\times \mathrm{Ext}^1(E\vert_{\{s_0\}\times X},E\vert_{\{s_0\}\times X}(-D))\\
&& \stackrel{\circ}{\longrightarrow} \mathrm{Ext}^2(E\vert_{\{s_0\}\times X},E\vert_{\{s_0\}\times X}(-2\,D))\stackrel{tr}{\longrightarrow} \mathrm{H}^2(X, \mathcal{O}_X(-2\,D))\stackrel{\cdot\, \omega}{\longrightarrow} \mathrm{H}^2(X,\omega_X)\cong k.
\end{eqnarray*}
\section{The tangent bundle of moduli spaces of framed sheaves}\label{sec:framedkodaira}
Let $\mathcal{M}^s(X;P)$ be the moduli space of Gieseker-stable torsion-free sheaves on $X$ with Hilbert polynomial $P.$ The open subset $\mathcal{M}_0(X;P)\subset \mathcal{M}^s(X;P)$ of points $[E]$ such that $\mathrm{Ext}_0^2(E,E)$ vanishes is smooth according to \cite[Theorem 4.5.4]{book:huybrechtslehn2010}. Suppose there exists a universal family $\tilde{E}$ on $\mathcal{M}_0(X;P)\times X.$ By using the Atiyah class of $\tilde{E}$, one can define the Kodaira-Spencer map for $\mathcal{M}_0(X;P)$:
\begin{equation*}
KS\colon T\mathcal{M}_0(X;P) \longrightarrow \mathcal{E}xt_p^1(\tilde{E},\tilde{E}),
\end{equation*}
where $p\colon \mathcal{M}_0(X;P)\times X\rightarrow \mathcal{M}_0(X;P)$ is the projection. Moreover, it is possible to prove that $KS$ is an isomorphism (this result holds also when a universal family for $\mathcal{M}_0(X;P)$ does not exist, cf. \cite[Theorem 10.2.1]{book:huybrechtslehn2010}). In this section we shall prove the framed analogue of this result for the moduli spaces of stable framed sheaves on $X.$
Let $\delta\in \mathbb{Q}[n]$ be a stability polynomial and $P$ a numerical polynomial of degree two. Let $\mathcal{M}^*_\delta(X;F_D,P)$ be the moduli space of framed sheaves on $X$ with Hilbert polynomial $P$ that are stable with respect to $\mathcal{O}_X(1)$ and $\delta.$ This is an open subset of the fine moduli space $\mathcal{M}_\delta^{s}(X;F_D,P)$ of stable $F_D$-framed modules with Hilbert polynomial $P.$
Let us denote by $\mathcal{M}^*_\delta(X;F_D,P)^{sm}$ the smooth locus of $\mathcal{M}^*_\delta(X;F_D,P)$ and by $\tilde{\mathcal{E}}=(\tilde{E},\tilde{\alpha})$ a \emph{universal framed sheaf} over $\mathcal{M}^*_\delta(X;F_D,P)^{sm}.$
\begin{theorem}\label{thm:tanbundle}
The framed Kodaira-Spencer map defined by $\tilde{\mathcal{E}}$ induces a canonical isomorphism
\begin{equation*}
KS_{fr}\colon T\mathcal{M}^*_\delta(X;F_D,P)^{sm}\stackrel{\sim}{\longrightarrow} \mathcal{E}xt_p^1(\tilde{E}, \tilde{E}\otimes p_X^*(\mathcal{O}_X(-D))),
\end{equation*}
where $p$ is the projection from $\mathcal{M}^*_\delta(X;F_D,P)^{sm}\times X$ to $\mathcal{M}^*_\delta(X;F_D,P)^{sm}.$
\end{theorem}
\proof
First note that $\mathcal{M}^*_\delta(X;F_D,P)^{sm}$ is a reduced separated scheme of finite type over $k.$ Hence it suffices to prove that the framed Kodaira-Spencer map is an isomorphism on the fibres over closed points.
Let $[(E,\alpha)]$ be a closed point. We want to show that the Kodaira-Spencer map $KS_{fr}([(E,\alpha)])$ on the fibre over $[(E,\alpha)]$ coincides with the isomorphism
\begin{equation*}
T_{[(E,\alpha)]} \mathcal{M}^*_\delta(X;F_D,P)^{sm}\stackrel{\sim}{\longrightarrow} \mathrm{Ext}^1(E,E(-D)),
\end{equation*}
coming from deformation theory (see proof of \cite[Theorem 4.1]{art:huybrechtslehn1995-II}).
Let $w$ be an element in $\mathrm{Ext}^1(E,E(-D)).$ Consider the long exact sequence
\begin{equation*}
\cdots \rightarrow \mathrm{Ext}^1(E,E(-D))\stackrel{j_*}{\longrightarrow} \mathrm{Ext}^1(E,E) \stackrel{\alpha_*}{\longrightarrow} \mathrm{Ext}^1(E,F_D)\rightarrow \cdots
\end{equation*}
obtained by applying the functor $\mathrm{Hom}(E,\cdot)$ to the exact sequence
\begin{equation*}
0\longrightarrow E(-D)\stackrel{j}{\longrightarrow} E\stackrel{\alpha}{\longrightarrow} F_D\longrightarrow 0.
\end{equation*}
Let $v=j_*(w)\in \mathrm{Ext}^1(E,E).$ We get a commutative diagram
\begin{equation*}
\begin{tikzpicture}[xscale=1.7,yscale=-1.2]
\node (A0_0) at (0, 0) {$0$};
\node (A0_1) at (1, 0) {$E(-D)$};
\node (A0_2) at (2, 0) {$\tilde{G}$};
\node (A0_3) at (3, 0) {$E$};
\node (A0_4) at (4, 0) {$0$};
\node (A1_0) at (0, 1) {$0$};
\node (A1_1) at (1, 1) {$E$};
\node (A1_2) at (2, 1) {$G$};
\node (A1_3) at (3, 1) {$E$};
\node (A1_4) at (4, 1) {$0$};
\path (A0_1) edge [->]node [auto] {$\scriptstyle{j}$} (A1_1);
\path (A0_0) edge [->]node [auto] {$\scriptstyle{}$} (A0_1);
\path (A0_1) edge [->]node [auto] {$\scriptstyle{\tilde{i}}$} (A0_2);
\path (A1_0) edge [->]node [auto] {$\scriptstyle{}$} (A1_1);
\path (A0_3) edge [thin, double distance=1.5pt]node [auto] {$\scriptstyle{}$} (A1_3);
\path (A1_1) edge [->]node [auto] {$\scriptstyle{i}$} (A1_2);
\path (A0_3) edge [->]node [auto] {$\scriptstyle{}$} (A0_4);
\path (A0_2) edge [->]node [auto] {$\scriptstyle{}$} (A1_2);
\path (A1_2) edge [->]node [auto] {$\scriptstyle{\pi}$} (A1_3);
\path (A0_2) edge [->]node [auto] {$\scriptstyle{\tilde{\pi}}$} (A0_3);
\path (A1_3) edge [->]node [auto] {$\scriptstyle{}$} (A1_4);
\end{tikzpicture}
\end{equation*}
where the first row is a representative for $w$ and the second one a representative for $v.$
Let $S=\mathrm{Spec}(k[\varepsilon])$ be the spectrum of the ring of dual numbers, where $\varepsilon^2=0.$ We can think of $G$ as a $S$-flat family by letting $\varepsilon$ act on $G$ as the morphism $i\circ \pi.$ Since $\varepsilon \tilde{G}=E(-D)$ and $\varepsilon G=E$, by applying the snake lemma to the previous diagram we get
\begin{equation*}
\begin{tikzpicture}[xscale=1.7,yscale=-1.2]
\node (A0_0) at (0, 0) {$0$};
\node (A0_1) at (1, 0) {$E(-D)$};
\node (A0_2) at (2, 0) {$\tilde{G}$};
\node (A0_3) at (3, 0) {$E$};
\node (A0_4) at (4, 0) {$0$};
\node (A1_0) at (0, 1) {$0$};
\node (A1_1) at (1, 1) {$E$};
\node (A1_2) at (2, 1) {$G$};
\node (A1_3) at (3, 1) {$E$};
\node (A1_4) at (4, 1) {$0$};
\node (A2_0) at (0, 2) {$0$};
\node (A2_1) at (1, 2) {$\varepsilon F_D$};
\node (A2_2) at (2, 2) {$\varepsilon F_D$};
\node (A2_3) at (3, 2) {$0$};
\node (A3_1) at (1, 3) {$0$};
\node (A3_2) at (2, 3) {$0$};
\path (A2_1) edge [->]node [auto] {$\scriptstyle{}$} (A3_1);
\path (A0_1) edge [->]node [auto] {$\scriptstyle{}$} (A1_1);
\path (A2_1) edge [->]node [auto] {$\scriptstyle{}$} (A2_2);
\path (A0_0) edge [->]node [auto] {$\scriptstyle{}$} (A0_1);
\path (A0_1) edge [->]node [auto] {$\scriptstyle{\tilde{i}}$} (A0_2);
\path (A1_0) edge [->]node [auto] {$\scriptstyle{}$} (A1_1);
\path (A0_3) edge [->]node [auto] {$\scriptstyle{\mathrm{id}_E}$} (A1_3);
\path (A1_1) edge [->]node [auto] {$\scriptstyle{i}$} (A1_2);
\path (A1_1) edge [->]node [auto] {$\scriptstyle{}$} (A2_1);
\path (A1_2) edge [->]node [auto] {$\scriptstyle{\beta}$} (A2_2);
\path (A2_2) edge [->]node [auto] {$\scriptstyle{}$} (A2_3);
\path (A0_3) edge [->]node [auto] {$\scriptstyle{}$} (A0_4);
\path (A2_2) edge [->]node [auto] {$\scriptstyle{}$} (A3_2);
\path (A0_2) edge [->]node [auto] {$\scriptstyle{}$} (A1_2);
\path (A2_0) edge [->]node [auto] {$\scriptstyle{}$} (A2_1);
\path (A1_2) edge [->]node [auto] {$\scriptstyle{\pi}$} (A1_3);
\path (A1_3) edge [->]node [auto] {$\scriptstyle{}$} (A2_3);
\path (A0_2) edge [->]node [auto] {$\scriptstyle{\tilde{\pi}}$} (A0_3);
\path (A1_3) edge [->]node [auto] {$\scriptstyle{}$} (A1_4);
\end{tikzpicture}
\end{equation*}
Moreover $\alpha_*(v)=0$, hence we have the commutative diagram
\begin{equation*}
\begin{tikzpicture}[xscale=1.7,yscale=-1.2]
\node (A0_0) at (0, 0) {$0$};
\node (A0_1) at (1, 0) {$E$};
\node (A0_2) at (2, 0) {$G$};
\node (A0_3) at (3, 0) {$E$};
\node (A0_4) at (4, 0) {$0$};
\node (A1_0) at (0, 1) {$0$};
\node (A1_1) at (1, 1) {$\varepsilon F_D$};
\node (A1_2) at (2, 1) {$E\oplus\varepsilon F_D$};
\node (A1_3) at (3, 1) {$E$};
\node (A1_4) at (4, 1) {$0$};
\path (A0_1) edge [->]node [auto] {$\scriptstyle{}$} (A1_1);
\path (A0_0) edge [->]node [auto] {$\scriptstyle{}$} (A0_1);
\path (A0_1) edge [->]node [auto] {$\scriptstyle{i}$} (A0_2);
\path (A1_0) edge [->]node [auto] {$\scriptstyle{}$} (A1_1);
\path (A0_3) edge [thin, double distance=1.5pt]node [auto] {$\scriptstyle{}$} (A1_3);
\path (A1_1) edge [->]node [auto] {$\scriptstyle{}$} (A1_2);
\path (A0_3) edge [->]node [auto] {$\scriptstyle{}$} (A0_4);
\path (A0_2) edge [->]node [auto] {$\scriptstyle{}$} (A1_2);
\path (A1_2) edge [->]node [auto] {$\scriptstyle{}$} (A1_3);
\path (A0_2) edge [->]node [auto] {$\scriptstyle{\pi}$} (A0_3);
\path (A1_3) edge [->]node [auto] {$\scriptstyle{}$} (A1_4);
\end{tikzpicture}
\end{equation*}
Thus we obtain a framing $\gamma\colon G\rightarrow F_D\oplus \varepsilon F_D$ induced by $\alpha$ and $\beta.$ Moreover $\gamma\vert_{\mathcal{D}}$ is an isomorphism. We denote by $\mathcal{G}$ the corresponding $S$-flat family of $(D,F_D)$-framed sheaves on $X.$
In the nonframed case, one define a relative Atiyah class for families of coherent sheaves parametrized by a scheme $S$ and takes its $S$-part (see \cite[Section 10.1.8]{book:huybrechtslehn2010}). As it is explained in \cite[Example 10.1.9]{book:huybrechtslehn2010}, since $S$ is affine, the relative $S$-part $\mathcal{A}t_S(G)$ of $G$ is an element of
\begin{equation*}
\mathrm{Ext}^1_{\mathcal{X}}(G,p_S^* \Omega^1_S\otimes G)\cong \mathrm{Ext}^1_\mathcal{X}(G,E).
\end{equation*}
Consider the short exact sequence of coherent sheaves over $\mathrm{Spec}\left(\sfrac{k[\varepsilon_1,\varepsilon_2]}{(\varepsilon_1,\varepsilon_2)^2}\right)\times X$
\begin{equation}\label{eq:s-part}
0\longrightarrow E\stackrel{i'}{\longrightarrow} G'\stackrel{\pi'}{\longrightarrow} G\longrightarrow 0,
\end{equation}
where $\varepsilon_1$ and $\varepsilon_2$ act trivially on $E$ and by $i\circ \pi$ on $G$, and $G'\cong \sfrac{k[\varepsilon_1]\otimes_{k} G}{\varepsilon_1\varepsilon_2 G}\cong G\oplus E$, with actions
\begin{equation*}
\varepsilon_1=
\left(
\begin{array}{cc}
0 & \pi\\
0 & 0
\end{array}
\right) \,\mbox{ and }
\varepsilon_2=
\left(
\begin{array}{cc}
i\pi & 0\\
0 & 0
\end{array}
\right).
\end{equation*}
By definition of Atiyah class, $\mathcal{A}t_S(G)$ is precisely the extension class of the short exact sequence \eqref{eq:s-part}, considered as a sequence of $k[\varepsilon_1]\otimes\mathcal{O}_X$-modules.
The morphism $\pi$ induces a pull-back morphism $\pi^*\colon \mathrm{Ext}_X^1(E,E) \rightarrow\mathrm{Ext}^1_\mathcal{X}(G,E)$, which is an isomorphism. As it is proved in \cite[Example 10.1.9]{book:huybrechtslehn2010}, $\pi^*(v)=\mathcal{A}t_S(G)$, indeed we have the commutative diagram
\begin{equation*}
\begin{tikzpicture}[xscale=1.7,yscale=-1.2]
\node (A0_0) at (0, 0) {$0$};
\node (A0_1) at (1, 0) {$E$};
\node (A0_2) at (2, 0) {$G'$};
\node (A0_3) at (3, 0) {$G$};
\node (A0_4) at (4, 0) {$0$};
\node (A1_0) at (0, 1) {$0$};
\node (A1_1) at (1, 1) {$E$};
\node (A1_2) at (2, 1) {$G$};
\node (A1_3) at (3, 1) {$E$};
\node (A1_4) at (4, 1) {$0$};
\path (A0_0) edge [->]node [auto] {$\scriptstyle{}$} (A0_1);
\path (A0_1) edge [thin, double distance=1.5pt]node [auto] {$\scriptstyle{}$} (A1_1);
\path (A1_0) edge [->]node [auto] {$\scriptstyle{}$} (A1_1);
\path (A0_3) edge [->]node [auto] {$\scriptstyle{\pi}$} (A1_3);
\path (A0_1) edge [->]node [auto] {$\scriptstyle{i'}$} (A0_2);
\path (A0_2) edge [->]node [auto] {$\scriptstyle{t'}$} (A1_2);
\path (A1_1) edge [->]node [auto] {$\scriptstyle{i}$} (A1_2);
\path (A1_2) edge [->]node [auto] {$\scriptstyle{\pi}$} (A1_3);
\path (A0_2) edge [->]node [auto] {$\scriptstyle{\pi'}$} (A0_3);
\path (A0_3) edge [->]node [auto] {$\scriptstyle{}$} (A0_4);
\path (A1_3) edge [->]node [auto] {$\scriptstyle{}$} (A1_4);
\end{tikzpicture}
\end{equation*}
Thus $G'$ is the sheaf of first jets of $G$ relative to the quotient $\Omega_\mathcal{X}^1\rightarrow p_S^*(\Omega_S^1)\rightarrow 0.$ By following Maakestad's construction of Atiyah classes of coherent sheaves relative to quotients of $\Omega_\mathcal{X}^1$ (cf. \cite[Section 3]{art:maakestad2010-I}) and by readapting to this particular case our construction of the framed sheaf of first jets given in Section \ref{sec:framedatiyah}, we can define a \textit{framed sheaf of first jets} $\tilde{G}'$ of the framed sheaf $\mathcal{G}$ relative to $p_S^*(\Omega_S^1).$ Thus we get a commutative diagram
\begin{equation*}
\begin{tikzpicture}[xscale=1.7,yscale=-1.2]
\node (A0_0) at (0, 0) {$0$};
\node (A0_1) at (1, 0) {$E(-D)$};
\node (A0_2) at (2, 0) {$\tilde{G}'$};
\node (A0_3) at (3, 0) {$G$};
\node (A0_4) at (4, 0) {$0$};
\node (A1_0) at (0, 1) {$0$};
\node (A1_1) at (1, 1) {$E$};
\node (A1_2) at (2, 1) {$G'$};
\node (A1_3) at (3, 1) {$G$};
\node (A1_4) at (4, 1) {$0$};
\path (A0_1) edge [->]node [auto] {$\scriptstyle{}$} (A1_1);
\path (A0_0) edge [->]node [auto] {$\scriptstyle{}$} (A0_1);
\path (A0_1) edge [->]node [auto] {$\scriptstyle{\tilde{i}'}$} (A0_2);
\path (A1_0) edge [->]node [auto] {$\scriptstyle{}$} (A1_1);
\path (A0_3) edge [thin, double distance=1.5pt]node [auto] {$\scriptstyle{}$} (A1_3);
\path (A1_1) edge [->]node [auto] {$\scriptstyle{i'}$} (A1_2);
\path (A0_3) edge [->]node [auto] {$\scriptstyle{}$} (A0_4);
\path (A0_2) edge [->]node [auto] {$\scriptstyle{}$} (A1_2);
\path (A1_2) edge [->]node [auto] {$\scriptstyle{\pi'}$} (A1_3);
\path (A0_2) edge [->]node [auto] {$\scriptstyle{\tilde{\pi}'}$} (A0_3);
\path (A1_3) edge [->]node [auto] {$\scriptstyle{}$} (A1_4);
\end{tikzpicture}
\end{equation*}
The first row is a representative for the $S$-part $\mathcal{A}t_S(\mathcal{G})$ of $\mathcal{G}$ in
\begin{equation*}
\mathrm{Ext}^1_{\mathcal{X}}(G,p_S^* \Omega^1_S(-\mathcal{D})\otimes G)\cong \mathrm{Ext}^1_\mathcal{X}(G,E(-D)).
\end{equation*}
Consider the diagram
\begin{equation*}
\begin{tikzpicture}[xscale=1.5,yscale=-1.2]
\node (A0_0) at (0, 0) {$0$};
\node (A1_1) at (1, 1) {$0$};
\node (A0_2) at (2, 0) {$E(-D)$};
\node (A1_3) at (3, 1) {$E(-D)$};
\node (A0_4) at (4, 0) {$\tilde{G}'$};
\node (A1_5) at (5, 1) {$\tilde{G}$};
\node (A0_6) at (6, 0) {$G$};
\node (A1_7) at (7, 1) {$E$};
\node (A0_8) at (8, 0) {$0$};
\node (A1_9) at (9, 1) {$0$};
\node (A2_0) at (0, 2) {$0$};
\node (A3_1) at (1, 3) {$0$};
\node (A2_2) at (2, 2) {$E$};
\node (A3_3) at (3, 3) {$E$};
\node (A2_4) at (4, 2) {$G'$};
\node (A3_5) at (5, 3) {$G$};
\node (A2_6) at (6, 2) {$G$};
\node (A3_7) at (7, 3) {$E$};
\node (A2_8) at (8, 2) {$0$};
\node (A3_9) at (9, 3) {$0$};
\node (A4_2) at (2, 4) {$F_D$};
\node (A5_3) at (3, 5) {$F_D$};
\node (A4_4) at (4, 4) {$F_D$};
\node (A5_5) at (5, 5) {$F_D$};
\path (A3_5) edge [->>]node [auto] {$\scriptstyle{}$} (A5_5);
\path (A4_2) edge [thin, double distance=1.5pt]node [auto] {$\scriptstyle{}$} (A5_3);
\path (A5_3) edge [thin, double distance=1.5pt]node [auto] {$\scriptstyle{}$} (A5_5);
\path (A4_4) edge [thin, double distance=1.5pt]node [auto] {$\scriptstyle{}$} (A5_5);
\path (A0_0) edge [->]node [auto] {$\scriptstyle{}$} (A0_2);
\path (A0_2) edge [->]node [auto] {$\scriptstyle{\tilde{i}'}$} (A0_4);
\path (A0_4) edge [->]node [auto] {$\scriptstyle{\tilde{\pi}'}$} (A0_6);
\path (A0_6) edge [->]node [auto] {$\scriptstyle{}$} (A0_8);
\path (A1_7) edge [->]node [auto] {$\scriptstyle{}$} (A1_9);
\path (A0_2) edge [thin, double distance=1.5pt]node [auto] {$\scriptstyle{}$} (A1_3);
\path (A0_6) edge [->]node [auto] {$\scriptstyle{\pi}$} (A1_7);
\path (A2_0) edge [->]node [auto] {$\scriptstyle{}$} (A2_2);
\path (A3_5) edge [->]node [auto] {$\scriptstyle{\pi}$} (A3_7);
\path (A3_7) edge [->]node [auto] {$\scriptstyle{}$} (A3_9);
\path (A2_2) edge [thin, double distance=1.5pt]node [auto] {$\scriptstyle{}$} (A3_3);
\path (A2_4) edge [->]node [auto] {$\scriptstyle{t'}$} (A3_5);
\path (A2_6) edge [->]node [auto] {$\scriptstyle{\pi}$} (A3_7);
\path (A0_2) edge [right hook->]node [auto] {$\scriptstyle{}$} (A2_2);
\path (A1_1) edge [-,draw=white,line width=4pt]node [auto] {$\scriptstyle{}$} (A1_3);
\path (A1_1) edge [->]node [auto] {$\scriptstyle{}$} (A1_3);
\path (A0_4) edge [right hook->]node [auto] {$\scriptstyle{}$} (A2_4);
\path (A1_3) edge [-,draw=white,line width=4pt]node [auto] {$\scriptstyle{}$} (A1_5);
\path (A1_3) edge [->]node [auto] {$\scriptstyle{\tilde{i}}$} (A1_5);
\path (A0_6) edge [thin, double distance=1.5pt]node [auto] {$\scriptstyle{}$} (A2_6);
\path (A1_5) edge [-,draw=white,line width=4pt]node [auto] {$\scriptstyle{}$} (A1_7);
\path (A1_5) edge [->,near start]node [auto] {$\scriptstyle{\tilde{\pi}}$} (A1_7);
\path (A2_2) edge [->>]node [auto] {$\scriptstyle{}$} (A4_2);
\path (A3_1) edge [-,draw=white,line width=4pt]node [auto] {$\scriptstyle{}$} (A3_3);
\path (A3_1) edge [->]node [auto] {$\scriptstyle{}$} (A3_3);
\path (A2_4) edge [->>]node [auto] {$\scriptstyle{}$} (A4_4);
\path (A3_3) edge [-,draw=white,line width=4pt]node [auto] {$\scriptstyle{}$} (A3_5);
\path (A3_3) edge [->,near start]node [auto] {$\scriptstyle{i}$} (A3_5);
\path (A2_2) edge [->,near start]node [auto] {$\scriptstyle{i'}$} (A2_4);
\path (A1_3) edge [-,draw=white,line width=4pt]node [auto] {$\scriptstyle{}$} (A3_3);
\path (A1_3) edge [right hook->]node [auto] {$\scriptstyle{}$} (A3_3);
\path (A2_4) edge [->,near start]node [auto] {$\scriptstyle{\pi'}$} (A2_6);
\path (A1_5) edge [-,draw=white,line width=4pt]node [auto] {$\scriptstyle{}$} (A3_5);
\path (A1_5) edge [right hook->]node [auto] {$\scriptstyle{}$} (A3_5);
\path (A2_6) edge [->]node [auto] {$\scriptstyle{}$} (A2_8);
\path (A1_7) edge [-,draw=white,line width=4.5pt]node [auto] {$\scriptstyle{}$} (A3_7);
\path (A1_7) edge [thin, double distance=1.5pt]node [auto] {$\scriptstyle{}$} (A3_7);
\path (A4_2) edge [thin, double distance=1.5pt]node [auto] {$\scriptstyle{}$} (A4_4);
\path (A3_3) edge [-,draw=white,line width=4.5pt]node [auto] {$\scriptstyle{}$} (A5_3);
\path (A3_3) edge [->>]node [auto] {$\scriptstyle{}$} (A5_3);
\end{tikzpicture}
\end{equation*}
By diagram chasing, one can define a morphism $\tilde{G}'\rightarrow \tilde{G}$ such that the corresponding diagram commutes. Thus the image of $w$ via the map $\mathrm{Ext}_X^1(E,E(-D)) \rightarrow\mathrm{Ext}^1_\mathcal{X}(G,E(-D))$ is exactly $\mathcal{A}t_S(\mathcal{G}).$ This completes the proof.
\endproof
\section{Closed two-forms on moduli spaces of framed sheaves}\label{sec:framedtwoforms}
In this section we show how to construct closed two-forms on the moduli space $\mathcal{M}^*_\delta(X;F_D,P)^{sm}$ by using global sections of the line bundle $\omega_X(2\,D).$ Moreover, we establish a nondegeneracy criterion for these two-forms.
Let $[(E,\alpha)]$ be a point in $\mathcal{M}^*_\delta(X;F_D,P)^{sm}.$ By Theorem \ref{thm:tanbundle}, the vector space $\mathrm{Ext^1}(E,E(-D))$ is naturally identified with the tangent space $T_{[(E,\alpha)]}\mathcal{M}^*_\delta(X;F_D,P).$
For any $\omega\in \mathrm{H}^0(X,\omega_X(2\,D))$, we can define a skew-symmetric bilinear form
\begin{eqnarray*}
&&\mathrm{Ext}^1(E,E(-D))\times \mathrm{Ext}^1(E,E(-D))\stackrel{\circ}{\longrightarrow} \mathrm{Ext}^2(E,E(-2\,D))\\
&&\stackrel{tr}{\longrightarrow} \mathrm{H}^2(X, \mathcal{O}_X(-2\,D))\stackrel{\cdot\, \omega}{\longrightarrow} \mathrm{H}^2(X,\omega_X)\cong k.
\end{eqnarray*}
By varying the point $[(E,\alpha)]$, these forms fit into a exterior two-form $\tau(\omega)$ on $\mathcal{M}^*_\delta(X;F_D,P)^{sm}.$
\begin{theorem}\label{thm:closed}
For any $\omega\in \mathrm{H}^0(X,\omega_X(2\,D))$, the two-form $\tau(\omega)$ is closed on $\mathcal{M}^*_\delta(X;F_D,P)^{sm}.$
\end{theorem}
\proof
It suffices to prove that given a smooth affine variety $S$, for any $S$-flat family $\mathcal{E}=(E,\alpha)$ of framed sheaves on $X$ defining a classifying morphism
\begin{eqnarray*}
\psi\colon S&\longrightarrow & \mathcal{M}^*_\delta(X;F_D,P)^{sm},\\
s&\longmapsto& [\mathcal{E}\vert_{\{s\}\times X}],
\end{eqnarray*}
the pullback $\psi^*(\tau(\omega))\in \mathrm{H}^0(S,\Omega^2_S)$ is closed. Since $\psi^{*}(\tau(\omega))=\tau_S(\omega)$ by construction, this follows from Proposition \ref{prop:closedness}.
\endproof
Thus we have constructed closed two-forms $\tau(\omega)$ on the moduli space $\mathcal{M}^*_\delta(X;F_D,P)^{sm}$ depending on a choice of $\omega\in \mathrm{H}^0(X,\omega_X(2\,D)).$ In general, these forms may be degenerate.
Now we give a criterion to check when the two-forms are non-degenerate.
\begin{proposition}\label{prop:non-dege}
Let $\omega\in \mathrm{H}^0(X,\omega_X(2\,D))$ and $[(E,\alpha)]$ a point in $\mathcal{M}^*_\delta(X;F_D,P)^{sm}.$ The closed two-form $\tau(\omega)_{[(E,\alpha)]}$ is non-degenerate at the point $[(E,\alpha)]$ if and only if the multiplication by $\omega$ induces an isomorphism
\begin{equation*}
\omega_*\colon \mathrm{Ext}^1(E,E(-D))\longrightarrow \mathrm{Ext}^1(E,E\otimes \omega_X(D)).
\end{equation*}
\end{proposition}
\proof
The proof is similar to that of \cite[Proposition 10.4.1]{book:huybrechtslehn2010}.
\endproof
Obviously, if the line bundle $\omega_X(2\, D)$ is trivial, for any point $[(E,\alpha)]$ in $\mathcal{M}^*_\delta(X;F_D,P)^{sm}$ the pairing
\begin{equation*}
\tau(1)\colon \mathrm{Ext}^1(E,E(-D))\times \mathrm{Ext}^1(E,E(-D))\longrightarrow k
\end{equation*}
is a non-degenerate alternating form, where $1\in \mathrm{H}^0(X,\omega_X(2\,D))\cong k.$
\section{Examples}\label{sec:example}
In this section we provide explicit examples of holomorphic symplectic structures on moduli spaces of $(D,F_D)$-framed sheaves with fixed Hilbert polynomial on some birationally ruled surfaces. In particular, if one compares our costruction of symplectic structures on the moduli spaces of framed sheaves on $\mathbb{CP}^1\times \mathbb{CP}^1$, $C\times \mathbb{CP}^1$ with $C$ elliptic curve, and on the second Hirzebruch surface $\mathbb{F}_2$ to that of Bottacin (cf. \cite{art:bottacin2000}), one can see that they are equivalent on the locally free part of the moduli space. On the other hand, we construct new examples of symplectic structures on the moduli spaces of $(D, F_D)$-framed sheaves over $\mathbb{CP}^2$ and the blowup of $\mathbb{CP}^2$ at a point, not covered by Bottacin's result.
These constructions are based upon the results of Theorem \ref{thm:bruzzomarkushevich} and Corollary \ref{cor:smoothness}. Useful references on birationally ruled surfaces are \cite[Chapter V]{book:hartshorne1977} and \cite{book:beauville1996}.
\subsection*{The complex projective plane}
Let us denote by $l_\infty$ a line in $\mathbb{CP}^2.$ Fix a positive integer $d$ and a smooth connected curve $D$ in the complete linear system $\vert d\, l_\infty\vert.$ The genus of $D$ is $\frac{(d-1)(d-2)}{2}.$
Let $c, n\in \mathbb{Z}$ and $F_D$ a Gieseker-semistable locally free $\mathcal{O}_D$-module of rank $r$ and degree $cd$ on $D.$ Let us denote by $\mathcal{M}^*(\mathbb{CP}^2;F_D,r,c\,l_\infty,n)$ the moduli space of $(D,F_D)$-framed sheaves on $\mathbb{CP}^2$ with rank $r$, first Chern class $c\,l_\infty$ and second Chern class $n.$ Since $K_{\mathbb{CP}^2}=-3\, l_\infty$, we get $K_{\mathbb{CP}^2}+2\, D=(-3+2d) \,l_\infty.$ Hence $\mathrm{H}^0(\mathbb{CP}^2,\omega_{\mathbb{CP}^2}(2\,D))$ is nonzero only for $2d-3\geq 0.$ Therefore, when $d=1$, that is, $D$ is a line, we cannot apply our method to construct closed two-forms on the moduli space $\mathcal{M}^*(\mathbb{CP}^2;F_D,r,c\,l_\infty,n)^{sm}.$
Fix $d=2.$ In this case $D$ is a nondegenerate conic in $\mathbb{CP}^2.$ Let $l$ be a line in $\mathbb{CP}^2$ of equation $\omega_l=0$, where $\omega_l\in\mathrm{H}^0(\mathbb{CP}^2,\omega_{\mathbb{CP}^2}(2\, D))=\mathrm{H}^0(\mathbb{CP}^2,\mathcal{O}_{\mathbb{CP}^2}(1)).$ Since for locally free sheaves on $\mathbb{CP}^1$ the triviality condition is equivalent to the semistability condition, we can define the open subset $\mathcal{M}_{lf}^*(\mathbb{CP}^2;F_D,r,c\,l_\infty,n)\subset \mathcal{M}^*(\mathbb{CP}^2;F_D,r,c\,l_\infty,n)$ consisting of isomorphism classes of $(D,F_D)$-framed vector bundles $[(E,\alpha)]$ on $\mathbb{CP}^2$, with $E$ trivial along $l.$
Let $[(E,\alpha)]$ be a point in $\mathcal{M}_{lf}^*(\mathbb{CP}^2;F_D,r,c\,l_\infty,n)^{sm}.$ Consider the induced map on the Ext-group given by the multiplication by $\omega_l$:
\begin{equation*}
(\omega_l)_*\colon \mathrm{Ext}^1(E,E(-D))\longrightarrow \mathrm{Ext}^1(E,E\otimes \omega_X(D)),
\end{equation*}
that is,
\begin{equation*}
(\omega_l)_*\colon \mathrm{Ext}^1(E,E(-2))\longrightarrow \mathrm{Ext}^1(E,E(-1)).
\end{equation*}
If we tensor by $E$ and then apply the functor $\mathrm{Hom}(E,\cdot)$ to the short exact sequence
\begin{equation*}
0\longrightarrow \mathcal{O}_{\mathbb{CP}^2}(-2)\stackrel{\cdot \omega_l}{\longrightarrow} \mathcal{O}_{\mathbb{CP}^2}(-1)\longrightarrow \mathcal{O}_l(-1)\longrightarrow 0,
\end{equation*}
we get the long exact sequence
\begin{equation*}
\cdots \rightarrow \mathrm{Hom}(E\vert_l,E\vert_l(-1))\rightarrow \mathrm{Ext}^1(E,E(-2))\stackrel{(\omega_l)_*}{\rightarrow} \mathrm{Ext}^1(E,E(-1))\rightarrow \mathrm{Ext}^1(E\vert_l,E\vert_l(-1))\rightarrow \cdots
\end{equation*}
Since $E$ is trivial along $l$ and $\mathrm{H}^0(l,\mathcal{O}_l(-1))=\mathrm{H}^1(l,\mathcal{O}_l(-1))=0$, we get that $(\omega_l)_*$ is an isomorphism, hence $\tau(\omega_l)_{[(E,\alpha)]}$ is non-degenerate at $[(E,\alpha)]$ by Proposition \ref{prop:non-dege}. Therefore $\tau(\omega_l)$ defines a holomorphic symplectic structure on $\mathcal{M}_{lf}^*(\mathbb{CP}^2;F_D,r,c\,l_\infty,n)^{sm}.$ For $F_D\cong \mathcal{O}_D^{\oplus r}$, we get $c=0$ and $\mathcal{M}_{lf}^*(\mathbb{CP}^2;\mathcal{O}_D^{\oplus r},r,0,n)$ is a holomorphic symplectic variety of dimension $2rn.$
\begin{remark}
In this case $\omega^{-1}_{\mathbb{CP}^2}(-2\,D)\cong \mathcal{O}_{\mathbb{CP}^2}(-1)$, which has not global sections. Hence it is not possible to use Bottacin's result to construct (nondegenerate) Poisson structures on $\mathcal{M}_{lf}^*(\mathbb{CP}^2;\mathcal{O}_D^{\oplus r},r,0,n).$ \triend
\end{remark}
\subsection*{The Hirzebruch surfaces}
Let $p$ be a nonnegative integer number. We denote by $\mathbb{F}_p$ the $p$-th Hirzebruch surface $\mathbb{F}_p:=\mathbb{P}(\mathcal{O}_{\mathbb{CP}^1}\oplus \mathcal{O}_{\mathbb{CP}^1}(-p))$, which is the projective closure of the total space of the line bundle $\mathcal{O}_{\mathbb{CP}^1}(-p)$ on $\mathbb{CP}^1.$ One can describe explicitly $\mathbb{F}_p$ as the divisor in $\mathbb{CP}^2\times \mathbb{CP}^1$
\begin{equation*}
\mathbb{F}_p=\{([z_0:z_1:z_2],[z:w])\in \mathbb{CP}^2\times \mathbb{CP}^1\,\vert\, z_1w^p=z_2 z^p\}.
\end{equation*}
Let us denote by $\pi\colon \mathbb{F}_p \rightarrow \mathbb{CP}^2$ the projection onto $\mathbb{CP}^2$ and by $l_\infty$ the inverse image of a generic line of $\mathbb{CP}^2$ through $\pi.$ Then $l_\infty$ is a smooth connected big and nef curve of genus zero. The Picard group of $\mathbb{F}_p$ is generated by $l_\infty$ and the fibre $F$ of the projection $\mathbb{F}_p\rightarrow \mathbb{CP}^1.$ One has
\begin{equation*}
l_\infty^2=p,\; l_\infty\cdot F=1,\; F^2=0.
\end{equation*}
In particular, the canonical divisor $K_p$ can be expressed as
\begin{equation*}
K_p=-2\,l_\infty +(p-2)\,F.
\end{equation*}
Let us consider the case $p=0.$ In this case $\mathbb{F}_0$ is the ruled surface $\mathbb{CP}^1\times \mathbb{CP}^1.$ Moreover, $K_0=-2\,l_\infty -2F.$ By \cite[Corollary V-2.18]{book:hartshorne1977}, there exists a smooth connected ample curve $D$ in the complete linear system $\vert l_\infty +F\vert.$ By the adjuction formula (see, e.g., \cite[Proposition V-1.5]{book:hartshorne1977}), $D$ has genus zero.
Let $n\in \mathbb{Z}$ and $F_D$ a Gieseker-semistable locally free $\mathcal{O}_D$-module of rank $r$ and degree $a+b$, for $a,b\in \mathbb{Z}$, on $D.$ Let $\mathcal{M}^*(\mathbb{F}_0;F_D,r,a\,l_\infty+b\,F,n)$ be the moduli space of $(D,F_D)$-framed sheaves on $\mathbb{F}_0$ of rank $r$, first Chern class $a\,l_\infty+b\,F$ and second Chern class $n.$ Since $K_{\mathbb{F}_0}+2\,D=0$, the line bundle $\omega_{\mathbb{F}_0}(2\,D)$ is trivial and for $1\in \mathrm{H}^0(\mathbb{F}_0,\omega_{\mathbb{F}_0}(2\,D))\cong \mathbb{C}$, the two-form $\tau(1)$ defines a holomorphic symplectic structure on $\mathcal{M}^*(\mathbb{F}_0;F_D,r,a\,l_\infty+b\,F,n)^{sm}.$ If $F_D\cong \mathcal{O}_D^{\oplus r}$, we have $b=-a$ and the moduli space $\mathcal{M}^*(\mathbb{F}_0;F_D,r,a(l_\infty-F),n)$ is a holomorphic symplectic variety of dimension $2(rn+(r-1)a^2).$
Consider $p=1.$ The first Hirzebruch surface $\mathbb{F}_1$ is isomorphic to the blowup of $\mathbb{CP}^2$ at a point. Consider the complete linear system $\vert l_\infty + F\vert.$ Again, by \cite[Corollary V-2.18]{book:hartshorne1977}, there exists a smooth connected curve $D$ on $\vert l_\infty + F\vert.$ By \cite[Theorem V-2.17]{book:hartshorne1977}, $\vert l_\infty + F\vert$ is base-point-free, hence $D$ is nef. Since $D^2=3$, $D$ is also big. By the adjuction formula, the genus of $D$ is zero. Moreover, from $K_1=-2\,l_\infty-F$, it follows that $K_1+2\,D=F.$
Let $n\in \mathbb{Z}$ and $F_D$ a Gieseker-semistable locally free $\mathcal{O}_D$-module of rank $r$ and degree $2a+b$, for $a,b \in \mathbb{Z}$, on $D.$ Let $\mathcal{M}^*(\mathbb{F}_1;F_D,r,a\,l_\infty+b\,F,n)$ be the moduli space of $(D,F_D)$-framed sheaves on $\mathbb{F}_1$ of rank $r$, first Chern class $a\,l_\infty +b\,F$ and second Chern class $n.$ Let $l$ be a smooth connected curve of genus zero in $\mathbb{F}_1$ defined by a nonzero section $\omega_l\in \mathrm{H}^0(\mathbb{F}_1,\omega_{\mathbb{F}_1}(2\, D)).$ As in the previous example, the two-form $\tau(\omega_l)$ defines a holomorphic symplectic structure on the smooth locus of the moduli space $\mathcal{M}_{lf}^*(\mathbb{F}_1;F_D,r,a\,l_\infty+b\,F,n)$ parametrizing isomorphism classes of $(D,F_D)$-framed vector bundles $[(E,\alpha)]$ on $\mathbb{F}_1$, with $E$ trivial along $l.$ For $F_D\cong \mathcal{O}_D^{\oplus r}$, $b=-2a$ and $\mathcal{M}_{lf}^*(\mathbb{F}_1;F_D,r,a\,l_\infty-2a\,F,n)$ is a holomorphic symplectic variety of dimension $2rn+3(r-1)a^2.$
\begin{remark}
As in the previous example, this case is not covered by Bottacin's result since $\omega_{\mathbb{F}_1}^{-1}(-2\, D)\cong \mathcal{O}_{\mathbb{F}_1}(-F)$, which has not global sections. \triend
\end{remark}
Finally, let $p=2.$ In this case $\mathbb{F}_2$ is the projective closure of the cotangent bundle $T^*\mathbb{CP}^1$ of the complex projective line $\mathbb{CP}^1.$ Let $D=l_\infty$, $n\in \mathbb{Z}$ and $F_D$ a Gieseker-semistable locally free $\mathcal{O}_D$-module of rank $r$ and degree $2a+b$, with $a,b \in \mathbb{Z}$, on $D.$
Let $\mathcal{M}^*(\mathbb{F}_2;F_D,r,a\,l_\infty+b\,F,n)$ be the moduli space of $(D,F_D)$-framed sheaves on $\mathbb{F}_2$ of rank $r$, first Chern class $al_\infty +bF$ and second Chern class $n.$ The canonical divisor of $\mathbb{F}_2$ is $K_2=-2\,D$, hence the line bundle $\omega_{\mathbb{F}_2}(2\, D)$ is trivial and, for $1\in \mathrm{H}^0(\mathbb{F}_2,\omega_{\mathbb{F}_2}(2\,D))\cong \mathbb{C}$, the two-form $\tau(1)$ defines a symplectic structure on $\mathcal{M}^*(\mathbb{F}_2;F_D,r,a\,l_\infty+b\,F,n)^{sm}.$
If $F_D\cong \mathcal{O}_D^{\oplus r}$, we have $b=-2a.$ Let us define $C=l_\infty-2F.$ This is the only irreducible curve in $\mathbb{F}_2$ with negative self intersection. We can normalize the value $a$ in the range $0\leq a \leq r-1$ upon twisting by $\mathcal{O}_{\mathbb{F}_2}(C).$ By \cite[Theorem 3.4]{phd:rava2012}, the moduli space $\mathcal{M}^*(\mathbb{F}_2;F_D,r,a\,C,n)$ is nonempty if and only if $n+a(a-1)\geq 0$, and if this is the case, $\mathcal{M}^*(\mathbb{F}_2;F_D,r,a\,C,n)$ is a holomorphic symplectic variety of dimension $2(rn+(r-1)a^2).$
\subsection*{A ruled surface over an elliptic curve}
Let $C$ be an elliptic curve. Then $C\times \mathbb{CP}^1$ with its first projection is a ruled surface. By \cite[Proposition V-2.8]{book:hartshorne1977}, there exists a section $C_0$ of $C\times \mathbb{CP}^1$, which is an ample divisor. Let $F$ denote the fibre of the projection $C\times \mathbb{CP}^1\rightarrow C.$ Then $\mathrm{Pic}(C\times \mathbb{CP}^1)=\mathbb{Z}C_0\oplus \mathbb{Z}F$, with $C_0^2=F^2=0$, and $C_0\cdot F=1.$ Moreover, $K_{C\times \mathbb{CP}^1}=-2\,C_0.$
Let $D$ be a smooth connected ample curve in the complete linear system $\vert C_0\vert.$ By the adjuction formula, the genus of $D$ is one. Let $n\in \mathbb{Z}$ and $F_D$ a Gieseker-semistable locally free $\mathcal{O}_D$-module of rank $r$ and degree $b\in \mathbb{Z}$ on $D.$
Let $\mathcal{M}^*(C\times \mathbb{CP}^1;F_D,r,a\,C_0+b\,F,n)$ be the moduli space of $(D,F_D)$-framed sheaves on $C\times \mathbb{CP}^1$ of rank $r$, first Chern class $a\,C_0 +b\,F$, with $a\in \mathbb{Z}$, and second Chern class $n.$ Also in this case, $K_{C\times \mathbb{CP}^1}+2\,D=0$, hence for $1\in \mathrm{H}^0(C\times \mathbb{CP}^1, \omega_{C\times \mathbb{CP}^1}(2\,D))\cong \mathbb{C}$, the two-form $\tau(1)$ defines a holomorphic symplectic structure on $\mathcal{M}^*(C\times \mathbb{CP}^1;F_D,r,a\,C_0+b\,F,n)^{sm}.$
|
\section{Introduction}
\vskip-.3cm
Jets are observed in Young Stellar Objects, post-AGB stars, X-ray
binaries, radio galaxies and other astrophysical objects. Models
suggest that jets are launched and collimated inside their ``central
engine'' by a symbiosis of accretion, rotation and magnetic mechanisms
(Pudritz et al.~\cite{pudritz}). The engines cannot be directly
observed however, because telescopes have insufficient resolution.
Recently, laboratory astrophysics experiments have
provided scale models of the launch and propagation of magnetized
jets with dimensionless parameters relevant for astrophysical systems (Lebedev
et~al.~\cite{leb5}; Ciardi et~al.~\cite{ciardi9}; Suzuki-Vidal
et~al.~\cite{suzuki}). These studies, when combined with numerical
simulations (Ciardi et~al.~\cite{ciardi7};
Huarte-Espinosa et~al.~\cite{we}), can help to resolve unanswered
questions in jet physics. A fundamental distinction can be made between the physics
of jet launch and that of jet propagation far from the engine. Since
we cannot resolve the former observationally, it is important to
identify distinct features of jets in the asymptotic propagation
regime that can distinguish different engine paradigms. While both
simulations and experiments now consistently reveal the promise,
if not essentiality, of dynamically significant magnetic fields for
jet launch, the correlation between the initial jet magnetic
configuration and the stability of the flow far from the launching
region is unclear. The importance of the magnetic field relative
to the flows' kinetic energy divides jets into (i) Poynting flux
dominated (PFD; Shibata \& Uchida~\cite{shibata}), in which magnetic
fields dominate the jet structure, (ii) magnetocentrifugal (Blandford
\& Payne~\cite{bland}), in which magnetic fields only dominate out
to the Alfv\'en radius. The observable differences between PFD and
magnetocentrifugal jets are unclear, as are the effects that cooling and
rotation have on PFD jets.
\section{Laboratory experiments}
\vskip-.3cm
Experiments and numerical modelling of astrophysically relevant
supersonic plasma jets have been performed by a number of authors
using high intensity lasers (see e.g. Farley et al.~\cite{laser1};
Shigemori et al.~\cite{laser2}; Foster et~al.~\cite{laser3};
see also review in Blackman 2007).
In this section however, we focus on magnetized jet experiments
which have dimensionless numbers (Reynolds, magnetic Reynolds and
Peclet) in reasonably appropriate regimes for astrophysics and have
used conical wire arrays and conducting foils on pulsed power
facilities (Lebedev et al. 2005; Ciardi et al. 2007,2009; Suzuki-Videl
et~al. 2010).
\subsection{Wire array, toroidal field}
\vskip-.3cm
Lebedev et~al.~(\cite{leb5}) applied a TW electrical pulse (1\,MA,
250\,ns) to an array of wires located inside a vacuum chamber. The
set up consisted of a pair of concentric electrodes connected radially
by tungsten wires of 13\,$\mu$m in diameter. The current causes
ablation of the wires which results in the formation of a
background ambient plasma (Fig.~1a).
This material is then pushed above the
wires by Lorentz forces, and resistive diffusion keeps the current
close to the wires. The current induces a toroidal magnetic field
which at this stage is confined below
the wires, around the central electrode. Then, after the complete ablation
of the wires near the central electrode, the current switches to
the plasma and creates a magnetic cavity (Fig.~1b) with a jet at its center.
The core of the jet is confined and accelerated upward by
the pressure of the toroidal field. The return current flows along
the walls of the magnetic cavity, which is in turn confined by both
the thermal pressure and the inertia of the ambient plasma. Next,
the magnetic cavity opens up, the jet becomes detached and it
propagates away from the source at a velocity of order 200\,km\,s$^{-1}$
(Fig.~1c).
\begin{figure}[hb]
\begin{center}
\includegraphics[width=\textwidth]{exp}
~~~~~~~~
(a)
~~~~~~~~~~~~~~
(b)
~~~~~~~~~~~~~~
(c)
~~~~~~~~~~~~~~
(d)
~~~~~~~~~~~~~~
(e)
~~~~~~~~~~~~~~
(f)
~~~~~~
\end{center}
\vspace{-15pt}
\caption{Time sequence of soft X-rays images showing the formation
of the background plasma (a), the expansion of the magnetic cavity (b), the
launch of the jet (c), the development of instabilities in the central
jet column (d) and the fragmentation of the jet (e,f).
The wire array shows in the bottom part of the figures.
Image taken from Suzuki-Vidal et~al.~(\cite{suzuki}).}
\end{figure}
Finally, instabilities, which resemble those of the kink mode ($m =\,$1), develop
within the body of the jet (Fig.~1d). The outflow is then fragmented into
well collimated structures with characteristic axial non-uniformities
(Fig.~1e,f; Lebedev et~al.~\cite{leb5}).
We note that
a critical ingredient of these experiments is the significant thermal
energy loss of both the jet and the ambient plasmas.
This is relevant because cooling plays a critical role in
many astrophysical jet environments, e.g. YSOs.
The experiment was reproduced a few years later by
Ciardi et~al.~(\cite{ciardi7}) using 3-D non-ideal MHD simulations.
A numerical model was carefully designed to simulate
the laboratory components (electrodes and wires) and all the plasma
evolution phases. Good agreement was found between the simulations
and the experiments (Ciardi et~al.~\cite{ciardi7}). The simulations showed
that during the final unstable phase of jet propagation, the magnetic
fields in the central jet adopted a twisted helical structure. This
confirmed that the jets are affected by normal mode, $m=\,$1, perturbations.
\subsection{Wire array, toroidal and axial fields}
\vskip-.3cm
A subsequent series of experiments were carried out by Suzuki-Vidal
et~al.~(\cite{suzuki}) who studied the effect of introducing an axial
magnetic field, $B_z$, into a radial wire array. Their experimental
configuration was the same as that of Lebedev et~al.~(\cite{leb5},
above), but they placed two solenoids below the wires, into the
path of the current. Suzuki-Vidal et~al.~(\cite{suzuki}) found that
the added $B_z$ affects the degree of compression of
the on-axis pinch of the jet. As they increased the magnitude of $B_z$,
the plasma column radially expanded after reaching a minimum radius.
They also saw that jets are more stable
when $B_z$ is present than otherwise. We note that
this stability effect can be explained analytically (with perturbation theory
on a plasma column) and numerically (section~3) as a result of
magnetic pressure form $B_z$ acting against hoop
stress compression from the toroidal field component.
\subsection{Conducting foil}
\vskip-.3cm
\begin{wrapfigure}[22]{r}{0.36\textwidth}
\vspace{-16pt}
\begin{center}
\includegraphics[width=.35\columnwidth]{ep.eps}
\vspace{-15pt}
\end{center}
\caption{Filtered XUV emission images of one of the episodic jet
experiments. The clumpy jet at the center is
contained inside three nested magnetic bubbles which were formed
by previously launched jets (Ciardi et~al.~\cite{ciardi9}).}
\end{wrapfigure}
Ciardi et~al.~(\cite{ciardi9}) extended the experiments of Lebedev
et~al.~(\cite{leb5}) by replacing the wire array (section~2.1) by a 6 $\mu$m
thick aluminum foil. Ciardi et~al.~(\cite{ciardi9}) found the following.
The conducting foil experiment produced very similar results than
the wire array one. However,
while the wire array experiment results in the launch of one jet only,
the conducting foil experiment results in a series of jets which
are launched sequentially (Fig.~2). The plasma flux caused by ablation of
the foil is much larger than that of the wires; the foil
provides an increased mass as a function of radius. Thus the current
gap produced by the magnetic cavity (section~2.1) is smaller in
the foil experiment than in the wire array one, and can be refilled
by the readily available plasma.
Closure of the gap, which does not happen in the wire array case,
allows the current to flow once again across the base of the magnetic
cavity, thus re-establishing the initial configuration. Once the
magnetic pressure is large enough to break through this newly
deposited mass, a new jet/bubble ejection cycle begins (Ciardi
et~al.~\cite{ciardi9}). This constituted the first laboratory study
to address time-dependent episodic jet launch.
\section{Simulations}
\vskip-.3cm
\subsection{Model}
\vskip-.3cm
We use the Adaptive Mesh Refinement code AstroBEAR2.0 (Cunningham
et~al.~\cite{bear}; Carroll-Nellenback et~al.~\cite{bear2}) to solve
the equations of radiative-MHD in 3D. The grid represents
160$\times$160$\times$400\,AU divided into 64$\times$64$\times$80
cells plus 2 adaptive refinement levels. We use periodic boundary
conditions at the four vertical faces of the domain, extrapolation
conditions at the top face and a combination of reflective and
inflow conditions at the bottom face. Initially, molecular gas
is static and has an ideal gas equation of state ($\gamma=\,$5$/$3),
a number density of 100\,cm\,s$^{-1}$ and a temperature of 10000\,K.
The magnetic field is helical, centrally localized and given by the
vector potential (in cylindrical coordinates)
\begin{equation}
{\bf A}(r,z) = \left\{
\begin{array}{c l}
\frac{r}{4} (cos(2\,r) + 1)( cos(2\,z) + 1 ) \hat{\phi} +
\frac{\alpha}{8} (cos(2\,r) + 1)( cos(2\,z) + 1 ) \hat{k},
& \mbox{for}~r,z < r_e; \\
0, & \mbox{for}~r,z \ge r_e,
\end{array} \right.
\label{apot}
\end{equation}
\noindent where $r_e=\,$30\,AU and
$\alpha=\,$40 has units of length and determines the ratio of
toroidal to poloidal magnetic fluxes.
The magnetic pressure exceeds the thermal pressure
inside the magnetized region.
Source terms continually inject magnetic or kinetic
energy at cells $r,z<r_e$. We carry out 4 simulations: an
adiabatic, a cooling, a rotating PFD jet, and a hydrodynamical
jet. The latter is constructed to have same time average propagation speed and energy
flux as the adiabatic PFD jet. The cooling tables of Dalgarno \&
McCray~(\cite{dm}) were used for the cooling case. For the rotating case
we allowed the gas and magnetic fields within $r,z<r_e$ to
move in Keplerian rotation.
\begin{wrapfigure}[21]{r}{0.66\textwidth}
\vspace{-42pt}
\begin{center}
\includegraphics[width=0.65\textwidth]{figure1} \\
\end{center}
\vspace{-20pt}
\caption{Logarithmic density maps of the adiabatic (1st column), rotating
(2nd column) and cooling (3rd column) PFD jets. Hydrodynamic jet
(4th column). From top to bottom the time is 42, 84 and 118\,yr.}
\vspace{-10pt}
\end{wrapfigure}
\subsection{Results}
\vskip-.3cm
Magnetic pressure pushes field lines and plasma upward, forming magnetic
cavities with low density (Fig.~3, all but right panel).
The adiabatic case is the most stable.
PFD jets decelerate more quickly relative to our hydro jet. This results because the PFD and hydro jets have the same injected energy flux, but the PFD case produces not only axial but radial expansion.
The pre-collimated hydro can only expand via a much lower thermal pressure. Thus all of the energy flux in the hyrdo-case for our set up is more efficiently directed to axial mechanical power.
In principle, our hydro case can emulate the asymptotic propagation regime of a jet that was magneto-centrifugally launched (e.g. Blackman 2007) which is distinct from a PFD jet.
The PFD jet cores are thin and unstable, whereas the hydro jet beam is thicker,
smoother and stable. The PFD jets are sub-Alfv\'enic. Their cores
are confined by magnetic hoop stress, while their surrounding
cavities are collimated by external thermal pressure. PFD jets carry
high axial currents which return along their outer contact discontinuity.
We see that the inner regions of the PFD jets, just outside the core, are low beta plasma columns in which
the axial magnetic field dominates over the toroidal one,
$|B_{\phi} / B_z | \ll 1$. The columns' instability condition is given by
\begin{equation}
\left| \frac{B_{\phi}}{B_z} \right| > | (\beta_z - 1)k r_{jet} |,
\label{insta}
\end{equation}
\noindent where $\beta_z=2 \mu_0 P / B_z^2$,
$\mu_0$ is the magnetic vacuum permeability,
$P$ is the plasma thermal pressure and $k^{-1}$ is the
characteristic wavelength of the current-driven perturbations.
We find that the cooling jet shows $\beta_z \sim\,$1
from early in the simulation (Fig.~4), and thus
does not have sufficient thermal energy to damp the
magnetic pressure kink perturbations. The latter grow
exponentially.
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=.32,angle=90]{velsLabel.ps}
\includegraphics[width=.315\columnwidth,bb=70 150 560 665,clip=]{1-vels-cho20-polarB.ps}
\hskip-.1cm
\includegraphics[width=.29\columnwidth,bb=125 150 560 645,clip=]{rota-vels-cho20-polarB.ps}
\hskip-.1cm
\includegraphics[width=.29\columnwidth,bb=125 150 560 645,clip=]{cool-vels-cho20-polarB.ps} \\
Adiabatic
~~~~~~~~~~~~~~~~~~~~~~~~
Rotating
~~~~~~~~~~~~~~~~~~~~~~~~
Cooling
\end{center}
\vspace{-15pt}
\caption{Plasma velocities along the axes of the PFD jets after expanding for 84\,yr.
The solid, dashed, dotted, dot-dash and dot-dash thick
lines represent the Alfv\'en speed, the sound speed, $v_z, v_x$ and $v_y$,
respectively. Each velocity unit represents 9.1\,km\,s$^{-1}$.}
\end{figure}
A different path to instability operates for the rotating (non-cooling) case.
For this case, rotation at the base of the jet causes a slow amplification
of the toroidal magnetic field. Hence the
left hand side of equation~(\ref{insta}) increases slowly, and so
do the kink mode perturbations. The rotating jet is not completely
destroyed by these perturbations, and their amplitude is about twice
the radius of the central jet (Fig~5), in agreement with the
Kruskal-Shafranov criterion (Kruskal et~al.~\cite{kruskal};
Shafranov~\cite{shafranov}). \\
In Figure~4 we show profiles of the relevant velocities of the PFD
jets along
the jet axis, as a function of cooling and rotation.
We also followed these in time. During the stable propagation
phase, we find that the jets are mostly sub-Alfv\'enic and trans-sonic,
independent of cooling or rotation. Fast-forward compressive MHD
(FF) and transmitted hydrodynamic shocks are evident in the ambient
medium, ahead of the jets. The FF shocks steepen in time,
whereas the hydrodynamic shocks are quickly dissipated in the cooling
case. In contrast, the adiabatic and rotating cases show regions
within the lower half of the jets, where the sound speed
is super-Alfv\'enic. Such regions are bounded by the reverse and
the forward slow-modes of compressive MHD waves, and characterized
by high thermal to magnetic pressure ratios.
\section{Conclusions}
\vskip-.3cm
\begin{wrapfigure}[29]{r}{0.69\textwidth}
\vspace{-39pt}
\begin{center}
Magnetic field strength [$\mu$G]\\
\includegraphics[width=.23\columnwidth,bb=5 605 375 670,clip=]{lines-scaleb.ps} \\
\includegraphics[width=.22\columnwidth,bb=160 120 450 700,clip=]{lines-128b.ps}
\includegraphics[width=.22\columnwidth,bb=160 120 450 700,clip=]{lines-rota28b.ps}
\includegraphics[width=.22\columnwidth,bb=160 120 450 700,clip=]{lines-cool28b.ps}
\includegraphics[width=.22\columnwidth]{linesUP-128b.ps}
\includegraphics[width=.22\columnwidth]{linesUP-rota28b.ps}
\includegraphics[width=.22\columnwidth]{linesUP-cool28b.ps}
\end{center}
\vspace{-10pt}
\caption{Central magnetic field lines at
$t=\,$118\,yr. From left to right these are the adiabatic, the
rotating and the cooling PFD jets, respectively. Bottom
panels show a pole-on view.}
\end{wrapfigure}
PFD jets can be produced in pulsed power laboratory facilities and
help to understand the physics of astrophysical jets. The PFD ``lab jets''
are collimated by hoop stress, and their outer magnetic cavities
are collimated by external pressure. The axial magnetic component
influences both the radial compression and the stability of the jets.
The outflows evolve to become corrugated, but still collimated, structures due to current-driven
instabilities. Thin conduction foil experiments produce episodic
jets and nested magnetic bubbles. Observations of multiple lobes
in the radio galaxy B0925+420, may be consistent with such processes (Brocksopp et
al. \cite{epis}).
Our simulations show that PFD jet beams are lighter, slower and
less stable than pre-collimated asymptotically hydrodynamic jets.
In practice the latter could represent the asymptotic propagation
regimes of magneto-centrifugally launched jets, which are distinct
from PFD in that PFD remain PFD out to much larger scales. We
find that current-driven perturbations in PFD jets are amplified
by both cooling and rotation for the regimes studied: Shocks and
thermal pressure support are weakened by cooling, making the jets
more susceptible to kinking. Rotation amplifies the toroidal magnetic
field which also exacerbates the kink instability. Our simulations
agree well with the models and experiments of Shibata \&
Uchida~(\cite{shibata}) and Lebedev et~al.~(\cite{leb5}), respectively.
\acknowledgements Financial support for this project was provided by the Space Telescope
Science Institute grants HST-AR-11251.01-A and HST-AR-12128.01-A;
by the National Science Foundation under award AST-0807363; by the
Department of Energy under award DE-SC0001063; by
NSF grant PHY0903797, and by Cornell
University grant 41843-7012.
|
\section{Introduction\label{sec:intro}}
In applying the idea of chiral effective field theory (EFT) to nuclear physics, a great deal of effort has been devoted to implementing Weinberg's original prescription~\cite{Weinberg:1990-1991} for the problems of few-nucleon systems, with the two-nucleon system as the starting point~\cite{Ordonez:1993-1995, Epelbaum:1998ka-1999dj, Epelbaum:2004fk, Entem:2001cg-2002sf, PavonValderrama:2005uj, YangOhio, saopaulo} (for more general reviews, see Refs.~\cite{vanKolck:1999mw, Beane:2000fx, Bedaque:2002mn, Epelbaum:2005pn, Epelbaum:2008ga, Machleidt:2010kb}). While the power counting of pion exchange diagrams follows the paradigm of chiral perturbation theory (ChPT), i.e., chiral EFT in the single-nucleon sector, estimating the size of $NN$ contact interactions often requires assumptions beyond chiral symmetry.
Assumed in Weinberg's power counting (WPC) is what we refer to as naive dimensional analysis (NDA): each derivative on or each power of pion mass dependence of the Lagrangian terms is always suppressed by the underlying scale of chiral EFT, $M_{\text{hi}} \sim m_\sigma$, where $m_\sigma$ is the mass of the $\sigma$ meson. While plausible, this assumption was questioned in a number of works~\cite{Kaplan:1996xu, Beane:2000wh, Beane:2001bc, Nogga:2005hy, Birse:2005um, Birse:2009my, Valderrama:2009ei, Valderrama:2011mv, Long:2011qx}, and was shown to be only partially correct from the perspective of renormalization group (RG) invariance; WPC does not have enough $NN$ contact interactions, or counterterms, to renormalize the nucleon-nucleon ($NN$) scattering amplitudes at a given order. One mechanism to spoil the renormalizability of WPC is the singular---diverging at least as fast as $1/r^2$---attraction of the tensor force of one-pion exchange (OPE): $-1/r^3$ at $r \to 0$. In $S$ and $P$ waves, the triplet channels subject to this singular attraction include one uncoupled, $\cp{3}{0}$, and two coupled, $\csd$ and $\cpf$. In Ref.~\cite{Long:2011qx}, we used $\cp{3}{0}$ to investigate the modification to WPC under the guidance of RG invariance. In this paper, we report a complete study of the triplet channels, in which we continue our efforts to modify WPC at the subleading orders, in a generalization of Ref.~\cite{Long:2007vp}. Interestingly, we reach conclusions that differ in some aspects from a parallel investigation in Refs.~\cite{Valderrama:2009ei, Valderrama:2011mv}.
Except for the attempts to treat OPE as perturbation~\cite{Kaplan:1996xu, Kaplan:1998tg, Fleming:1999ee, Beane:2008bt}, it is well accepted that in $S$ and $P$ waves the leading order (LO) amplitude requires the full iteration of OPE~\cite{Birse:2005um, Nogga:2005hy}. Even though the nonperturbative unitarity requires any nonperturbative, nonrelativistic $T$-matrix to scale as $Q^{-1}$, we \emph{choose} to label the LO as $\mathcal{O}(1)$ so that one does not need to change the standard ChPT notation for power counting pion exchange diagrams. What is more consequential is that we denote different orders of the EFT expansion by its relative correction to the LO, i.e., the next-to-leading order (NLO) by $\mathcal{O}(Q/M_{\text{hi}})$ or $\mathcal{O}(Q)$ for short, and next-to-next-to-leading order (NNLO) by $\mathcal{O}(Q^{2}/M_{\text{hi}}^{2})$ or $\mathcal{O}(Q^2)$, and so on.\footnote{Here NLO and NNLO are defined differently from a more conventional notation~\cite{Epelbaum:1998ka-1999dj, Entem:2001cg-2002sf}, where NLO is $\mathcal{O}(Q^{2})$ and NNLO is $\mathcal{O}(Q^{3})$.} RG invariance, or, more specifically, invariance of the amplitude with respect to the ultraviolet (UV) momentum cutoff inherent in the Lippmann-Schwinger equation, demands as well one counterterm to be fully iterated in the partial waves where the tensor OPE is attractive~\cite{Nogga:2005hy}. In $P$ waves, these singular attractive channels include $\cp{3}{0}$ and $\cpf$. However, WPC considers the counterterms in these channels to be subleading, for they are second-order polynomials in momenta. Stated differently, the leading counterterms in $\cp{3}{0}$ and $\cpf$ are underestimated in WPC, and RG invariance requires them to be enhanced by $\mathcal{O}(M_{\text{hi}}^2/M_{\text{lo}}^2)$. Here, $M_{\text{lo}}$ refers to a cluster of infrared mass scales that include the pion decay constant $f_\pi \simeq 92$ MeV, the pion mass $m_\pi \simeq 140$ MeV, and certain combinations of them.
Although not one of the centerpieces of WPC, the indiscriminate, full iteration of different order potentials as a whole has been the standard practice in its implementations. However, the ordering of the potentials according to their matrix elements for small momenta is not always valid in the nonperturbative treatment in which the intermediate states could reach $\mhi$, where the higher-order potentials usually have larger matrix elements. In the nonperturbative setup, it seems to take intricate cancellations for higher-order potentials to eventually contribute less to the low-energy, on-shell amplitude. But it is far beyond the scope of our paper to decide whether or under what conditions these cancellations will happen. We refer readers to Refs.~\cite{Epelbaum:2009sd, Entem:2007jg} for discussion regarding renormalization and power counting in the nonperturbative treatment.
To minimize the interference between lower- and higher-order potentials in the UV region, we choose the natural way to go beyond the LO, that is, to treat the subleading interactions as perturbations. Now that the potentials from different orders are no longer on an equal footing, it is, as we will see, much easier to separate in the UV region the contributions of higher-order potentials from those of lower-order ones. If the subleading interactions are too strong to be perturbative, they are simply not subleading in a \textit{bona fide} EFT. Reference~\cite{Long:2007vp} explained the perturbative formalism with a toy model: $-1/r^{2}$ as LO and $\pm 1/r^{4}$ as $\mathcal{O}(Q^2)$ long-range potentials. More importantly, the general lesson drawn from the study of Ref.~\cite{Long:2007vp}, referred to in the paper as modified NDA (\,$\mnda$\,), is that in the case of the LO long-range potential being singular and attractive, the subleading counterterms (SCTs) are enhanced relative to NDA by the same amount as the LO counterterms; as the long-range force gets an $\mathcal{O}(Q^2)$ correction, so do the contact operators that have two more derivatives than the LO counterpart.
The validity of $\mnda$ is confirmed by renormalization of uncoupled $\cp{3}{0}$ up to $\mathcal{O}(Q^3)$~\cite{Valderrama:2011mv, Long:2011qx}, which shows that the leading long-range potential being $-1/r^2$ is not essential to the applicability of $\mnda$. Now $\csd$ and $\cpf$ pose interesting questions as to the extension of $\mnda$ to the coupled-channel problems.
Take $\csd$ as an example. The LO counterterm is a constant, $C_\cs{3}{1}$, and, according to $\mnda$, two second-derivative terms will turn up at $\mathcal{O}(Q^2)$. Therefore, $\mnda$ suggests a total of three counterterms up to $\mathcal{O}(Q^3)$ in each of $\csd$ and $\cpf$, which will be discussed in more details in Sec.~\ref{sec:cc}. However, using a coordinate space setup, Refs.~\cite{Valderrama:2009ei, Valderrama:2011mv} concluded that there must be six counterterms in $\csd$ or $\cpf$ up to $\mathcal{O}(Q^3)$. If this proliferation of counterterms in the coupled channels is true, then the predictive power of nuclear EFT is further weakened.
We carry out in the paper a momentum-space calculation to verify the power counting based on $\mnda$. In particular, we are interested to see which can be confirmed in the coupled channels: the proliferation of six counterterms or three counterterms inferred from $\mnda$.
Our principle of establishing power counting is summarized as follows:
\begin{itemize}
\item[(i)] The size of pion exchanges is decided by the non-analytic part of the corresponding Feynman diagram, which is rightly captured by WPC.
\item[(ii)] We promote counterterms over WPC only if RG invariance requires it. That is, when a counterterm is not needed for renormalization, its counting will follow NDA.
\end{itemize}
The rationale for the second point is that RG analysis in terms of the floating momentum cutoff of the $NN$ intermediate states touches upon only (nonrelativistic) nucleon momenta and it does not ``know'' anything about the contributions of heavy mesons that are integrated out in the first place~\cite{bira-private}.
A study of power counting of chiral $NN$ forces is not complete without the singlet channels. But the drastically different short-range behavior of OPE in the singlet ($1/r$) and triplet ($1/r^3$) channels signals different structures of counterterms. Therefore, we leave the singlet channels to a further study~\cite{BwLCJY-singlet}.
We also leave out $D$ and higher waves in our analysis, except for $\cd{3}{1}$ and $\cf{3}{2}$ which are, respectively, coupled to $\cs{3}{1}$ and $\cp{3}{2}$. The reasons are as follows. First, the conceptual issues concerning renormalization and power counting can be well illustrated by $S$ and $P$ waves. Second, it is debatable whether or to what extent OPE is perturbative in $D$ waves~\cite{Kaiser:1997mw, Kaiser:1998wa, Entem:2001cg-2002sf, Birse:2005um}, and answering this question is beyond the scope of our paper.
After briefly reviewing the LO and establishing our notation in Sec.~\ref{sec:lo}, we will establish in Sec.~\ref{sec:subcon} the SCTs in $\cp{3}{1}$, $\cp{3}{0}$, $\csd$ and $\cpf$ by examining the cutoff dependence of the subleading amplitudes. Finally, we offer a discussion and a conclusion in Sec.~\ref{sec:conclusion}.
\section{Leading Order\label{sec:lo}}
OPE is the leading long-range $NN$ interaction,
\begin{equation}
V_{L}^{(0)}(\vec{q}\,)=V_{1\pi}(\vec{q}\,)\equiv -\frac{g_{A}^{2}}{4f_{\pi}^{2}}\,\bm{\tau}_{1}\bm{\cdot}\bm{\tau}_{2}\,\frac{\vec{\sigma}_{1}\cdot\vec{q}\,\vec{\sigma}_{2}\cdot\vec{q}}{{\vec{q}}\,^{2}+m_{\pi}^{2}}\,,
\end{equation}
where $\vec{q}\equiv \vec{p}\,'-\vec{p}$ is the difference between the outgoing ($\vec{p}\,'$) and the incoming ($\vec{p}\,$) momenta in the center-of-mass frame, the axial vector coupling constant $g_{A}=1.29$, the pion decay constant $f_{\pi}=92.4$ MeV, and the pion mass $m_{\pi}=138$ MeV. Its coordinate space version will be useful in the discussion,
\begin{equation}
V_{1\pi}(\vec{r}\,) = \lambda_\pi \bm{\tau}_{1}\bm{\cdot}\bm{\tau}_{2}\, \left[T(r) S_{12} + Y(r) \vec{\sigma}_1\cdot\vec{\sigma}_2 \right] \, ,
\end{equation}
where
\begin{align}
\lambda_\pi &= \frac{m_\pi^3}{12\pi}\left(\frac{g_{A}^{2}}{4f_{\pi}^{2}}\right) \, , \\
T(r) &= \frac{e^{-m_\pi r}}{m_\pi r} \left[1 + \frac{3}{m_\pi r} + \frac{3}{(m_\pi r)^2} \right] \, , \\
Y(r) &= \frac{e^{-m_\pi r}}{m_\pi r} \, ,
\end{align}
and
\begin{equation}
S_{12} = 3(\vec{\sigma}_1\cdot\hat{r})(\vec{\sigma}_2\cdot\hat{r}) - \vec{\sigma}_1\cdot\vec{\sigma}_2 \, .
\end{equation}
The tensor force $T(r)$ has an inverse cubic short-range core, $1/r^3$, but it contributes to only triplet channels.
In the lower partial waves, the LO amplitude $T^{(0)}$ is obtained by the full iteration of OPE and necessary counterterms, through solving the Lippmann-Schwinger equation. For the coupled channel with total angular momentum $j$, the off-shell Lippmann-Schwinger equation reads
\begin{equation}
T^{(0)}_{l' l}(p', p; k) = V^{(0)}_{l' l}(p', p) + \frac{2}{\pi} m_N \sum_{l''} \int^\Lambda d\kappa\, \kappa^2\, V^{(0)}_{l' l''}(p', \kappa) \frac{T^{(0)}_{l'' l}(\kappa, p; k)}{k^2 - \kappa^2 + i\epsilon} \, ,\label{eqn:LSE0}
\end{equation}
with $l$, $l'$, and $l''$ running over $j-1$ and $j+1$, $k$ as the center-of-mass momentum, and $\Lambda$ as the momentum cutoff. Extension to the uncoupled channels is straightforward. The product of the Schr\"odinger propagator and the integral measure scales as $m_N Q$. Since OPE scales, more or less casually, as $(m_N \mlo)^{-1}$, with $\mlo$ a certain combination of $f_\pi^2$ and $m_N$, OPE must become nonperturbative when $Q \sim \mlo$.
Depending on the sign of the matrix element of $S_{12}$, the OPE tensor force drives $NN$ contact interactions in very different ways~\cite{Nogga:2005hy}. This is best elucidated in the uncoupled channels. When $\langle lsj|S_{12}|lsj \rangle$ is positive, where $l$ is the orbital angular momentum, $s=1$ is the spin, and $j$ is the total angular momentum, the OPE tensor force is $\sim +1/r^3$. If one picks up the regular solution to this repulsive potential, the wave function dies off exponentially near the origin. As a consequence, the sensitivity to the UV cutoff vanishes very quickly; thus, there is no need for an extra counterterm to absorb the cutoff dependence~\cite{PavonValderrama:2005uj}. This is in agreement with WPC because, according to WPC, the first counterterm in $\cp{3}{1}$---the lowest repulsive, uncoupled triplet channel---appears at $\mathcal{O}(Q^2)$.
When $\langle lsj|S_{12}|lsj \rangle$ is negative, the OPE tensor force overpowers the kinetic energy and the centrifugal barrier, causing the $NN$ system to ``collapse''~\cite{landau}. The mathematical origin of this pathology is the simultaneous existence of two equally good solutions to the Schr\"odinger equation~\cite{PavonValderrama:2005uj}, which eventually lead to ambiguity in predicting physical observables. Or, in terms of the Lippmann-Schwinger equation, the scattering amplitude is very sensitive to the UV cutoff.
The modern-day interpretation of renormalization offers a cure to this sort of pathological potential with singular attraction near the origin: supplementing short-range interactions $V_S$ rather than naively extrapolating the long-range interaction to short distances. The model-independent treatment involves arranging $V_S$ as counterterms that run with the UV cutoff $\Lambda$ in such a way that physical observables do not depend on $\Lambda$~\cite{Birse:1998dk, Beane:2000wh, Barford:2002je, Birse:2005um}. Although this is not a complete innovation in the context of chiral EFT since $NN$ contact terms are always part of the chiral Lagrangian, the consideration of RG invariance offers \textit{a priori} insights into gauging the importance of counterterms at a given order.
In $\cp{3}{0}$ where $l > 0$, the singular attraction of OPE requires at $\mathcal{O}(1)$ a counterterm, which is, however, considered $\mathcal{O}(Q^2)$ in WPC for it is a second order polynomial in momenta~\cite{Nogga:2005hy},
\begin{equation}
\langle\, \cp{3}{0} | V_S^{(0)} |\, \cp{3}{0} \rangle = C_\cp{3}{0} p' p \, ,
\label{eqn:LOuncpldCC}
\end{equation}
where $p'$ ($p$) is the magnitude of $\vec{p}\,'$ ($\vec{p}\,$). Now that $C_\cp{3}{0} p' p$ is at LO with OPE $\sim 4\pi/(m_N \mlo)$, $C_\cp{3}{0}$ must scale as
\begin{equation}
C_\cp{3}{0} \sim \frac{4\pi}{m_N} \frac{1}{\mlo^3} \, ,
\end{equation}
where $4\pi/m_N$ is introduced to cancel a common factor that usually concurs with loop integrals involving $NN$ intermediate states. For comparison, WPC considers $C_\cp{3}{0} p' p$ to be of the same order as the leading two-pion exchange, which in turn is counted as an $\mathcal{O}(Q^2/\mhi^2)$ correction to OPE; therefore, WPC prescribes~\cite{vanKolck:1999mw}
\begin{equation}
C^\text{WPC}_\cp{3}{0} \sim \frac{4\pi}{m_N} \frac{1}{\mlo\mhi^2} \, .
\end{equation}
In the coupled channels, it is convenient to write the tensor OPE in coordinate space as a $2\times 2$ matrix in the basis of two coupled orbital angular momentum states, $l = j \pm 1$~\cite{goldberger}. In $\cs{3}{1} - \cd{3}{1}$ and $\cp{3}{2} - \cf{3}{2}$,
\begin{align}
V_{T 1\pi}(\cs{3}{1} - \cd{3}{1}) &= \lambda_\pi
\begin{pmatrix}
0 & -6\sqrt{2} \\
-6\sqrt{2} & 6
\end{pmatrix}
T(r) \, , \label{eqn:VT1pi3s1} \\
V_{T 1\pi}(\cp{3}{2} - \cf{3}{2}) &= \frac{\lambda_\pi}{5}
\begin{pmatrix}
-2 & -8 \\
-8 & 6\sqrt{6}
\end{pmatrix}
T(r) \, . \label{eqn:VT1pi3p2}
\end{align}
There is a domain of $r$ near the origin in which $T(r)$ dominates over the centrifugal barrier and in which diagonalizing $V_{T 1\pi}$ also diagonalizes the Schr\"odinger equation.
OPE matrix elements in all of the coupled channels
share one property: there always is one attractive and one repulsive eigen subchannel. Therefore, one needs to summon at LO a short-range input to counter the singular attraction in one of the subchannels.
This accords with WPC in $\cs{3}{1} - \cd{3}{1}$ but calls for amendment in $\cp{3}{2} - \cf{3}{2}$~\cite{Nogga:2005hy}.
Although the counterterms are easily formulated as polynomials in momentum space, the above-mentioned diagonalization cannot be trivially realized therein. The counterterms in $\csd$ and $\cpf$ have the generic momentum space form
\begin{align}
\langle \csd | V_S | \csd \rangle &=
\begin{pmatrix}
C_\cs{3}{1} + D_\cs{3}{1}({p'}^2 + p^2) & E_\text{SD}\,p^2 \\
E_\text{SD}\, {p'}^2 & F_\cd{3}{1}\, {p'}^2p^2
\end{pmatrix} + \cdots \, , \label{eqn:3s1gnrVS} \\
\langle \cpf | V_S | \cpf \rangle &= p' p
\begin{pmatrix}
C_\cp{3}{2} + D_\cp{3}{2}({p'}^2 + p^2) & E_\text{PF}\,p^2 \\
E_\text{PF}\, {p'}^2 & F_\cf{3}{2}\, {p'}^2p^2
\end{pmatrix} + \cdots \, . \label{eqn:3p2gnrVS}
\end{align}
As shown in Ref.~\cite{Nogga:2005hy}, it is not necessary to design a counterterm that exclusively acts on the attractive subchannel of $V_{T 1\pi}$ \eqref{eqn:VT1pi3s1}; the $C$ term alone
will properly renormalize all of the $T$-matrix elements.
While the $C$ term in $\csd$ does not violate WPC, the promotion of the $C$ term in $\cpf$
asks for an enhancement of $\mathcal{O}(\mhi^2/\mlo^2)$:
\begin{align}
\langle\, \csd | V_S^{(0)} |\, \csd \rangle &=
\begin{pmatrix}
C_\cs{3}{1}^{(0)} & 0 \\
0 & 0
\end{pmatrix} \, , \\
\langle\, \cpf | V_S^{(0)} |\, \cpf \rangle &= p' p
\begin{pmatrix}
C_\cp{3}{2}^{(0)} & 0 \\
0 & 0
\end{pmatrix} \, .
\label{eqn:c3p2lo}
\end{align}
Stated differently, the renormalized $C_\cs{3}{1}$ and $C_\cp{3}{2}$ scale as
\begin{equation}
C_\cs{3}{1} \sim \frac{4\pi}{m_N} \frac{1}{\mlo}\, , \qquad C_\cp{3}{2} \sim \frac{4\pi}{m_N} \frac{1}{\mlo^3} \, ,
\end{equation}
whereas WPC differs for $C_\cp{3}{2}$,
\begin{equation}
C_\cp{3}{2}^\text{WPC} \sim \frac{4\pi}{m_N} \frac{1}{\mlo\mhi^2} \, .
\end{equation}
\section{Subleading Orders\label{sec:subcon}}
\subsection{Generalities}
It has been long known that the two-pion exchanges (TPEs) with chiral index $\nu =0$ vertices (TPE0) give an $\mathcal{O}(Q^{2})$ correction to OPE, and that the TPEs with one insertion of $\nu =1$ vertices (TPE1) lead to $\mathcal{O}(Q^{3})$ long-range potentials. TPEs are not uniquely defined because different regularization schemes---such as dimensional and spectral function regularizations~\cite{Epelbaum:2003gr2003xx}---applied to two-pion-exchange Feynman diagrams may lead to different expressions. To remove the arbitrariness in the choice of regulator, one should always pair TPE expressions with a momentum polynomial.
We refer to this polynomial as the ``primordial'' counterterm for TPEs. By definition, the primordial counterterm is power counted as the same order as the corresponding TPEs, which is exactly the content of WPC.
In addition to the general rationale given at the end of Sec.~\ref{sec:intro}, the concept of primordial counterterms serves as a complementary argument against demoting counterterms in the cases where they are not needed for renormalization in the context of the Lippmann-Schwinger equation.
TPEs, as the two-pion-exchange diagrams evaluated in the plane-wave basis, are not the only sources driving contact interactions when high momentum modes are integrated out. There are two other classes of diagrams that contribute to the evolution of the SCTs. One is insertions of TPEs into the LO $T$-matrix, which, after being properly renormalized, generates an $\mathcal{O}(Q^{2})$ or $\mathcal{O}(Q^{3})$ correction to the LO [see Eq.~\eqref{eqn:LSE23}]. The structure of divergence in these diagrams has been used as the primary tool to gauge the SCTs~\cite{Long:2007vp, Valderrama:2009ei, Valderrama:2011mv, Long:2011qx}.
The other mechanism driving the SCTs is the LO amplitude itself. With the necessary counterterms, if any, the cutoff dependence of the LO amplitude vanishes at $\Lambda \to \infty$; there normally is residual cutoff dependence at finite $\Lambda$s.
When $\Lambda$ is rescaled to a smaller value, $\Lambda'$, the integrated-out momentum modes in the interval of $(\Lambda', \Lambda)$ will, for the most part, contribute to the evolution of the leading counterterm, but a perfect RG invariance would request the SCTs to evolve as well, so that there would not even be a small cutoff dependence. References~\cite{Birse:2005um, Birse:2009my} studied this mechanism using Wilson's RG equation, but their assumption about the fixed-point solutions to the RG equation seems to be at odds with the limit-cycle-like behavior of the LO counterterms of the attractive channels~\cite{Nogga:2005hy}.
A simpler way to gauge the LO-amplitude-induced SCTs is to consider how the LO residual cutoff dependence scales against $\Lambda$. According to Ref.~\cite{PavonValderrama:2007nu}, the LO residual cutoff dependence in the attractive triplet channels is $\mathcal{O}(\Lambda ^{-5/2})$, which means our ignorance of the LO-amplitude-induced SCTs is smaller than $\mathcal{O}(Q^2/\mhi^2)$---the corrections brought by TPE0 and its primordial counterterms. As a consequence, there is no need to have a nonvanishing $\mathcal{O}(Q)$ counterterm that is not accompanied by any long-range force.
Yet another evidence of vanishing $\mathcal{O}(Q)$, though \textit{a posteriori}, is that the LO error of the $\csd$ mixing angle scales as $k^2$ instead of $k$, $|\epsilon^{\text{EFT}}_1 - \epsilon^{\text{PWA}}_1| \propto k^2$, as seen in Fig. 6 of Ref.~\cite{Beane:2001bc}.
In conclusion, the counting of the SCTs in the triplet channels will be decided by the larger one of (i) their primordial size $\mathcal{O}(Q^2)$, and (ii) what the divergence of one insertion of TPE requires.
With $\mathcal{O}(Q)$ vanishing, the on-shell $\mathcal{O}(Q^{2})$ and $\mathcal{O} (Q^{3})$ $T$-matrices, $T^{(2)}$ and $T^{(3)}$, are calculated, respectively, by one insertion of $\mathcal{O}(Q^{2})$ and $\mathcal{O}(Q^{3})$ potentials into $T^{(0)}$,
\begin{equation}
\begin{split}
T^{(2,\, 3)}(k, k) & = V^{(2,\, 3)}(k, k) + \frac{4}{\pi} m_N \int^\Lambda d\kappa\, \kappa^2\, V^{(2,\, 3)}(k, \kappa) \frac{T^{(0)}(\kappa, k)}{k^2 - \kappa^2 + i\epsilon} \\
&{} \quad + \frac{4}{\pi^2}m_N^2 \int^\Lambda\int^\Lambda d\kappa\, d\kappa'\, \kappa^2\, {\kappa'}^2 \frac{T^{(0)}(k, \kappa)}{k^2 - \kappa^2 + i\epsilon} V^{(2,\, 3)}(\kappa, \kappa') \frac{T^{(0)}(\kappa', k)}{k^2 - {\kappa'}^2 + i\epsilon} \, ,
\end{split}\label{eqn:LSE23}
\end{equation}
where $T$-matrices and $V$s are understood as $2\times 2$ matrices for the coupled channels. This is of course nothing more than the first-order distorted-wave expansion.
Treating the subleading potentials as perturbations will no doubt break the exact unitarity of the $S$-matrix, as does any perturbation-theory-based calculation. But in a consistent power counting scheme, the violation of unitarity is of higher order. This makes it slightly nontrivial to extract the phase shifts and the mixing angles from the expanded EFT $T$-matrix. Although it has been covered in the literature, we list in the Appendix~\ref{sec:conversion} the useful formulas for convenience of reference.
\subsection{Uncoupled Channels: $\cp{3}{1}$ and $\cp{3}{0}$}
In $\cp{3}{1}$, OPE (with short-distance behavior $+1/r^3$) leads to an LO wave function exponentially suppressed near the origin: $\sim \exp[-(\alpha r)^{1/2}]$, with $\alpha$ as a positive mass scale~\cite{Vald:ope}. The exponential damping of the LO wave function would eliminate the singularity of TPE0 ($\sim 1/r^5$) and TPE1 ($\sim 1/r^6$), even without any counterterm. But, as we argued, we will not demote any counterterm with respect to WPC. Therefore, the $\mathcal{O}(Q^2)$ and $\mathcal{O}(Q^3)$ $\cp{3}{1}$ counterterms are
\begin{equation}
\begin{split}
\langle\, {\cp{3}{1}}|V_{S}^{(2,\, 3)}|\,{\cp{3}{1}}\rangle = C_{{\cp{3}{1}}
}^{(0,\, 1)}\,p'p\,.
\end{split}
\label{eqn:3p1ccpresI}
\end{equation}
The splitting of $C_\cp{3}{1}$ into different orders does not mean that we will take more than one input for $C_\cp{3}{1}$; it only reflects the possibility that the ``bare'' values of the counterterms could be modified by the short-range core of TPEs.
Perturbative renormalization at subleading orders in $\cp{3}{0}$ was first studied in Ref.~\cite{Valderrama:2011mv} and in a parallel work of ours~\cite{Long:2011qx}. Although we have reached the same conclusion about the uncoupled channels as Ref.~\cite{Valderrama:2011mv}, we include here, for completeness, our analysis of $\cp{3}{0}$~\cite{Long:2011qx}.
After the LO amplitude is renormalized with $C_\cp{3}{0}$, the LO wave function in $\cp{3}{0}$ can be approximated near the origin in powers of $k^2$ up to a normalization factor~\cite{frank, Beane:2001bc, PavonValderrama:2007nu},
\begin{equation}
\psi_{k}^{(0)}(r) \sim \left(\frac{\lambda}{r}\right) ^{\frac{1}{4}}\left[u_{0} + k^2r^2 \sqrt{\frac{r}{\lambda}}u_{1} +
\mathcal{O}(k^{4}) \right] \, ,
\label{eqn:psi0}
\end{equation}
where $\lambda =\frac{3g_{A}^{2}m_{N}}{32\pi f_{\pi }^{2}}$, $u_{0}$ and $u_{1}$ are oscillatory functions in terms of $r/\lambda$ and $\phi$ with amplitudes $\sim 1$, and $\phi$ is the phase between the two independent solutions and is related to $C_\cp{3}{0}$. Combined with the short-range behavior of TPE0, $V_{2\pi}^{(0)} \sim 1/r^5$, and TPE1, $V_{2\pi}^{(1)} \sim 1/r^6$, the superficial divergence of one insertion of TPE is estimated on a dimensional ground~\cite{Valderrama:2011mv, Long:2011qx}:
\begin{align}
T^{(0)}_{2\pi,\,\cp{3}{0}} &= \langle \psi ^{(0)}|V_{2\pi}^{(0)}|\psi ^{(0)}\rangle_\cp{3}{0} \sim
\int_{\sim 1/\Lambda }drr^{2}|\psi ^{(0)}(r)|^{2}\frac{1}{r^{5}} \notag \\
& \sim \alpha _{0}(\Lambda )\Lambda ^{5/2}+\beta
_{0}(\Lambda )k^{2}+\mathcal{O}(k^{4}\Lambda ^{-5/2}) ,
\label{eqn:supT2} \\
T^{(1)}_{2\pi,\,\cp{3}{0}} &= \langle \psi ^{(0)}|V_{2\pi}^{(1)}|\psi ^{(0)}\rangle_\cp{3}{0} \sim
\int_{\sim 1/\Lambda }drr^{2}|\psi ^{(0)}(r)|^{2}\frac{1}{r^{6}} \notag \\
& \sim \alpha _{1}(\Lambda )\Lambda ^{7/2}+\beta
_{1}(\Lambda )\Lambda k^{2}+\mathcal{O}(k^{4}\Lambda ^{-3/2}) \, ,
\label{eqn:supT3}
\end{align}
where $\alpha_{0,1}(\Lambda)$ and $\beta_{0,1}(\Lambda )$ are oscillatory functions diverging slower than $\Lambda$.
The presence of two divergent terms suggests that (i) the running of the LO counterterm $C_\cp{3}{0}(\Lambda)$ needs to be corrected at higher orders, and (ii) one SCT, $D_\cp{3}{0}p' p({p'}^2 + p^2)$, needs to be enlisted.
The fact that $D_\cp{3}{0}$ arises at the same order as TPE0 leads us to arrange counterterms at $\mathcal{O}(Q^2)$ and $\mathcal{O}(Q^3)$ as follows:
\begin{equation}
\begin{split}
\langle\, {\cp{3}{0}}|V_{S}^{(2,\, 3)}|\,{\cp{3}{0}}\rangle = C_{{\cp{3}{0}}
}^{(2,\, 3)}\,p^{\prime }p+D_{{\cp{3}{0}}}^{(0,\, 1)}\,p^{\prime }p({p^{\prime}}
^{2}+p^{2})\,.
\end{split}
\label{eqn:3p0ccpresII}
\end{equation}
We see that the enhancement of $D_\cp{3}{0}$ is the same as that of $C_\cp{3}{0}$: $\mathcal{O}(\mhi^2/\mlo^2)$.
The lesson of power counting learned here coincides with the conclusion of Ref.~\cite{Long:2007vp}. Although NDA fails to prescribe a counterterm at LO, we could use $\mnda$ to determine how the SCTs scale when subleading long-range potentials are taken into account, which states that the enhancement of each SCT is the same as the LO counterterm.
\subsection{Coupled channels: $\csd$ and $\cpf$\label{sec:cc}}
In the coupled channels, the LO wave functions are dominated at short distances by the attractive subchannel. Because there are three independent on-shell $T$-matrix elements, calculating the superficial divergence of TPEs using Eqs.~\eqref{eqn:supT2} and \eqref{eqn:supT3} gives rise to six divergent terms, with two for each $T$-matrix element,
as shown in detail in Ref.~\cite{Valderrama:2011mv}. Reference~\cite{Valderrama:2011mv} proposes to use six counterterms to cancel these divergent pieces on a one-to-one basis. Although it guarantees RG invariance, lost is the regularity of power counting enjoyed by $\mnda$ in the uncoupled channels.
On the other hand, the presence of six divergent terms does not necessarily mean that one must have six counterterms to achieve RG invariance. $\mnda$ suggests that when the LO long-range potential gets $\mathcal{O}(Q^2)$ correction going from OPE to TPE0, so do the SCTs; therefore, the $\mathcal{O}(Q^2)$ SCTs in the coupled channels should be the $D$ and $E$ terms in Eqs.~\eqref{eqn:3s1gnrVS} and \eqref{eqn:3p2gnrVS}, which have two more derivatives than the LO counterterm. We propose the following power counting based on $\mnda$ for the SCTs of the coupled channels:
\begin{itemize}
\item[(i)] In $\csd$, we do not change WPC,
\begin{equation}
\begin{split}
\langle\, \csd | V_S^{(2,\, 3)} |\, \csd \rangle &=
\begin{pmatrix}
C_\cs{3}{1}^{(2,\, 3)} + D^{(0,\, 1)}_\cs{3}{1}({p'}^2 + p^2) & E^{(0,\, 1)}_\text{SD}\, p^2 \\
E^{(0,\, 1)}_\text{SD}\, {p'}^2 & 0
\end{pmatrix} \, . \\
\label{eqn:sub3s1}
\end{split}
\end{equation}
\item[(ii)] In $\cpf$, an enhancement of $\mathcal{O}(\mhi^2/\mlo^2)$ leads to
\begin{equation}
\begin{split}
\langle\, \cpf | V_S^{(2,\, 3)} |\, \cpf \rangle &= p' p
\begin{pmatrix}
C_\cp{3}{2}^{(2,\, 3)} + D^{(0,\, 1)}_\cp{3}{2}({p'}^2 + p^2) & E^{(0,\, 1)}_\text{PF}\, p^2 \\
E^{(0,\, 1)}_\text{PF}\, {p'}^2 & 0
\end{pmatrix} \, .\\
\label{eqn:sub3p2}
\end{split}
\end{equation}
\end{itemize}
An analytical proof of renormalizability with the above counterterms is difficult because the closed form of the LO $T$-matrix is not available, so we will resort to numerical experiments in Sec.~\ref{sec:numerics} to test RG invariance, or the lack thereof.
At a given energy, there are three scattering parameters---two phase shifts and one mixing angle---to be extracted from the $2\times2$ $T$-matrix \eqref{eqn:stapp}. With $T^{(2)}$ determined by three inputs from partial-wave analysis (PWA) and Eq.~\eqref{eqn:S2abc}, Eq.~\eqref{eqn:LSE23} provides a group of linear equations to solve for $C^{(2)}$, $D^{(0)}$, and $E^{(0)}$. Though not obvious, it is straightforward to show that the three linear equations built from the three PWA inputs at the same energy are linearly dependent. Therefore, the needed three PWA inputs must be incorporated from at least two different energies. Not surprisingly, the same also applies to $T^{(3)}$.
\subsection{Numerics\label{sec:numerics}}
We will use TPEs without the explicit delta-isobar to demonstrate renormalization. There are a few versions of TPEs~\cite{Ordonez:1993-1995, Kaiser:1997mw, Kaiser:1998wa, Epelbaum:1998ka-1999dj, Rentmeester:1999vw, Entem:2001cg-2002sf} in the literature with slight differences in how double counting is avoided~\cite{Friar:1999sj}. For definitiveness, we use the version in Ref.~\cite{Epelbaum:1998ka-1999dj}, i.e., delta-less TPE expressions with dimensional regularization. We adopt the following low-energy constants for the $\nu =1$ $\pi \pi NN$ seagull couplings (GeV$^{-1}$): $c_{1}=-0.81$, $c_{3}=-4.7$, and $c_{4}=3.4$~\cite{Buettiker:1999ap}. With the delta integrated out, we expect the EFT expansion to break down around $Q \sim \delta$, where $\delta \simeq 300$ MeV is the delta-nucleon mass splitting. Such a breakdown scale is noticed in, e.g., Ref.~\cite{Birse:2007sx} through ``deconstruction.''
Figure~\ref{fig:3p1} shows the $\cp{3}{1}$ phase shifts as a function of laboratory kinetic energy $T_\text{lab}$ for two different cutoffs and a function of the cutoff at $T_\text{lab} = 100$ MeV. The sharp momentum cutoff is chosen as the regulator throughout the paper: $\theta(\Lambda - \kappa)$, where $\kappa$ is the magnitude of the loop momentum, as defined in Eq.~\eqref{eqn:LSE0}. The values of $C_\cp{3}{1}^{(0,\, 1)}$ are solved for by a fit of the phase shift to the Nijmegen PWA~\cite{Stoks:1993tb} for $T_\text{lab} = 50$ MeV. The cutoff independence is not surprising since the LO wave function is exponentially suppressed at short distances. The large shift from $\mathcal{O}(Q^2)$ to $\mathcal{O}(Q^3)$ indicates the influence of the uncertainties of $\pi \pi NN$ coupling constants $c_i$. Nevertheless, a decent agreement with the PWA up to $T_\text{lab} = 100$ MeV is obtained.
\begin{figure}
\includegraphics[scale=0.25, clip=true]{3p1-1500_mod.eps}
\includegraphics[scale=0.25, clip=true]{fxk-2-3p1_mod.eps}
\caption{(Color online) With the SCTs \eqref{eqn:3p1ccpresI}, the $\cp{3}{1}$ phase shifts (a) as a function of $T_\text{lab}$ up to $\mathcal{O}(Q^3)$, and (b) as a function of $\Lambda$ for $T_\text{lab} = 100$ MeV. In the legend of (a), the number following the symbols is the cutoff value in GeV. The red dots are from the Nijmegen PWA.}
\label{fig:3p1}
\end{figure}
$\cp{3}{0}$ has been thoroughly studied in our momentum-space framework in Ref.~\cite{Long:2011qx}, which confirmed the RG invariance of $\mnda$ for the uncoupled channels; see \eqref{eqn:3p0ccpresII}. We refer the reader to Ref.~\cite{Long:2011qx} for the numerical results.
Now we move on to the coupled channels. Figure~\ref{fig:tlab3s1} shows the phase shifts of $\csd$ and the mixing angle $\epsilon_1$ as functions of $T_\text{lab}$ at $\mathcal{O}(Q^2)$ and $\mathcal{O}(Q^3)$. The values of the counterterms are determined such that the EFT curves reproduce the Nijmegen PWA for $\delta_\cs{3}{1}$ at $T_\text{lab} = 30$ and $50$ MeV and $\epsilon_1$ at $50$ MeV. Consequently, the $\cd{3}{1}$ EFT phase shifts are predictions. The first indication of the cutoff independence is the closeness of two EFT curves with $\Lambda = 1.5$ and $2.5$ GeV. The plot of $\epsilon_1$ shows a larger cutoff dependence toward higher energies, but the fact that the $\Lambda = 2.0$ GeV curve is closer to $\Lambda = 2.5$ GeV than $\Lambda = 1.5$ GeV suggests that the cutoff independence is finally achieved at larger $\Lambda$s.
\begin{figure}
\includegraphics[scale = 0.45, clip = true]{tlab_3s1_pres3_mod.eps}
\includegraphics[scale = 0.45, clip = true]{tlab_3d1_pres3_mod.eps}
\includegraphics[scale = 0.45, clip = true]{tlab_e1_pres3_mod.eps}
\caption{(Color online) With the SCTs \eqref{eqn:sub3s1}, the $\cs{3}{1}$, $\cd{3}{1}$ phase shifts and the mixing angle $\epsilon_1$ as functions of $T_\text{lab}$ at $\mathcal{O}(Q^2)$ and $\mathcal{O}(Q^3)$. In the legend, the number in front of the line symbols is the cutoff value in GeV.}
\label{fig:tlab3s1}
\end{figure}
The cutoff independence is more clearly demonstrated in Fig.~\ref{fig:fxk-3s1-pres3_012_all}, which shows the phase shifts and the mixing angle at $T_\text{lab} = 40$ and $100$ MeV as functions of $\Lambda$. The residual cutoff dependence is still visible at lower $\Lambda$s, but it is much smaller than the size of the corresponding EFT correction. That is, the corrections are meaningful even at the lower cutoffs because they are not washed out by the cutoff uncertainties.
\begin{figure}[tbp]
\centering
\begin{tabular}{rr}
\includegraphics[scale=0.53, clip=true]{fxk_3s1_137_mod.eps} & \hspace{5mm}
\includegraphics[scale=0.53, clip=true]{fxk_3s1_216_mod.eps}\\
\includegraphics[scale=0.53, clip=true]{fxk_3d1_137_mod.eps} & \hspace{5mm}
\includegraphics[scale=0.53, clip=true]{fxk_3d1_216_mod.eps}\\
\includegraphics[scale=0.53, clip=true]{fxk_e1_137_mod.eps} & \hspace{5mm}
\includegraphics[scale=0.53, clip=true]{fxk_e1_216_mod.eps}
\end{tabular}
\caption{(Color online) With the SCTs \eqref{eqn:sub3s1}, the $\cs{3}{1}$, $\cd{3}{1}$ phase shifts and the mixing angle $\epsilon_1$ at $T_\text{lab} = 40$ and $100$ MeV, as functions of the momentum cutoff. The dashed, dot-dashed and solid lines are $\mathcal{O}(1)$, $\mathcal{O}(Q^2)$ and $\mathcal{O}(Q^3)$, respectively.}
\label{fig:fxk-3s1-pres3_012_all}
\end{figure}
Figures~\ref{fig:tlab3p2} and \ref{fig:fxk-3p2-pres3_012_all} show the $\cpf$ phase shifts and the mixing angle $\epsilon_2$ at $\mathcal{O}(Q^2)$ and $\mathcal{O}(Q^3)$ as functions of $T_\text{lab}$ and the cutoff, respectively, with the SCTs \eqref{eqn:sub3p2}. Similar to the case of $\csd$, we fit $\delta_\cp{3}{2}$ at $T_\text{lab} = 30$ and $50$ MeV and $\epsilon_2$ at 50 MeV to the PWA values.
\begin{figure}[tbp]
\includegraphics[scale = 0.45, clip = true]{tlab_3p2_pres3_mod.eps}
\includegraphics[scale = 0.45, clip = true]{tlab_3f2_pres3_mod.eps}
\includegraphics[scale = 0.45, clip = true]{tlab_e2_pres3_mod.eps}
\caption{(Color online) With the SCTs \eqref{eqn:sub3p2}, the $\cp{3}{2}$, $\cf{3}{2}$ phase shifts and the mixing angle $\epsilon_2$ as functions of $T_\text{lab}$ at $\mathcal{O}(Q^2)$ and $\mathcal{O}(Q^3)$. The symbols are explained in the legend of Fig.~\ref{fig:tlab3s1}.}
\label{fig:tlab3p2}
\end{figure}
\begin{figure}[tbp]
\centering
\begin{tabular}{rr}
\includegraphics[scale=0.53, clip=true]{fxk_3p2_137_mod.eps} & \hspace{5mm}
\includegraphics[scale=0.53, clip=true]{fxk_3p2_216_mod.eps}\\
\includegraphics[scale=0.53, clip=true]{fxk_3f2_137_mod.eps} & \hspace{5mm}
\includegraphics[scale=0.53, clip=true]{fxk_3f2_216_mod.eps}\\
\includegraphics[scale=0.53, clip=true]{fxk_e2_137_mod.eps} & \hspace{5mm}
\includegraphics[scale=0.53, clip=true]{fxk_e2_216_mod.eps}
\end{tabular}
\caption{(Color online) With SCTs \eqref{eqn:sub3p2}, the $\cp{3}{2}$, $\cf{3}{2}$ phase shifts and the mixing angle $\epsilon_2$ at $T_\text{lab} = 40$ and $100$ MeV, as functions of the momentum cutoff. The symbols are explained in the caption of Fig.~\ref{fig:fxk-3s1-pres3_012_all}.}
\label{fig:fxk-3p2-pres3_012_all}
\end{figure}
The main goal of this paper is to verify the RG invariance of our power counting in the UV region, so the cutoff window was chosen such that the EFT curves start to show the cutoff independence. For smaller cutoffs not shown in the plots ($1.2 \gtrsim \Lambda \gtrsim 0.6\,\text{GeV}$), the general trend is similar to $\cp{3}{0}$ (Fig.~2 of Ref.~\cite{Long:2011qx}): the $\mathcal{O}(Q^2)$ EFT curve is the first to become cutoff independent, while the LO is the latest.
A good fit to the PWA up to $100$ MeV is presented at $\mathcal{O}(Q^3)$ in $\csd$. Without special effort to improve the fits, the EFT result agrees less well with the PWA in $\cpf$. We think that this is largely owing to a disappointing LO, which departs quickly from the PWA as the energy increases. The unusually small $\cp{3}{2}$ scattering volume, $\alpha_\cp{3}{2} \simeq -0.28$ fm$^3$, compared with $\alpha$ of other $P$ waves, $|\alpha| \simeq 1.5 - 2.8$ fm$^3$~\cite{PavonValderrama:2005ku}, is suggestive of a certain amount of fine tuning, which calls for a more sophisticated fitting strategy.
To assess the feasibility of improving the fitting quality, we refit the counterterms to the PWA inputs at higher energies: $\delta_\cp{3}{2}$ at $50$ and $100$ MeV, and $\epsilon_2$ at 50 MeV. The updated $\cpf$ EFT phases are plotted in Fig.~\ref{fig:tlab3p2_better}. Although the EFT convergence still breaks down at lower energies than in $\csd$, a good fit to the PWA until $130$ MeV is achieved at $\mathcal{O}(Q^2)$ and $\mathcal{O}(Q^3)$, and it is comparable to the WPC-based calculation with the same TPEs, which is shown in Ref.~\cite{Epelbaum:1998ka-1999dj}. We notice another fitting strategy used in Ref.~\cite{Valderrama:2011mv}, which sacrifices the LO near threshold in order to facilitate better agreements with the PWA at higher orders. This amounts to tuning the LO counterterm to further reduce the attraction of OPE.
Overall, the breakdown scale implied in the numerical results is consistent with our expectation for the delta-less theory, $k \sim \delta \sim 300$ MeV, with the exceptions of $\cp{3}{1}$ and $\cp{3}{2}-^3F_2$. Even for these two channels, one cannot help wondering whether the delta can bring some attraction from $\mathcal{O}(Q^3)$ to $\mathcal{O}(Q^2)$ and improve the convergence of EFT expansion~\cite{daniel-private}.
\begin{figure}[tbp]
\includegraphics[scale = 0.45, clip = true]{tlab_3p2_pres3_better_mod.eps}
\includegraphics[scale = 0.45, clip = true]{tlab_3f2_pres3_better_mod.eps}
\includegraphics[scale = 0.45, clip = true]{tlab_e2_pres3_better_mod.eps}
\caption{(Color online) With the SCTs \eqref{eqn:sub3p2}, the $\cp{3}{2}$, $\cf{3}{2}$ phase shifts and the mixing angle $\epsilon_2$ as functions of $T_\text{lab}$ with $\Lambda = 1.5$ GeV. The values of the counterterms are adjusted to reproduce the PWA values for $\delta_{\cp{3}{2}}$ at $T_\text{lab} = 50$, and $100$ MeV and $\epsilon_2$ at $50$ MeV. The dots are from the Nijmegen PWA. The dashed, dot-dashed and solid lines are $\mathcal{O}(1)$, $\mathcal{O}(Q^2)$, and $\mathcal{O}(Q^3)$, respectively.}
\label{fig:tlab3p2_better}
\end{figure}
\section{Discussion and Conclusion\label{sec:conclusion}}
We have studied up to $\mathcal{O}(Q^3)$ the structure of counterterms of chiral $NN$ contact interactions in the triplet channels, with $S$ and $P$ waves as the examples. The essential guideline we have followed is to promote counterterms over WPC when RG invariance requires it.
We found that the scaling of SCTs are mainly driven by the interplay between TPEs and the LO wave functions. A direct consequence is that $\mathcal{O}(Q)$ by contact interactions alone vanishes. The resulting arrangement of the counterterms in the studied channels are given by Eqs.~\eqref{eqn:3p1ccpresI}, \eqref{eqn:3p0ccpresII}, \eqref{eqn:sub3s1}, and \eqref{eqn:sub3p2}, which can be very nicely summarized by $\mnda$: the SCTs are enhanced by the same amount as the LO counterterm so that the whole tower of counterterms with the same quantum number is shifted uniformly. While this means that WPC remains intact in $\csd$ and $\cp{3}{1}$, it requires an enhancement of $\mathcal{O}(\mhi^2/\mlo^2)$ to all counterterms in $\cp{3}{0}$ and $\cpf$.
It is interesting to compare the chiral $NN$ forces with the three-body system that has only contact interactions. When the two-body $S$-wave scattering length $a_2 \to \infty$, the three-body ``pionless'' theory can be mapped onto a dual two-body theory with $-1/r^2$ long-range force and contact interactions that represent three-body operators in the original system~\cite{efimov}. Reference~\cite{Barford:2004fz} proposed a power counting similar to $\mnda$ for the three-body contact interactions. However, the resemblance between chiral EFT forces and the system investigated in Ref.~\cite{Barford:2004fz} is not perfect because the long-range interactions beyond the leading $-1/r^2$ in the dual two-body system, if any, are resummed nonperturbatively instead of being treated as perturbations. Therefore, it is not clear to us whether the scaling of SCTs obtained in Ref.~\cite{Barford:2004fz} is driven by the LO or subleading long-range interactions.
Our finding that three counterterms are needed in the coupled channels up to $\mathcal{O}(Q^3)$ differs from that of Refs.~\cite{Valderrama:2009ei, Valderrama:2011mv}, which also adopted the perturbative approach on top of the nonperturbative LO, but
concluded instead that six counterterms are necessary for renormalization purpose.
(Although it is speculated in Ref.~\cite{Valderrama:2011mv} that it might be possible to reduce the number of short-range parameters, an alternative is not offered unless the higher-wave component is treated in perturbation theory.) However, that there are only two second-derivative terms---$D_\cs{3}{1}$ and $E_\text{SD}$ in \eqref{eqn:3s1gnrVS}---suggests that one could correlate these six divergent pieces in a model-independent way, which is justified by the numerical evidence of RG invariance shown in Figs.~\ref{fig:fxk-3s1-pres3_012_all} and \ref{fig:fxk-3p2-pres3_012_all}.
Without a dedicated effort to fine tune the fits, our results show a good agreement with the Nijmegen PWA up to $T_\text{lab} \sim 100$ MeV. Regardless of the comparison with the PWA, the relatively large deviation from $\mathcal{O}(Q^2)$ to $\mathcal{O}(Q^3)$ in $\cp{3}{1}$ and $\cp{3}{2}$ encourages one to
hope that a delta-ful EFT can improve the convergence by including the delta-isobar as explicit degrees of freedom. Aside from the debatable issues of power counting counterterms, the delta-ful nuclear forces have been shown to achieve a more rapid convergence in the two-nucleon~\cite{Valderrama:2008kj_2010fb, Entem:2007jg} and, on a more qualitative level, the three-nucleon~\cite{Pandharipande:2005sx} sectors.
Although there have been many efforts to derive the delta-ful TPEs~\cite{Ordonez:1993-1995, Kaiser:1998wa, Krebs:2007rh}, it appears desirable to update the extraction of low-energy constants in the delta-ful chiral Lagrangian from $\pi N$ scattering through the chiral EFT description around the delta peak~\cite{Long:2009wq, Pascalutsa:2002pi} where the effects of the delta are most prominent. In addition, the formulation of the delta-ful chiral Lagrangian may need to be reexamined in light of the discussion in Ref.~\cite{Long:2010kt}.
\acknowledgments We thank Bira van Kolck and Daniel Phillips for their encouragement in the early stage of this work and thoughtful comments on the manuscript. BwL thanks Jos\'e Goity for useful discussion and for the term ``primordial counterterm.'' CJY thanks Bruce Barrett for valuable support. We are grateful for hospitality to the National Institute for Nuclear Theory (INT) at the University of Washington and the organizers of the INT program ``Simulations and Symmetries: Cold Atoms, QCD, and Few-hadron Systems,'' in which the work was stimulated.
This work is supported by the US Department of Energy under Contracts No.DE-AC05-06OR2317 7 (BwL), No. DE-FG02-04ER41338 (CJY), and the NSF under Grant No. PHYS-0854912 (CJY), and is coauthored by Jefferson Science Associates, LLC under U.S. Department of Energy Contract No. DE-AC05-06OR23177.
|
\section{Introduction} Given a reductive algebraic group $G$, a reductive
subgroup $H$ and some irreducible $G$-module $V$, then $V$ is also a $H$-module
in a natural way. An obvious problem is to find branching rules that describe
the decomposition of the $H$-module $V$ into irreducible components.
We will deal with this problem in
the situation where $G$ is a complex simply connected simple algebraic group of
exceptional type. The subgroup structure of these groups has been studied in
great detail and we want to consider maximal reductive subgroups of $G$. The
maximal closed connected subgroups are listed in Theorem 1 of \cite{Sei91}.
These groups are either semisimple or parabolic. So the maximal reductive
subgroups are easily obtained by adding the Levi factors of the maximal
parabolic groups which are maximal reductive in $G$ to the list of maximal
semisimple subgroups. The modules $V$ that we consider are those having as
highest weights a multiple of a fundamental weight.
We will approach this problem by working with spherical varieties. We consider
the flag variety $G/P$ where $P$ is a maximal parabolic subgroup of~$G$.
Of special interest to us are the
flag varieties of that form, that are $H$-spherical, i.e.\ they contain an open
orbit for a Borel subgroup of $H$. The property of being
spherical can also be described in a representation-theoretic way. Namely a
normal affine $G$-variety is spherical if and only if its coordinate ring is a
multiplicity-free $G$-module \cite{Vin78}. Let $\widehat Y$ denote the
affine cone over $G/P$. Then the flag variety is
$H$-spherical if and only if all restrictions of the homogeneous
components of the coordinate ring of $\widehat Y$ to $H$ are multiplicity-free.
These homogeneous comonents are exactly the irreducible submodules of the
coordinate ring $\mathbb{C}[\widehat Y]$ and they are of shape $V_{k\omega_i}^*$. In the
case of sphericity we can derive branching rules for these modules.
So the content of this paper is twofold. We classify the
spherical $H$-varieties $G/P$ and furthermore we derive branching rules for the
simple $G$-submodules of the coordinate ring of the affine
cones in the spherical cases. The results are summarized in Table~1. A flag
variety $G/P$ is $H$-spherical if and only if the branching rules for the
corresponding modules $V$ are given in the table.
\section{Notation}
We work over the field of
complex numbers throughout the article. $G$ always denotes a simply connected
simple algebraic group of exceptional type. Within $G$ we choose a Borel
subgroup $B$, a maximal torus $T$ and thereby define a set $\{\alpha_1,\ldots,\alpha_r\}$ of
simple roots which are labeled according to Bourbaki-notation. The system of
roots of $G$ is denoted by $\Phi$, the system of positive roots of $G$ is
denoted by $\Phi^+$ and $(a_1,\ldots,a_r)$ stands for the root
$\sum_{i=1}^r a_i \alpha_i$. Further $X_\alpha$ denotes a non-trivial
element of the root space associated to $\alpha$. Let $\Lambda^+$ be the set of
dominant weights related to $B$ and $T$. The irreducible $G$-module of highest
weight $\lambda \in \Lambda^+$ is denoted by $V_\lambda$. The fundamental weights of $G$ are
$\omega_1,\ldots,\omega_r$ and $\omega_1^*,\ldots,\omega_r^*$ are the
fundamental weights such that $(V_{\omega_i})^*= V_{\omega_i^*}$, where
$(V_{\omega_i})^*$ is the dual of $V_{\omega_i}$. If we write $k \omega_i$, then
$k \in \mathbb{N}$.
Let $H$ denote a reductive subgroup of $G$ with root system $\Phi_H$
and analogous to $G$ we use the notation $(b_1,\ldots,b_s)_H:=
\sum_{i=1}^s b_i \beta_i$ where $\{\beta_1,\ldots,\beta_s\}$ is a set of
simple roots of $\Phi_H$ given by the Borel subgroup $B_H=B\cap H$. The
fundamental weights of $H$ are denoted by $\lambda_1,\ldots,\lambda_s$, if $H$
is semisimple. When $H$ is a Levi subgroup, $\lambda_1,\ldots,\lambda_s$ denote
the fundamental weights of the semisimple part of $H$.
Lastly $\mathfrak b$ denotes the Lie algebra of $B_L$, $\mathfrak u$ the
Lie algebra of $U_L$ the unipotent radical of $B_L$ and $\mathfrak h$ the Lie
algebra of the maximal torus $T$ of $B_L$.
\section{Main results and outline of proof}
We will now summarize the results and give an outline of the proof.
In this paper we will derive the branching rules stated in the following table.
Further we show that if $\textnormal{res}^G_H(V_{k\omega_i})$ is given in the table, then
$G/P_{\omega_i^*}$ is a spherical $H$-variety.
Conversely, if a maximal reductive subgroup $H\subset G$ does not appear
in the table, then the varieties $G/P_{\omega_i}$ are not $H$-spherical.
Note that for the subgroups $D_5\times \mathbb{C}^* \subset E_6$ and $E_6 \times \mathbb{C}^*
\subset E_7$ the weight of the $\mathbb{C}^*$-action depends on the embedding of $\mathbb{C}^*$.
The embedding that we chose is given in the corresponding sections.
\needspace{4cm}
\input{table_arxiv.tex}
To obtain the previous table we shall adapt the proof of Proposition 4.4 in
\cite{FL10} by Feigin and Littelmann. But first we will introduce some
additional notation.
Let $P_i \supset B$ denote the maximal parabolic subgroup of $G$ associated
to the fundamental weight $\omega_i$. We shall consider the natural action of
$H$ on the projective varieties $Y=G/P_i$. The affine cone over $Y$ is denoted by
$\widehat Y$ and the stabilizer of $\overline 1 \in G/P_i$ is denoted by
$H_{\overline 1}$. The group $H_{\overline 1}$ is a parabolic subgroup of $H$.
Its opposite parabolic subgroup in $H$ is denoted by $Q$. Furthermore let $Q^u$
be its unipotent radical and let $L$ be the Levi-subgroup $H_{\overline 1} \cap
Q$ with Borel subgroup $B_L$ defined by the simple roots of $H$ that appear in
$L$. If we consider the orbit $O= H. \overline 1 \simeq H/H_{\overline 1}$ with
normal bundle $\mathcal N$ having fiber $N$ at~$\overline 1$ then $N$ has the
structure of an $L$-module since $L \subset H_{\overline 1}$.
If no confusion can arise we will write $P$ instead of $P_i$ from now on.
The proof is divided into two parts. First we
will determine in which cases $Y$ is a spherical $H$-variety. This part of
the proof is conducted in four steps.
{\itshape Step 1:} We apply the Brion-Luna-Vust Local
Structure Theorem \cite{BLV86} to get the following proposition.
\begin{prp}
There exists a locally closed affine subvariety $Z \subset Y$ such that
$\overline 1 \in Z$, $Z$ is stable under the action of $L$, $Q^u.Z$ is open in
$Y$ and the canonical map $Q^u\times Z \rightarrow Q^u.Z$ is an isomorphism of
varieties.
\end{prp}
\begin{proof}
Note that since the Borel subgroup $B_H$ is a subgroup of $P$, it is
contained in the stabilizer $H_{\overline 1}$ of $\overline 1 \in Y$. Thus
$H_{\overline 1}$ is a parabolic subgroup of~$H$.
Now we can apply the Local Structure Theorem to this situation and obtain the
proposition.
\end{proof}
{\itshape Step 2:} We have the following proposition.
\begin{prp}
The variety $Y$ is $H$-spherical if and only if $Z$ is a
spherical $L$-variety.
\end{prp}
\begin{proof}
Assume $Z$ is spherical, i.e.\ a Borel subgroup of $L$
has a dense orbit in $Z$. Let $B_L$ be the Borel subgroup $B_H \cap L
\subset L$ and let $B_{L}^-$ be the opposite Borel subgroup. Then
$B^-_H=Q^u B^-_{L}$ is a Borel subgroup of $H$. Let $z \in Z$ be an element
such that $B^-_{L}. z$ is dense in $Z$. Since $Q^u.Z$ is dense in $Y$, so is
$B^-_H.z = Q^u(B^-_L.z)$. Hence $Y$ is a spherical $H$-variety.
If on the other
hand $Y$ is $H$-spherical, then $B_H^-.y=Q^u(B^-_L).y$ is open in $Y$ for some
$y \in Y$. Since $Q^u.Z$ is open in $Y$ we can assume that $y \in Z$. Now if
$Q^u (B_L^-.y)$ is dense in $Y$ it follows that $B_L^-.y$ is dense in $Z$.
\end{proof}
{\itshape Step 3:} Now $N$ is isomorphic to the tangent space $T_{\overline 1}
Z$ and thanks to Luna's Slice Theorem $Y$ is $H$-spherical if and only if $N$ is
$L$-spherical.
{\itshape Step 4:}
It remains to compute $N$ and to check in which cases it
is a spherical $L$-module. Note that we have
\begin{equation*}
N \simeq
(\text{Lie}\,G/\text{Lie}\,P_i) /
(\text{Lie}H/\text{Lie}\,H_{\overline 1}).
\end{equation*}
So if $\Phi_H \subset \Phi$, then we can describe $N$ as the root spaces that occur in
$T_{\overline 1}Y=\text{Lie}G/\text{Lie}P_i$ but not in
$T_{\overline 1}(H/H_{\overline 1})$.
These are
all the root spaces $\mathbb{C} X_\alpha$ such that $\alpha$ is negative and $\mathbb{C}
X_\alpha \not \subset \text{Lie}P_i$ as well as $\mathbb{C} X_\alpha \not \subset
\text{Lie} H$.
\begin{rem}
There is an algorithm by F.\ Knop \citep[Thm.\ 3.3]{Kno97} to check whether a
given $L$-module is spherical. But in order for this paper to be self-contained
we compute an explicit $X \in N$ such that $B_L.X$ is a dense orbit in $N$ in
the spherical cases.
\end{rem}
The second part is to compute the restrictions of the $G$-modules
$V_{k\omega_i^*}$ to~$H$. It is well-known that
\begin{equation*}
\mathbb{C}[\widehat Y]= \bigoplus_{k \geq 0} V_{k\omega^\ast_i}
\end{equation*}
where $V_{k\omega^\ast_i}$ corresponds to the homogeneous functions of degree
$k$ on $\widehat Y$. In order to derive branching rules for $V_{k \omega_i^*}$
we need to determine the $U_H$-invariants of $V_{k\omega^\ast_i}$.
Because $\widehat Y$ is a spherical $(H\times \mathbb{C}^*)$-variety and because
$U_H=U_{H\times \mathbb{C}^*}$, we know from Lemma 1 in \cite{Lit94} that the ring
$\mathbb{C}[\widehat Y]^{U_H}$ is a polynomial ring with some set of generators
$f_j$ of degree $d_j$, $1\leq j \leq s$, where $s$ is the number of generators.
Thus we have the following branching rules in this situation.
\begin{thm}
Let $\eta_j$ denote the weight of $f_j$ with respect to $H$ and suppose $G/P_i$
is a spherical $H$-variety. Then we get
\begin{equation*}
\text{res}^{G}_{H}(V_{k\omega^\ast_i}) = \bigoplus_{a_1 d_1 + \ldots +a_s d_s =
k} V_{a_1\eta_1+\ldots+a_s \eta_s}.
\end{equation*}
\end{thm}
We need to compute the number of generators, i.e.\ the dimension of
$\mathbb{C}[\widehat Y]^{U_H}$.
\begin{prp}\label{prp:codim}
We have
\begin{equation*}
\dim \mathbb{C} [\widehat Y]^{U_H}= \dim N -\dim (\textnormal{generic } U_L
\textnormal{-orbit}) +1.
\end{equation*}
\end{prp}
\begin{proof}
We know that $\dim \mathbb{C} [\widehat Y]^{U_H}=
\text{trdeg}\;\mathbb{C}(\widehat Y)^{U_H}$ and by a theorem of Rosenthal we know that
$\textnormal{trdeg}\,\, \mathbb{C}(\widehat Y)^{U_H}=\dim\widehat Y - \dim(\textnormal{generic
}U_H\textnormal{-orbit})$ (paragraph II.4.3.E in \citep[p.\ 143]{Kra84}).
So the proposition is an immediate corollary of the following lemma.
\end{proof}
\begin{lem}
Let $Y$, $N$, $U_L$ and $U_H$ be defined as above. Let $O_1$ be
a generic $U_H$-orbit in $Y$ and $O_2$ be a generic $U_L$-orbit in $N$.
Then
\begin{equation*}
\dim Y - \dim O_1 = \dim N - \dim O_2.
\end{equation*}
\end{lem}
\begin{proof}
Let $O\subset Y$ be the open subset of $X$ such that $\dim U_H.x$ is maximal for
all $x \in O$ (i.e.\ $U_H.x$ is an generic orbit). We have $O \cap Q^u.Z \ne
\emptyset$, because $Q^u . Z$ is open and dense in~$Y$.
Let $x=qz$ be an element in $O \cap Q^u.Z$. We know that
$U_H=U_L.Q^u=Q^u.U_L$.
So we have
$U_H.x=U_H.(qz)= U_L Q^u (qz)= U_L Q^u .z = U_H.z$ and we can assume that $U_H
.x$ is a generic $U_H$-orbit in $Y$ with $x \in Z$.
Suppose $y$ is an element of the stabilizer $(U_H)_x$ of $x$.
Then we have $y=q.u$ for some $q\in Q^u$, $u\in U_L$.
So it follows from the Local Structure Theorem that $q= \text{id}$ and $ux=x$.
Thus we get $(U_H)_x=(U_L)_x$.
With $\dim Y = \dim Z + \dim Q^u$ (Local
Structure Theorem) we get
\begin{equation*}
\begin{split}
\dim Y - \dim U_H.x
&= \dim Q^u + \dim Z - \dim U_H.z\\
&= \dim Z -(\dim U_H.x - \dim Q^u)\\
&= \dim Z -(\dim U_H- \dim (U_H)_x - \dim Q^u)\\
&= \dim Z -(\dim U_H- \dim Q^u - \dim (U_L)_x)\\
&= \dim Z -(\dim U_L - \dim (U_L)_x)\\
&= \dim Z -\dim U_L.x.
\end{split}
\end{equation*}
\end{proof}
\section{The maximal reductive subgroups of the exceptional groups}
We want to list all maximal reductive subgroups of the exceptional
algebraic groups. G.\ Seitz listed all maximal closed connected subgroups in
arbitrary characteristics. We recall his results for the case that the ground
field is $\mathbb{C}$ (\cite{Sei91}, Thm.\ 1).
\begin{thm}
Let $G$ be a simple algebraic group of exceptional type and let $X$ be maximal
among the proper closed connected subgroups of $G$. Then either $X$ contains a
maximal torus of $G$ or $X$ is semisimple and the pair $(G,X)$ is given below.
Moreover, maximal subgroups of each type exist and are unique up to conjugacy in
$\text{Aut}(G)$.
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
$G$ & $X$ simple & $X$ not simple\\\hline\hline
$G_2$ & $A_1$ &\\\hline
$F_4$ & $A_1$ & $A_1\times G_2$\\ \hline
$E_6$ & $A_2$, $G_2$, $F_4$, $C_4$ & $A_2 \times G_2$\\ \hline
$E_7$ & $A_1$, $A_2$ & $A_1 \times A_1$, $A_1 \times G_2$, $A_1 \times F_4$,
$G_2 \times C_3$\\ \hline
$E_8$ & $A_1$, $B_2$ & $A_1 \times A_2$, $G_2 \times F_4$\\ \hline
\end{tabular}
\end{center}
\end{thm}
Since the maximal subgroups that do not contain a maximal torus are semisimple
they are also maximal reductive subgroups of $G$.
It remains to identify the maximal reductive subgroups that are contained in a
maximal subgroup of maximal rank. These groups fall in two categories. Some are
the maximal parabolic subgroups of $G$ and the others are so called subsystem
subgroups. There is an algorithm (cf.\ paragraph no.\ 17 of \cite{Dyn57.1} or \cite{BS49})
that determines these subgroups: Start with the Dynkin diagram of
$G$ and adjoin the smallest root $\delta$ to obtain the extended Dynkin diagram.
By removing a node from the extended diagram you arrive at the Dynkin diagram
of a subgroup of $G$. By Theorem\ 5.5 and the subsequent remark in
\cite{Dyn57.1} these groups are maximal. Since they are semisimple they are also maximal
reductive.
To complete the list we need to consider the maximal parabolic subgroups of $G$.
Any reductive subgroup of a parabolic can be assumed to be a subgroup of its
Levi factor by Theorem\ 1 in \cite{LS96}. By considering the Dynkin diagrams it
is transparent that the Levi subgroups need not be maximal reductive but can be
subgroups of a subsystem subgroup. A simple case by case check shows that there
are only two Levi groups, that are maximal reductive.
Summarizing this we have the following maximal reductive subgroups containing a
maximal torus.
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
$G$ & subsystem subgroups & Levi subgroups\\\hline\hline
$G_2$ & $A_2$, $A_1 \times A_1$ & \\ \hline
$F_4$ & $A_1 \times C_3$, $A_2 \times A_2$, $A_3 \times A_1$, $B_4$ &\\ \hline
$E_6$ & $A_5 \times A_1$, $A_2 \times A_2 \times A_2$ & $D_5 \times \mathbb{C}^*$\\
\hline $E_7$ & $D_6 \times A_1$ $A_5 \times A_2$, $A_3 \times A_3 \times A_1$,
$A_7$ & $E_6 \times \mathbb{C}^*$ \\\hline
$E_8$ & $A_1 \times E_7$, $A_2 \times E_6$, $A_3 \times D_5$, $A_4
\times A_4$ &\\
& $A_5 \times A_2 \times A_1$, $A_7 \times A_1$, $D_8$, $A_8$&\\ \hline
\end{tabular}
\end{center}
\section{\texorpdfstring{The exceptional group of type $G_2$}{The
exceptional group of type G2}}
We will now consider the simply connected simple algebraic group $G$ of type
$G_2$. The long roots of its root system form a subsystem of type $A_2$ and we
will consider the subsystem subgroup $H$ obtained in this way. The simple roots
of $H$ are given by
\begin{equation*}
(1,0)_{A_2}=(3,1) \text{ and } (0,1)_{A_2}=(0,1).
\end{equation*}
Using the same methods as before we can prove:
\begin{thm}
The varieties $G/P_1$ and $G/P_2$ are $H$-spherical.
\end{thm}
\begin{proof}\mbox{}\\
\noindent \underline{Case $G/P_1$:} We compute
\begin{equation*}
L=\langle T, U_{\pm (0,1)}\rangle.
\end{equation*}
and
\begin{equation*}
N= \mathbb{C} X_{-(1,0)_{G_2}} \oplus \mathbb{C} X_{-(1,1)_{G_2}} \oplus \mathbb{C}
X_{-(2,1)_{G_2}}.
\end{equation*}
If we define $X:= X_{-(1,1)} + X_{-(2,1)}$ we have $[\mathfrak b, X]=N$, which
shows that $N$ is $L$-spherical. It follows that $G/P_1$ is a spherical
$H$-variety.
\underline{Case $G/P_2$:} In this case we can compute that $L=T$ and
\begin{equation*}
N = \mathbb{C} X_{-(1,1)} \oplus \mathbb{C} X_{-(2,1)}.
\end{equation*}
The module $N$ consists of two linearly independent root spaces and since $T$ is
2-dimensional $N$ is obviously $L$-spherical. That implies that $G/P_2$
is a spherical $H$-variety.
\end{proof}
\begin{thm}
Let $G$ be of type $G_2$ and $H$ of type $A_2$. Then we have the following
branching rules:
\begin{alignat*}{4}
\text{i)}\quad & &\textnormal{res}^G_H (V_{k \omega_1}) &= &
&\,\,\,\,\bigoplus_{a_1+a_2\leq k} &&V_{a_1\lambda_1+a_2
\lambda_2},\\
\text{ii)}\quad & &\textnormal{res}^G_H (V_{k \omega_2}) &= &
&\bigoplus_{a_1+a_2+a_3=k} &&V_{(a_1+a_3)\lambda_1+(a_2 + a_3)\lambda_2}.
\end{alignat*}
\end{thm}
\begin{rem}
In $G_2$ the fundamental weights are self-dual.
\end{rem}
\begin{proof}
\underline{i)} We use ``LiE'' to compute the restriction of $V_{\omega_1}$ and
get
\begin{equation*}
\textnormal{res}^G_H(V_{\omega_1})= \mathbb{C} \oplus V_{\lambda_1} \oplus V_{\lambda_2}.
\end{equation*}
Let $f_0,f_1,f_2$ be highest weight vectors of these representations. We
need to show that $\mathbb{C}[\widehat Y]^{U_H}$ is generated by these elements, i.e.\
we need to show that the dimension of $\mathbb{C}[\widehat Y]^{U_H}$ is~3.
By considering $X_{-(1,0)}\in N$ we immediately see that the $U_L$-orbit of this
element is of codimension~2. Thus $\dim\mathbb{C}[\widehat Y]^{U_H}=3$ and since we
have already found three algebraically independent elements the branching rules
follow immediately.
\underline{ii)} We use ``LiE'' to compute
\begin{equation*}
\textnormal{res}^G_H(V_{\omega_2})= V_{\lambda_1} \oplus V_{\lambda_2} \oplus
V_{\lambda_1+\lambda_2}.
\end{equation*}
Let $f_1,f_2,f_3$ be highest weight vectors of these modules. We know that $U_L$
is the maximal torus in this case and so the unipotent radical is just the
identity. A generic orbit in $N$ is of dimension~0. And since $N$ is
2-dimensional, its codimension is~2. That means a generic $U_H$-orbit has
codimension 3 in $\widehat Y$ and that is also the dimension of $\mathbb{C}[\widehat
Y]^{U_H}$. We have already found three linearly independent elements which form
a generating set. The branching rules follow immediately.
\end{proof}
\begin{prp}
The varieties $G/P_i$ are not spherical $H$-varieties if $H$ is any other
maximal reductive subgroup of $G_2$.
\end{prp}
\begin{proof}
We have the following maximal reductive subgroups besides $A_2$: $A_1\times
A_1$ and $A_1$. If we compute the dimensions of a Borel subgroup in each case
and the dimensions of $G/P_i$ we obtain:
\begin{equation*}
\begin{array}{l|cc}
& G/P_1 & G/P_2 \\\hline
\dim & 5 & 5
\end{array}
\end{equation*}
and
\begin{equation*}
\begin{array}{l|cc}
H & A_1\times A_1 & A_1\\\hline
\dim B_H & 4 & 2
\end{array}
\end{equation*}
So $\dim B_H < \dim G/P_i$, $i=1,2$ for these subgroups.
\end{proof}
\section{\texorpdfstring{The exceptional group of type $F_4$}{The
exceptional group of type F4}}
In this section let $G$ be the group of type $F_4$.
Let $H$ be the subgroup of type $B_4$ in $G$. This is a subsystem subgroup so
from the Dynkin-diagram of $F_4$ we pass on to the extended Dynkin-diagram by
adding the smallest root $\delta$ to the system of simple roots.
\begin{center}
\begin{tikzpicture}[vertex/.style={circle,fill,thick, inner sep=0pt,minimum
size=5pt}, node distance=3.5mm and 7 mm]
\node[vertex, label= below:$\delta$] (delta){};
\node[vertex, right =of delta, label= below:$1$] (1){};
\node[vertex, right =of 1,label= below: $2$] (2) {};
\node[vertex, right =of 2,label= below: $3$] (3) {};
\node[vertex, right =of 3,label= below: $4$] (4) {};
\node at ($(2.east)!.5!(3.west)$) {$\rangle$};
\path
(delta.east) edge (1.west);
\path
(1.east) edge (2.west);
\path
($(2.east)+(-0.9mm,0.8mm)$) edge
($(3.west)+(0.9mm,0.8mm)$);
\path
($(2.east)+(-0.9mm,-0.8mm)$) edge
($(3.west)+(0.9mm,-0.8mm)$);
\path
(3.east) edge (4.west);
\end{tikzpicture}
\end{center}
By removing the simple root $\alpha_4$ we obtain a root-subsystem of type
$B_4$ and thus we find the corresponding subgroup $H\subset G$.
Explicitly we can choose the roots
\begin{alignat*}{2}
(1,0,0,0)_{B_4} &= (0,1,2,2), & \quad (0,1,0,0)_{B_4} &= (1,0,0,0),\\
(0,0,1,0)_{B_4} &= (0,1,0,0), & \quad (0,0,0,1)_{B_4} &= (0,0,1,0),\\
\end{alignat*}
which form a set of simple roots of a root subsystem of type $B_4$ in $F_4$.
We have the following theorem:
\begin{thm}
The varieties $G/P_i$, $i=1,\ldots,4$, are spherical $H$-varie\-ties.
\end{thm}
\begin{proof}
We need to check that $N$ is a spherical $L$-module in each case.
\underline{Case $G/P_1$:}
In this case we have
\begin{align*}
L =\langle T,
& U_{\pm(0,1,2,2)}, U_{\pm(0,1,0,0)}, U_{\pm(0,0,1,0)},\\
& U_{\pm(0,1,1,0)}, U_{\pm(0,1,2,0)}\rangle
\end{align*}
and
\begin{equation*}
N=\mathbb{C} X_{-(1,2,3,1)} \oplus \mathbb{C} X_{-(1,2,2,1)} \oplus
\mathbb{C} X_{-(1,1,2,1)} \oplus \mathbb{C} X_{-(1,1,1,1)}.
\end{equation*}
The Borel subgroup $B_{L}$ of $L$ obviously contains the maximal torus $T$
of $G$. Since $N$ consists of four root spaces with linearly independent roots
and $T$ is 4-dimensional we know that there is a dense $B_{L}$-orbit in $N$.
Hence $N$ is $L$-spherical and that implies that $G/P_1$ is $H$-spherical.
\underline{Case $G/P_2$:}
Here we have
\begin{align*}
L=\langle T, U_{\pm(1,0,0,0)}, U_{\pm(0,0,1,0)} \rangle.
\end{align*}
We compute $N$ in the same way as in the previous case and get
\begin{align*}
N=\; &\mathbb{C} X_{-(0,1,1,1)} \oplus \mathbb{C} X_{-(0,1,2,1)} \oplus
\mathbb{C} X_{-(1,1,1,1)} \oplus \mathbb{C} X_{-(1,1,2,1)} \oplus\\
&\mathbb{C} X_{-(1,2,2,1)} \oplus \mathbb{C} X_{-(1,2,3,1)}.
\end{align*}
We check the sphericity on the level of Lie algebras. Consider the element
\begin{equation*}
X:=X_{-(1,1,2,1)} + X_{-(0,1,2,1)} +
X_{-(1,1,1,1)} + X_{-(1,2,3,1)}
\end{equation*}
in $N$.
Then $[\mathfrak b, X]= N$. That means that $N$ is a
spherical $L$-variety and therefore $G/P_2$ is a spherical $H$-variety.
\underline{Case $G/P_3$:}
We get
\begin{align*}
N = & \mathbb{C} X_{-(0,0,1,1)} \oplus \mathbb{C} X_{-(0,1,1,1)} \oplus \mathbb{C}
X_{-(0,1,2,1)} \oplus\\
& \mathbb{C} X_{-(1,1,1,1)} \oplus \mathbb{C} X_{-(1,1,2,1)} \oplus \mathbb{C}
X_{-(1,2,2,1)} \oplus \\
& \mathbb{C} X_{-(1,2,3,1)}.
\end{align*}
If we consider
\begin{equation*}
X:=X_{-(1,2,3,1)}+ X_{-(1,2,2,1)} +
X_{-(1,1,1,1)} + X_{-(0,1,2,1)} \in N
\end{equation*}
we have that $[\mathfrak b,X]=N$,
i.e.\ $N$ is a spherical $L$-variety and that means that $G/P_3$ is a
spherical $H$-variety.
\underline{Case $G/P_4$:}
In this case we have
\begin{align*}
L=\langle T,
&U_{\pm(1,0,0,0)}, U_{\pm(0,1,0,0)}, U_{\pm(0,0,1,0)},\\
&U_{\pm(1,1,0,0)}, U_{\pm(0,1,1,0)}, U_{\pm(1,1,1,0)},\\
&U_{\pm(0,1,2,0)}, U_{\pm(1,1,2,0)},
U_{\pm(1,2,2,0)}\rangle
\end{align*}
and
\begin{align*}
N= & \mathbb{C} X_{-(0,0,0,1)} \oplus \mathbb{C} X_{-(0,0,1,1)} \oplus \mathbb{C}
X_{-(0,1,1,1)} \oplus\\
& \mathbb{C} X_{-(0,1,2,1)} \oplus \mathbb{C} X_{-(1,1,1,1)} \oplus \mathbb{C}
X_{-(1,1,2,1)} \oplus\\
& \mathbb{C} X_{-(1,2,2,1)} \oplus \mathbb{C} X_{-(1,2,3,1)}.
\end{align*}
The module $N$ has the following structure.
\begin{equation*}
\xymatrix {
& & & X_{-(0,1,2,1)}\ar[dr]^{(0,0,1,0)}\\
X_{-(1,2,3,1)} \ar[r]^{(0,0,1,0)} & X_{-(1,2,2,1)} \ar[r]^{(0,1,0,0)} &
X_{-(1,1,2,1)} \ar[dr]_{(0,0,1,0)} \ar[ur]^{(1,0,0,0)} & &
X_{-(0,1,1,1)} \ar@{-}[r]& \cdots\\
& & & X_{-(1,1,1,1)} \ar[ur]_{(1,0,0,0)} & &\\
\cdots \ar[r]^{(0,1,0,0)} & X_{-(0,0,1,1)} \ar[r]^{(0,0,1,0)}&
X_{-(0,0,0,1)}& & }
\end{equation*}
We have $L=\mathbb{C}^* \times SO_7$ and $N$ is an irreducible $L$-module of
dimension~8. There exists only one such module which is the
$\text{Spin}_7$-module. That $N$ is a spherical $L$-module was proven by Victor
Kac \citep[][Thm. 3, p.~208]{Kac80}. It follows that $G/P_4$ is a spherical
$H$-module.
\end{proof}
The spherical cases imply the following branching rules.
\begin{thm}
Let $G$ be of type $F_4$ and $H$ of type $B_4$. Then we have the following branching rules:
\begin{alignat*}{4}
\text{i)}\quad & &\textnormal{res}^G_H (V_{k \omega_1}) &= &
&\,\,\,\,\bigoplus_{a_1+a_2=k} &&V_{a_1\lambda_2+a_2
\lambda_4},\\
\text{ii)}\quad & &\textnormal{res}^G_H (V_{k \omega_2}) &= &
&\bigoplus_{a_1+\ldots+a_5=k} &&V_{(a_1+a_2)\lambda_1+(a_3 + a_4)\lambda_2 + (a_1+a_5)\lambda_3 + (a_2 +
a_4)\lambda_4},\\
\text{iii)}\quad & &\textnormal{res}^G_H (V_{k \omega_3}) &=
&&\bigoplus_{a_1+\ldots+a_5=k} &&V_{(a_1+a_5)\lambda_1 + a_2 \lambda_2 + a_3 \lambda_3 + (a_4+a_5)
\lambda_4},\\
\text{iv)}\quad & &\textnormal{res}^G_H (V_{k \omega_4}) &= &&
\,\,\,\,\bigoplus_{a_1+a_2\leq k} &&V_{a_1\lambda_2+a_2 \lambda_4}.\\
\end{alignat*}
\end{thm}
\begin{rem}
In $F_4$ the fundamental weights are self-dual.
\end{rem}
\begin{proof}\mbox{}\\
\underline{i):}
Standard computations yield
\begin{equation*}
\textnormal{res}^G_H (V_{\omega_1}) = V_{\lambda_2} \oplus V_{\lambda_4}.
\end{equation*}
Let now $f_1,f_2 \in V_{\omega_1}$ be highest weight vectors of $V_{\lambda_2}$
and $V_{\lambda_4}$ respectively. We will show that $\mathbb{C}[\widehat Y]^{U_H}$
is generated by these degree 1 elements. We know that $\mathbb{C}[\widehat Y]^{U_H}$
is a polynomial ring. The grading and weights of $f_1$ and $f_2$ imply that
they are algebraically independent. To rule out the possibility that
there are generators of degree two or higher we need to show that the Krull
dimension of $\mathbb{C}[\widehat Y]^{U_H}$ is $2$.
Thus we need to find a generic $U_L$-orbit in $N$ and compute its
codimension. Since we have found 2 algebraically independent elements in
$\mathbb{C}[\widehat Y]^{U_H}$, we already know that the codimension must be at least
2.
Consider the Lie algebra $\mathfrak l$ of $L$. From above we know that the
Lie algebra $\mathfrak u$ of
$U_L$, is
\begin{equation*}
\mathfrak u = \mathbb{C} X_{(0,1,2,2)}\oplus \mathbb{C} X_{(0,1,0,0)} \oplus
\mathbb{C} X_{(0,0,1,0)} \oplus \mathbb{C} X_{(0,1,1,0)} \oplus \mathbb{C}
X_{(0,1,2,0)}.
\end{equation*}
Define $X:= X_{-(1,2,3,1)} \in N$. Then
\begin{align*}
[X_{(0,1,2,2)},X]&= 0, & [X_{(0,1,0,0)},X]&=0, \\
[X_{(0,0,1,0)},X]&=X_{-(1,2,2,1)},
& [X_{(0,1,1,0)},X]&=X_{-(1,1,2,1)},\\
[X_{(0,1,2,0)},X]&=
X_{-(1,1,1,1)},
\end{align*}
which shows that the orbit of $X$ is of dimension 3. Thus a generic orbit has
dimension at least 3 with codimension at most~1. By Proposition \ref{prp:codim}
we know that in this case $\dim \mathbb{C} [\widehat Y]^{U_H}\leq 2$. But since we
have found two generators the dimension is exactly~2 and the restriction rules
follow.
\underline{ii):}
In this case we need to find generators of $\mathbb{C}[\widehat
Y]^{U_H}$. One can use the software ``LiE'' to compute
\begin{equation*}
\textnormal{res}^G_H (V_{\omega_2}) = V_{\lambda_1 + \lambda_3} \oplus
V_{\lambda_1 + \lambda_4} \oplus V_{\lambda_2} \oplus V_{\lambda_2 + \lambda_4}
\oplus V_{\lambda_3}.
\end{equation*}
Let $f_1,\ldots,f_5$ be highest weight vectors of these irreducible modules.
Consider $X:=X_{-(1,1,2,1)}+X_{-(1,2,3,1)} \in N$ and let $\mathfrak u$ be the
Lie-algebra of $U_L$ the unipotent radical of $L$. The stabilizer of this
element is just 0, which means that the dimension of a generic $U_L$-orbit is~2
with codimension~4. This implies that the codimension of a generic $U_H$-orbit
in $\widehat Y$ is 5. Thus $\mathbb{C}[\widehat Y]^{U_H}$ is generated by its degree 1
elements and the assertion follows.
\underline{iii):}
We need to find generators of $\mathbb{C}[\widehat Y]^{U_H}$. One can use
``LiE'' to compute
\begin{equation*}
\textnormal{res}^G_H (V_{\omega_3}) = V_{\lambda_1} \oplus
V_{\lambda_2} \oplus V_{\lambda_3} \oplus V_{\lambda_4} \oplus V_{\lambda_1 +
\lambda_4}.
\end{equation*}
Let $f_1,\ldots,f_5$ be highest weight vectors of these irreducible modules.
Consider $X:=X_{-(1,1,1,1)}+X_{-(1,2,2,1)} \in N$ and take an element $u \in
\mathfrak u$ with $u = a X_{(1,0,0,0)} + bX_{(0,1,0,0)} + c
X_{(1,1,0,0)}$. Then
\begin{alignat*}{2}
&&\quad [u,X]&=0\\
&\Rightarrow \quad & &= aX_{-(0,1,1,1)} + b X_{-(1,1,2,1)} + c
(X_{-(0,1,2,1)}+X_{-(0,0,1,1)})\\
&\Rightarrow & &a=b=c=0 \Rightarrow u=0
\end{alignat*}
and hence a generic $U_L$-orbit has dimension 3 with codimension 4. That means
that $\mathbb{C}[\widehat Y]^{U_H}$ is of dimension 5 and generated by the elements
$f_i$.
\underline{iv):}
In this case we need to find generators of $\mathbb{C}[\widehat
Y]^{U_H}$. We use ``LiE'' to compute
\begin{equation*}
\textnormal{res}^G_H (V_{\omega_4}) = \mathbb{C} \oplus V_{\lambda_1} \oplus
V_{\lambda_4}.
\end{equation*}
Let $f_1,\ldots,f_3$ be highest weight vectors of these irreducible modules.
Consider $X:= X_{-(1,2,3,1)}$. We know that for
\begin{equation*}X_{(1,0,0,0)}, X_{(0,1,0,0)},X_{(1,1,0,0)}
\in \mathfrak u
\end{equation*}
we have
\begin{align*}
[X_{(1,0,0,0)},X]&=[X_{(0,1,0,0)},X]=[X_{(1,1,0,0)},X]=0\\
\end{align*}
and
\begin{align*}
[X_{(0,0,1,0)},X]&=X_{-(1,2,2,1)}, &
[X_{(0,1,1,0)},X]&=X_{-(1,1,2,1)},\\
[X_{(0,1,2,0)},X]&=X_{-(1,1,1,1)},&
[X_{(1,1,1,0)},X]&=X_{-(0,1,2,1)},\\
[X_{(1,1,2,0)},X]&=X_{-(0,1,1,1)},&
[X_{(1,2,2,0)},X]&=X_{-(0,0,1,1)}\\
\end{align*}
and thus the generic stabilizer is at most of dimension 3. The generic orbit is
at least of dimension 6 and thus its codimension is at most 2. This means that a
generic $U_H$-orbit in $\widehat Y$ is of dimension less or equal to 3.
Since we have found 3 algebraically independent elements the dimension of
$\mathbb{C}[\widehat Y]^{U_H}$ is exactly 3 and this finishes the proof.
\end{proof}
\begin{prp}
The varieties $G/P_i$ are not spherical $H$-varieties if $H$ is any other
maximal reductive subgroup of $F_4$.
\end{prp}
\begin{proof}
We have the following maximal reductive subgroups besides $B_4$: $A_1 \times
C_3$, $A_2 \times A_2$, $A_3 \times A_1$, $A_1\times G_2$ and $A_1$. If we
compute the dimensions of a Borel subgroup in each case and the dimensions of
$G/P_i$ we obtain:
\begin{equation*}
\begin{split}
&\begin{array}{l|cccc}
& G/P_1 & G/P_2 & G/P_3 & G/P_4\\\hline
\dim \phantom{B_H}& 15 & 20 & 20 & 15
\end{array}\\
&\begin{array}{l|ccccc}
H & A_1\times C_3 & A_2 \times A_2 & A_3 \times A_1 & A_1\times
G_2 & A_1\\\hline
\dim B_H & 14 & 10 & 11 & 10 & 2
\end{array}
\end{split}
\end{equation*}
So we have $\dim B_H< \dim G/P_i$ for $i=1,\ldots,4$ in each case.
\end{proof}
\section{\texorpdfstring{The exceptional group of type $E_6$}{The exceptional
group of type E6}}
We will now turn to the group of type $E_6$. First we calculate the dimensions
of the Borel subgroups of the maximal reductive subgroups as well as the
dimensions of $G/P_i$ for $i=1,\ldots,6$.
\begin{align*}
&\begin{array}{l|c c c c c c c c}
H &A_5\times A_1& A_2\!\times\! A_2 \times\! A_2 & D_5 \times \mathbb{C}^*
& A_2 \times G_2 & G_2 & A_2 & F_4 & C_4\\\hline
\dim B_H & 22 & 15 & 26 & 13 & 8 & 5 & 28 & 20
\end{array}
\intertext{and}
&\begin{array}{l|cccccc}
& G/P_1 & G/P_2 & G/P_3 & G/P_4 & G/P_5 & G/P_6\\\hline
\dim \phantom{B_H}& 16 & 21 & 25 & 29 & 25 & 16
\end{array}\quad.
\end{align*}
Thus we get the following proposition.
\begin{prp}
Let $G$ be the simply connected simple algebraic group of type
$E_6$ and let $H$ be a maximal reductive subgroup of type $A_2\times A_2 \times
A_2$, $A_2 \times G_2$, $G_2$ or $A_2$.\\
Then $G/P_i$ is not $H$-spherical for $i=1,\ldots,6$.
\end{prp}
\begin{proof}
In these cases we have $\dim B_H < \dim G/P_i$ for $i=1,\ldots,6$.
\end{proof}
Now we will consider the remaining groups and first we start with the
subsystem subgroup of type $A_5 \times A_1$.
\begin{thm} Let $G$ be the simply connected simple algebraic group of
type $E_6$ and let $H$ be the maximal reductive subgroup of type $A_5\times A_1$.
Then $G/P_1$ and $G/P_6$ are spherical $H$-varieties. The varieties
$G/P_2,\ldots,G/P_5$ are not $H$-spherical.
\end{thm}
\begin{proof}
The dimension of a Borel subgroup of a group of type $A_5\times A_1$ is~$22$.
Since we have $\dim G/P_3=25$, $\dim G/P_4=29$, $\dim G/P_5=25$ these varieties
cannot be spherical.
We know that $\omega^*_2=\omega_2$ in type $E_6$. Now if $G/P_2$ was a spherical
$H$-variety, $\text{res}^G_H(V_{k\omega_2})$ would be multiplicity-free for all
$k \in \mathbb{N}$ by what has been said above. But with ``LiE'' we compute
\begin{equation*}
\text{res}^G_H(V_{4\omega_2})=\ldots\oplus2 (V_{2\lambda_3}\otimes
V_{3\lambda_6})\oplus\ldots
\end{equation*}
which means that there are multiplicities in this case.
To prove that $G/P_1$ and $G/P_6$ are spherical $H$-varieties we proceed as in
the cases above. We will show how $H$ is embedded in $G$. For doing so
we consider the extended Dynkin-diagram of type $E_6$ again by adding the
smallest root $\delta$ to the simple roots. Now omitting the root $\alpha_2$ we
obtain the embedding of $A_5 \times A_1$ in $E_6$.
\begin{center}
\begin{tikzpicture}
{
[start chain, node
distance=3.5mm and 7 mm,vertex/.style={circle,fill,thick, inner sep=0pt,minimum
size=5pt}]
\node[vertex,on chain,join,label= below: $1$] (1) {};
\node[vertex,on chain,join,label= below: $3$] (3) {};
\node[vertex,on chain,join,label= below: $4$] (4) {};
{[start branch=4]
\node[vertex,on chain=going above,join,label=left: $2$] (2) {};
\node[vertex,on chain=going above,join,label=left: $\delta$] (2) {};
}
\node[vertex,on chain,join,label= below: $5$] (5) {};
\node[vertex,on chain,join,label= below: $6$] (6) {};
}
\end{tikzpicture}
\end{center}
Explicitly we get the following set of simple roots:
\begin{alignat*}{2}
(1,0,0,0,0,0)_{A_5\times A_1} &= (1,0,0,0,0,0) & \quad
(0,1,0,0,0,0)_{A_5\times A_1} &= (0,0,1,0,0,0)\\
(0,0,1,0,0,0)_{A_5\times A_1} &= (0,0,0,1,0,0) & \quad
(0,0,0,1,0,0)_{A_5\times A_1} &= (0,0,0,0,1,0)\\
(0,0,0,0,1,0)_{A_5\times A_1} &= (0,0,0,0,0,1) & \quad
(0,0,0,0,0,1)_{A_5\times A_1} &= (1,2,2,3,2,1)
\end{alignat*}
\underline{Case $G/P_1$:}
We compute
\begin{equation*}
\begin{split}
L =\langle &T, U_{\pm(0,0,1,0,0,0)}, U_{\pm(0,0,0,1,0,0)},
U_{\pm(0,0,0,0,1,0)}, U_{\pm(0,0,0,0,0,1)},\\
&U_{\pm(0,0,1,1,0,0)},U_{\pm(0,0,0,1,1,0)},
U_{\pm(0,0,0,0,1,1)},\\
&U_{\pm(0,0,1,1,1,0)}, U_{\pm(0,0,0,1,1,1)},
U_{\pm(0,0,1,1,1,1)}\rangle
\end{split}
\end{equation*}
and
\begin{equation*}
\begin{split}
N=\;&\mathbb{C} X_{-(1,1,1,1,0,0)} \oplus \mathbb{C} X_{-(1,1,1,1,1,0)} \oplus \mathbb{C}
X_{-(1,1,1,2,1,0)} \oplus \\
&\mathbb{C} X_{-(1,1,1,1,1,1)} \oplus \mathbb{C} X_{-(1,1,2,2,1,0)} \oplus \mathbb{C}
X_{-(1,1,1,2,1,1)} \oplus\\
& \mathbb{C} X_{-(1,1,2,2,1,1)} \oplus \mathbb{C} X_{-(1,1,1,2,2,1)} \oplus \mathbb{C}
X_{-(1,1,2,2,2,1)} \oplus\\
&\mathbb{C} X_{-(1,1,2,3,2,1)}.
\end{split}
\end{equation*}
Now let $X:= X_{-(1,1,2,3,2,1)}+X_{-(1,1,1,1,1,1)}$.
We have
\begin{equation*}
[\mathfrak h, X]=
\langle
X_{-(1,1,2,3,2,1)},\, X_{-(1,1,1,1,1,1)}
\rangle,
\end{equation*}
since the roots are linearly independent.
Next we compute
{\allowdisplaybreaks
\begin{alignat*}{2}
[X_{(0,0,0,1,0,0)},X]&=X_{-(1,1,2,2,2,1)}\quad &
[X_{(0,0,0,1,1,0)},X]&=X_{-(1,1,2,2,1,1)}\\
[X_{(0,0,1,1,0,0)},X]&=X_{-(1,1,1,2,2,1)}\quad &
[X_{(0,0,1,1,1,0)},X]&=X_{-(1,1,1,2,1,1)}\\
[X_{(0,0,0,1,1,1)},X]&=X_{-(1,1,2,2,1,0)}\quad &
[X_{(0,0,1,1,1,1)},X]&=X_{-(1,1,1,2,1,0)}\\
[X_{(0,0,0,0,0,1)},X]&=X_{-(1,1,1,1,1,0)}\quad &
[X_{(0,0,0,0,1,1)},X]&=X_{-(1,1,1,1,0,0)}\\
\end{alignat*}
}%
and these computations show that we have ten linearly independent vectors in
$[\mathfrak b,X] \Rightarrow [\mathfrak b,X]=N \Rightarrow$ $N$ is a spherical
$L$-module. Hence $G/P_1$ is a spherical $H$-variety.
\underline{Case $G/P_6$:} The $H$-sphericity of $G/P_6$ is an immediate
corollary of the following theorem which states that
$\mathbb{C}[\widehat Y]$ is multiplicity free.
\end{proof}
\begin{thm}
Let $G$ be the simply connected simple algebraic group of type $E_6$ and let
$H\subset G$ be the maximal reductive subgroup of type $A_5 \times A_1$.
Then we have the following branching rules:
\begin{alignat*}{4}
\text{i)}\quad & &\textnormal{res}^G_H (V_{k \omega_1}) &= &
&\,\,\,\,\bigoplus_{a_1+2a_2+a_3=k} &&
V_{a_1\lambda_2+a_2\lambda_4 + a_3 \lambda_5}\otimes V_{a_3\lambda_6},\\
\text{ii)}\quad & &\textnormal{res}^G_H (V_{k \omega_6}) &= &
&\,\,\,\,\bigoplus_{a_1+2a_2+a_3=k} &&
V_{a_1\lambda_1+a_2\lambda_2 + a_3 \lambda_4} \otimes V_{a_1 \lambda_6}.
\end{alignat*}
\end{thm}
\begin{rem}
In $E_6$ we have $\omega^*_1=\omega_6$, $\omega_2^*= \omega_2$,
$\omega_3^*=\omega_5$ and $\omega_4^*=\omega_4$.
\end{rem}
\begin{proof}
\underline{ii)}
With ``LiE'' we compute
\begin{equation*}
\begin{split}
\textnormal{res}^G_H(V_{\omega_6})&=(V_{\lambda_4} \otimes \mathbb{C}) \oplus
(V_{\lambda_1} \otimes V_{\lambda_6}),\\
\textnormal{res}^G_H(V_{2\omega_6})&=(V_{2\lambda_4}\otimes \mathbb{C}) \oplus
(V_{\lambda_1+\lambda_4}\otimes V_{\lambda_6})\oplus (V_{2\lambda_1}\otimes
V_{2\lambda_6}) \oplus
(V_{\lambda_2}\otimes \mathbb{C}).
\end{split}
\end{equation*}
There are at least two generators of degree~1 and of weights $(\lambda_4,0)$ and
$(\lambda_1,\lambda_6)$ and one generator of degree~2 and of weight
$(\lambda_2,0)$ for $\mathbb{C} [\widehat Y]^{U_H}$ with $Y=G/P_1$. In the proof of the
previous theorem we have found an element $X \in N$ with a $U_L$-orbit of codimension~2. So it
follows that $\dim \mathbb{C}[\widehat Y]^{U_H}=3$ and the branching rules
follow immediately.
\underline{i)} Theses branching rules follow directly from ii) by noting that
$\omega_1=\omega_6^*$, $\lambda_1^*=\lambda_5$, $\lambda_2^*=\lambda_4$ and
$\lambda_6^*=\lambda_6$.
\end{proof}
\begin{thm}\label{thm:E6-F4}
Let $G$ be the simply connected simple algebraic group of type
$E_6$ and let $H$ be the maximal reductive subgroup of type $F_4$.
Then $G/P_i$, $i \not =4$, are
spherical $H$-varieties. The variety $G/P_4$ is not
$H$-spherical.
\end{thm}
\begin{proof}
If we have the Dynkin diagrams
\begin{center}
\begin{tikzpicture}[vertex2/.style={circle,fill,thick, inner sep=0pt,minimum
size=5pt}, node distance=3.5mm and 7 mm]
{[start chain, node
distance=3.5mm and 7 mm,vertex/.style={circle,fill,thick, inner sep=0pt,minimum
size=5pt}]
\node[vertex,on chain,join,label= below: $1$] (1) {};
\node[vertex,on chain,join,label= below: $3$] (3) {};
\node[vertex,on chain,join,label= below: $4$] (4) {};
{[start branch=4]
\node[vertex,on chain=going above,join,label=left: $2$] (2) {};
}
\node[vertex,on chain,join,label= below: $5$] (5) {};
\node[vertex,on chain,join,label= below: $6$] (6) {};
}
\node[right= of 6] (and) {and};
\node[vertex2, right= of and, label= below:$x$] (x){};
\node[vertex2, right =of x,label= below: $y$] (y) {};
\node[vertex2, right =of y,label= below: $z$] (z) {};
\node[vertex2, right =of z,label= below: $u$] (u) {};
\node at ($(y.east)!.5!(z.west)$) {$\rangle$};
\path
(x.east) edge (y.west);
\path
($(y.east)+(-0.9mm,0.8mm)$) edge
($(z.west)+(0.9mm,0.8mm)$);
\path
($(y.east)+(-0.9mm,-0.8mm)$) edge
($(z.west)+(0.9mm,-0.8mm)$);
\path
(z.east) edge (u.west);
\end{tikzpicture}
\end{center}
of $E_6$ and $F_4$, then we have an embedding of the simple Lie-algebra $F_4$ in
$E_6$ by choosing the following root vectors
\begin{alignat*}{2}
X_x&:=X_{(0,1,0,0,0,0)},\quad & X_z&:=\frac 1 {\sqrt 2}
(X_{(0,0,1,0,0,0)}+X_{(0,0,0,0,1,0)})\\
X_y &:= X_{(0,0,0,1,0,0)},& X_u&:=\frac 1 {\sqrt 2}
(X_{(1,0,0,0,0,0)}+X_{(0,0,0,0,0,1)})
\end{alignat*}
(\citep[p.\ 258, Table 24]{Dyn57.1} with different numbering of the Dynkin
diagrams). Now we consider the associated algebraic
subgroup of $E_6$.
\underline{Case $G/P_1$:}
We compute
\begin{equation*}
N= \mathbb{C} X_{-(1,1,1,2,2,1)}.
\end{equation*}
So $N$ is obviously $L$-spherical and thus $G/P_1$ is $H$-spherical.
\underline{Case $G/P_6$:} The $H$-sphericity of $Y=G/P_6$ is an immediate
corollary of the following theorem which states that
$\mathbb{C}[\widehat Y]$ is multiplicity free.
\underline{Case $G/P_2$:} In this case we get
\begin{equation*}
\begin{split}
N= &\mathbb{C} X_{-(0,1,0,1,1,0)} \oplus \mathbb{C} X_{-(0,1,0,1,1,1)} \oplus \mathbb{C}
X_{-(0,1,1,1,1,1)} \oplus\\
&\mathbb{C} X_{-(0,1,1,2,1,1)} \oplus \mathbb{C} X_{-(0,1,1,2,2,1)} \oplus \mathbb{C}
X_{-(1,1,1,2,2,1)_{E_6}}.
\end{split}
\end{equation*}
If we define $X:=X_{-(1,1,1,2,2,1)}$ then we have:
\begin{alignat*}{2}
[X_{(0,0,0,1)_{F_4}}, X]&= X_{-(0,1,1,2,2,1)}, &\quad
[X_{(0,0,1,1)_{F_4}},X]&=X_{-(0,1,1,2,1,1)},\\
[X_{(0,1,1,1)_{F_4}}, X]&= X_{-(0,1,1,1,1,1)}, &\quad
[X_{(0,1,2,1)_{F_4}},X]&=X_{-(0,1,0,1,1,1)},\\
[X_{(0,1,2,2)_{F_4}},X]&=X_{-(0,1,0,1,1,0)}.
\end{alignat*}
With $[\mathfrak h, X]=\mathbb{C} X$ we get $[\mathfrak b ,X]=N$ and it follows that $N$
is a spherical $L$-module.
\underline{Case $G/P_3$:} In this case we get
\begin{equation*}
\begin{split}
N=&\mathbb{C} X_{-(0,0,1,1,1,1)}\oplus \mathbb{C} X_{-(0,1,1,1,1,1)} \oplus \mathbb{C}
X_{-(0,1,1,2,1,1)} \oplus\\
&\mathbb{C} X_{-(0,1,1,2,2,1)} \oplus \mathbb{C} X_{-(1,1,1,2,2,1)}.
\end{split}
\end{equation*}
Set $X:= X_{-(1,1,1,2,2,1)}+ X_{-(0,1,1,2,1,1)}$. Then
we have
\begin{equation*}
[\mathfrak h, X] = \mathbb{C} X_{-(1,1,1,2,2,1)} \oplus \mathbb{C}
X_{-(0,1,1,2,1,1)},
\end{equation*} since the roots of the root vectors defining $X$ are linearly
independent. Furthermore we have
\begin{alignat*}{2}
[X_{(0,0,0,1)_{F_4}},X]&= X_{-(0,1,1,2,2,1)}, \quad &
[X_{(1,0,0,0)_{F_4}},X]&= X_{-(0,1,1,1,1,1)},\\
[X_{(1,1,0,0)_{F_4}},X]&= X_{-(0,0,1,1,1,1)}.
\end{alignat*}
So $[\mathfrak b, X]=N \Rightarrow$ $N$ is a spherical $L$-module and this
implies that $G/P_3$ is $H$-spherical.
\underline{Case $G/P_5$:} The $H$-sphericity of $Y=G/P_5$ is an immediate
corollary of the following theorem which states that
$\mathbb{C}[\widehat Y]$ is multiplicity free.
\end{proof}
We can derive branching rules in the cases where $G/P_i$ is
a spherical $H$-variety.
\begin{thm}
Let $G$ be the simple simply connected algebraic group of type $E_6$ and $H$ be
the subgroup of type $F_4$.
Then we have the branching rules:
\begin{equation*}
\begin{array}{r r c c l}
\text{i)} &\textnormal{res}^G_H (V_{k \omega_1}) &\!\!\!=\!\!\!
&\displaystyle\bigoplus\limits_{a_1 \leq k} & V_{a_1 \lambda_4},\\
\text{ii)} &\textnormal{res}^G_H (V_{k \omega_2})&\!\!\!=\!\!\!
&\displaystyle\bigoplus\limits_{a_1+a_2=k} & V_{a_1\lambda_1+a_2 \lambda_4},\\
\text{iii)} &\textnormal{res}^G_H (V_{k \omega_3}) &\!\!\!=\!\!\!
&\displaystyle\bigoplus\limits_{a_1 + a_2 + a_3 = k} &
V_{a_1 \lambda_1 + a_2 \lambda_3+ a_3 \lambda_4},\\
\text{iv)} &\textnormal{res}^G_H (V_{k \omega_5})&\!\!\!=\!\!\!
&\displaystyle\bigoplus\limits_{a_1+a_2+a_3=k} &
V_{a_1\lambda_1+a_2\lambda_3 + a_3 \lambda_4},\\
\text{v)} &\textnormal{res}^G_H (V_{k \omega_6}) &\!\!\!=\!\!\!
&\displaystyle\bigoplus\limits_{a_1 \leq k} &
V_{a_1 \lambda_4}.
\end{array}
\end{equation*}
\end{thm}
\begin{proof}
\underline{v)} In this case we work with $Y=G/P_1$. With ``LiE'' we compute
\begin{equation*}
\textnormal{res}^G_H(V_{\omega_6})= \mathbb{C} \oplus V_{\lambda_4}.
\end{equation*}
Since $N$ is 1-dimensional in this case, each $U_L$-orbit is 0-dimensional
with codimension~1. So $\dim \mathbb{C}[\widehat Y]^{U_H} = 2$ and $\mathbb{C} [\widehat
Y]^{U_H}$ is generated by its degree-1-elements. The branching rules follow.
\underline{i)} Theses branching rules follow directly from v) by noting that
$\omega_1=\omega_6^*$ and $\lambda_i^*=\lambda_i$.
\underline{ii)} In this case we work with $Y=G/P_2$. With ``LiE'' we compute
\begin{equation*}
\textnormal{res}^G_H(V_{\omega_2})= V_{\lambda_1} \oplus V_{\lambda_4},
\end{equation*}
so there are two generators of degree 1. The module $N$ is of dimension 6 and
we have seen that $X_{-(1,1,1,2,2,1)}\in N$ is an element such that $U_L.X$ is
of dimension 5. So $\dim \mathbb{C}[\widehat Y]^{U_H}\leq 2$ and hence $\mathbb{C} [\widehat
Y]^{U_H}$ is generated by its degree-1-elements. The branching rules follow immediately.
\underline{iv)} In this case we work with $G/P_3$. With ``LiE'' we compute
\begin{equation*}
\textnormal{res}^G_H(V_{\omega_5})= V_{\lambda_1} \oplus V_{\lambda_3} \oplus V_{\lambda_4},
\end{equation*}
so again there are 3 generators of degree 1. The module $N$ is of
dimension~5 and $X_{-(1,1,1,2,2,1)_{E_6}}+ X_{-(0,1,1,2,1,1)_{E_6}}$ is an
element of $N$ with a 3-dimensional $U_L$-orbit (cf.\ proof of previous
theorem). So $\dim \mathbb{C} [\widehat Y]^{U_H}\leq 3$. It follows that $\mathbb{C} [\widehat
Y]^{U_H}$ is generated by its degree-1-elements and so the branching rules follow.
\underline{iii)} These branching rules follow directly from v) by noting that
$\omega_3=\omega_5^*$ and $\lambda_i^*=\lambda_i$.
\end{proof}
\begin{thm}
Let $G$ be the simply connected simple algebraic group of type
$E_6$ and let $H$ be the maximal reductive subgroup of type $C_4$.
Then $G/P_1$ and $G/P_6$ are
spherical $H$-varieties. The varieties $G/P_2,\ldots,G/P_5$ are not
$H$-spherical.
\end{thm}
\begin{proof}
That $G/P_2,\ldots,G/P_5$ are not $H$-spherical follows by dimension reasons.
For the other two cases we consider the Dynkin diagrams
\begin{center}
\begin{tikzpicture}[vertex2/.style={circle,fill,thick, inner sep=0pt,minimum
size=5pt}, node distance=3.5mm and 7 mm]
{[start chain, node
distance=3.5mm and 7 mm,vertex/.style={circle,fill,thick, inner sep=0pt,minimum
size=5pt}]
\node[vertex,on chain,join,label= below: $1$] (1) {};
\node[vertex,on chain,join,label= below: $3$] (3) {};
\node[vertex,on chain,join,label= below: $4$] (4) {};
{[start branch=4]
\node[vertex,on chain=going above,join,label=left: $2$] (2) {};
}
\node[vertex,on chain,join,label= below: $5$] (5) {};
\node[vertex,on chain,join,label= below: $6$] (6) {};
}
\node[right= of 6] (and) {and};
\node[vertex2, right= of and, label= below:$x$] (x){};
\node[vertex2, right =of x,label= below: $y$] (y) {};
\node[vertex2, right =of y,label= below: $z$] (z) {};
\node[vertex2, right =of z,label= below: $u$] (u) {};
\node at ($(z.east)!.5!(u.west)$) {$\langle$};
\path
(x.east) edge (y.west);
\path
(y.east) edge (z.west);
\path
($(z.east)+(-0.9mm,0.8mm)$) edge
($(u.west)+(0.9mm,0.8mm)$);
\path
($(z.east)+(-0.9mm,-0.8mm)$) edge
($(u.west)+(0.9mm,-0.8mm)$);
\end{tikzpicture}
\end{center}
of $E_6$ and $C_4$ respectively. Then the simple Lie-algebra of type $C_4$ is
embedded into the simple Lie-algebra of type $E_6$ by choosing the following
root vectors:
\begin{alignat*}{2}
X_x&\!:=\!\frac 1 {\sqrt 2} (X_{(0,1,1,1,0,0)}\!+\!X_{(0,1,0,1,1,0)}),\quad
& X_y&\!:=\!\frac 1 {\sqrt 2} (X_{(1,0,0,0,0,0)}+X_{(0,0,0,0,0,1)})\\
X_z\! &:=\!\frac 1 {\sqrt 2} (X_{(0,0,1,0,0,0)}\! +\! X_{-(0,0,0,0,1,0}),&
X_u&\!:=\!X_{(0,0,0,1,0,0)}
\end{alignat*}
(cf.\ \citep[p.\ 258, Table 24]{Dyn57.1}).
Now we consider the associated subgroup $H$ of $G$.
\underline{Case $G/P_1$:} We compute
\begin{equation*}
\begin{split}
N= &\mathbb{C} X_{-(1,1,1,1,1,1)}\oplus \mathbb{C} X_{-(1,1,1,2,1,1)} \oplus \mathbb{C}
X_{-(1,1,2,2,1,1)} \oplus\\
& \mathbb{C} X_{-(1,1,2,2,2,1)} \oplus \mathbb{C} X_{-(1,1,2,3,2,1)}.
\end{split}
\end{equation*}
We define $X:= X_{-(1,1,2,3,2,1)}+X_{-(1,1,1,1,1,1)}$.
Then we have
\begin{equation*}
[\mathfrak h, X]= \mathbb{C} X_{-(1,1,2,3,2,1)}\oplus \mathbb{C}
X_{-(1,1,1,1,1,1)}.
\end{equation*}
Further we get
\begin{alignat*}{2}
&[X_{(0,0,0,1)_{C_4}},X]=X_{-(1,1,2,2,2,1)},\quad &
[X_{(0,0,1,1)_{C_4}},X]&=X_{-(1,1,2,2,1,1)}\\
& [X_{(0,0,2,1)_{C_4}},X]=X_{-(1,1,1,2,1,1)}.
\end{alignat*}
This implies that $[\mathfrak b, X]$ contains five linearly independent vectors
of $N$ $\Rightarrow [\mathfrak b, X]=N$. Hence $N$ is $L$-spherical.
\underline{Case $G/P_6$:} The $H$-sphericity of $Y=G/P_6$ is an immediate
corollary of the following theorem which states that
$\mathbb{C}[\widehat Y]$ is multiplicity free.
\end{proof}
From the spherical cases we can derive the following branching rules:
\begin{thm}
Let $G$ be the simply connected simple algebraic group of type $E_6$ and $H$ be
the subgroup of type $C_4$.
Then we have the following branching rules:
\begin{equation*}
\begin{array}{r r c c l}
\text{i)} &\textnormal{res}^G_H (V_{k \omega_1})&\!\!\!=\!\!\!
&\displaystyle\bigoplus\limits_{a_1+2a_2+2a_3 = k} & V_{a_1 \lambda_2 +a_2
\lambda_4},\\
\text{ii)} &\textnormal{res}^G_H (V_{k \omega_6}) &\!\!\!=\!\!\!
&\displaystyle\bigoplus\limits_{a_1+2a_2+2a_3 = k} & V_{a_1 \lambda_2 +a_2
\lambda_4}.
\end{array}
\end{equation*}
\end{thm}
\begin{proof}\mbox{}\\
\underline{ii)} Here we are in the case $Y=G/P_1$. With ``LiE'' we compute
\begin{align*}
\textnormal{res}^G_H (V_{\omega_6}) = V_{\lambda_2} \quad \text{and}\quad
\textnormal{res}^G_H (V_{2\omega_6}) = \mathbb{C} \oplus V_{2\lambda_2} \oplus V_{\lambda_4}.
\end{align*}
So there is one generator of degree 1 and two of degree 2 in $\mathbb{C}[\widehat
Y]^{U_H}$.
From the calculations in the proof of the previous
theorem we know that $X_{-(1,1,2,3,2,1)}+ X_{-(1,1,1,1,1,1)}$ is an element of
$N$ whose $U_L$-orbit is of codimension~2. Hence $\dim \mathbb{C}[\widehat Y]^{U_H}
\leq 3$. But since we have already found three generators we know that $\dim
\mathbb{C}[\widehat Y]^{U_H} = 3$. The branching rules follow immediately.
\underline{i)} Theses branching rules follow directly from ii) by noting
that $\omega_1=\omega_6^*$, $\lambda_2^*=\lambda_2$ and $\lambda_4^*=\lambda_4$.
\end{proof}
Next we will consider the Levi subgroup $H$ of $G$ that is obtained by omitting
the simple root $\alpha_1$. From the Dynkin diagram of $E_6$ we see that $H$
is the group $D_5 \times \mathbb{C}^*$.
\begin{center}
\begin{tikzpicture}[vertex2/.style={circle,fill,thick, inner sep=0pt,minimum
size=5pt}, node distance=3.5mm and 7 mm]
{[start chain, node
distance=3.5mm and 7 mm,vertex/.style={circle,fill,thick, inner sep=0pt,minimum
size=5pt}]
\node[vertex,on chain,join,label= below: $1$] (1) {};
\node[vertex,on chain,join,label= below: $3$] (3) {};
\node[vertex,on chain,join,label= below: $4$] (4) {};
{[start branch=4]
\node[vertex,on chain=going above,join,label=left: $2$] (2) {};
}
\node[vertex,on chain,join,label= below: $5$] (5) {};
\node[vertex,on chain,join,label= below: $6$] (6) {};
}
\end{tikzpicture}
\end{center}
\begin{thm}
Let $G$ be the simply connected simple algebraic group of type
$E_6$ and let $H$ be the Levi subgroup $D_5 \times \mathbb{C}^*$.\\
Then $G/P_i$ is a spherical $H$-variety for $i \ne 4$. The variety $G/P_4$ is
not $H$-spherical.
\end{thm}
\begin{proof}
This is proven in \cite{Lit94}.
\end{proof}
\begin{thm}
Let $G$ be the simply connected simple algebraic groups of type $E_6$ let
$H\subset G$ be the Levi subgroup $D_5 \times \mathbb{C}^*$. Then we have the following
branching rules.
{\allowdisplaybreaks
\begin{equation*}
\begin{array}{r r c c l}
\text{i)} &\textnormal{res}^G_H (V_{k \omega_1})&\!\!\!=\!\!\!
&\displaystyle\bigoplus\limits_{a_1+a_2 +a_3= k} &
V_{a_1 \lambda_1 + a_2 \lambda_4} \otimes V_{-2a_1+a_2+4a_3},\\
\text{ii)} &\textnormal{res}^G_H (V_{k \omega_2})&\!\!\!=\!\!\!
&\displaystyle\bigoplus\limits_{a_1+a_2+a_3 + a_4 = k} &
V_{a_1 \lambda_2 + a_2 \lambda_4 + a_3 \lambda_5} \otimes V_{-3a_2+3a_3},\\
\text{iii)} &\textnormal{res}^G_H (V_{k \omega_3})&\!\!\!=\!\!\!
&\displaystyle\bigoplus\limits_{a_1+\ldots+a_6 = k} &
\begin{minipage}{6.5cm}
$V_{(a_1+a_6) \lambda_1 +a_2 \lambda_2+ a_3 \lambda_3 + (a_4+a_6) \lambda_4 +
a_5 \lambda_5} \otimes$\\
$\phantom{VV \otimes}V_{2a_1-4a_2+2a_3+5a_4 - a_5-3a_6}$
\end{minipage}
,\\
\text{iv)} &\textnormal{res}^G_H (V_{k\omega_5})&\!\!\!=\!\!\! &
\displaystyle\bigoplus\limits_{a_1+\ldots+a_6 = k} &
\begin{minipage}{6.5cm}
$V_{(a_1+a_6) \lambda_1 +a_2 \lambda_2+ a_3 \lambda_3 + a_4+\lambda_4 +
(a_5+a_6) \lambda_5} \otimes$\\
$\phantom{VV \otimes}V_{-2a_1+4a_2-2a_3+a_4-5a_5+a_6}$
\end{minipage}
,\\
\text{v)} &\textnormal{res}^G_H (V_{k \omega_6}) &\!\!\!=\!\!\!
&\displaystyle\bigoplus\limits_{a_1+a_2 +a_3 = k} &
V_{a_1 \lambda_1 + a_2 \lambda_5}\otimes V_{2a_1-a_2-4a_3}.
\end{array}
\end{equation*}
}
\end{thm}
\begin{proof} From paragraph 1.4 in \cite{Lit94} we get the following branching
rules.
{\allowdisplaybreaks
\begin{equation*}
\begin{array}{r r c c l}
\text{i)} &\textnormal{res}^G_H (V_{k \omega_1})&\!\!\!=\!\!\!
&\displaystyle\bigoplus\limits_{a_1+a_2 +a_3= k} & V_{(a_3-a_1-a_2) \omega_1
+a_2 \omega_3 + a_1 \omega_6},\\
\text{ii)} &\textnormal{res}^G_H (V_{k \omega_2})&\!\!\!=\!\!\!
&\displaystyle\bigoplus\limits_{a_1+a_2+a_3 + a_4 = k} & V_{-(a_1+2a_2)
\omega_1 +a_3 \omega_2 + a_2 \omega_3 + a_1 \omega_5},\\
\text{iii)} &\textnormal{res}^G_H (V_{k \omega_3})&\!\!\!=\!\!\!
&\displaystyle\bigoplus\limits_{a_1+\ldots+a_6 = k} &
V_{\substack{-(2a_2+a_3+a_5+2a_6) \omega_1 +a_5 \omega_2+ (a_4+a_6) \omega_3\\ +
a_3\omega_4 + a_2 \omega_5+(a_1+a_6)\omega_6}},\\
\text{iv)} &\textnormal{res}^G_H (V_{k
\omega_5})&\!\!\!=\!\!\! &\displaystyle\bigoplus\limits_{a_1+\ldots+a_6 = k} &
V_{\substack{-(a_1+2a_3+a_4+2a_5+a_6) \omega_1 +(a_5+a_6)\omega_2+ a_4
\omega_3\\ +a_3 \omega_4 + a_2 \omega_5 +(a_1+a_6) \omega_6}},\\
\text{v)} &\textnormal{res}^G_H (V_{k \omega_6}) &\!\!\!=\!\!\!
&\displaystyle\bigoplus\limits_{a_1+a_2 +a_3 = k} & V_{-(a_2+a_3) \omega_1 +a_2
\omega_2 + a_1 \omega_6}.
\end{array}
\end{equation*}
}
We would like to write these highest weights in terms of the fundamental weights
of $D_5$ and $\mathbb{C}^*$. We have $\omega_6= \lambda_1$, $\omega_5=\lambda_2$,
$\omega_4= \lambda_3$, $\omega_3=\lambda_4$ and $\omega_2=\lambda_5$, where $\lambda_i$ are the
fundamental weights of $D_5$ and we fix the coweight
$3\omega_1^\vee=4\alpha_1^\vee + 3\alpha_2^\vee + 5\alpha_3^\vee + 6
\alpha_4^\vee + 4 \alpha_5^\vee+ 2 \alpha_6^\vee$ which determines the highest
weights for $\mathbb{C}^*$. Thus we get the branching rules in the theorem.
\end{proof}
\section{\texorpdfstring{The exceptional group of type $E_7$}{The exceptional
group of type E7}}
Let $G$ be of type $E_7$ with the following
Dynkin-Diagram.
\begin{center}
\begin{tikzpicture}[vertex2/.style={circle,fill,thick, inner sep=0pt,minimum
size=5pt}, node distance=3.5mm and 7 mm]
{[start chain, node
distance=3.5mm and 7 mm,vertex/.style={circle,fill,thick, inner sep=0pt,minimum
size=5pt}]
\node[vertex,on chain,join,label= below: $1$] (1) {};
\node[vertex,on chain,join,label= below: $3$] (3) {};
\node[vertex,on chain,join,label= below: $4$] (4) {};
{[start branch=4]
\node[vertex,on chain=going above,join,label=left: $2$] (2) {};
}
\node[vertex,on chain,join,label= below: $5$] (5) {};
\node[vertex,on chain,join,label= below: $6$] (6) {};
\node[vertex,on chain,join,label= below: $7$] (7) {};
}
\end{tikzpicture}
\end{center}
For this group there are only a few cases of sphericity as we will see. As we
did in the last section we start by calculating the dimensions of the
Borel subgroups of the maximal reductive subgroups as well as the dimensions of
$G/P_i$ for $i=1,\ldots,7$.
We have
\begin{equation*}
\begin{array}{l | c c c c c c c}
& G/P_1 & G/P_2 & G/P_3 & G/P_4 & G/P_5 & G/P_6 & G/P_7\\\hline
\dim & 33 & 42 & 47& 53 & 50 & 42 & 27
\end{array}.
\end{equation*}
For the Borel subgroups $B_H$ we have:
\begin{equation*}
\begin{split}
&\begin{array}{l | c c c c c c}
H & A_7 & E_6\!\times \! \mathbb{C}^* & A_3 \times A_3 \times A_1 & A_5\times A_2 &
D_6\times A_1 & A_1 \times A_1\\\hline
\dim B_H & 35 & 43 & 20 & 25 & 38 &
\end{array}\\
&\begin{array}{l| c c c c c}
H& A_1 \times G_2 & G_2 \times C_3 & A_1 \times F_4 & A_1 & A_2\\\hline
\dim B_H & 10 & 20 & 30 & 2 &5
\end{array}
\end{split}
\end{equation*}
So we can rule out a lot of cases by dimension comparison.
\begin{prp}
Let $G$ be the simply connected simple algebraic group of type $E_7$. If $H$ is
a maximal reductive subgroup of type $A_3 \times A_3 \times A_1$, $A_5\times
A_2$, $A_1\times A_1$, $A_1\times G_2$, $G_2\times C_3$, $A_1$ or $A_2$, then
$G/P_i$ is not a spherical $H$-variety for $i=1,\ldots,7$. \qed
\end{prp}
\begin{proof}
In these cases we have $\dim B_H < \dim G/P_i$ for $i=1,\ldots,7$.
\end{proof}
Now we turn to the remaining subgroups and start with the subgroup of type
$A_7$. This is a subsystem subgroup so we add the smallest root $\delta$
to the simple roots and consider the extended Dynkin diagram.
\begin{center}
\begin{tikzpicture}[vertex2/.style={circle,fill,thick, inner sep=0pt,minimum
size=5pt}, node distance=3.5mm and 7 mm]
{[start chain, node
distance=3.5mm and 7 mm,vertex/.style={circle,fill,thick, inner sep=0pt,minimum
size=5pt}]
\node[vertex,on chain,join,label= below: $\delta$] (delta) {};
\node[vertex,on chain,join,label= below: $1$] (1) {};
\node[vertex,on chain,join,label= below: $3$] (3) {};
\node[vertex,on chain,join,label= below: $4$] (4) {};
{[start branch=4]
\node[vertex,on chain=going above,join,label=left: $2$] (2) {};
}
\node[vertex,on chain,join,label= below: $5$] (5) {};
\node[vertex,on chain,join,label= below: $6$] (6) {};
\node[vertex,on chain,join,label= below: $7$] (7) {};
}
\end{tikzpicture}
\end{center}
By omitting the simple root $\alpha_2$ we obtain the embedding of the root
system $A_7$ into $E_7$. Explicitly we get
{\allowdisplaybreaks
\begin{alignat*}{2}
(1,0,0,0,0,0,0)_{A_7}\!&=(1,0,0,0,0,0,0),\;&
(0,1,0,0,0,0,0)_{A_7}\!&=(0,0,1,0,0,0,0),\\
(0,0,1,0,0,0,0)_{A_7}\!&=(0,0,0,1,0,0,0),&
(0,0,0,1,0,0,0)_{A_7}\!&=(0,0,0,0,1,0,0),\\
(0,0,0,0,1,0,0)_{A_7}\!&=(0,0,0,0,0,1,0),&
(0,0,0,0,0,1,0)_{A_7}\!&=(0,0,0,0,0,0,1),\\
(0,0,0,0,0,0,1)_{A_7}\!&=(1,2,2,3,2,1,0).&
\end{alignat*}
}
Now we consider the corresponding subsystem subgroup $H$.
\begin{thm}\label{thm:E7-A7}
Let $G$ be the simply connected simple algebraic group of type $E_7$ and $H$ the
maximal reductive subgroup of type $A_7$. Then $G/P_7$ is a spherical
$H$-variety whereas $G/P_i$ is not $H$-spherical for $i \ne 7$.
\end{thm}
\begin{proof}
By dimension comparison $G/P_i$ can only be spherical for $i=1$ or $i=7$.
We know that for $E_7$ we have $\omega_i^*=\omega_i$. And with LiE we compute
\begin{equation*}
\text{res}^G_H(V_{4\omega_1})=\ldots \oplus 2 V_{\lambda_4}\oplus\ldots\,.
\end{equation*}
This shows that we have multiplicities in this case and $G/P_1$ is not a
spherical $H$-variety.
For $G/P_7$ we use the same methods as above. We compute
\begin{equation*}
\begin{split}
N
=
&\mathbb{C} X_{-(0,1,0,1,1,1,1)} \oplus \mathbb{C} X_{-(0,1,1,1,1,1,1)} \oplus
\mathbb{C} X_{-(1,1,1,1,1,1,1)} \oplus\\
&\mathbb{C} X_{-(0,1,1,2,1,1,1)} \oplus \mathbb{C} X_{-(1,1,1,2,1,1,1)} \oplus \mathbb{C}
X_{-(0,1,1,2,2,1,1)} \oplus\\
&\mathbb{C} X_{-(1,1,2,2,1,1,1)} \oplus \mathbb{C} X_{-(1,1,1,2,2,1,1)} \oplus \mathbb{C}
X_{-(0,1,1,2,2,2,1)} \oplus\\
&\mathbb{C} X_{-(1,1,2,2,2,1,1)} \oplus \mathbb{C} X_{-(1,1,1,2,2,2,1)} \oplus \mathbb{C}
X_{-(1,1,2,3,2,1,1)} \oplus\\
&\mathbb{C} X_{-(1,1,2,2,2,2,1)} \oplus \mathbb{C} X_{-(1,1,2,3,2,2,1)} \oplus \mathbb{C}
X_{-(1,1,2,3,3,2,1)}.
\end{split}
\end{equation*}
Define $X:= X_{-(1,1,2,3,3,2,1)} + X_{-(1,1,2,2,1,1,1)}+ X_{-(0,1,0,1,1,1,1)}$.
The roots of the root-vectors in $X$ are linearly independent. Thus we get that
\begin{equation*}
[\mathfrak h, X] := \langle X_{-(1,1,2,3,3,2,1)_{E_7}}, X_{-(1,1,2,2,1,1,1)_{E_7}},
X_{-(0,1,0,1,1,1,1)_{E_7}}\rangle
\end{equation*}
and further
\begin{alignat*}{2}
[X_{(0,0,1,0,0,0,0)},X]&= X_{-(1,1,1,2,1,1,1)}, & \quad
[X_{(0,0,0,0,1,0,0)},X]&= X_{-(1,1,2,3,2,2,1)},\\
[X_{(1,0,1,0,0,0,0)},X]&= X_{-(0,1,1,2,1,1,1)},&
[X_{(0,0,1,1,0,0,0)},X]&= X_{-(1,1,1,1,1,1,1)},\\
[X_{(0,0,0,1,1,0,0)},X]&= X_{-(1,1,2,2,2,2,1)},&
[X_{(0,0,0,0,1,1,0)},X]&= X_{-(1,1,2,3,2,1,1)},\\
[X_{(1,0,1,1,0,0,0)},X]&= X_{-(0,1,1,1,1,1,1)},&
[X_{(0,0,1,1,1,0,0)},X]&= X_{-(1,1,1,2,2,2,1)},\\
[X_{(0,0,0,1,1,1,0)},X]&= X_{-(1,1,2,2,2,1,1)},&
[X_{(1,0,1,1,1,0,0)},X]&= X_{-(0,1,1,2,2,2,1)},\\
[X_{(0,0,1,1,1,1,0)},X]&= X_{-(1,1,1,2,2,1,1)},&
[X_{(1,0,1,1,1,1,0)},X]&= X_{-(0,1,1,2,2,1,1)}.
\end{alignat*}
This shows that $\dim [\mathfrak b,X]=15 = \dim N \Rightarrow $
$[\mathfrak b,X]=N$ $\Rightarrow$ $N$ is a spherical $L$-module.
And thus $G/P_7$ is a spherical $H$-variety.
\end{proof}
Since $G/P_7$ is a spherical $H$-variety we can derive branching rules for
$V_{k\omega^*_7}=V_{k\omega_7}$.
\begin{thm}
Let $G$ be the simply connected simple algebraic group of type $E_7$ and $H$
the maximal reductive subgroup of type $A_7$. Then
\begin{equation*}
\textnormal{res}^G_H(V_{k\omega_7})= \bigoplus_{2a_1+a_2+2a_3+a_4=k} V_{a_2\lambda_2 + a_3
\lambda_4 + a_4 \lambda_6}.
\end{equation*}
\end{thm}
\begin{proof}
With ``LiE'' we compute
\begin{equation*}
\textnormal{res}^G_H (V_{\omega_7})= V_{\lambda_2} \oplus V_{\lambda_6}.
\end{equation*}
So there are two generators of degree 1 of weight $\lambda_2$ and
$\lambda_6$. Further we have
\begin{equation*}
\textnormal{res}^G_H (V_{2\omega_7})=\mathbb{C} \oplus V_{2\lambda_2} \oplus V_{2\lambda_6} \oplus
V_{\lambda_2 + \lambda_6} \oplus V_{\lambda_4},
\end{equation*}
which shows that there are 2 generators of degree 2 which are of weight $0$ and
$\lambda_4$. This shows that $\dim \mathbb{C} [\widehat Y]^{U_H}\geq 4$.
In the proof of the previous theorem we have found an $X \in N$
such that $U_L.X$ is of codimension~3. It follows that $\dim \mathbb{C}
[\widehat Y]^{U_H}=4$ and we have found four generators. The branching rules
follow immediately.
\end{proof}
Next we will consider the Levi subgroup $E_6 \times \mathbb{C}^*$, which is
obtained by omitting the simple root $\alpha_7$ in the
Dynkin-diagram.
\begin{thm}
Let $G$ be the simply connected simple algebraic group of type $E_7$ and $H
\subset G$ the Levi subgroup of type $E_6 \times \mathbb{C}^*$. Then $G/P_1$ and
$G/P_7$ are spher\-ical $H$-varieties whereas $G/P_i$, $i=2,\ldots,6$ are not
spherical $H$-varieties.
\end{thm}
\begin{proof} This was proven in \cite{Lit94}.
\end{proof}
We get the following branching rules from the spherical cases.
\begin{thm}
Let $G$ be the simply connected simple algebraic group of type $E_7$ and $H$
the Levi subgroup of type $E_6\times \mathbb{C}^*$. Then we have the following branching
rules.
{\allowdisplaybreaks
\begin{equation*}
\begin{array}{r r c c l}
\text{i)} &\textnormal{res}^G_H (V_{k \omega_1})&\!\!\!=\!\!\!
&\displaystyle\bigoplus\limits_{a_1+ a_2 + a_3 +a_4 = k}
&V_{a_1 \lambda_1+ a_2 \lambda_2+ a_3 \lambda_6}\otimes V_{2a_1-2a_3},\\
\text{ii)} &\textnormal{res}^G_H (V_{k \omega_2})&\!\!\!=\!\!\!
&\displaystyle\bigoplus\limits_{\substack{a_1+a_2+a_3+2a_4+\\a_5+a_6+a_7 = k}} &
\begin{minipage}{6.2cm}
$V_{a_1 \lambda_1 +(a_2+a_7) \lambda_2 + a_3 \lambda_3 + a_4 \lambda_4+ a_5
\lambda_5 + a_6 \lambda_6} \otimes$\\
$\phantom{VV\otimes} V_{-a_1+3a_2+a_3-a_5-2a_6}$
\end{minipage}
,\\
\text{iii)} &\textnormal{res}^G_H (V_{k \omega_7})&\!\!\!=\!\!\!
&\displaystyle\bigoplus\limits_{a_1+a_2+a_3+a_4 = k} & V_{a_1 \lambda_1 + a_2
\lambda_6} \otimes V_{-a_1+a_2+3a_3-3a_4}.\\
\end{array}
\end{equation*}
}
\end{thm}
\begin{proof} From paragraph 1.4 in \cite{Lit94} we get the following branching
rules.
{\allowdisplaybreaks
\begin{equation*}
\begin{array}{r r c c l}
\text{i)} &\textnormal{res}^G_H (V_{k \omega_1})&\!\!\!=\!\!\!
&\displaystyle\bigoplus\limits_{a_1+ a_2 + a_3 +a_4 = k}
&V_{a_1 \omega_1+ a_2 \omega_2+ a_3 \omega_6 -(a_2+2a_3)\omega_7},\\
\text{ii)} &\textnormal{res}^G_H (V_{k \omega_2})&\!\!\!=\!\!\!
&\displaystyle\bigoplus\limits_{\substack{a_1+a_2+a_3+2a_4+\\a_5+a_6+a_7 = k}} &
V_{\substack{a_1 \omega_1 +(a_2+a_7) \omega_2 + a_3 \omega_3 + a_4 \omega_4+ a_5
\omega_5 + a_6 \omega_6\\ \quad -(a_1+a_3+2a_4+2a_5+2a_6+a_7)\omega_7}}\;\;,\\
\text{iii)} &\textnormal{res}^G_H (V_{k \omega_7})&\!\!\!=\!\!\!
&\displaystyle\bigoplus\limits_{a_1+a_2+a_3+a_4 = k} & V_{a_1 \omega_1 + a_2
\omega_6+(a_3-a_1-a_2-a_4)\omega_7}.\\
\end{array}
\end{equation*}
}
We have $\omega_i = \lambda_i$ for $i=1,\ldots,6$ and we fix the coweight
$2\omega_7^\vee=2\alpha_1^\vee + 3 \alpha_2^\vee + 4 \alpha_3^\vee + 6
\alpha_4^\vee + 5 \alpha_5^\vee + 4 \alpha_6^\vee + 3 \alpha_7^\vee$ which
determines the highest weights for $\mathbb{C}^*$. Thus we get the branching rules
in the theorem.
\end{proof}
Now we will turn to the subgroup of $E_7$ of type $D_6\times A_1$. We will
consider the extended Dynkin-diagram of $E_7$ again by adding the smallest root
$\delta$ to the simple roots.
\begin{center}
\begin{tikzpicture}[vertex2/.style={circle,fill,thick, inner sep=0pt,minimum
size=5pt}, node distance=3.5mm and 7 mm]
{[start chain, node
distance=3.5mm and 7 mm,vertex/.style={circle,fill,thick, inner sep=0pt,minimum
size=5pt}]
\node[vertex,on chain,join,label= below: $\delta$] (delta) {};
\node[vertex,on chain,join,label= below: $1$] (1) {};
\node[vertex,on chain,join,label= below: $3$] (3) {};
\node[vertex,on chain,join,label= below: $4$] (4) {};
{[start branch=4]
\node[vertex,on chain=going above,join,label=left: $2$] (2) {};
}
\node[vertex,on chain,join,label= below: $5$] (5) {};
\node[vertex,on chain,join,label= below: $6$] (6) {};
\node[vertex,on chain,join,label= below: $7$] (7) {};
}
\end{tikzpicture}
\end{center}
If we omit the simple root $\alpha_6$ we have a sub-diagram of type $D_6\times
A_1$ and consider the the corresponding subsystem subgroup.
Explicitly we can choose the following simple roots:
{\allowdisplaybreaks
\begin{alignat*}{2}
(1,0,0,0,0,0,0)_H&= (0,1,1,2,2,2,1),& \;\,
(0,1,0,0,0,0,0)_H&= (1,0,0,0,0,0,0),\\
(0,0,1,0,0,0,0)_H&= (0,0,1,0,0,0,0),&
(0,0,0,1,0,0,0)_H&= (0,0,0,1,0,0,0),\\
(0,0,0,0,1,0,0)_H&= (0,1,0,0,0,0,0),&
(0,0,0,0,0,1,0)_H&= (0,0,0,0,1,0,0),\\
(0,0,0,0,0,0,1)_H&= (0,0,0,0,0,0,1).
\end{alignat*}
}
\begin{thm}
Let $G$ be the simply connected simple algebraic group of type $E_7$. If $H$ is
the subgroup of type $D_6\times A_1$ then $G/P_7$ is a spherical $H$-variety and
$G/P_i$ is not a spherical $H$-variety for $i=1,\ldots,6$.
\end{thm}
\begin{proof}
Dimension comparison shows that $G/P_2,\ldots,G/P_6$ are not $H$-spher\-ical.
For $G/P_1$ we can compute the restriction of $V_{k\omega_1}$ (note that
$\omega^*_i=\omega_i$ for $E_7$) with LiE and get
\begin{equation*}
\text{res}^G_H(V_{4\omega_1})=\ldots\oplus 2(V_{2\lambda_6} \otimes
V_{2\lambda_7})\oplus\ldots\;\;.
\end{equation*}
Thus there are multiplicities in this case and we know that the $H$-variety
$G/P_1$ is not $H$-spherical.
\underline{Case $G/P_7$:} We compute
{\allowdisplaybreaks
\begin{align*}
\quad N
=
&\mathbb{C} X_{-(0,0,0,0,0,1,1)} \oplus \mathbb{C} X_{-(0,0,0,0,1,1,1)} \oplus \mathbb{C}
X_{-(0,0,0,1,1,1,1)}\oplus\\
&\mathbb{C} X_{-(0,1,0,1,1,1,1)} \oplus \mathbb{C} X_{-(0,0,1,1,1,1,1)} \oplus \mathbb{C}
X_{-(1,0,1,1,1,1,1)} \oplus\\
&\mathbb{C} X_{-(0,1,1,1,1,1,1)}\oplus \mathbb{C} X_{-(1,1,1,1,1,1,1)} \oplus \mathbb{C}
X_{-(0,1,1,2,1,1,1)} \oplus\\
&\mathbb{C} X_{-(1,1,1,2,1,1,1)} \oplus \mathbb{C} X_{-(0,1,1,2,2,1,1)}\oplus \mathbb{C}
X_{-(1,1,2,2,1,1,1)} \oplus \\
&\mathbb{C} X_{-(1,1,1,2,2,1,1)} \oplus \mathbb{C} X_{-(1,1,2,2,2,1,1)} \oplus \mathbb{C}
X_{-(1,1,2,3,2,1,1)} \oplus\\
&\mathbb{C} X_{-(1,2,2,3,2,1,1)}.
\end{align*}
}
Now define $X:= X_{-(1,2,2,3,2,1,1)}+ X_{-(1,0,1,1,1,1,1)}$. The roots of these
two root vectors are linearly independent and we have
\begin{equation*}
[\mathfrak h, X]=\langle X_{-(1,2,2,3,2,1,1)}, X_{-(1,0,1,1,1,1,1)}\rangle
\end{equation*}
Further we have
{\allowdisplaybreaks
\begin{alignat*}{2}
[X_{(1,0,0,0,0,0,0)},X]&= X_{-(0,0,1,1,1,1,1)}, \;\; &
[X_{(0,1,0,0,0,0,0)},X]&= X_{-(1,1,2,3,2,1,1)},\\
[X_{(1,0,1,0,0,0,0)},X]&= X_{-(0,0,0,1,1,1,1)},&
[X_{(0,1,0,1,0,0,0)},X]&= X_{-(1,1,2,2,2,1,1)},\\
[X_{(1,0,1,1,0,0,0)},X]&= X_{-(0,0,0,0,1,1,1)},&
[X_{(0,1,1,1,0,0,0)},X]&= X_{-(1,1,1,2,2,1,1)},\\
[X_{(0,1,0,1,1,0,0)},X]&= X_{-(1,1,2,2,1,1,1)},&
[X_{(1,1,1,1,0,0,0)},X]&= X_{-(0,1,1,2,2,1,1)},\\
[X_{(1,0,1,1,1,0,0)},X]&= X_{-(0,0,0,0,0,1,1)},&
[X_{(0,1,1,1,1,0,0)},X]&= X_{-(1,1,1,2,1,1,1)},\\
[X_{(1,1,1,1,1,0,0)},X]&= X_{-(0,1,1,2,1,1,1)},&
[X_{(0,1,1,2,1,0,0)},X]&= X_{-(1,1,1,1,1,1,1)},\\
[X_{(1,1,1,2,1,0,0)},X]&= X_{-(0,1,1,1,1,1,1)},&
[X_{(1,1,2,2,1,0,0)},X]&= X_{-(0,1,0,1,1,1,1)}.
\end{alignat*}
}
So we have $\dim [\mathfrak b,X]=16=\dim N$. This implies that $N$ is a
spherical $L$-module and thus $G/P_7$ is a spherical $H$-variety.
\end{proof}
From the sphericity of $G/P_7$ we can derive branching rules for $V_{k
\omega^*_7}= V_{k \omega_7}$.
\begin{thm}
Let $G$ be the simply connected simple algebraic group of type $E_7$ and let $H$
be a maximal reductive subgroup of type $D_6 \times A_1$.
Then
\begin{equation*}
\textnormal{res}^G_H (V_{k \omega_7}) = \bigoplus_{a_1+ 2 a_2 + a_3= k}
V_{a_1 \lambda_1 + a_2 \lambda_2 + a_3 \lambda_6} \otimes V_{a_1 \lambda_7}.
\end{equation*}
\end{thm}
\begin{proof}
With ``LiE'' we compute
\begin{equation*}
\textnormal{res}^G_H (V_{\omega_7})= (V_{\lambda_1} \otimes V_{\lambda_7}) \oplus
(V_{\lambda_6} \otimes \mathbb{C}).
\end{equation*}
So there are two generators of degree 1 with weights $(\lambda_1, \lambda_7)$
and $(\lambda_6,0)$.
Further we have
\begin{equation*}
\textnormal{res}^G_H (V_{2\omega_7})= (V_{2\lambda_1}\otimes V_{2\lambda_7}) \oplus
(V_{\lambda_1 + \lambda_6}\otimes V_{\lambda_7}) \oplus (V_{2\lambda_6}\otimes
\mathbb{C}) \oplus (V_{\lambda_2}\otimes \mathbb{C}).
\end{equation*}
Thus there is a further generator of degree 2 and weight $\lambda_2$ and we know
that $\dim \mathbb{C} [\widehat Y]^{U_H}\geq 3$.
In the proof of the previous theorem we have seen that there is an $X \in N$
such that $\dim U_H.X$ is of codimension 2.
It follows that $\dim \mathbb{C} [\widehat Y]^{U_H}= 3$. The branching rules follow.
\end{proof}
The last maximal reductive subgroup of $G$ where a sphericity of $G/P_i$ can
occur is the group $H$ of type $A_1 \times F_4$. From the table with dimensions of
$G/P_i$ we know that only $G/P_7$ can be a spherical $H$-variety. But with LiE
we compute
\begin{equation*}
\text{res}^G_H(V_{4\omega_7})= \ldots \oplus 2 (V_{4\lambda_1}\otimes
V_{\lambda_5})
\oplus
\ldots
\end{equation*}
and thus there are multiplicities in this case. We have shown:
\begin{thm}
Let $G$ be the simply connected simple group of type $E_7$ and $H$ the maximal
subgroup of type $A_1\times F_4$.
Then $G/P_i$ ($i=1,\ldots,7$) is not a spherical variety.\qed
\end{thm}
\section{\texorpdfstring{The exceptional group of type $E_8$}{The exceptional
group of type E8}}
We start our computations again by calculating the dimensions of the Borel
subgroups of the maximal reductive subgroups and the dimensions of $G/P_i$ for
$i=1,\ldots,8$.
\begin{equation*}
\begin{split}
&\begin{array}{l| c c c c c c c c c}
H
&E_7\!\!\times\!\! A_1
&E_6 \!\! \times\!\! A_2
&A_3 \!\! \times\!\! D_5
&A_4 \!\! \times\!\! A_4
&A_5 \!\! \times\!\! A_2 \!\! \times\!\! A_1\\ \hline
\dim B_H
&72
&47
&34
&28
&27
\end{array}\\
&
\begin{array}{l| ccccccc}
H
&A_7 \!\! \times \!\! A_1
& D_8
& A_8
& G_2 \!\! \times \!\! F_4
& A_2 \!\! \times \!\! A_1
& C_2
& A_1\\ \hline
\dim B_H
& 37
& 72
& 44
& 36
& 6
& 6
& 2
\end{array}
\end{split}
\end{equation*}
The dimensions of the varieties $G/P_i$ ($i=1,\ldots,8$) are:
\begin{equation*}
\begin{array}{l|c c c c c c c c}
&G/P_1
&G/P_2
&G/P_3
&G/P_4
&G/P_5
&G/P_6
&G/P_7
&G/P_8\\\hline
\dim
& 78
& 92
& 98
& 106
& 104
& 97
& 83
& 57
\end{array}
\end{equation*}
By dimension comparison there are only two possibilities of sphericity. If we
take the maximal reductive subgroup $H_1$ of type $E_7 \times A_1$ or the
maximal reductive subgroup $H_2$ of type $D_8$, then the variety $G/P_8$ can be
spherical for $H_1$ or $H_2$. But we can compute the following restrictions by using LiE
\begin{equation*}
\begin{split}
\text{res}^G_{H_1}(V_{5\omega_8})&= \ldots \oplus 2 (V_{1\lambda_1 +
2\lambda_7} \otimes V_{2 \lambda_8}) \oplus \ldots,\\
\text{res}^G_{H_2}(V_{4\omega_8})&=\ldots \oplus 2 V_{\lambda_8} \oplus
\ldots,\\
\end{split}
\end{equation*}
which show that there are multiplicities in these cases. So there are no
spherical cases for~$G$. We have shown:
\begin{thm}
Let $G$ be the simply connected simple algebraic groups of type~$E_8$. Let $H$
be one of its maximal reductive subgroups.
Then $G/P_i$ ($i=1,\ldots,8$) is not a spherical variety.\qed
\end{thm}
|
\section*{ACKNOWLEDGMENTS}
This work was supported by the National Science Foundation under CAREER Grant No. DMR-0846426.
|
\section{Introduction}
The diversity-multiplexing trade-off (DMT) introduced by \cite{zheng2003diversity} studies the diversity function of the multiplexing gain in the high SNR regime.
\cite{kumar2009asymptotic} showed that the MMSE linear receivers, widely used for their simplicity, exhibit a largely suboptimal DMT in flat fading MIMO channels.
Nonetheless, for a finite data rate (i.e. when the rate does not increase with the signal to noise ratio), the MMSE receivers take several diversity values, depending on the aimed rate, as noticed earlier in \cite{hedayat2005linear}, and also in \cite{hedayat2004outage,tajer2007diversity} for frequency-selective MIMO channels.
In particular they achieve full diversity for sufficiently low data rates, hence their great interest.
This behavior was partially explained in \cite{kumar2009asymptotic,mehana2010diversity} for flat fading MIMO channels and in \cite{mehana2011diversity} for frequency-selective MIMO channels.
Indeed the proof of the upper bound on the diversity order for the flat fading case given in \cite{mehana2010diversity} contains a gap, and
the approach of \cite{mehana2010diversity} based on the Specht bound seems to be unsuccessfull.
As for MIMO frequency selective channels with cyclic prefix, \cite{mehana2011diversity} only derives the diversity in the particular case of a number of channel taps equal to the transmission data block length, and claims that this value provides an upper bound in more realistic cases, whose expression is however not explicitly given.
In this paper we provide a rigorous proof of the diversity for MMSE receivers in flat fading MIMO channels for finite data rates.
We also derive the diversity in MIMO frequency selective channels with cyclic prefix for finite data rates if the transmission data block length is large enough.
Simulations corroborate our derived diversity in the frequency selective channels case.
\section{Problem statement}
We consider a MIMO system with $M$ transmitting, $N \geq M$ receiving antennas,
with coding and ideal interleaving at the transmitter, and with
a MMSE linear equalizer at the receiver, followed by a de-interleaver and a decoder (see Fig. \ref{fig:scheme}).
We evaluate in the following sections the achieved diversity by studying the outage probability, that is the probability that the capacity does not support the target data rate, at high SNR regimes.
We denote $\rho$ the SNR, $I$ the capacity and $R$ the target data rate.
We use the notation $\doteq$ for {\it exponential equality} \cite{zheng2003diversity}, i.e.
\begin{equation}
f(\rho) \doteq \rho^d
\Leftrightarrow
\lim_{\rho \to \infty} \frac{\log f(\rho)}{\log \rho}= d,
\label{eq:exp-equ}
\end{equation}
and the notations $\dot\leq$ and $\dot\geq$ for exponential inequalities, which are similarly defined.
We note $\log$ the logarithm to base $2$.
\begin{figure*}[t]
\centering
\includegraphics[width=6.5in]{div_sch3}
\caption{Considered MIMO system}
\label{fig:scheme}
\end{figure*}
\section{Flat fading MIMO channels}
In this section we consider a flat fading MIMO channel.
The output of the MIMO channel is given by
\begin{equation}
\y= \sqrt{\frac{\rho}{M}} \H \x + \n,
\end{equation}
where $\n \sim \mathcal{CN}(\bs{0},\I_N)$ is the additive white Gaussian noise and
$\x$ the channel input vector,
$\H$ the $N \times M$ channel matrix with i.i.d. entries $\sim \mathcal{CN}(0,1)$.
\medskip
\begin{theorem}
For a rate $R$ such that
$\log \frac{M}{m} < \frac{R}{M} < \log \frac{M}{m-1}$,
with $m \in \{ 1, \ldots, M \}$,
the outage probability verifies
\begin{equation}
\PP(I<R) \doteq \rho^{-m(N-M+m)},
\end{equation}
that is, a diversity of $m(N-M+m)$.
\end{theorem}
\medskip
Note that for a rate $R < M \log \frac{M}{M-1}$ (i.e. $m=M$) full diversity $MN$ is attained, while for a rate $R > M \log M$ the diversity corresponds to the one derived by DMT approach.
This result was stated by \cite{mehana2010diversity}.
Nevertheless the proof of the outage lower bound in \cite{mehana2010diversity} omits that the event noted $\mathcal{B}_a$ is not independent from the eigenvalues of $\H^H\H$, hence questioning the validity of the given proof.
We thus provide an alternative proof based on an approach suggested by the analysis of \cite{kumar2009asymptotic} in the case
where $R = r \log \rho$ with $r > 0$.
\medskip
\begin{proof}
The capacity $I$ of the MIMO MMSE considered system is given by
\[
I = \sum_{j=1}^M \log ( 1 + \beta_j),
\]
where $\beta_j$ is the SINR for the $j$th stream:
\[
\beta_j= \frac{1}{\left( \left[ \I + \frac{\rho}{M} \H^*\H \right]^{-1} \right)_{jj} } - 1.
\]
We lower bound in the first place $\PP(I<R)$ and prove in the second place that the bound is tight by upper bounding $\PP(I<R)$ with the same bound.
\subsection{Lower bound of the outage probability}
\label{sec:lowB_flat}
We here assume that $R/M>\log (M/m$).
In order to lower bound $\PP(I<R)$ we need to upper bound the capacity $I$. Using Jensen's inequality on function $x \mapsto \log x$ yields
\begin{align}
I
&\leq M \log \Bigg[ \frac{1}{M} \sum_{j=1}^M \left( 1+\beta_j \right) \Bigg] \label{ineq:jensen1}
\\
&= M \log \Bigg[ \frac{1}{M} \sum_{j=1}^M \bigg( \left[ \left( \I + \frac{\rho}{M} \H^*\H \right)^{-1} \right]_{jj} \bigg)^{-1} \Bigg]. \label{ineq:logconcave}
\end{align}
We note $\H^*\H= \U^*\Lambda\U$ the SVD of $\H^*\H$ with $\Lambda=\mathrm{diag}(\lambda_1,\ldots,\lambda_M)$, $\lambda_1 \leq \lambda_2 \ldots \leq \lambda_M$.
We recall that the $(\lambda_k)_{k=1, \ldots, M}$ are independent from the entries of matrix $\U$ and that $\U$ is a Haar distributed unitary random matrix, i.e. the probability distribution of $\U$ is invariant by left (or right) multiplication by deterministic matrices.
Using this SVD we can write
\begin{equation}
\frac{1}{M} \sum_{j=1}^M \bigg( \left[ \left( \I + \frac{\rho}{M} \H^*\H \right)^{-1} \right]_{jj} \bigg)^{-1}
= \frac{1}{M} \sum_{j=1}^M \bigg( \sum_{k=1}^M \frac{|\U_{kj}|^2}{1+ \frac{\rho}{M} \lambda_k } \bigg)^{-1}. \label{eq:sumSINR}
\end{equation}
\subsubsection{\texorpdfstring{Case $m=1$}{Case m=1}}
In order to better understand the outage probability behavior, we first consider the case $m=1$.
In this case $R/M>\log M$.
We review the approach of \cite[III]{kumar2009asymptotic},
which consists in upper bounding \eqref{eq:sumSINR} by
$\left( 1+ \frac{\rho}{M} \lambda_1 \right) \frac{1}{M} \sum_{j=1}^M \frac{1}{|\U_{1j}|^2}$,
as
$\sum_{k=1}^M \frac{|\U_{kj}|^2}{1+ \frac{\rho}{M} \lambda_k } \geq \frac{|\U_{1j}|^2}{1+\frac{\rho}{M}\lambda_1}$.
Using this bound in \eqref{ineq:logconcave} gives
\[
I \leq M \log \bigg[ \left( 1 + \frac{\rho}{M} \lambda_1 \right) \frac{1}{M} \sum_{j=1}^M \frac{1}{|\U_{1j}|^2} \bigg].
\]
Therefore
\[
\Big( \left( 1 + \frac{\rho}{M} \lambda_1 \right) \frac{1}{M} \sum_{j=1}^M \frac{1}{|\U_{1j}|^2} < 2^{R/M} \Big) \subset (I<R).
\]
In order to lower bound $\PP(I<R)$, \cite{kumar2009asymptotic} introduced the set
\[
\mathcal{A}_1 = \bigg\{ \frac{1}{M} \sum_{j=1}^M \frac{1}{|\U_{1j}|^2} < M + \eps \bigg\}
\]
for $\eps > 0$.
Then,
\begin{align*}
\PP(I <R)
& \geq \PP\left( (I<R) \cap \mathcal{A}_1 \right)
\\
& \geq \PP \bigg[ \bigg( \left( 1 + \frac{\rho}{M} \lambda_1 \right) \frac{1}{M} \sum_{j=1}^M \frac{1}{|\U_{1j}|^2} < 2^{R/M} \bigg) \cap \mathcal{A}_1\bigg]
\\
& \geq \PP \left[ \left( 1 + \frac{\rho}{M} \lambda_1 < \frac{2^{R/M}}{M+\eps} \right) \cap \mathcal{A}_1\right]
\\
& = \PP(\mathcal{A}_1) \cdot \PP\left[ 1 + \frac{\rho}{M} \lambda_1 < \frac{2^{R/M}}{M+\eps} \right],
\end{align*}
where the last equality comes from the independence between eigenvectors and eigenvalues of Gaussian matrix $\H^*\H$.
It is shown in \cite[Appendix A]{kumar2009asymptotic} that
$\PP(\mathcal{A}_1) \neq 0$.
Besides, as we supposed
$2^{R/M}>M$,
we can take $\eps$ such that
$\frac{2^{R/M}}{M+\eps}>1$, ensuring that $\PP\Big[ \left( 1 + \frac{\rho}{M} \lambda_1 \right) < \frac{2^{R/M}}{M+\eps} \Big] \neq 0$.
Hence there exists $\kappa>0$ such that
\[
\PP(I<R) \ \dot\geq \ \PP\left( \lambda_1 < \frac{\kappa}{\rho} \right),
\]
which is asymptotically equivalent to $\rho^{-(N-M+1)}$ in the sense of \eqref{eq:exp-equ} (see, e.g., \cite[Th. II.3]{jiang2011performance}).
\subsubsection{\texorpdfstring{General case $1 \leq m\leq M$}{General case 1<=m<= M}}
By the same token as for $m=1$ we now consider the general case --
we recall that we assumed that $\log (M/m) < R/M$.
We first lower bound $\sum_k\frac{|\U_{kj}|^2}{1+ \frac{\rho}{M} \lambda_k}$ which appears in \eqref{eq:sumSINR} by the $m$ first terms of the sum and then use Jensen's inequality applied on $x \mapsto x^{-1}$, yielding
\begin{align*}
\sum_{k=1}^M \frac{|\U_{kj}|^2}{1+ \frac{\rho}{M} \lambda_k}
&\geq \sum_{k=1}^m \frac{|\U_{kj}|^2}{1+ \frac{\rho}{M} \lambda_k}
\\
&\geq \frac{\left( \sum_{l=1}^m |\U_{lj}|^2 \right)^2}{\sum_{k=1}^m |\U_{kj}|^2 \left( 1+ \frac{\rho}{M} \lambda_k\right)}.
\end{align*}
Using this inequality in \eqref{eq:sumSINR}, we obtain that
\begin{align}
\frac{1}{M} \sum_{j=1}^M \bigg( \left[ \left( \I + \frac{\rho}{M} \H^*\H \right)^{-1} \right]_{jj} \bigg)^{-1}
\leq \frac{1}{M} & \sum_{j=1}^M \frac{\sum_{k=1}^m |\U_{kj}|^2\left( 1+ \frac{\rho}{M} \lambda_k\right)}{\left( \sum_{l=1}^m |\U_{lj}|^2 \right)^2} \notag
\\
& = \sum_{k=1}^m \left( 1+ \frac{\rho}{M} \lambda_k\right) \delta_k(\U), \label{ineq:intr_delta}
\end{align}
where
$\delta_k(\U) = \frac{1}{M} \sum_{j=1}^M \frac{|\U_{kj}|^2}{\left( \sum_{l=1}^m |\U_{lj}|^2 \right)^2}$.
Equation \eqref{ineq:intr_delta}, together with \eqref{ineq:logconcave}, yields the following inclusion:
\begin{align*}
\Bigg( \sum_{k=1}^m \delta_k(\U) \left( 1+ \frac{\rho}{M} \lambda_k\right) < 2^{R/M} \Bigg)
\subset (I<R).
\end{align*}
Similarly to the case $m=1$, we introduce the set $\mathcal{A}_m$ defined by
\[
\mathcal{A}_m = \left\{ \delta_k(\U) < \frac{M}{m^2} + \eps, \
k=1,\ldots,m \right\}
\]
for $\eps > 0$.
We now use this set to lower bound $\PP(I<R)$.
\begin{align*}
\PP(I<R) & \geq \PP\left( (I<R) \cap \mathcal{A}_m \right)
\\
& \geq \PP \bigg[ \Bigg( \sum_{k=1}^m \delta_k(\U) \left( 1+ \frac{\rho}{M} \lambda_k\right) < 2^{R/M} \Bigg) \cap \mathcal{A}_m\bigg]
\\
& \geq \PP \left[ \Bigg( \sum_{k=1}^m \left( 1+ \frac{\rho}{M} \lambda_k\right) < \frac{2^{R/M}}{\frac{M}{m^2} + \eps} \Bigg) \cap \mathcal{A}_m\right]
\\
&= \PP(\mathcal{A}_m) \cdot \PP\left[ \sum_{k=1}^m \left( 1 + \frac{\rho}{M} \lambda_k \right) < \frac{2^{R/M}}{\frac{M}{m^2}+\eps} \right].
\end{align*}
The independence between eigenvectors and eigenvalues of Gaussian matrix $\H^*\H$ justifies the last equality.
As we assumed that
$\log(M/m) < R/M$, that is $m < \frac{2^{R/M}}{M/m^2}$,
we can choose $\eps$ such that
$m < \frac{2^{R/M}}{M/m^2 +\eps}$.
That ensures that $\PP\left[ \sum_{k=1}^m \left( 1 + \frac{\rho}{M} \lambda_k \right) < \frac{2^{R/M}}{ M/m^2 +\eps} \right] \neq 0$.
We show in Appendix \ref{apx:prob_sum_first_ev} that this probability is asymptotically equivalent to $\rho^{-m(N-M+m)}$ in the sense of \eqref{eq:exp-equ}, leading to
\begin{equation}
\PP(I<R) \ \dot\geq \ \frac{\PP(\mathcal{A}_m)}{\rho^{m(N-M+m)}}.
\label{ineq:P_I<R}
\end{equation}
We still need to prove that $\PP(\mathcal{A}_m) \neq 0$.
Any Haar distributed random unitary matrix can be parameterized by $M^2$ independent angular random variables
$(\alpha_1, \ldots, \alpha_{M^2})=\bs\alpha$ whose probability distributions are almost surely positive (see \cite{dita2003factorization, lundberg2004haar} and Appendix \ref{apx:ang_par}).
We note $\Phi_m$ the functions such that $\U=\Phi_m(\bs\alpha)$.
Consider a deterministic unitary matrix $\U_*$ such that $|(\U_*)_{ij}|^2 = \frac{1}{M} \ \forall i,j$,
and denote by $\bs\alpha_*$ a corresponding $M^2$ dimensional vector.
It is straightforward to check that $\delta_k \circ \Phi_m (\alpha_*) = M / m^2$.
Functions $\bs\alpha \mapsto (\delta_k \circ \Phi_m)(\bs\alpha)$ are continuous at point $\bs\alpha_*$ for $1 \leq k \leq m$ and therefore there exists $\eta>0$ such that the ball
$\mathcal{B} \left( \bs\alpha_*, \eta \right)$ is included in the set $\left\{ \bs\alpha, \ (\delta_k \circ \Phi_m)(\alpha) < \frac{M}{m^2}+\eps, \ k=1, \ldots, m \right\}$.
We have therefore $\PP(\mathcal{A}_m) \neq 0$ as
\begin{align*}
\PP(\mathcal{A}_m)
& = \int_{\left\{ (\delta_k \,\circ\, \Phi_m) (\bs\alpha) < \frac{M}{m^2} + \eps, \, k=1, \ldots, m \right\}} p(\bs\alpha) d\bs\alpha
\\
& > \int_{\mathcal{B} \left( \bs\alpha_*, \eta \right)} p(\bs\alpha) d\bs\alpha > 0
\end{align*}
Coming back to \eqref{ineq:P_I<R}, we eventually have
\[
\PP(I<R) \ \dot\geq \ \frac{1}{\rho^{m(N-M+m)}},
\]
that is the diversity of the MMSE receiver is upper bounded by $m(N-M+m)$.
\subsection{Upper bound of the outage probability}
\label{sec:upper-flat}
We now conclude by studying the upper bound of the outage probability, showing that $m(N-M+m)$ is also a lower bound for the diversity.
Note that this lower bound has been derived in \cite{kumar2009asymptotic, mehana2010diversity} using however rather informal arguments; we provide a more rigorous proof here for the sake of completeness.
We now assume that
$R/M < \log (M/(m-1))$, i.e. $m-1 < M 2^{-R/M}$.
Using Jensen inequality on function $y \mapsto \log(1/y)$, the capacity $I$ can be lower bounded:
\begin{align*}
I
&= - \sum_{j=1}^M \log \left( \left[ \left( \I + \frac{\rho}{M} \H^*\H \right)^{-1} \right]_{jj} \right)
\\
&\geq -M \log \left( \frac{1}{M} \Tr \left[ \left( \I + \frac{\rho}{M} \H^*\H \right)^{-1} \right] \right),
\end{align*}
which leads to an upper bound for the outage probability:
\begin{equation}
\PP(I<R)
\leq \PP \left[ \Tr \left[ \left( \I + \frac{\rho}{M} \H^*\H \right)^{-1} \right] > M 2^{-R/M} \right].
\label{ineq:PP-B0}
\end{equation}
We need to derive the probability in the right-hand side of the above inequality.
Noting
$\mathcal{B}_0= \left\{ \lambda_1 \leq \lambda_2 \ldots \leq \lambda_M, \ \sum_{k=1}^M \left( 1 + \frac{\rho}{M} \lambda_k \right)^{-1} > M 2^{-R/M} \right\}$,
\begin{align}
\PP \bigg[ \Tr \Big[ \Big( \I + & \frac{\rho}{M} \H^*\H \Big)^{-1} \Big] > M 2^{-R/M} \bigg] \notag
\\
& = \int_{\mathcal{B}_0} p(\lambda_1, \ldots, \lambda_M) d\lambda_1 \ldots d\lambda_M.
\label{eq:int_B0_PP}
\end{align}
We now introduce $\mu_m = \sup_{(\lambda_1, \ldots, \lambda_M) \in \mathcal{B}_0}\{ \rho \,\lambda_m \}$
and prove by contradiction that $\mu_m < +\infty$.
If $\mu_m = +\infty$, there exists a sequence $(\lambda_1^{(n)}, \lambda_2^{(n)}, \ldots, \lambda_M^{(n)})_{n\in\mathbb{N}}$ such that
$\lambda_k^{(n)} \rightarrow +\infty$ for any $k \geq m$.
Besides,
\[
M 2^{-R/M} \mhs < \sum_{k=1}^M \mhs \Big( 1+\frac{\rho}{M}\lambda^{(n)}_k \Big)^{-1} \mhs \leq (m-1) + \sum_{k=m}^M \mhs \Big(1+\frac{\rho}{M}\lambda^{(n)}_k\Big)^{-1}.
\]
In particular $M 2^{-R/M} < (m-1) + \sum_{k=m}^M \mhs \big(1+\frac{\rho}{M}\lambda^{(n)}_k\big)^{-1} $,
which, taking the limit when $n \rightarrow +\infty$, leads to $m-1 \geq M 2^{-R/M}$, a contradiction with the assumption
$m-1 < M 2^{-R/M}$. Hence, $\mu_m < +\infty$.
We introduce the set $\mathcal{B}_1= \{ \lambda_1 \leq \lambda_2 \ldots \leq \lambda_M, \, 0 < \lambda_k \leq \frac{\mu_m}{\rho}, \, k=1,\ldots,m \}$,
which verifies $\mathcal{B}_0 \subset \mathcal{B}_1$.
Using \eqref{ineq:PP-B0} and \eqref{eq:int_B0_PP}, this implies that
\[
\PP(I<R) \leq \int_{\mathcal{B}_1} p(\lambda_1, \ldots, \lambda_M) d\lambda_1 \ldots d\lambda_M,
\]
which is shown to be asymptotically smaller than
$\rho^{-m(N-M+m)}$ in the sense of \eqref{eq:exp-equ}
in Appendix \ref{apx:prob_first_ev_bounded}.
The diversity is thus lower bounded by $m(N-M+m)$, ending the proof.
\end{proof}
\section{Frequency selective MIMO channels with cyclic prefix}
\label{sec:freqsel}
We consider a frequency selective MIMO channel with $L$ independent taps.
We consider a block transmission cyclic prefix scheme, with a block length of $K$.
The output of the MIMO channel at time $t$ is given by
\begin{align*}
\y_t =\sqrt{\frac{\rho}{ML}}\,\sum_{l=0}^{L-1} \H_l \x_{t-l} +\n_t
= \sqrt{\frac{\rho}{ML}}\,[\H(z)] \x_t +\n_t
\end{align*}
where
$\x_t$ is the channel input vector at time $t$,
$\n_t \sim \mathcal{CN}(\bs{0}, \I_N)$ the additive white Gaussian noise,
$\H_l$ is the $N \times M$ channel matrix associated to $l^{\mathrm{th}}$ channel tap, for $l \in \{0,\ldots, L-1\}$,
and $\H(z)$ denotes the transfer function of the discrete-time equivalent channel defined by
\[
\H(z) = \sum_{l=0}^{L-1} \H_l \, z^{-l}.
\]
We make the common assumption that the entries of $\H_l$ are i.i.d and $\mathcal{CN}(0,1)$ distributed.
We can now state the second diversity theorem of the paper.
\medskip
\begin{theorem}
Assume that the non restrictive condition
$K > {M^{2}(L-1)}$
holds, ensuring that $\log \frac{M}{m} < -\log\big(\frac{m-1}{M}+\frac{(L-1)(M-(m-1))}{K}\big)$ for any $m=1, \ldots, M$.
Then, for a rate $R$ verifying
\begin{equation}
\hspace{-11pt}
\textstyle
\log \frac{M}{m}
< \frac{R}{M} <
-\log \left(\frac{m-1}{M} + \frac{(L-1)(M-(m-1))}{K} \right),
\label{eq:bounds_R_fsel}
\end{equation}
$m \in \{ 1, \ldots, M \}$, the outage probability verifies
\begin{equation}
\PP(I<R) \doteq \rho^{-m(LN-M+m)},
\end{equation}
that is a diversity of $m(LN-M+m)$.
\end{theorem}
\medskip
The diversity of the MMSE receiver is thus $m(LN-M+m)$, corresponding to a flat fading MIMO channel with $M$ transmit antennas and
$LN$ receive antennas. For a large block length $K$, the upper bound for rate $R$ is close to the bound of the previous flat fading case $\log \frac{M}{m-1}$.
Concerning data rates verifying
$
-\log\big(\frac{m-1}{M} + \frac{L-1}{K} (M-(m-1))\big)
< \frac{R}{M} <
\log \frac{M}{m-1},
$
the $m(LN-M+m)$ diversity is only an upper bound;
nevertheless the diversity is also lower bounded by $(m-1)(LN-M+(m-1))$.
\medskip
\begin{proof}
Similarly to previous section the capacity of the MIMO MMSE system is written
$
I = \sum_{j=1}^M \log ( 1 + \beta_j),
$
where $\beta_j$ is the SINR for the $j$th stream of $\x_t$.
It is standard material that in MIMO frequency selective channel with cyclic prefix the SINR of the MMSE receiver is given by
\begin{equation}
\beta_j = \frac{1}{ \frac{1}{K}\sum_{k=1}^K \left[ \left( \S\left(\frac{k-1}{K}\right) \right)^{-1} \right]_{jj} } -1,
\label{eq:SINR_freqsel}
\end{equation}
where $\S(\nu)= \I_N + \frac{\rho}{M} \H(e^{2 i \pi \nu})^*\H(e^{2 i \pi \nu}) $.
\subsection{Lower bound for the outage probability}
We assume that $R/M > \log(M/m)$.
One can show that function $\A \mapsto (\A^{-1})_{jj}$, defined over the set of positive-definite matrices, is convex.
Using Jensen's inequality then yields
\begin{align*}
\frac{1}{K}\sum_{k=1}^K {\textstyle \left[ \left( \S\left(\frac{k-1}{K}\right) \right)^{-1} \right]_{jj} }
& \geq \bigg( \bigg[ \frac{1}{K} \sum_{k=1}^K {\textstyle \S\left(\frac{k-1}{K}\right) }\bigg]^{-1} \bigg)_{jj}
\\
& = \bigg( \bigg[ \I_N + \sum_{l=0}^{L-1} \frac{\rho}{M} \H_l^*\H_l \bigg]^{-1} \bigg)_{jj}.
\end{align*}
The last equality follows from the fact that $\frac{1}{K}\sum_{k=1}^K e^{2 i\pi \frac{k-1}{K}(l-n)}=\delta_{ln}$.
Using this inequality in the SINR expression \eqref{eq:SINR_freqsel} gives
\[
1+\beta_j \leq \bigg( \bigg( \bigg[ \I_N + \sum_{l=0}^{L-1} \frac{\rho}{M} \H_l^*\H_l \bigg]^{-1} \bigg)_{jj} \bigg)^{-1}.
\]
We now come back to the capacity $I$ of the system;
similarly to \eqref{ineq:jensen1}, using Jensen's inequality yields
\begin{align*}
I
& \leq M \log \Bigg[ \frac{1}{M} \sum_{j=1}^M \left( 1+\beta_j \right) \Bigg]
\\
& \leq M \log \Bigg[ \frac{1}{M} \sum_{j=1}^M \bigg( \bigg( \bigg[ \I_N + \frac{\rho}{M} \sum_{l=0}^{L-1} \H_l^*\H_l \bigg]^{-1} \bigg)_{jj} \bigg)^{-1} \,\Bigg].
\end{align*}
We can now use the results of section \ref{sec:lowB_flat} by simply replacing $N \times M$ matrix $\H$ in \eqref{ineq:logconcave} by $LN \times M$ matrix $\Hb=[\H_0^T, \H_1^T, \ldots, \H_{L-1}^T]^T$.
They lead to the following lower bound for the outage capacity, for a rate $R$ verifying $R/M > \log (M/m)$:
\[
\PP(I<R) \ \dot\geq \ \frac{1}{\rho^{m(LN-M+m)}}.
\]
\subsection{Upper bound for the outage probability}
We assume that $\frac{R}{M} < -\log\big(\frac{m-1}{M}+\frac{(L-1)(M-(m-1))}{K}\big)$, that is
$2^{-R/M} < \frac{m-1}{M} \mhs + \mhs \frac{L-1}{K} (M \mhs - \mhs (m-1))$.
We first derive a lower bound for the capacity $I$.
\begin{align*}
I & = - \sum_{j=1}^M \log \left( \frac{1}{K}\sum_{k=1}^K \left( \left[ {\textstyle \S\left(\frac{k-1}{K}\right)} \right]^{-1} \right)_{jj} \right)
\\
& \geq - M \log \left( \frac{1}{KM} \sum_{k=1}^K \Tr\left( \left[ {\textstyle \S\left(\frac{k-1}{K}\right)} \right]^{-1} \right) \right)
\end{align*}
The latter inequality follows once again from Jensen's inequality on function $x \mapsto \log x$.
We now analyze $\Tr\left(\S(\nu)^{-1}\right)$.
To that end, we write $LN \times M$ matrix $\Hb=[\H_0^T, \ldots, \H_{L-1}^T]^T$ under the form
$\Hb=\bs\Theta (\Hb^*\Hb)^{1/2}$,
where $\bs\Theta=[\bs\Theta_0^T, \ldots, \bs\Theta_{L-1}^T]^T$ and $\bs\Theta^*\bs\Theta=\I_M$.
Besides, we note $\U^*\Lambda\U$ the SVD of $\Hb^*\Hb$ with
$\Lambda=\mathrm{diag}(\lambda_1,\ldots,\lambda_M)$,
$\lambda_1 \leq \ldots \leq \lambda_M$.
Hence,
\[
\H(e^{2 i\pi \nu})=\bs\Theta(e^{2 i\pi \nu})\U^*\Lambda^{1/2}\U,
\]
where
$\bs\Theta(z)=\sum_{l=0}^{L-1}\bs\Theta_l z^{-l}$.
Using this parametrization,
\begin{align*}
\Tr\left(\S(\nu)^{-1}\right)
&= \Tr \left[ \left( \I + \frac{\rho}{M} \U \bs\Theta^*(e^{2 i \pi \nu})\bs\Theta(e^{2 i \pi \nu}) \U^* \Lambda \right)^{-1} \right]
\\
&\leq \Tr \left[ \left( \I + \frac{\rho}{M} \gamma(e^{2 i \pi \nu}) \Lambda \right)^{-1} \right],
\end{align*}
where $\gamma(\nu)=\lambda_{\mathrm{min}}(\bs\Theta^*(e^{2 i \pi \nu})\bs\Theta(e^{2 i \pi \nu}))$.
Coming back to the outage probability,
\begin{align}
\PP(I<R) \leq & \PP \mhs \left[ \frac{1}{K} \sum_{k=0}^{K-1} \sum_{j=1}^M \mhs \left( \mhs 1 + \frac{\rho \lambda_j }{M} \gamma \mhs\left(\frac{k}{K}\right) \mhs \right)^{\mhs-1} \mhs > M2^{-R/M}\right] \notag
\\
&= \PP \mhs \left[ \Hb \in \mathcal{B}_0 \right], \label{ineq:PP-fs}
\end{align}
where $\mathcal{B}_0 = \big\{ \Hb, \frac{1}{K} \sum_{k=0}^{K-1} \sum_{j=1}^M \mhs \big( 1 + \frac{\rho \lambda_j }{M} \gamma \mhs\left(\frac{k}{K}\right) \big)^{-1} \mhs > M2^{-R/M} \big\}$.
We now prove by contradiction that
$\mu_m < +\infty$, where $\mu_m =\sup_{\Hb \in \mathcal{B}_0} \{ \rho \lambda_m \}$.
If $\mu_m = +\infty$
there exists a sequence of matrices $\Hb^{(n)} \mhs \in \mhs \mathcal{B}_0$ such that $\rho \lambda_m^{(n)} \rightarrow +\infty$.
Besides,
\begin{align}
M 2^{-\frac{R}{M}}
& < \frac{1}{K} \sum_{k=0}^{K-1} \sum_{j=1}^M \bigg( 1 + \frac{\rho \lambda_j^{(n)} }{M} \gamma^{(n)} \mhs\left(\frac{k}{K}\right) \bigg)^{\mhs-1} \notag
\\
& \leq (m-1) + \frac{1}{K} \sum_{k=0}^{K-1} \sum_{j=m}^M \bigg( 1 + \frac{\rho \lambda_j^{(n)} }{M} \gamma^{(n)} \mhs\left(\frac{k}{K}\right) \bigg)^{\mhs-1}
\label{ineq:mat-seq}
\end{align}
As $\bs\Theta^{(n)}$ belongs to a compact we can extract a subsequence $\bs\Theta^{(\psi(n))}$ which converges towards a matrix $\bs\Theta_\infty$.
For this subsequence, inequality \eqref{ineq:mat-seq} becomes
\begin{equation}
M 2^{-\frac{R}{M}}
\leq (m-1) + \frac{1}{K} \sum_{k=0}^{K-1} \sum_{j=m}^M \bigg( 1 + \frac{\rho \lambda_j^{(\psi(n))} }{M} \gamma^{(\psi(n))} \mhs\left(\frac{k}{K}\right) \bigg)^{-1}.
\label{ineq:sumsum}
\end{equation}
Let $\gamma_\infty$ be the function defined by $\gamma_\infty(\nu)=\lambda_{\mathrm{min}}(\bs\Theta^*_\infty(e^{2 i \pi \nu})\bs\Theta_\infty(e^{2 i \pi \nu}))$ and $k_1, \ldots, k_p$ be the integers for which $\gamma_{\infty}(k_j/K) = 0$.
Then $\det \bs\Theta_\infty(z) = \det \big( \sum_{l=0}^{L-1} \bs\Theta_{\infty,l} z^{-l} \big)=0$ for all $z \in \big\{ e^{2 i \pi k_j/K}, j=1,\ldots,p \big\}$.
Nevertheless, polynomial $z \mapsto \sum_{l=0}^{L-1} \bs\Theta_{\infty,l} z^{-l}$ has a maximum degree of $M(L-1)$, therefore $p \leq M(L-1)$.
Inequality \eqref{ineq:sumsum} then leads to
\begin{equation}
M 2^{-\frac{R}{M}}
\leq (m-1) + \frac{M(L-1)}{K}
+\frac{1}{K} \sum_{k\notin\{k_1, \ldots, k_p\}} \sum_{j=m}^M \bigg( 1 + \frac{\rho \lambda_j^{(\psi(n))} }{M} \gamma^{(\psi(n))} \left( \frac{k}{K} \right) \bigg)^{\mhs-1} \label{ineq:M2-RMp}
\end{equation}
Moreover, if $k \notin \{k_1, \ldots, k_p\}$,
$\lambda_j^{(\psi(n))} \gamma^{(\psi(n))}(\frac{k}{K}) \rightarrow +\infty$ for $j \geq m$,
as $\gamma^{(\psi(n))} \mhs\left(\frac{k}{K}\right) \rightarrow \gamma_\infty \mhs\left(\frac{k}{K}\right) \neq 0$ for $k \notin \left\{ k_1, \ldots, k_p \right\}$.
Therefore taking the limit of \eqref{ineq:M2-RMp} when $n \rightarrow +\infty$ gives
\[
M 2^{-\frac{R}{M}}
\leq (m-1) + \frac{M(L-1)}{K},
\]
which is in contradiction with the original assumption
$2^{-R/M} < \mhs \frac{m-1}{M} \mhs + \mhs \frac{L-1}{K} (M \mhs - \mhs (m-1))$.
Hence $\mu_m < +\infty$, and
$\mathcal{B}_0 \subset \mathcal{B}_1= \{ \Hb, \rho \lambda_m(\Hb^*\Hb) < \mu_m \}$.
Using \eqref{ineq:PP-fs}, we thus have
\[
\PP(I<R) \leq \PP(\Hb \in \mathcal{B}_1),
\]
which, by Appendix \ref{apx:prob_first_ev_bounded}, is asymptotically smaller than
$\rho^{-m(NL-M+m)}$ in the sense of \eqref{eq:exp-equ}, therefore ending the proof.
\end{proof}
\section{Numerical Results}
We here illustrate the derived diversity in the frequency selective case.
In the conducted simulation we took a block length of $K=64$,
a number of transmitting and receiving antennas $M=N=2$,
$L=2$ channel taps
and an aimed data rate $R=3$~bits/s/Hz.
Rate $R$ then verifies \eqref{eq:bounds_R_fsel} with $m=1$, therefore the expected diversity is $LN-M+1=3$.
The outage probability is displayed on Fig. \ref{fig:Pout} as a function of SNR.
We observe a slope of $-10^{-3}$ per decade, hence a diversity of $3$, confirming the result stated in part \ref{sec:freqsel}.
\begin{figure}[t]
\centering
\includegraphics[width=3in]{div_40M}
\caption{Outage probability of the MMSE receiver, L=2, K=64, M=N=2}
\label{fig:Pout}
\end{figure}
\section{Conclusion}
In this paper we provided rigorous proofs regarding the diversity of the MMSE receiver at fixed rate, in both flat fading and frequency selective MIMO channels.
The higher the aimed rate the less diversity is achieved;
in particular, for sufficiently low rates, the MMSE receiver achieves full diversity in both MIMO channel cases, hence its great interest.
Nonetheless, in frequency selective channels, the diversity bounds are not tight for some specific rates; this could probably be improved.
Simulations corroborated our results.
\appendices
\section{}
\label{apx:prob_sum_first_ev}
We prove in this appendix that, for $b>0$, $\PP(\sum_{k=1}^m \rho \lambda_k < b) \ \dot\geq \ \rho^{-m(N-M+m)}$.
We note $\mathcal{C}_m$ the set defined by
$\mathcal{C}_m = \{ \lambda_1, \ldots, \lambda_m: \ 0 < \lambda_1 \leq \ldots \leq \lambda_m, \ \sum_{k=1}^m \rho \lambda_k < b \}$.
As the $\lambda_i$ verify
$0 < \lambda_1 \leq \ldots \leq \lambda_M$,
we can write
\begin{equation}
\PP \Bigg( \sum_{k=1}^m \rho \lambda_k < b \Bigg)
= \int_{(\lambda_1, \ldots, \lambda_m) \in \mathcal{C}_m} \int_{\lambda_m}^{+\infty} \bhs\ldots \int_{\lambda_{M-1}}^{+\infty} p_{M,N}(\lambda_1, \ldots, \lambda_M) \ d\lambda_1 \ldots d\lambda_M,
\label{eq:prob_sum_first_ev}
\end{equation}
where
$p_{M,N}: \mathbb{R}^M \rightarrow \mathbb{R}$
is the joint probability density function of the ordered eigenvalues of a $M \times M$ Wishart matrix with scale matrix $\I_M$ and $N$ degrees of freedom, given by (see, e.g., \cite{zheng2003diversity}):
\begin{equation}
p_{M,N} = K_{M,N}^{-1} \prod_{i=1}^M \left( \lambda_i^{N-M} e^{-\lambda_i} \right) \prod_{i<j}(\lambda_i-\lambda_j)^2,
\label{eq:p-wish}
\end{equation}
where $K_{M,N}$ is a normalizing constant.
We now try to separate the integral in \eqref{eq:prob_sum_first_ev} in two integrals, one over $\lambda_1, \ldots, \lambda_m$, the other over $\lambda_{m+1}, \ldots, \lambda_M$.
As we have $(\lambda_1, \ldots, \lambda_m) \in \mathcal{C}_m$ in \eqref{eq:prob_sum_first_ev}, $\lambda_m < b / \rho$ and thus
\begin{equation}
\begin{split}
\int_{\lambda_m \leq \lambda_{m+1} \leq \ldots \leq \lambda_M} & p_{M,N}(\lambda_1, \ldots, \lambda_M) \ d\lambda_{m+1} \ldots d\lambda_M
\\
& \geq \int_{(\lambda_{m+1}, \ldots, \lambda_M) \in \mathcal{D}} p_{M,N}(\lambda_1, \ldots, \lambda_M) \ d\lambda_{m+1} \ldots d\lambda_M
\end{split}
\label{ineq:2intg}
\end{equation}
where $\mathcal{D}=\{(\lambda_{m+1}, \ldots, \lambda_M) \in \mathbb{R}_+^{M-m}; \ b/\rho \leq \lambda_{m+1}\leq \ldots \leq \lambda_M \}$.
This integral can be simplified by noticing that $p_{M,N}(\lambda_1, \ldots, \lambda_M)$ explicit expression \eqref{eq:p-wish} is invariant by permutation of its parameters
$\lambda_1, \ldots, \lambda_M$,
in particular by permutation of its parameters
$\lambda_{m+1}, \ldots, \lambda_M$.
Therefore, noting
$\mathcal{S}=\mathrm{Sym}(\{\lambda_{m+1}, \ldots, \lambda_M\})$ the group of permutations over the finite set $\{\lambda_{m+1}, \ldots, \lambda_M\}$,
we get
\begin{align}
\int_{b/\rho}^{+\infty} \bhs\ldots \int_{b/\rho}^{+\infty} & p_{M,N}(\lambda_1, \ldots, \lambda_M) \ d\lambda_{m+1} \ldots d\lambda_M \notag
\\
& = \sum_{s \in \mathcal{S}} \int_{s(\lambda_{m+1}, \ldots, \lambda_M) \in \mathcal{D}} p_{M,N}(\lambda_1, \ldots, \lambda_M) \ d\lambda_{m+1} \ldots d\lambda_M \notag
\\
& = \mathrm{Card}(\mathcal{S}) \int_{ (\lambda_{m+1}, \ldots, \lambda_M) \in \mathcal{D}} p_{M,N}(\lambda_1, \ldots, \lambda_M) \ d\lambda_{m+1} \ldots d\lambda_M \notag
\\
& = (M-m)! \int_{ (\lambda_{m+1}, \ldots, \lambda_M) \in \mathcal{D}} p_{M,N}(\lambda_1, \ldots, \lambda_M) \ d\lambda_{m+1} \ldots d\lambda_M.
\label{eq:intg-symm}
\end{align}
Using \eqref{ineq:2intg} and \eqref{eq:intg-symm} in \eqref{eq:prob_sum_first_ev}, we obtain
\[
\PP \Bigg( \sum_{k=1}^m \rho \lambda_k < b \Bigg)
\geq \frac{1}{(M-m)!} \int_{\mathcal{C}_m} \int_{b/\rho}^{+\infty} \bhs\ldots \int_{b/\rho}^{+\infty} p_{M,N}(\lambda_1, \ldots, \lambda_M) \ d\lambda_1 \ldots d\lambda_M.
\]
We now replace
$p_{M,N}$
by its explicit expression \eqref{eq:p-wish} and then try to separate the $m$ first eigenvalues from the others.
Note that we can drop the constants $(M-m)!$ and $K_{M,N}$ as we only need an asymptotic lower bound.
\begin{align*}
\PP \Bigg( \sum_{k=1}^m \rho \lambda_k < b \mhs \Bigg) \mhs
\ \dot\geq & \int_{\mathcal{C}_m} \int_{b/\rho}^{+\infty} \bhs \ldots \int_{b/\rho}^{+\infty} \prod_{i=1}^M \left( \lambda_i^{N-M} e^{-\lambda_i} \right) \prod_{i<j}(\lambda_i-\lambda_j)^2 \,
d\lambda_1 \ldots d\lambda_M
\\
= & \ \int_{\mathcal{C}_m} \int_{b/\rho}^{+\infty} \bhs \ldots \int_{b/\rho}^{+\infty}
\Bigg( \prod_{i=1}^m \left( \lambda_i^{N-M} e^{-\lambda_i} \right) \prod_{i<j\leq m}(\lambda_i-\lambda_j)^2 \, \Bigg)
\\
& \cdot \Bigg( \prod_{i=m+1}^M \left( \lambda_i^{N-M} e^{-\lambda_i} \right) \prod_{i\leq m<j}(\lambda_i-\lambda_j)^2 \prod_{m<i<j}(\lambda_i-\lambda_j)^2 \, \Bigg)
\ d\lambda_1 \ldots d\lambda_M
\end{align*}
For $i\leq m<j$, we have that $\lambda_i \leq b/\rho$ and thus $(\lambda_i-\lambda_j)^2 \geq \big(\lambda_j-\frac{b}{\rho}\big)^2$.
Hence,
\begin{align}
\PP \Bigg( \sum_{k=1}^m \rho \lambda_k < b \mhs \Bigg)
\dot\geq & \ \bigg( \int_{\mathcal{C}_m} \prod_{i=1}^m \left( \lambda_i^{N-M} e^{-\lambda_i} \right) \prod_{i<j\leq m}(\lambda_i-\lambda_j)^2 \ d\lambda_1 \ldots d\lambda_m \bigg)
\label{eq:2sep-intg}
\\
& \cdot \mhs \bigg( \mhs \int_{b/\rho}^{+\infty} \bhs\mhs\mhs \ldots \int_{b/\rho}^{+\infty}
\mhs\mhs \prod_{i=m+1}^M \mhs\mhs\mhs \left( \lambda_i^{N-M}e^{-\lambda_i} \right)
\mhs \prod_{j=m+1}^M \mhs\mhs\mhs \left(\lambda_j-\frac{b}{\rho}\right)^{\,\mhs\mhs 2m}
\mhs\mhs \prod_{m<i<j} \mhs\mhs\mhs (\lambda_i-\lambda_j)^2 \ d\lambda_{m+1} \ldots d\lambda_M \mhs\mhs\bigg) \notag
\end{align}
We now have two separate integrals.
We first consider the second one, in which we make the substitution
$\beta_i= \lambda_i - b/\rho$ for $i=m+1, \ldots, M$.
\begin{align}
\int_{b/\rho}^{+\infty} & \bhs \ldots \int_{b/\rho}^{+\infty}
\prod_{i=m+1}^M \mhs\mhs\mhs \left( \lambda_i^{N-M}e^{-\lambda_i} \right)
\mhs \prod_{j=m+1}^M \mhs\mhs\mhs \left(\lambda_j-\frac{b}{\rho}\right)^{\,\mhs\mhs 2m}
\mhs\mhs \prod_{m<i<j} \mhs\mhs\mhs (\lambda_i-\lambda_j)^2 \ d\lambda_{m+1} \ldots d\lambda_M
\notag
\\
& = e^{- {(M-m)b/\rho}} \int_0^{+\infty} \bhs \ldots \int_0^{+\infty}
\mhs\mhs \prod_{i=m+1}^M \mhs\mhs \left( \mhs \left( {\textstyle \beta_i + \frac{b}{\rho}} \right)^{N-M} e^{- \beta_i} \beta_i ^{2m} \mhs \right)
\mhs \prod_{m<i<j}(\beta_i-\beta_j)^2 \ d\beta_{m+1} \ldots d\beta_M
\notag
\\
& \geq \frac{1}{2} \int_0^{+\infty} \bhs \ldots \int_0^{+\infty} \prod_{i=m+1}^M \left( \beta_i ^{N-M+2m} e^{- \beta_i} \right) \prod_{m<i<j}(\beta_i-\beta_j)^2 \ d\beta_{m+1} \ldots d\beta_M
\label{ineq:2nd-intg}
\end{align}
for $\rho$ large enough, i.e. such that
$e^{- (M-m) b/\rho} > 1/2$.
It is straightforward to see that the integral in \eqref{ineq:2nd-intg} is nonzero, finite, independent from $\rho$ and therefore asymptotically equivalent to $1$ in the sense of \eqref{eq:exp-equ}.
Hence, we can drop the second integral in \eqref{eq:2sep-intg}, leading to:
\begin{equation}
\PP \Bigg( \sum_{k=1}^m \rho \lambda_k < b \mhs \Bigg)
\dot\geq \ \int_{\mathcal{C}_m} \prod_{i=1}^m \left( \lambda_i^{N-M} e^{-\lambda_i} \right) \prod_{i<j\leq m}(\lambda_i-\lambda_j)^2 \ d\lambda_1 \ldots d\lambda_m.
\label{ineq:only1-intg}
\end{equation}
Making the substitution $\alpha_i= \rho \lambda_i$ for $i=1, \ldots, m$ in \eqref{ineq:only1-intg}
and
noting
$\mathcal{C}'_m=\{ \alpha_1, \ldots, \alpha_m: \ 0 < \alpha_1 \leq \ldots \leq \alpha_m, \ \sum_{k=1}^m \alpha_k < b \}$
we then have
\begin{align}
\PP \Bigg(\sum_{k=1}^m \rho \lambda_k < b \Bigg)
& \, \dot\geq \, \bigg( \rho^{-m-m(N-M)-m(m-1)} \int_{\mathcal{C}'_m} \prod_{i=1}^m \left( \alpha_i^{N-M} e^{-\alpha_i/\rho} \right) \prod_{i<j\leq m}(\alpha_i-\alpha_j)^2 \ d\alpha_1 \ldots d\alpha_m \bigg) \notag
\\
& \geq \rho^{-m(N-M+m)} \int_{\mathcal{C}'_m} \prod_{i=1}^m \left( \alpha_i^{N-M} e^{-\alpha_i} \right) \prod_{i<j\leq m}(\alpha_i-\alpha_j)^2 \ d\alpha_1 \ldots d\alpha_m
\label{ineq:no-intg}
\end{align}
for $\rho \geq 1$, as we have then $e^{- \alpha_i/\rho} \geq e^{- \alpha_i}$ for $i=1, \ldots, m$.
As $b>0$ it is straightforward to see that the integral in \eqref{ineq:no-intg} is nonzero but also finite and independent from $\rho$; it is therefore asymptotically equivalent to $1$ in the sense of \eqref{eq:exp-equ}, yielding
\[
\PP \Bigg( \sum_{k=1}^m \rho \lambda_k < b \Bigg) \dot\geq \ \rho^{-m(N-M+m)},
\]
which concludes the proof.
\section{}
\label{apx:prob_first_ev_bounded}
We prove in this section that $\PP\left( \mathcal{B}_1 \right) \dot\leq \, \rho^{-m(M-N+m)}$, where the set $\mathcal{B}_1$ is defined by
\[
\mathcal{B}_1= \{ \lambda_1 \leq \lambda_2 \ldots \leq \lambda_M, \, 0 < \lambda_k \leq b, \, k=1,\ldots,m \},
\]
with $b>0$ and $\lambda_1, \ldots, \lambda_M$ the ordered eigenvalues of the Wishart matrix $\H^*\H$.
We use the same approach as in Appendix \ref{apx:prob_sum_first_ev}.
For we note $p_{M,N}$ the joint probability density function of the ordered eigenvalues of a $M \times M$ Wishart matrix with scale matrix $\I_M$ and $N$ degrees of freedom,
the probability $\PP(\mathcal{B}_1)$ can be written as
\[
\PP(\mathcal{B}_1)
= \int_{(\lambda_1, \ldots, \lambda_M) \in \mathcal{B}_1}
p_{M,N}(\lambda_1, \ldots, \lambda_M) \ d\lambda_1 \ldots d\lambda_M.
\]
Similarly to Appendix \ref{apx:prob_sum_first_ev} we try to upper bound $\PP(\mathcal{B}_1)$ by the product of two integrals, one containing the $m$ first eigenvalues and the other the $M-m$ remaining eigenvalues.
We first replace $p_{M,N}$ by it explicit expression \eqref{eq:p-wish}:
\begin{align*}
\PP(\mathcal{B}_1)
& = K_{M,N}^{-1} \int_{(\lambda_1, \ldots, \lambda_M) \in \mathcal{B}_1}
\prod_{i=1}^M \lambda_i^{N-M} e^{-\lambda_i} \prod_{i<j}(\lambda_i-\lambda_j)^2 \ d\lambda_1 \ldots d\lambda_M
\\
& \doteq \int_{(\lambda_1, \ldots, \lambda_M) \in \mathcal{B}_1}
\Bigg( \prod_{i=1}^m \left( \lambda_i^{N-M} e^{-\lambda_i} \right) \prod_{i<j\leq m}(\lambda_i-\lambda_j)^2 \Bigg)
\\
& \fhs \cdot \Bigg( \prod_{i=m+1}^M \left( \lambda_i^{N-M} e^{-\lambda_i} \right) \prod_{i\leq m<j}(\lambda_i-\lambda_j)^2 \prod_{m<i<j}(\lambda_i-\lambda_j)^2 \Bigg)
\ d\lambda_1 \ldots d\lambda_M.
\end{align*}
Note that we dropped the normalizing constant $K_{M,N}$, as $K_{M,N}^{-1} \doteq 1$.
For $i \leq m < j$, we have
$|\lambda_i - \lambda_j| \leq \lambda_j$
and thus
$\prod_{i\leq m<j}(\lambda_i-\lambda_j)^2 \leq \prod_{j=m+1}^M \lambda_j^{2m}$, yielding
\begin{align*}
\PP(\mathcal{B}_1)
& \,\dot\leq\,
\int_0^{b/\rho} \int_{\lambda_1}^{b/\rho} \bhs \ldots \int_{\lambda_{m-1}}^{b/\rho} \int_{\lambda_m}^{+\infty} \bhs \ldots \int_{\lambda_{M-1}}^{+\infty}
\Bigg( \prod_{i=1}^m \left( \lambda_i^{N-M} e^{-\lambda_i} \right) \prod_{i<j\leq m}(\lambda_i-\lambda_j)^2 \Bigg)
\\
& \fhs \cdot \Bigg( \prod_{i=m+1}^M \left( \lambda_i^{N+2m-M} e^{-\lambda_i} \right) \prod_{m<i<j}(\lambda_i-\lambda_j)^2 \Bigg)
\ d\lambda_1 \ldots d\lambda_M
\end{align*}
In order to obtain two separate integrals we discard the $\lambda_m$ in the integral bound simply by noticing that $\lambda_m>0$, therefore
\begin{align*}
\PP(\mathcal{B}_1)
& \,\dot\leq\,
\Bigg( \int_0^{b/\rho} \int_{\lambda_1}^{b/\rho} \bhs \ldots \int_{\lambda_{m-1}}^{b/\rho}
\prod_{i=1}^m \left( \lambda_i^{N-M} e^{-\lambda_i} \right) \prod_{i<j\leq m}(\lambda_i-\lambda_j)^2
\ d\lambda_1 \ldots d\lambda_m \Bigg)
\\
& \fhs \cdot \Bigg( \int_0^{+\infty} \int_{\lambda_{m+1}}^{+\infty} \bhs \ldots \int_{\lambda_{M-1}}^{+\infty}
\prod_{i=m+1}^M \left( \lambda_i^{N+2m-M} e^{-\lambda_i} \right) \prod_{m<i<j}(\lambda_i-\lambda_j)^2
\ d\lambda_{m+1} \ldots d\lambda_M \Bigg)
\end{align*}
As the second integral (in $\lambda_{m+1}$, \ldots, $\lambda_M$) is nonzero, finite and independent of $\rho$ it is asymptotically equivalent to $1$ in the sense of \eqref{eq:exp-equ}.
Hence,
\begin{equation}
\PP(\mathcal{B}_1)
\,\dot\leq\,
\int_0^{b/\rho} \int_{\lambda_1}^{b/\rho} \bhs \ldots \int_{\lambda_{m-1}}^{b/\rho}
\prod_{i=1}^m \left( \lambda_i^{N-M} e^{-\lambda_i} \right) \prod_{i<j\leq m}(\lambda_i-\lambda_j)^2
\ d\lambda_1 \ldots d\lambda_m.
\label{ineq:PB1-1intg}
\end{equation}
We now make the substitutions $\alpha_i=\rho\lambda_i$ for $i=1, \ldots, m$ inside the remaining integral.
\begin{align}
& \int_0^{b/\rho} \int_{\lambda_1}^{b/\rho} \bhs \ldots \int_{\lambda_{m-1}}^{b/\rho}
\prod_{i=1}^m \left( \lambda_i^{N-M} e^{-\lambda_i} \right) \prod_{i<j\leq m}(\lambda_i-\lambda_j)^2
\ d\lambda_1 \ldots d\lambda_m \notag
\\
& \hspace{15pt} = \ \rho^{-m(N-M+m)}
\int_0^{b} \int_{\alpha_1}^{b} \bhs \ldots \int_{\alpha_{m-1}}^{b}
\prod_{i=1}^m \left( \alpha_i^{N-M} e^{-\alpha_i/\rho} \right) \prod_{i<j\leq m}(\alpha_i-\alpha_j)^2
\ d\alpha_1 \ldots d\alpha_m \notag
\\
& \hspace{15pt} \leq \rho^{-m(N-M+m)}
\mhs \int_0^{b} \int_{\alpha_1}^{b} \bhs \ldots \int_{\alpha_{m-1}}^{b}
\prod_{i=1}^m \alpha_i^{N-M} \mhs \prod_{i<j\leq m} \mhs\mhs (\alpha_i-\alpha_j)^2
\ d\alpha_1 \ldots d\alpha_m , \label{ineq:rem_intg}
\end{align}
as $e^{-\alpha_i/\rho} \leq 1$.
The remaining integral in \eqref{ineq:rem_intg} is nonzero ($b > 0$), finite and does not depend on $\rho$; therefore, \eqref{ineq:rem_intg} is asymptotically equivalent to $\rho^{-m(N-M+m)}$ in the sense of \eqref{eq:exp-equ}.
Coming back to \eqref{ineq:PB1-1intg} we obtain
\[
\PP(\mathcal{B}_1) \ \dot\leq \ \rho^{-m(N-M+m)}.
\]
\section{}
\label{apx:ang_par}
In this appendix, we review the results of \cite{dita2003factorization, lundberg2004haar} for the reader's convenience.
It has been shown in \cite{dita2003factorization} that any $n \times n$ unitary matrix $A_n$ can be written as
\begin{equation}
A_n = d_n {\mathcal{O}}_n \begin{bmatrix} 1 & 0 \\ 0 & A_{n-1} \end{bmatrix},
\label{eq:unitary_dec}
\end{equation}
with
$A_{n-1}$ a $(n-1) \times (n-1)$ unitary matrix,
$d_n$ a diagonal phases matrix, that is
$d_n = \mathrm{diag}(e^{i\varphi_1}, \ldots, e^{i\varphi_n})$ with $\varphi_1, \ldots, \varphi_n \in [0, 2\pi]$,
and
$\mathcal{O}_n$ an orthogonal matrix (the angles matrix).
Matrix $\mathcal{O}_n$ can be written in terms of parameters $\theta_1, \ldots, \theta_n \in [0, \frac{\pi}{2}]$ thanks to the following decomposition:
$\mathcal{O}_n = J_{n-1,n} J_{n-2,n-1} \ldots J_{1,2}$,
where
\[
J_{i,i+1}= \begin{bmatrix} \I_{i-1} & 0 & 0 & 0 \\ 0 & \cos\theta_i & -\sin\theta_i & 0 \\ 0 & \cos\theta_i & -\sin\theta_i & 0 \\ 0 & 0 & 0 & \I_{n-i-1} \end{bmatrix}.
\]
Let $\U_M$ be a $M \times M$ unitary Haar distributed matrix.
Then, using decomposition \eqref{eq:unitary_dec},
\[
\U_M= \D_M(\bs\varphi_1) \V_M(\bs\theta_1) \begin{bmatrix} 1 & 0 \\ 0 & \U_{M-1} \end{bmatrix},
\]
with
$\bs\varphi_1=(\varphi_{1,1},\ldots,\varphi_{1,M}) \in [0,2\pi]^M$,
$\bs\theta_1=(\theta_{1,1},\ldots,\theta_{1,M-1}) \in [0,\frac{\pi}{2}]^{M-1}$,
$\D_M(\bs\varphi_1)$ the diagonal matrix defined by $\D_M(\bs\varphi_1)=\mathrm{diag}(e^{i\varphi_{1,1}},\ldots,e^{i\varphi_{1,M}})$,
$\V_M(\bs\theta_1)$ the orthogonal matrix defined by $\V_M(\bs\theta_1)=J_{M-1,M} J_{M-2,M-1} \ldots J_{1,2}$
and
$\U_M$ a $M-1 \times M-1$ unitary matrix.
Matrix $\U_{M-1}$ can naturally be similarly factorized.
Similarly to \cite{lundberg2004haar}, we can show that,
in order $\U_M$ to be a Haar matrix it is sufficient
that $(\varphi_{1,i})_{i=1,\ldots,M}$ are i.i.d. random variables uniformly distributed over interval $[0, 2\pi[$,
that $\theta_{1,1},\ldots,\theta_{1,M-1}$ are independent with densities respectively equal to $(\sin \theta_1)^{M-2}, (\sin \theta_2)^{M-3}, \ldots, (\sin \theta_{M-2}), 1$ and independent from $\bs\varphi_1$
and that $\U_{M-1}$ is Haar distributed and independent from $\bs\varphi_1$ and $\bs\theta_1$.
The proof consists in first showing, by a simple variable change, that if the $(\varphi_{1,i})_{i=1,\ldots,M}$ and the $\theta_{1,1},\ldots,\theta_{1,M-1}$ follow the mentioned distributions then $\D_M(\bs\varphi_1) \V_M(\bs\theta_1)$ is uniformly distributed over the unity sphere of $\mathbb{C}^M$.
The proof is then completed by showing that if $\U_{M-1}$ is a Haar matrix independent from $\bs\varphi_1$ and $\theta_1$ then $\U_M$ is Haar distributed.
Finally one can parameterize a Haar matrix $\U_M$ by $\bs\varphi_1$, $\theta_1$ and $\U_{M-1}$.
Repeating the same parametrization for $\U_{M-1}$ we obtain that $\U_M$ can be parameterized by the $M^2$ following independent variables
\begin{align*}
&(\varphi_{1,1}, \ldots, \varphi_{1,M}), (\theta_{1,1}, \ldots, \theta_{1,M-1}), (\varphi_{2,1}, \ldots, \varphi_{2,M-1}), (\theta_{2,1}, \ldots, \theta_{2,M-2}), \ldots,
\\
&(\varphi_{M-2,1}, \varphi_{M-2,2}), \theta_{M-2,1}, \varphi_{M-1,1},
\end{align*}
whose probability laws are almost surely positive.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec:Introduction}
The decay $\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace$ is described in the Standard Model by loop (penguin)
diagrams that contain a $b \ensuremath{\rightarrow}\xspace s$ transition. The partial width of the decay
arises from three helicity amplitudes that, assuming no additional contributions
from physics beyond the Standard Model, are determined by the chiral structure
of the quark operators. Predictions obtained within the framework of QCD
factorization~\cite{beneke} are ${\ensuremath{\cal B}\xspace}(\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace) = (9.1^{+11.3}_{-6.8}) \times
10^{-6}$ for the branching fraction and ${0.63}^{+0.42}_{-0.29}$ for the \ensuremath{\kaon^{*0}}\xspace
longitudinal polarization fraction. Predictions improve to
$({7.9}^{+4.3}_{-3.9}) \times 10^{-6}$ and ${0.72}^{+0.16}_{-0.21}$,
respectively, when experimental input is used from $\ensuremath{\PB}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^*}\xspace\phi$
\cite{babar_kstphi,belle_kstphi}. The possibility to use $\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace$ for
precision \ensuremath{C\!P}\xspace-violation studies to determine the phases \ensuremath{\beta_{\ensuremath{\Ps}\xspace}}\xspace and $\gamma$ of
the CKM matrix \cite{pdg} has been emphasized by several authors
\cite{matias,ciuchini,fleischer,fleischer99}. The U-spin related channel, $\ensuremath{\B^0}\xspace
\ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace \ensuremath{\Kbar^{*0}}\xspace$, a $b \ensuremath{\rightarrow}\xspace d$ transition, has been observed by
BaBar~\cite{bbar}, reporting a branching fraction of $({1.28}^{+0.35}_{-0.30}
\pm 0.11) \times 10^{-6}$ and $f_L = {0.80}^{+0.10}_{-0.12} \pm 0.06$ with a
signal yield of ${33.5}^{+9.1}_{-8.1}$ events. An upper limit for the
$\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace$ branching fraction of $1.68 \times {10}^{-3}$ with $90\%$
confidence level was reported by the SLD experiment \cite{sld}.
We present in this Letter the first observation of the $\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace$ decay using
$pp$ collisions at $\sqrt{s} = 7$ TeV at the LHC. The data were collected
during 2010 and corresponds to 35 \ensuremath{\mbox{\,pb}^{-1}}\xspace of integrated luminosity. LHCb has
excellent capabilities to trigger and reconstruct beauty and charm hadrons, and
covers the pseudorapidity region $2<\eta<5$. The tracking system consists of a
21 station, 1-metre long array of silicon strip detectors placed within 8 mm of
the LHC beams. This is followed by a four layer silicon strip detector upstream
of a 4 Tm dipole magnet. Downstream of the magnet are three tracking stations,
each composed of a four-layer silicon strip detector in the high occupancy
region near the beam pipe, and an eight layer straw tube drift chamber composed
of 5 mm diameter straws outside this high occupancy region. Overall, the
tracking system provides an impact parameter (IP) \footnote{The impact parameter
is the distance of closest approach between a particle's trajectory and its
assumed production point.} resolution of 16 $\mu$m + 30 $\mu$m/$p_{\rm{T}}
(\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c}}\xspace)$, and a momentum resolution $\sigma_p/p$ below 8 per mille up to 100
\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c}}\xspace. Two ring imaging Cherenkov detectors, one upstream of the magnet, and a
second just downstream of the tracking stations, together provide a typical kaon
identification efficiency of 90\%. The pion fake rate, over the momentum range
from $3-100$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c}}\xspace, is between 5 and 10 percent. Further downstream is a
Preshower/Scintillating Pad Detector, an electromagnetic calorimeter, and a
hadronic calorimeter. The LHCb spectrometer also features a large, five station
muon system used for triggering on and identifying muons. A more detailed
description of the LHCb detector can be found in~\cite{jinst}.
To reduce the data rate from the LHC crossing rate to about 2 kHz for permanent
storage, LHCb uses a two-level trigger system. The first level of the trigger,
implemented in hardware,
searches for either a large transverse energy ($E_{\rm{T}}$) cluster in the
calorimeters ($E_{\rm{T}}>3.6$ GeV is a representative value during the 2010 run), or a
single high transverse momentum ($p_{\rm{T}}$) muon or di-muon pair in the muon
stations.
Events passing the hardware trigger are read out and sent to a large computing
farm, where they are analyzed using a software-based trigger~\cite{HLT2}. The
first stage of the software trigger relies on the selection of a single track
with IP larger than 125 $\mu$m, $p_{\rm{T}}> 1.8$ \ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c}}\xspace, $p > 12.5$ \ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c}}\xspace, along
with other track quality requirements. Events are subsequently analyzed by a
second software stage, where the event is searched for 2, 3, or 4-particle
vertices that are consistent with originating from $b$-hadron decays. The
impact parameter $\chi^2$ of the selected tracks (IP$\chi^2$), defined as the
difference between the $\chi^2$ of the primary vertex (PV) built with and
without the considered track, is required to be greater than 16 with respect to
any PV. The tracks are also required to have $p>5$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c}}\xspace and $p_{\rm{T}} >
0.5$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c}}\xspace. The \ensuremath{\B^0_s}\xspace decay vertex must have at least one track with
$p_{\rm{T}}>1.5$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c}}\xspace, a scalar $p_{\rm{T}}$ sum of at least 4 \ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c}}\xspace, and a
corrected mass\footnote{The corrected mass is related to the invariant mass $m$,
as $m_{corr} = \sqrt{m^2 + |p_{T miss}|^2} + |p_{T miss}|$ , where $p_{T miss}$
is the missing momentum transverse to the \ensuremath{\B^0_s}\xspace direction.} between 4 and 7
\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace. Additional track and vertex quality cuts are also applied.
Events with large occupancy are slow to reconstruct and were suppressed by
applying global event cuts to hadronically triggered decays. These included
limits on the number of hits in the tracking detectors and scintillating pad
detector.
\section{Selection procedure and signal yield}
\label{sec:selyield}
To search for the decay process $\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace(\ensuremath{\kaon^+}\xspace\ensuremath{\pion^-}\xspace) \ensuremath{\Kbar^{*0}}\xspace(\ensuremath{\kaon^-}\xspace\ensuremath{\pion^+}\xspace)$ we
applied a number of offline selection criteria. When a four-track secondary
vertex is found, the reconstructed momentum of the \ensuremath{\B^0_s}\xspace candidate is used to
calculate the smallest impact parameter with respect to all primary vertices in
the event. Tracks are required to have $p_{\rm{T}} > 500$~\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c}}\xspace, and a large
impact parameter (IP$\chi^2>9$) with respect to the PV. The difference in the
natural logarithm of the likelihoods of the kaon and pion hypotheses must be
greater than 2 for $K^{+}$ and $K^{-}$ candidates, and less than 0 for $\pi^{+}$
and $\pi^{-}$ candidates. In addition, the $K^+ \pi^-$
combinations\footnote{This expression refers hereafter to both charge
combinations: $K^+ \pi^-$ and $K^- \pi^+$.} must form an acceptable quality
common vertex ($\chi^{2}$/ndf$~<9$), where ndf is the number of degrees of
freedom in the vertex fit) and must have an invariant mass within $\pm150$
\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace of the nominal \ensuremath{\kaon^{*0}}\xspace mass (this is around $\pm$3 times its physical
width~\cite{pdg}). The \ensuremath{\kaon^{*0}}\xspace and \ensuremath{\Kbar^{*0}}\xspace candidates must have $p_{\rm{T}} >
900$~\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c}}\xspace and the distance of closest approach between their trajectories must
be less than 0.3 mm. The secondary vertex must be well fitted
($\chi^{2}/$ndf$<5$). Finally, the \ensuremath{\B^0_s}\xspace candidate momentum is required to point
to the PV.
\begin{figure}[htbp]
\begin{center}
\ifthenelse{\boolean{pdflatex}}{
\includegraphics*[width=\columnwidth]{Fig1.pdf}
}{
\includegraphics*[width=\columnwidth]{Fig1.eps}
}
\end{center}
\caption{Fit to the $\ensuremath{\kaon^+}\xspace\ensuremath{\pion^-}\xspace\ensuremath{\kaon^-}\xspace\ensuremath{\pion^+}\xspace$ mass distribution of selected candidates.
The fit model (dashed pink curve) includes a signal component that has two
Gaussian components corresponding to the \ensuremath{\B^0_s}\xspace and \ensuremath{\B^0}\xspace decays. The background is
described as an exponential component (dotted blue) plus the parametrization
indicated in the text (dash-dotted green).}
\label{figBsmass}
\end{figure}
To improve the signal significance, a multivariate analysis is used that takes
into account the properties of the $\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace(\ensuremath{\kaon^+}\xspace\ensuremath{\pion^-}\xspace) \ensuremath{\Kbar^{*0}}\xspace(\ensuremath{\kaon^-}\xspace\ensuremath{\pion^+}\xspace)$
signal, as well as those of the background. It is based on a geometrical
likelihood (GL)~\cite{Karlen,DiegoThesis} that uses the following set of
variables
as input:
\begin{itemize}
\item \ensuremath{\B^0_s}\xspace candidate impact parameter with respect to the closest primary vertex.
\item Decay time of the \ensuremath{\B^0_s}\xspace candidate.
\item Minimum impact parameter \ensuremath{\chi^2}\xspace of the four tracks with respect to all
primary vertices in the event.
\item Distance of closest approach between the two \ensuremath{\kaon^{*0}}\xspace trajectories
reconstructed from the pion and kaon tracks.
\item $p_{\rm{T}}$ of the \ensuremath{\B^0_s}\xspace candidate.
\end{itemize}
For a given input sample, the above distributions are converted into a set of
uncorrelated, Gaussian-distributed variables. Two vectors are defined for each
event indicating its distance to the signal $\{S_i\}$ and to the background
$\{B_i\}$ hypotheses by means of $\chi^2_S = \sum S_i^2$ and $\chi^2_B = \sum
B_i^2$, {where the index $i$ runs over the five discriminating variables indicated above.}
The quantity $\Delta\chi^2 = \chi^2_S-\chi^2_B$ is found to be a good
discriminant between the two hypotheses and is used to construct the GL function
in such a way that it is uniformly distributed in the range $[0,1]$ for signal
events and tends to have low values for the background. {The signal input is generated
by EvtGen \cite{evtgen} and Pythia 6.4 \cite{pythia} for the kinematic spectra, and the full detector simulation is based on GEANT4 \cite{geant4}.}
The GL selection requirement was determined by maximizing the signal
significance. The GL was trained using a fully reconstructed $\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace$
simulation sample for the signal, and a selected background sample from the
first 2 \ensuremath{\mbox{\,pb}^{-1}}\xspace of data, which is not used in the analysis. The requirement
GL$>$0.24, together with the above selection criteria, resulted in the mass
spectrum in Fig.~\ref{figBsmass} for the selected $\ensuremath{\kaon^+}\xspace\ensuremath{\pion^-}\xspace\ensuremath{\kaon^-}\xspace\ensuremath{\pion^+}\xspace$ candidates.
It is observed that the events with masses below the signal region have on
average slightly higher GL values than those with masses above. This indicates
the presence of a background from partially reconstructed \ensuremath{\PB}\xspace decays.
To describe the data, we have used a fit including two Gaussian probability
density functions (PDFs) centered at the \ensuremath{\B^0}\xspace and \ensuremath{\B^0_s}\xspace masses respectively, a
decreasing exponential and the following parametrization for partially
reconstructed \mbox{\ensuremath{\PB}\xspace-decays}
\begin{equation}
\begin{small}
\label{eq:physbkg}
A M^{\prime}
\left(1-\frac{M^{\prime 2}}{M_p^2}\right)
\Theta(M_p-M^{\prime})
e^{-k_p \cdot M^{\prime}}
\otimes
G(M-M^{\prime};\sigma_p),
\end{small}
\end{equation}
where $\Theta$ is the Heaviside-step function, $\otimes$ represents the
convolution, $M^{\prime}$ is the variable over which the convolution integral is
calculated, $G(M - M^{\prime}; \sigma_p)$ is a Gaussian PDF with standard
deviation $\sigma_p$ and $M_p$ and $k_p$ are free parameters. The fit results
are given in Table \ref{tab:massfitresults}.
The measured signal yield in a window of $\pm 50 \ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace$ around the \ensuremath{\B^0_s}\xspace mass is
$N_{\ensuremath{\kaon^+}\xspace\ensuremath{\pion^-}\xspace\ensuremath{\kaon^-}\xspace\ensuremath{\pion^+}\xspace} = 49.8 \pm 7.5 ({\rm stat.})$. The width of the \ensuremath{\B^0_s}\xspace peak is
in good agreement with the LHCb\xspace resolution measured in decays with similar
kinematics such as $\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\phi$. The significance of the \ensuremath{\B^0_s}\xspace signal was
determined to be $10.9~\sigma$ by comparing the log of the likelihood between
the models with and without signal. When doing this test, the mass and width of
the \ensuremath{\B^0_s}\xspace and \ensuremath{\B^0}\xspace mesons were fixed to those obtained from independent LHCb
measurements of $\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\phi$ and $\ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace$, respectively. The peak at
the \ensuremath{\B^0}\xspace mass, though not significant, is compatible with the $\ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\ensuremath{\Kbar^{*0}}\xspace$
branching fraction measured by BaBar~\cite{bbar}.
\begin{table}
\caption{Fitted values of the model parameters for the mass spectrum, as
described in the text. $N_s$ and $N_d$ are the number of events for the \ensuremath{\B^0_s}\xspace and
\ensuremath{\B^0}\xspace signals, $\mu_s$ is the fitted mass value for the \ensuremath{\B^0_s}\xspace signal and $\sigma$ is
the Gaussian width. The mass difference between \ensuremath{\B^0_s}\xspace and \ensuremath{\B^0}\xspace was fixed to its
nominal value~\cite{pdg}. $N_{b}$ is the number of background events in the full
mass range (4900-5800 \ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace), and $c_{b}$ is the exponential parameter in the
fit. $M_p$, $\sigma_p$ and $k_p$ are the parameters of Eq.~(\ref{eq:physbkg}).
Finally, $f_{p}$ is the fraction of the background associated with
Eq.~(\ref{eq:physbkg}). }
\label{tab:massfitresults}
\begin{center}
\newcolumntype{+}{D{/}{\ \pm\ }{-1}}
\begin{tabular}{|l|+|}
\hline
Parameter & \multicolumn{1}{c|}{Value}\\
\hline
$N_s$ & 50.1/7.5 \\
$N_d$ & 11.2/4.3 \\
$\mu_s$ {\small(\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace)} & 5362.5/4.8 \\
$\sigma$ {\small(\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace)} & 21.2/3.3 \\
$N_{b}$ & 90/10 \\
$c_{b}$ {\small($10^{-3}(\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace)^{-1}$)} & -3.37/0.55 \\
$k_p$ {\small($10^{-2}(\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace)^{-1}$)}& 5.5/5.3 \\
$f_{p}$ & \multicolumn{1}{c|}{${}_{\ \ }0.06_{\ -\ 0.05}^{\ +\ 0.24}$} \\
$M_p$ {\small(\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace)} & 5170/170 \\
$\sigma_p$ {\small(\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace)} & 37/23 \\
\hline
\end{tabular}
\end{center}
\label{tab:massfit}
\end{table}
As the \ensuremath{\kaon^{*0}}\xspace meson is light compared to the \ensuremath{\B^0_s}\xspace meson, the invariant masses of
the three-body systems $\ensuremath{\kaon^+}\xspace\ensuremath{\kaon^-}\xspace\ensuremath{\pion^\pm}\xspace$ and $\ensuremath{\kaon^+}\xspace\ensuremath{\pion^-}\xspace\ensuremath{\pion^\pm}\xspace$ are rather high, above
those of the charmed hadrons. This kinematically excludes the possibility of
contamination from $b \ensuremath{\rightarrow}\xspace c$ decays with very short charm flight distance, in
particular $\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace D_s^- \pi^+$.
After subtraction of the {non-\ensuremath{\B^0_s}\xspace component}, the $\ensuremath{\kaon^+}\xspace\ensuremath{\pion^-}\xspace$ mass combinations were studied, within a
$\pm$50 \ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace window of the \ensuremath{\B^0_s}\xspace signal, by means of a maximum likelihood fit
in the $(m_{\ensuremath{\kaon^+}\xspace\ensuremath{\pion^-}\xspace},m_{\ensuremath{\kaon^-}\xspace\ensuremath{\pion^+}\xspace})$ plane. Three components are included in the
fit, namely a double Breit-Wigner distribution describing $\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace$
production, a symmetrized product of a Breit-Wigner and a nonresonant linear
model adjusted for phase space in the $\ensuremath{\kaon^+}\xspace\ensuremath{\pion^-}\xspace$ mass, and a double nonresonant
component. The fit result, as shown in Fig.~\ref{m1m2fit}, gives (62$\pm$18)\%
$\ensuremath{\kaon^{*0}}\xspace\ensuremath{\Kbar^{*0}}\xspace$ production. The remainder is the symmetrized Breit-Wigner/
nonresonant model.
The shape of the background mass distribution was extracted from a fit to the
$\ensuremath{\kaon^+}\xspace\ensuremath{\pion^-}\xspace$ mass spectrum observed in two 400 \ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace wide sidebands below and
above the \ensuremath{\B^0_s}\xspace mass. The number of background events to be subtracted was
determined from the results in Table~\ref{tab:massfitresults}. The sizable
\ensuremath{\kaon^{*0}}\xspace contribution present in this background was taken into account.
\begin{figure}[htbp]
\begin{center}
\ifthenelse{\boolean{pdflatex}}{
\includegraphics*[width=\columnwidth]{Fig2.pdf}
}{
\includegraphics*[width=\columnwidth]{Fig2.eps}
}
\end{center}
\caption{Background subtracted $K^+\pi^-$ and $K^-\pi^+$ combinations for
selected candidates within a $\pm$50 \ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace window of the \ensuremath{\B^0_s}\xspace mass. The solid
blue line shows the projection of the 2D fit model described in the text,
indicating the \ensuremath{\kaon^{*0}}\xspace\ensuremath{\Kbar^{*0}}\xspace yield (dashed-dotted red line) and a nonresonant
component (blue dotted line), assumed to be a linear function times the two-body
phase space. The dashed red line indicates the overall $\ensuremath{\B^0_s}\xspace\rightarrow\ensuremath{\kaon^{*0}}\xspace
X$ contribution.}
\label{m1m2fit}
\end{figure}
A model for $\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\ensuremath{\Kbar^{*0}}\xspace$(1430), representing a broad scalar state
interfering with \ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace was also studied in the available $K^+\pi^-$ mass
range of $\pm 150\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace$ around the \ensuremath{\kaon^{*0}}\xspace mass. The small number of events
made it impossible to measure precisely the size of such a contribution for all
values of the interfering phase. However, for values of the phase away from
$\pi/2$ and $3\pi/2$ it was determined to be below 12\%. Further study of this
issue requires a larger data sample.
\section{Selection of the control channel}
The branching fraction measurement of $\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace$ is based upon the use of a
normalization channel with a well measured branching fraction, and knowledge of
the selection and trigger efficiencies for both the signal and normalization
channels. We chose $\ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace$, with $\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace \ensuremath{\rightarrow}\xspace \mu^+ \mu^-$, for this
purpose. This decay has a similar topology to the signal, allowing the selection
cuts to be harmonized, and it is copiously produced in the LHCb acceptance. The
presence of two muons in the final state means that $\ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace$ tends to be
triggered by a muon rather than a hadron, leading to a higher efficiency than
for $\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace$. The differences in the trigger can be mitigated by only
considering $\ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace$ candidates where the trigger decision was not allowed
to be based on muon triggers that use tracks from the decay itself.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{Fig3.pdf}
\caption{Fit to the mass distribution of selected $\ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace$ events. The
dashed red curve is the Gaussian component for the \ensuremath{\PB}\xspace signal. The green
dashed-dotted line accounts for partially reconstructed \ensuremath{B\ensuremath{\rightarrow}\xspace J/\psi X}\xspace (see
Eq.~\ref{eq:BJpsiX}). The pink hatched region accounts for a possible $\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace \phi$
contamination, parametrized as a sum of two Crystal-Ball functions
\cite{cb}. The combinatorial background is parametrized as an exponential and
indicated as a blue dotted line.}
\label{fig:jpsikstarnew}
\end{center}
\end{figure}
The offline selection criteria for $\ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace$ were designed to mimic those of
\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace. In particular, all cuts related to the \ensuremath{\B^0_s}\xspace vertex definition were
kept the same. We also used the same GL as for the signal.
The overall detection efficiency was factorized as
$\epsilon^{sel}\epsilon^{trig}$. The first factor $\epsilon^{sel}$ is the
probability of the generated tracks being accepted in the LHCb\xspace angular
coverage, reconstructed, and selected. The second factor $\epsilon^{trig}$
defines the efficiency of the trigger on the selected events. Both are
indicated in Table~\ref{tab:BR_numbers}, as calculated from Monte Carlo
simulation, along with the number of selected events. Note that our measurement
depends only on the ratios of efficiencies between signal and control channels.
\begin{table*}[ht]
\begin{center}
\caption{Selection and trigger efficiencies obtained from simulation. The
observed yield found for the signal and control channels in the full mass range
are also indicated. The efficiency errors are statistical, derived from the size
of the simulated samples.}
\newcolumntype{+}{D{/}{\ \pm\ }{5}}
\begin{tabular}{|c|+|+|+|}
\cline{2-4}
\multicolumn{1}{c|}{} & \multicolumn{1}{c|}{$\epsilon^{sel}$ $(\%)$} & \multicolumn{1}{c|}{$\epsilon^{trig}$ $(\%)$} &
\multicolumn{1}{c|}{\begin{minipage}{0.9in} \begin{center}
\medskip
Yield
\medskip \end{center} \end{minipage}} \\
\hline
\begin{minipage}{1.3in} \begin{center} \smallskip $\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace$
\smallskip \end{center} \end{minipage} & 0.370/0.005 & 37.12/0.39 & 42.5/6.7 \\
\begin{minipage}{1.3in} \begin{center} \smallskip $\ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace$ \smallskip
\end{center} \end{minipage} & 0.547/0.007 & 31.16/0.63 & 657/27 \\
\hline
\begin{minipage}{1.3in} \begin{center} \smallskip ratio \smallskip
\end{center} \end{minipage} & 0.678/0.013 & 1.191/0.027 & 0.065/0.011\\
\hline
\end{tabular}
\label{tab:BR_numbers}
\end{center}
\end{table*}
The event yield for the selected data was determined from a fit to the
$\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\ensuremath{\kaon^+}\xspace\ensuremath{\pion^-}\xspace$ invariant mass spectrum as shown in Fig.~\ref{fig:jpsikstarnew}.
In this fit, a constrained \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace mass was used in order to improve the \ensuremath{\B^0}\xspace mass
resolution and therefore background rejection.
{A component for the particular background source $\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace \phi$, with $\phi \ensuremath{\rightarrow}\xspace K^+K^-$,
was included in the fit, with a parametrization defined from simulation, yielding the result 8$\pm$8 events.
The complete suppression of this background was subsequently confirmed using }
the Armenteros-Podolanski \cite{armenteros} plot
for the \ensuremath{\kaon^{*0}}\xspace kinematics.
The fit model also includes a Gaussian signal for the \ensuremath{\B^0}\xspace meson, and a combinatorial
background component parameterized with an exponential function and an
additional component to account for partially reconstructed \ensuremath{B\ensuremath{\rightarrow}\xspace J/\psi X}\xspace~\cite{PK}.
This partially reconstructed component can be described as
\begin{equation}
\begin{tiny}
\rho(M, \overline{M}, \mu,\kappa) \propto \begin{cases}
e^{-\frac{1}{2}(\frac{M-\overline{M}}{\kappa})^2} & \text{if } M > \mu ; \\
e^{-\frac{1}{2}(\frac{\mu-\overline{M}}{\kappa})^2 +
\frac{(\overline{M}-\mu)(M-\mu)}{\kappa^{2}}} & \text{if } M \leq \mu. \\
\end{cases}
\label{eq:BJpsiX}
\end{tiny}
\end{equation}
where the parameters $\mu$, $\kappa$ and $\overline{M}$ are allowed to float.
{The fitted signal according to this model is indicated in the third column of
Table~\ref{tab:BR_numbers}.}
A small fraction of the selected sample contains two alternative candidates for
the reconstructed event, which share three of the particles but differ in the
fourth one. Those events, which amount to 3.8 \% (3.7\%) in the signal
(control) channels, were retained for the determination of the branching
fraction.
\section{Analysis of \ensuremath{\kaon^{*0}}\xspace polarization}
\label{sec:Kpolarization}
The four-particle $K^+\pi^-K^-\pi^+$ angular distribution describing the decay
of \ensuremath{\B^0_s}\xspace into two vector mesons ($\ensuremath{\kaon^{*0}}\xspace \ensuremath{\rightarrow}\xspace K^+\pi^-$ and $\ensuremath{\Kbar^{*0}}\xspace \ensuremath{\rightarrow}\xspace
K^-\pi^+$) is determined by three transversity amplitudes $\mathcal{A}_L$,
$\mathcal{A}_{\parallel}$ and $\mathcal{A}_{\perp}$. The relative fraction of
these can be determined from the distribution of the decay products in three
angles $\theta_1$, $\theta_2$ and $\varphi$. Here $\theta_1$ ($\theta_2$) is the
$K^+$ ($K^-$) emission angle with respect to the direction opposite to the \ensuremath{\B^0_s}\xspace
meson momentum in the \ensuremath{\kaon^{*0}}\xspace (\ensuremath{\Kbar^{*0}}\xspace) rest frame, and $\varphi$ is the angle between
the normals to the \ensuremath{\kaon^{*0}}\xspace and \ensuremath{\Kbar^{*0}}\xspace decay planes in the \ensuremath{\B^0_s}\xspace rest frame
\cite{matias}. We will refer generically to the $\theta$ angle from now on,
unless differences between $\theta_1$ and $\theta_2$ become relevant for the
discussion. In a time-integrated and flavor-averaged analysis, and assuming the
\ensuremath{\B^0_s}\xspace mixing phase $\beta_s \approx 0$ as in the Standard Model, the angular
distribution is given by~\cite{matias,dunietz}
\begin{multline}
I(\theta_1,\theta_2,\varphi) = \frac{{\rm d}^3\Gamma}{{\rm d}\cos\theta_1\, {\rm
d} \cos\theta_2\, {\rm d}\varphi} = \\ \left( \frac{1}{\Gamma_L} \right.
|\mathcal{A}_L|^2 \cos^2\theta_1 \cos^2\theta_2 \\ + \frac{1}{\Gamma_L}
|\mathcal{A}_\parallel|^2 \frac{1}{2} \sin^2\theta_1 \sin^2\theta_2
\cos^2\varphi \\ + \frac{1}{\Gamma_H} |\mathcal{A}_\perp|^2 \frac{1}{2}
\sin^2\theta_1 \sin^2\theta_2 \sin^2\varphi \\ + \frac{1}{\Gamma_L}
|\mathcal{A}_L| \left.|\mathcal{A}_\parallel| \cos\delta_\parallel
\frac{1}{2\sqrt{2}} \sin 2\theta_1 \sin 2\theta_2 \cos\varphi \right)
\label{eq:angular}
\end{multline}
\noindent
We denote the polarization fractions by
\begin{equation}
\begin{small}
f_k = \frac{|\mathcal{A}_k|^2}{|\mathcal{A}_L|^2+|\mathcal{A}_\parallel|^2+
|\mathcal{A}_\perp|^2} \ \ \ ,\ \ \ k=L, \parallel, \perp.
\end{small}
\label{eq:amplitudes}
\end{equation}
\noindent and consequently $ f_L+f_{\parallel}+f_{\perp}=1 $. No \ensuremath{C\!P}\xspace violation
in the mixing or in the decay has been considered. The interference terms
related to the $\mathcal{A}_{\perp}$ amplitude, both proportional to
sin$\phi_s$, have been neglected. $\Gamma_{L,H}$ are the total widths of the
low and high mass eigenstates of the \ensuremath{\B^0_s}\xspace meson, respectively, and
$\delta_\parallel$ is the phase difference between $\mathcal{A}_L$ and
$\mathcal{A}_{\parallel}$. The total decay width is defined as $\Gamma =
(\Gamma_L + \Gamma_H)/2$ and $\Delta\Gamma=\Gamma_L-\Gamma_H$. Note that as a
consequence of time integration the relative normalization acquired by the
\ensuremath{C\!P}\xspace-even and \ensuremath{C\!P}\xspace-odd terms is different. The values $\Delta\Gamma =
(0.062^{+0.034}_{-0.037}) \times 10^{12}\ \rm{s}^{-1}$ and $\Gamma =
(0.679^{+0.012}_{-0.011}) \times 10^{12}\ \rm{s}^{-1}$~\cite{pdg} were used.
The detector acceptance is compatible with being constant in $\varphi$. In
contrast, it has a significant dependence on the \ensuremath{\kaon^{*0}}\xspace polarization angle
$\theta$. The two-dimensional angular acceptance function
$\epsilon$($\cos\theta_1$,$\cos\theta_2$) was studied with a full detector
simulation. It drops to nearly zero asymmetrically as $\cos \theta_{1,2}$
becomes close to $\pm 1$, as a consequence of the minimum \mbox{$p$}\xspace and \mbox{$p_T$}\xspace of the
tracks imposed by the reconstruction.
The Monte Carlo simulation of the \ensuremath{\kaon^{*0}}\xspace acceptance was extensively
cross-checked using the \ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace control channel, taking advantage of the fact
that the \ensuremath{\kaon^{*0}}\xspace polarization in this channel was measured at the \ensuremath{\PB}\xspace-factory
experiments \cite{jpsikstar_pol,swave-bbar}. The function
$\epsilon$($\cos\theta_1$,$\cos \theta_2$) has been projected onto the \ensuremath{\kaon^{*0}}\xspace
and \ensuremath{\Kbar^{*0}}\xspace axes separately, showing no appreciable difference, and a small
average correlation, given the size of the simulated sample. We have then used
the one-dimensional acceptance $\epsilon_{\theta}(\cos\theta)$ as the basis of
our analysis, and determined it in five bins of $\cos\theta$. Since the
longitudinal polarization fraction for the \ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace channel is well measured,
a comparison between data and simulation is possible. Agreement was found
including variations of the angular distribution with longitudinal and
transverse \ensuremath{\kaon^{*0}}\xspace momentum. In the region $\cos\theta > 0.6$ these variations
were four times larger than for lower values of $\cos\theta$.
The background $\cos \theta$ distribution was studied in two 200\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace
sidebands, defined below and above the \ensuremath{\B^0_s}\xspace signal region. Like the signal, it
showed a dip close to $\cos \theta = +1$ and it was parameterized as
$\epsilon_{\theta} \cdot (1+\beta \cos\theta)$. A one parameter fit for $\beta$
gives the result $\beta=-0.18\pm0.13$.
An unbinned maximum likelihood fit was then performed to the data in a $\pm 50$
\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace window around the \ensuremath{\B^0_s}\xspace mass, in the region $\cos\theta < 0.6 $, according
to the PDF
\begin{eqnarray}
\lefteqn{F(\theta_1,\theta_2,\varphi) = (1-\alpha) \epsilon_{\theta}(\theta_1) \epsilon_{\theta}(\theta_2) I(\theta_1,\theta_2,\varphi)} \nonumber\\
& & + \alpha (1+\beta \cos\theta_1) (1+\beta\cos\theta_2) \epsilon_{\theta}(\theta_1) \epsilon_{\theta}(\theta_2). \nonumber \\
& &
\label{eq:angularPDF}
\end{eqnarray}
The background fraction $\alpha$ was determined from the fit to the \ensuremath{\B^0_s}\xspace mass
spectrum described in Sec. \ref{sec:selyield}. Only three parameters were
allowed to vary in the fit, namely $f_L$, $f_{\parallel}$ and the phase
difference $\delta_{\parallel}$.
One-dimensional projections of the fit results are shown in
Fig.~\ref{fig_final}. The consistency of the measurement in various regions of
the \ensuremath{\kaon^{*0}}\xspace phase space, and of the impact parameter of the daughter particles,
was checked. The experimental systematic error on $f_L$ was estimated from the
variation of the measurements amongst those regions to be $\pm$0.03.
The acceptance for \ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace is not uniform as a function of proper decay time
due to the cuts made on the IP of the kaons and pions, and a small correction to
the polarization fractions, of order 3\%, was applied in order to take into
account this effect. It was calculated from the variation in the measured
polarization amplitudes induced by including a parametrization of the time
acceptance in Eq.~\ref{eq:angularPDF}. Note the different correction sign for
each polarization fraction, as a consequence of the assumption $\Delta\Gamma \ne
0$.
The sensitivity of the $f_L$ measurement with respect to small variations of the
$\cos\theta$ distribution has been tested. These variations could be attributed
to experimental errors not accounted for in the simulation or to interference
with other partial waves in the $K\pi$ system. A high statistics study using
\ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace muon triggers revealed a small systematic difference between data and
simulation in $\epsilon_{\theta}(\cos\theta)$ as $\cos\theta$ approaches +1,
which was taken into account as a correction in our analysis. When this
correction in varied by $\pm 100$\%, $f_L$ varies by $\pm$0.02 which we
consider as an additional source of systematic error. The total systematic on
$f_L$ is thus $\pm$0.04.
We finally measure the \ensuremath{\kaon^{*0}}\xspace longitudinal polarization fraction $f_L =
0.31\pm 0.12 ({\rm stat.}) \pm 0.04 ({\rm syst.})$, as well as the transverse
components $f_{\parallel}$ and $f_{\perp}$. In the small sample available, the
\ensuremath{C\!P}\xspace-odd component $f_{\perp}$ appears to be sizable $f_{\perp} = 0.38 \pm 0.11
({\rm stat}.)\pm 0.04 ({\rm syst}.)$. A significant measurement of
$\delta_{\parallel}$ could not be achieved ($\delta_{\parallel}$ = 1.47 $\pm$
1.85).
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{Fig4.pdf}
\caption{$\cos\theta$ (above) and $\varphi$ (below) acceptance corrected
distributions for events in the narrow window around the \ensuremath{\B^0_s}\xspace mass. The blue line
is the projection of the fit model given by Eq.~\ref{eq:angular} for the
measured values of the parameters $f_L$, $f_{\parallel}$ and
$\delta_{\parallel}$. The dotted lines indicate $\pm 1\sigma$ variation of the
$f_L$ central value.}
\label{fig_final}
\end{figure}
As seen in Eq.~(\ref{eq:angular}), due to a nonzero $\Delta\Gamma$ time
integration changes the relative proportion between the various terms of the
angular distribution, with respect to their values at $t=0$. If we call $
f_k^0$ the polarization fractions we would have measured under the assumption
$\Delta\Gamma = 0$, it can be derived from Eq. \ref{eq:angular} that our
measured values are
\begin{equation}
f_k = f_k^0 \left(1 + \eta_k \frac{\Delta\Gamma}{2\Gamma}\right)
\end{equation}
with CP eigenvalue $\eta_k = +1,+1,-1$ for $k=L,\parallel, \perp$. Given the
current knowledge of $\Delta\Gamma$/$\Gamma$ \cite{pdg}, the magnitude of the
correction to $f_k$ amounts to 4.6\%, and the associated systematic error
related to $\Delta\Gamma$ error is 2.6\%, which we have neglected in comparison
to other sources.
\section{Determination of the branching fraction}
\label{sec:branchingF}
The results of the previous sections can be brought together to provide a
determination of the branching fraction of the $\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace$ decay based upon
the use of the normalization channel $\ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace$ through the expression
\begin{eqnarray}
\lefteqn{{\ensuremath{\cal B}\xspace}\left( \ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace\right) = \lambda_{f_L} \times
\frac{\epsilon^{sel}_{\ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace}}{\epsilon^{sel}_{\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace}}} \nonumber \\
& & \times \frac{\epsilon^{trig}_{\ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace}}{\epsilon^{trig}_{\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace}} \times \frac{N_{\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace}}{N_{\ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace}} \nonumber\\
& & \times {\ensuremath{\cal B}\xspace}_{vis}{(\ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace)} \times \frac{f_{d}}{f_s} \times \frac{9}{4},
\label{eqn:BR_BR_JPsiKstar}
\end{eqnarray}
where ${\ensuremath{\cal B}\xspace}_{vis}{(\ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace)}$, the visible branching ratio, is the product
${\ensuremath{\cal B}\xspace}(\ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace) \times {\ensuremath{\cal B}\xspace}(\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\Pmu^+\Pmu^-}\xspace) \times {\ensuremath{\cal B}\xspace}(\ensuremath{\kaon^{*0}}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^+}\xspace \ensuremath{\pion^-}\xspace)$.
The numerical value of ${\ensuremath{\cal B}\xspace}(\ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace) = (1.33 \pm 0.06) \times 10^{-3}$ is
taken from the world average in~\cite{pdg}, ${\ensuremath{\cal B}\xspace}(\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\Pmu^+\Pmu^-}\xspace) = 0.0593
\pm 0.0006$ \cite{pdg} and ${\ensuremath{\cal B}\xspace}(\ensuremath{\kaon^{*0}}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^+}\xspace \ensuremath{\pion^-}\xspace) = 2/3$ \cite{pdg}. The
ratio of $b$-quark hadronization factors that accounts for the different
production rate of \ensuremath{\B^0}\xspace and \ensuremath{\B^0_s}\xspace mesons is $f_s/f_d = 0.253 \pm
0.031$~\cite{fsfdlhcb}. The factor 9/4 is the inverse square of the 2/3
branching fraction of $\ensuremath{\kaon^{*0}}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^+}\xspace \ensuremath{\pion^-}\xspace$. The number of candidate events in
the signal and control channel data samples are designated by $N_{\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace}$
and $N_{\ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace}$.
The correction factor $\lambda_{f_L}$ is motivated by the fact that the overall
efficiency of the LHCb\xspace detector is a linear function of the \ensuremath{\kaon^{*0}}\xspace
longitudinal polarization $f_L$. Taking into account the measured value and
errors reported in section \ref{sec:Kpolarization}, Monte Carlo simulation was
used to estimate $\lambda_{f_L}=0.812\pm 0.059$.
We have considered two sources of systematic uncertainty associated to the ratio
of selection efficiencies. The first source results from discrepancies between
data and simulation in the variables related to track and vertex quality, and
the second is related to particle identification. A small difference observed in
the average impact parameter of the particles was corrected for by introducing
an additional smearing to the track parameters in the
simulation~\cite{lhcbmumu}. While the absolute efficiencies vary significantly
as a function of vertex resolution, the ratio of efficiencies remains stable. We
have assigned a $2\%$ uncertainty to the ratio, after comparison between
simulation and the \ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace data. The \ensuremath{\PK}\xspace/\ensuremath{\Ppi}\xspace identification efficiency
was determined using a sample of \ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace events selected without making use
of the RICH detectors. As the signal channel contains one more kaon than the
control channel, a correction factor of $1.098\pm0.019$ was applied to the
branching fraction, and a 2\% error was assigned to it. The efficiency of muon
identification agrees with simulation within $1.1\%$~\cite{Jpsi_1}. All these
factors are combined to produce an overall systematic uncertainty of 3.4\% in
the ratio of selection efficiencies. The uncertainty in the background model in
the \ensuremath{\B^0_s}\xspace mass fit ($\pm 2$ events) contributes an additional systematic error of
$4.7\%$.
Trigger efficiencies can be determined, for particular trigger paths in LHCb,
using the data driven algorithm described in~\cite{Jpsi_1}. This algorithm
could be applied for the specific hadronic triggers used for \ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace, but not
for the small \ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace signal. The efficiency related to cuts on global event
properties, applied during the 2010 data taking, is determined from $J/\psi$
minimum bias triggers~\cite{Jpsi_1}. The result indicates a trigger efficiency
of $(26.8 \pm 3.8)\%$, smaller than the simulation result of $(31.16\pm 0.63)\%$
shown in Table~\ref{tab:BR_numbers}. Although these are consistent within
uncertainties, we nonetheless apply a $-9$\% correction to the ratio of trigger
efficiencies between \ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace and \ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace channels, taking into account
correlations in the trigger probability. A systematic error of 11\% was
assigned to uncertainty on the trigger efficiency, entirely limited by
statistics, both in the signal and control channels. Detector occupancies,
estimated by the average number of reconstructed tracks, are larger by 10\% in
the data than in the simulation. This implies an additional correction of +4.5\%
to the ratio of efficiencies, since the control channel is observed to be more
sensitive to occupancy than the signal channel.
An $\sim$ 8\% S-wave contribution under the \ensuremath{\kaon^{*0}}\xspace resonance in the \ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace
channel has been observed by BaBar \cite{swave-bbar}, and the data in a $\pm$70
\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace mass interval around the \ensuremath{\kaon^{*0}}\xspace mass \cite{note2} yields a
($9.0\pm3.6$)\% extrapolation to the $\pm$150 \ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace mass window. The S-wave
background doubles for the \ensuremath{\kaon^{*0}}\xspace\ensuremath{\Kbar^{*0}}\xspace final state, and it may certainly
have a different coupling for both channels. Our direct measurement reported
in section~\ref{sec:selyield} of (19$\pm$9)\% is still lacking precision to be
used for this purpose. When evaluating the branching fraction, we have assumed a
9\% S-wave contribution, and assigned a systematic error of 50\% to this
hypothesis. A summary of the various contributions to the systematic error can
be seen in Table \ref{tab:systematicsBR}.
\begin{table}
\centering
\caption{Estimated systematic error sources in the ${\ensuremath{\cal B}\xspace}\left( \ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace\right)$ measurement.}
\begin{tabular}{|l|c|}
\hline
Systematic effect & Error ($\%$) \\
\hline
Trigger efficiency & 11.0 \\
Global angular acceptance & 7.2 \\
S-wave fraction & 5.0 \\
Background subtraction & 4.7 \\
\begin{minipage}{1.8in} \ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\!\ensuremath{\kaon^{*0}}\xspace and $\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}\xspace\ensuremath{\rightarrow}\xspace\mu\mu$ BR uncertainty \end{minipage}& 4.6 \\
Selection efficiency & 3.4 \\
\hline
Total & 15.9 \\
\hline
\end{tabular}
\label{tab:systematicsBR}
\end{table}
Our final result is
\begin{equation}
\begin{split}
{\ensuremath{\cal B}\xspace}(\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace) = ( 2.81 & \pm 0.46\ ({\rm stat.}) \\
& \pm 0.45\ ({\rm syst.}) \\
& \pm 0.34\ (f_s/f_d) ) \times 10^{-5}. \nonumber
\end{split}
\end{equation}
As we have seen at the end of section \ref{sec:Kpolarization}, unequal
normalization factors arise upon time integration of individual polarization
amplitudes with well-defined \ensuremath{C\!P}\xspace-eigenvalues. This has the interesting
implication that the time-integrated flavor-averaged branching fraction ($B_1$)
as determined above cannot be directly compared with theoretical predictions
solely formulated in terms of the decay amplitudes
${\mathcal{A}_L}^2+{\mathcal{A}_{\parallel}}^2+{\mathcal{A}_{\perp}}^2$
($B_0$). Meson oscillation needs to be taken into account, since two distinct
particles with different lifetimes are involved. Owing to the fact that
$\mathcal{A}_{\perp}$ is \ensuremath{C\!P}\xspace-odd, the relationship between these quantities
reads as follows
\begin{equation}
B_0 = B_1 \left(1 +
\frac{\Delta\Gamma}{2\Gamma}(f_L+f_\parallel-f_{\perp}) \right).
\end{equation}
According to our measurements of $f_L+f_\parallel-f_{\perp}$, the correction is
small (3\% if current values are taken for $\Delta\Gamma$), and we do not apply
it to our measurement.
\section{Conclusion}
The $b\rightarrow s$ penguin decay $\ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace$ has been observed for the first
time. Using $35$ \ensuremath{\mbox{\,pb}^{-1}}\xspace of $pp$ collisions at 7 \ensuremath{\mathrm{\,Te\kern -0.1em V}}\xspace centre-of-mass energy,
LHCb\xspace has found $49.8 \pm 7.5 $ signal events in the mass interval $\pm 50
\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace$ around the \ensuremath{\B^0_s}\xspace mass. Analysis of the $K^+\pi^-$ mass distributions
shows that most of the signal comes from \ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace, with some S-wave
contribution. The branching fraction has been measured, with the result
{${\ensuremath{\cal B}\xspace}\left( \ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace\right) = (2.81 \pm 0.46 ({\rm stat.}) \pm 0.45 ({\rm
syst.}) \pm 0.34\, (f_s/f_d) )\times10^{-5}$.} The \ensuremath{C\!P}\xspace-averaged longitudinal
\ensuremath{\kaon^{*0}}\xspace polarization fraction has also been measured to be $f_L = 0.31 \pm 0.12
({\rm stat.}) \pm 0.04 ({\rm syst.})$, as well as the \ensuremath{C\!P}\xspace-odd component
$f_{\perp} = 0.38 \pm 0.11 ({\rm stat.}) \pm 0.04 ({\rm syst.})$.
When we consider our measurement in association with that of~\cite{bbar},
it is remarkable that the longitudinal polarization of the \ensuremath{\kaon^{*0}}\xspace mesons seems
to be quite different between \ensuremath{\B^0_s}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\!\ensuremath{\Kbar^{*0}}\xspace ($f_L = 0.31 \pm 0.12 ({\rm stat.})
\pm 0.04 ({\rm syst.})$) and \ensuremath{\B^0}\xspace \ensuremath{\rightarrow}\xspace \ensuremath{\kaon^{*0}}\xspace\ensuremath{\Kbar^{*0}}\xspace ($f_L = {0.80}^{+0.10}_{-0.12} ({\rm
stat.}) \pm 0.06 ({\rm syst.})$), despite the fact that the two decays are
related by a U-spin rotation. However, the ratio of the branching ratios of \ensuremath{\B^0_s}\xspace
and \ensuremath{\B^0}\xspace decays is consistent with $1/\lambda^2$ where $\lambda$ is the
Wolfenstein parameter, as expected.
\section*{Acknowledgements}
\noindent We would like to thank J. Mat\'{\i}as for useful discussions.
We express our gratitude to our colleagues in the CERN accelerator
departments for the excellent performance of the LHC. We thank the
technical and administrative staff at CERN and at the LHCb institutes,
and acknowledge support from the National Agencies: CAPES, CNPq,
FAPERJ and FINEP (Brazil); CERN; NSFC (China); CNRS/IN2P3 (France);
BMBF, DFG, HGF and MPG (Germany); SFI (Ireland); INFN (Italy); FOM and
NWO (The Netherlands); SCSR (Poland); ANCS (Romania); MinES of Russia and
Rosatom (Russia); MICINN, XuntaGal and GENCAT (Spain); SNSF and SER
(Switzerland); NAS Ukraine (Ukraine); STFC (United Kingdom); NSF
(USA). We also acknowledge the support received from the ERC under FP7
and the Region Auvergne.
|
\section{Introduction}
An independent set in a graph $G=(V,E)$ is a set $I \subset V$ of vertices such that no two vertices in $I$ are adjacent. The independence
number of $G$, denoted $\alpha(G)$, is the size of the
largest independent set in $G$. Determining the independence number of a graph is one of the most pervasive and fundamental problems in graph theory. The independence number naturally arises when studying other fundamental graph parameters like the chromatic number (minimum size of a partition of $V$ into independent sets), clique number (independence number of the complementary graph), minimum vertex cover (complement of a maximum independent set), matching number (independence number in the line graph) and many others.
Throughout this paper, we suppose that $G$ is a graph with $n$ vertices and average degree $t$. Turan's \cite{turan} basic theorem of extremal graph theory, in complementary form, states that $\alpha(G) \geq \lceil n/(t+1)\rceil $
for any graph $G$. This bound is tight, as demonstrated by the complement of Turan's graph $G=\overline{T(n, r)}$ which, in the case $n=kr$ is the disjoint union of $r$ cliques, each with $k$ vertices (then $\alpha(G)=r$ and $t=k-1$).
Since $G$ contains large cliques it is natural to ask whether Tur\'an's bound on $\alpha(G)$ can be improved if we prohibit cliques of a prescribed (small) size in $G$.
In \cite{trifreeaks}, Ajtai, Koml\'os, and Szemer\'edi showed that if $G$ contains no $K_3$, then this is indeed the case, by improving Tur\'an's bound by a factor that is logarithmic in $t$. More precisely, they proved that if $G$ is triangle-free, then
$$\alpha(G) \geq \frac{n}{100t}\log{t}.$$
Shortly after, Shearer \cite{trifreeshearer} improved this to
$\alpha(G) \geq (1-o(1))\frac{n}{t}\log{t}$ (assume for convenience throughout this paper that $\log=\log_2$). Random graphs \cite{trifreeshearer} show that for infinitely many $t$ and $n$ with $t=t(n)\rightarrow \infty$ as $n \rightarrow \infty$, there are $n$-vertex triangle-free graphs
with average degree $t$ and independence number $(2-o(1))((n/t)\log t)$. Consequently, the results of \cite{trifreeaks, trifreeshearer} cannot be improved apart from the multiplicative constant.
There is a tight connection between the problem of determining $\alpha(G)$ and questions in Ramsey theory. More precisely, determining the minimum possible $\alpha(G)$ for a triangle-free $G$ is equivalent to determining the Ramsey number $R(3, k)$, which is the minimum $n$ so that
every graph on $n$ vertices contains a triangle or an independent set of size $k$. Moreover, the above lower bounds for $\alpha(G)$ are equivalent to the upper bound $R(3,t) = O(t^2/\log t)$. It was a major open problem, dating back to the 1940's, to determine the order of magnitude of $R(3,t)$, and this was achieved by
Kim \cite{trifreekim} who showed that for every $n$
sufficiently large, there exists an $n$-vertex triangle-free graph $G$ with $\alpha(G)<9\sqrt{n \log n}$. As a consequence, the upper bound $R(3,t)=O(t^2/\log t)$ from \cite{trifreeaks} is of the correct order of magnitude.
In this paper, our goal is to take the result of Ajtai, Koml\'os, and Szemer\'edi~\cite{trifreeaks} further by not only finding an independent set of the size guaranteed by their result, but by showing that many of the vertex subsets of approximately that size are independent sets.
\begin{definition}
Given a graph $G$, let $i(G)$ denote the number of independent sets in $G$.
\end{definition}
Upper bounds for $i(G)$ have been motivated by combinatorial group theory.
In \cite{upperalon}, Alon showed that if $G$ is a $d$-regular graph, then $i(G) \leq 2^{(1/2+o(d))n}$;
he also conjectured that
$$i(G) \leq (2^{d+1}-1)^{n/2d}.$$
Kahn~\cite{upperkahn} proved this conjecture for $d$-regular bipartite graphs.
Galvin~\cite{uppergalvin} obtained a similar bound for any $d$-regular graph $G$, namely
$$i(G) \leq 2^{n/2(1+1/d+c/d\sqrt{\log{d}/d})}$$ for some constant $c$ .
Finally, Zhao~\cite{upperzhao} recently resolved Alon's conjecture.
In this paper, we consider lower bounds for $i(G)$.
This problem is fundamental in extremal graph theory, indeed, the Erd\H os-Stone theorem ~\cite{linearerdos} gives a lower bound for $i(G)$ that is the correct order of magnitude {\em provided $n/t$ is a constant}. More recently, the problem in the range $t=\Theta(n)$ has been investigated by Razborov~\cite{tridensityrazborov},
Nikiforov~\cite{tridensitynikiforov}, and Reiher.
For example, the results of Razborov and Nikiforov determine $g(\rho,3)$,
the minimum triangle density of an $n$-vertex graph with edge density $\frac{1}{2}<\rho<1$.
Looking at the complementary graph, this gives tight lower bounds on the number of independent
sets of size three in a graph with density $1-\rho\sim\frac{t}{n}$.
Lower bounds
for $i(G)$ appear not to have been studied with the same intensity when $t$ is much smaller than $n$, in particular, when $t \rightarrow \infty$ and $t/n \rightarrow 0$.
Let us make some easy observations that are relevant for our work here. We assume that $\alpha:=\alpha(G) \le n/4$.
Since every subset of an independent set is also independent, Turan's theorem implies
\[
i(G) \geq 2^{\alpha} \geq 2^{n/(t+1)}.
\]
In Section \ref{general}, we will improve this to
\begin{equation} \label{easyprop}
i(G) \geq 2^{\frac{1}{250}\frac{n}{t}\log{t}}.
\end{equation}
Our proof uses the standard probabilistic argument which establishes the order of magnitude given by Tur\'an's bound on $\alpha(G)$. This result is certainly not new, and we present it only to serve as a warm-up for our main result in Section \ref{trifree}. Let us observe below that the result is essentially tight.
As no subset of size more than $\alpha(G)$ is independent, an easy upper bound on $i(G)$ (using $\alpha \le n/4$)
is
\begin{equation}\label{easyup}
i(G) \leq \sum_{i=0}^{\alpha}{\binom{n}{i}} \leq
2\binom{n}{\alpha}.
\end{equation}
Since $\alpha(\overline{T(kr,r)}) = r = n/(t+1)$ (recall that $n=kr$ and $t=k-1$), this bound implies that as $n \rightarrow \infty$
\[
i(\overline{T(kr,r)}) \leq 2\binom{kr}{r} \leq 2(ek)^{r} = 2^{1+r\log{ek}} = 2^{(1+o(1))\frac{n}{t}\log{t}}.
\]
Thus, apart from the constant, the exponent in \eqref{easyprop} cannot be improved.
Our main result addresses the case where $G$ contains no triangles. As in the case
of the independence number, prohibiting triangles improves the bound in \eqref{easyprop}.
\begin{thm} {\bf (Main Result)} \label{mr}
Suppose that $G$ is a triangle-free graph on $n$ vertices with average degree
$t$, where $t$ is sufficiently large. Then
\begin{equation} \label{formula}i(G) \geq 2^{\frac{n}{2400t}\log^2{t}}.\end{equation}
\end{thm}
\noindent
Suitable modifications of Random graphs provide constructions of $n$-vertex triangle-free graphs $G$ with average degree $t=t(n)\rightarrow \infty$ as $n \rightarrow \infty$, and $\alpha(G)=O((n/t) \log t)$. Plugging this into (\ref{easyup}), we see that Theorem \ref{mr}
is tight (apart from the constant) for infinitely many $t$. However, it remains open if the theorem is sharp for all $t$ where $t=n^{1/2+o(1)}$. Indeed, the open problem that remains is to obtain a sharp lower bound on $i(G)$ for triangle-free graphs with no restriction on degree. Since all subsets of the neighborhood of a vertex of maximum degree are independent, $i(G) > 2^t$. Combining this with (\ref{formula}) we get
$$i(G)> \max\{2^t, 2^{\frac{n}{2400t}\log^2{t}} \} > 2^{c n^{1/2} \log n}$$
for some constant $c>0$. We conjecture that this can be improved as follows.
\begin{conj} There is an absolute positive constant $c$ such that every $n$-vertex triangle-free graph $G$ satisfies
$$i(G) > 2^{c n^{1/2} (\log n)^{3/2}}.$$\end{conj}
The conjecture, if true, is sharp (apart from the constant in the exponent) by the graphs (due to Kim~\cite{trifreekim} and more recently Bohman~\cite{trifreebohman})
which show that $R(3,t)=\Omega(t^2/\log t)$. Indeed, their graphs are triangle-free and have independence number
\[
\alpha(G) = \Theta(t) = \Theta(\frac{n}{t}\log{t}) = \Theta(\sqrt{n\log{n}}),
\]
so $i(G) \leq 2^{O(\sqrt{n}\log^{3/2}{n})}$ by \eqref{easyup}.
As mentioned before, throughout the paper, all logarithms are base 2. For a graph $G$,
let $n(G), e(G)$ and $t(G)$ denote the number of vertices, edges, and average degree of $G$.
\section{General case}\label{general}
In this section, we give the simple proof of (\ref{easyprop}). Our purpose in doing this is to familiarize the reader with
the general approach to the proof of Theorem \ref{mr} in the next section.
\begin{prop}
If G is a graph on n vertices with average degree $t$, where $2 \leq t \leq \frac{n}{800}$, then
$i(G) \geq 2^{\frac{1}{250}\frac{n}{t}\log{t}}$.
\end{prop}
\begin{proof}
Set $k=\lfloor \frac{1}{100}\frac{n}{t} \rfloor$. Pick a $k$-set uniformly at random from all $k$-sets
in $V(G)$. Let $H$ be the subgraph induced by the $k$ vertices.
Then
\begin{equation*}
\Ex[e(H)] = \frac{1}{2}nt\frac{\binom{n-2}{k-2}}{\binom{n}{k}}
= \frac{1}{2}nt\frac{k(k-1)}{n(n-1)} <\frac{1}{2}\frac{tk^2}{n}.
\end{equation*}
Recall that Markov's inequality states that if $X$ is a positive random variable and $a>0$,
then $\Pr[X \geq a] \leq \Ex[X]/a$; hence $\Pr[e(H) \geq 2\frac{1}{2}\frac{tk^2}{n}] \leq 1/2$.
So for at least half of the choices for $H$, $e(H) \leq \frac{tk^2}{n}$. Therefore,
the number of choices of $H$ for which $e(H) \leq \frac{tk^2}{n}$ is at least
\begin{equation}\label{hct}
\frac{1}{2}\binom{n}{k} \geq \frac{1}{2}(\frac{n}{k})^k >
2^{\frac{k}{2}\log{n/k}} =
2^{\frac{k}{2}(\log{n}-\log{k})}.
\end{equation}
Now, if $e(H) < \frac{tk^2}{n} = \frac{1}{100}k$, then at most $\frac{1}{50}k$ of the
vertices in $H$ have degree at least one. This in turn implies that $H$ contains an independent
set $I$ of size at least $\frac{49}{50}k$. The set $I$ can be obtained from any $H$
which contains it; the number of ways to pick the $\frac{1}{50}k$ vertices of $H-I$ is at most
\begin{equation}
\binom{n}{k/50} \leq (\frac{50ne}{k})^{k/50} =
2^{\frac{k}{50}\log{50ne}-\frac{k}{50}\log{k}} \leq
2^{\frac{k}{50}(\log{n}-\log{k}) + \frac{k}{50}\log{100t}}.
\end{equation}
Combining this with \eqref{hct} and using $\frac{20}{23}\frac{1}{100}\frac{n}{t} < k \leq \frac{1}{100}\frac{n}{t} $ for $t \leq \frac{n}{800}$,
\begin{equation*}
i(G) \geq
2^{k(\frac{1}{2}-\frac{1}{50})(\log{n}-\log{k})-\frac{k}{50}\log{100t}} \geq
2^{k\frac{24}{50}\log{100t}-\frac{k}{50}\log{100t}} =
2^{k\frac{23}{50}\log{100t}} >
2^{\frac{1}{250}\frac{n}{t}\log{t}}.
\end{equation*}
\end{proof}
\section{Triangle-free graphs}\label{trifree}
In this section we prove our main result, Theorem \ref{mr}. We begin with some modifications of a
lemma from \cite{trifreeaks} (see the proof of Lemma 4 in \cite{trifreeaks}).
\begin{lemma}\label{akslemma} {\bf (Ajtai-Koml\'os-Szemer\'edi~\cite{trifreeaks})}
Suppose that $G$ is a triangle-free graph on $n$ vertices with average degree $t$, and
let $k \leq n/100t$. Let $H$ be the subgraph consisting of $k$ vertices chosen
uniformly at random from all the $k$-sets contained in $\{v \in V(G): deg(v) \leq 10t\}$.
Let $M$ be the subgraph of $G$ consisting of vertices adjacent to no vertex in $H$.
Let $n'$ and $t'$ denote the number of vertices and average degree of $M$. Then the
random variables $H$ and $M$ satisfy:
\begin{align}
\Ex[n(M)] &> n(1-\frac{k}{n-t})^{t+1} > \frac{9n}{10} \label{exvm} \\
\Ex[e(M)] &> \frac{nt}{2}(1-\frac{k}{n-20t})^{20t+1} > \frac{nt}{10} \label{exem} \\
\Ex[e(H)] &= \frac{1}{2}nt\frac{k(k-1)}{n(n-1)} \leq \frac{tk^2}{n} \label{exeh} \\
\Var[n(M)] &< \frac{2nk(t+1)(10t+1)}{n-k-20t-2} < nt \label{varvm} \\
\Var[e(M)] &< 2400kt^4 < 40nt^3 \label{varem} \\
\Var[e(H)] &\leq \frac{tk^2(10k+n)}{n^2} \label{vareh}
\end{align}
Further, if $e(M) < (1+\delta)\Ex[e(M)]$ and $n(M) > (1-\delta)\Ex[n(M)]$, then
$n'/t' > \nu n/t$, where $\delta = 800\sqrt{t/n}$ and $\nu=1-1/t-c_{10}\sqrt{t/n}$
for some positive constant $c_{10}$.
\end{lemma}
\bigskip
\noindent
{\bf Remark.}
Ajtai-Koml\'os-Szemer\'edi state their lemma for $k = n/100t$ and prove each of the
first inequalities in \eqref{exvm}, \eqref{exem}, \eqref{varvm}, and \eqref{varem} for all $k$.
They prove each of the second inequalities for $k = n/100t$, but it is easily observed that
they continue to hold for $k < n/100t$.
\bigskip
The next lemma is implied by the computation in the proof of Lemma 4 from \cite{trifreeaks}.
However, the last statement of Lemma \ref{thelemma} is crucial to our proof of Theorem \ref{mr}, so we
make the computations in \cite{trifreeaks} explicit.
\begin{lemma}\label{thelemma}
Suppose $G$ is a triangle-free graph on $n>2^{50}$ vertices with average degree $t \leq 2\sqrt{n}\log{n}$ and
$k \leq n/100t$. Then $G$ contains a subgraph $H$ with $n(H) = k$
$e(H) \leq k/50$.
Moreover, if $M$ is the subgraph of $G$ consisting of vertices adjacent to no vertex
in $H$, then
\begin{enumerate}
\item \label{vertcond} $n(M) > n(G)/2$ and
\item \label{ratiocond} $n(M)/t(M) > \nu n/t$, where $\nu = 1-1/t-c_{10}\sqrt{t/n}$.
\end{enumerate}
Further, if the vertices in $H$ are chosen uniformly at random from all the $k$-sets
contained in $\{v \in V(G): deg(v) \leq 10t\}$, then at least half of the choices
for $H$ satisfy $e(H) \leq k/50$, along with conditions \ref{vertcond} and \ref{ratiocond}.
\end{lemma}
\bigskip
\begin{proof}
Recall that for a random variable $X$ and $a>0$, Chebyshev's inequality states that
$\Pr[|X-\Ex[X]| \geq a] \leq \Var[X]/a^2$. Thus, with $a = k/50-\Ex[e(H)]$,
Lemma \ref{akslemma} implies
\begin{align*}
\Pr[e(H) \geq k/50] &\leq \frac{tk^2(10k+n)}{n^2(k/50-\frac{k^2t}{2n})^2} \\
&=\frac{t(10k+n)}{n^2(1/50 - \frac{kt}{2n})^2} \\
&\leq \frac{t(n/10t +n)}{n^2(1/50-1/200)^2} \\
&= \frac{1/10 + t}{n(3/200)^2} \\
&< 5000\frac{t}{n} \\
&\leq 5000\frac{2\log{n}}{\sqrt{n}} \\
&\leq 1/1000.
\end{align*}
So with probability at most 1/1000, the condition $e(H) \leq k/50$ fails.
Set $\delta = 800\sqrt{t/n}$. Again by Lemma \ref{akslemma} and Chebyshev,
with $a=\delta\Ex[n(M)]$,
\begin{equation*}
\Pr[n(M) \leq n/2] \leq \Pr[n(M) \leq (1-\delta)\Ex[n(M)]] \leq
\frac{nt}{(\frac{9n}{10})^2800^2\frac{t}{n}} \leq 1/1000.
\end{equation*}
Thus the probability that condition \eqref{vertcond} fails is at most 1/1000.
With $a=\delta\Ex[e(M)]$,
\begin{equation*}
\Pr[e(M) \geq (1+\delta)\Ex[e(M)]] \leq
\frac{40nt^3}{\frac{nt}{10}800^2\frac{t}{n}} = 1/160.
\end{equation*}
Since $\Pr[e(M) \geq (1+\delta)\Ex[e(M)] \text{ or } n(M) \leq n/2] <
1/160 + 1/1000$, the last assertion of Lemma \ref{akslemma} implies
that the probability of condition \eqref{ratiocond} failing is at most
$1/160 + 1/1000$. Therefore, the probability that condition $e(H) \leq k/50$
fails or condition \eqref{vertcond} fails or condition \eqref{ratiocond} fails
is at most $1/1000 + 1/1000 + 1/160 + 1/1000 < 1/2$.
\end{proof}
Our proof of Theorem \ref{mr} is achieved by analyzing Algorithm \ref{algorithm} below.
The algorithm is a slight modification of the algorithm from \cite{trifreeaks} that
yields an independent set of size $\frac{1}{100}\frac{n}{t}\log{t}$. Recall that $c_{10}$
is the constant that appear in Lemma \ref{akslemma}.
\begin{algorithm} [ht]
\KwIn{Triangle-free graph $G$ with $n$ vertices, average degree $t$}
\KwOut{Independent set $I$}
$M_o = G$\;
$R = \lfloor (\log{t})/2 \rfloor$\;
\For{$i\leftarrow 0$ \KwTo $R$} {
$n_i = $ number of vertices in $M_i$\;
$t_i = $ average degree in $M_i$\;
$\nu_i = 1 - 1/t_{i-1} - c_{10}\sqrt{t_{i-1}/n_{i-1}}$\;
\uIf{$i = 0$ \emph{or} $\nu_i > 1-1/\log{t}$} { \nllabel{condnu}
Apply Lemma \ref{thelemma} with $G = M_i$ and $k = \lfloor \frac{1}{200}\frac{n}{t} \rfloor$\;
$M_{i+1}, H_{i+1} = M, H$ from Lemma \ref{thelemma}\; \nllabel{getmh}
}
\Else{
$I = $ Independent set in $M_{i-1}$ of size $\lceil n_{i-1}/(t_{i-1}+1) \rceil$\;
\Return{$I$}\; \nllabel{retturan}
}
}
$H = H_1 \cup \dots \cup H_R$\; \nllabel{lineh}
$I = $ Independent set in $H$ of size $\lceil \frac{48}{50}kR \rceil$\; \nllabel{linei}
\Return{$I$}\; \nllabel{reth}
\caption{Independent set algorithm}
\label{algorithm}
\end{algorithm}
\begin{thm}\label{algthm}
Suppose Algorithm \ref{algorithm} is run on a triangle-free graph $G$ with
$n$ vertices and average degree $t$, where $2^{100} < t < \sqrt{n}\log{n}$ and $n > (3c_{10})^{12}$.
If Algorithm \ref{algorithm} terminates at line \ref{retturan},
then $|I| > \frac{1}{2}\frac{n}{t}\log^2{t}$. Otherwise, for each iteration $i=0,...,R-1$,
Algorithm \ref{algorithm} successfully applies Lemma \ref{thelemma} to the graph
$M_i$ to obtain a graph $H_{i+1}$ with $k=\lfloor \frac{1}{200}\frac{n}{t} \rfloor$ vertices.
Moreover, the graph $H$ in line \ref{lineh} is the disjoint union of the $H_i$, and
the independent set $I$ in line \ref{linei} consists of $\frac{48}{50}kR$ vertices from $H$.
\end{thm}
\begin{proof}
We break our proof into two cases, depending on whether Algorithm \ref{algorithm}
terminates at line \ref{reth} or \ref{retturan}.
\noindent {\bf Line \ref{reth}}: We need to show that Lemma \ref{thelemma} can be applied
at every iteration and that the graph $H$ in line \ref{lineh} contains an
independent set of size at least
$\frac{48}{50}kR \geq \frac{1}{500}\frac{n}{t}\log{t}$.
If $i=0$, then $k \leq n/100t$, $t \leq \sqrt{n}\log{n}$,
and $n > t > 2^{50}$, so we may apply Lemma
\ref{thelemma} to obtain graphs $M_1$ and $H_1$, where $|V(H_1)|=k$.
Suppose that $i > 0$ and that Lemma \ref{thelemma} was successfully applied at all previous iterations.
Using \ref{vertcond} of Lemma \ref{thelemma}, $i<R$, and $R=\lfloor (\log{t})/2 \rfloor <(\log{n})/2$,
\begin{equation}\label{nibound}
n_i \geq n/2^i > n/2^R > \sqrt{n} > \sqrt{t} > 2^{50}.
\end{equation}
By the condition in line \ref{condnu}, $\nu_i > 1-1/\log{t}$ for each
iteration $i$. So by \ref{ratiocond} of Lemma \ref{thelemma},
$\frac{n_i}{t_i} \geq \frac{n}{t}\nu_1\nu_2\dots\nu_i > \frac{n}{t}(1-1/\log{t})^R$. Thus
\begin{equation*
\frac{n_i}{t_i} > \frac{n}{t}(1-1/\log{t})^R >
\frac{n}{t}(1-\frac{R}{\log{t}}) \geq \frac{1}{2}\frac{n}{t}.
\end{equation*}
In particular,
\begin{equation}\label{lemmak}
k \leq \frac{1}{200}\frac{n}{t} \leq \frac{1}{100}\frac{n_i}{t_i},
\end{equation}
and also, $t < \sqrt{n}\log{n}$ and $n_i < n$ yield
\begin{equation}\label{lemmat}
t_i < 2n_i\frac{t}{n} \leq 2n_i\frac{\log{n}}{\sqrt{n}}
\leq 2n_i\frac{\log{n_i}}{\sqrt{n_i}} = 2\sqrt{n_i}\log{n_i}.
\end{equation}
The inequalities \eqref{nibound}, \eqref{lemmak}, and \eqref{lemmat} ensure that we may
again apply Lemma \ref{thelemma} with $M_i, M_{i+1}, H_{i+1}, $ and $k$ playing the roles
of $G, M, H, $ and $k$, respectively. Applying Lemma \ref{thelemma} $R$ times yields a collection of
sparse graphs $H_1, H_2, \dots, H_R$, each with $k$ vertices. Each
$H_i$ contains at most $\frac{2}{50}k$ vertices of degree at least one,
so each $H_i$ contains an independent set of size at least $\frac{48}{50}k$.
By definition of $M_i$, these independent sets may be combined into one
independent set of size at least $\frac{48}{50}kR$.
\noindent {\bf Line \ref{retturan}}: Suppose that the algorithm terminates at line
\ref{retturan} during iteration $i+1$. Then $n_i$, $t_i$ (and $n_{i+1}$, $t_{i+1}$)
have been defined and $\nu_{i+1} = 1-1/t_i-c_{10}\sqrt{t_i/n_i}$.
If $t_i \leq (\frac{3}{2})^{2/3}$, then $1/t_i > 1/\log^3{t}$.
Assume $t_i > (\frac{3}{2})^{2/3}$. Then
\begin{equation}\label{ti1}
\frac{2}{3}t_i^{-1/3} > 1/t_i.
\end{equation}
By \eqref{nibound}, $n_i > \sqrt{n}$, so for $n > (3c_{10})^{12}$,
\[
n_i^3 > n^{3/2} > (3c_{10})^6n > (3c_{10})^6t_i.
\]
This implies
\begin{equation}\label{ti2}
\frac{1}{3}t_i^{1/3} > c_{10}\sqrt{t_i/n_i}.
\end{equation}
Combining \eqref{ti1} and \eqref{ti2} yields
\begin{equation*}
1-t_i^{-1/3} < 1-1/t_i - c_{10}\sqrt{\frac{t_i}{n_i}} = \nu_{i+1} \leq 1-1/\log{t}.
\end{equation*}
Thus $1/t_i > 1/\log^3{t}$. Since $t \geq 2^{100}$ (which implies that $t > (\log^5{t}+\log^2{t})^2$)
and $i < R \leq \frac{\log{t}}{2}$,
\begin{equation*}
\frac{n_i}{t_i+1} > \frac{n}{2^i}\frac{1}{(\log^3{t}+1)} \geq
\frac{n}{\sqrt{t}}\frac{1}{(\log^3{t}+1)} = \frac{n}{t}\frac{\sqrt{t}}{(\log^3{t}+1)} >
\frac{n}{t}\log^2{t}.
\end{equation*}
Turan's theorem now implies that $M_i$ contains an independent set of size at least
$\frac{n_i}{t_i+1} > \frac{1}{2}\frac{n}{t}\log^2{t}$.
\end{proof}
\bigskip
We now complete the proof of Theorem \ref{mr} by obtaining a lower bound on the number of
outcomes given by line \ref{reth} of Algorithm \ref{algorithm}.
\noindent
{\bf Proof of Theorem \ref{mr}.}
Recall that we are to show that if $G$ is a triangle-free graph on $n$ vertices with
average degree $t$ sufficiently large, then $i(G) \geq 2^{\frac{n}{2400t}\log^2{t}}$.
Assume $t > \max\{(3c_{10})^{12}, 2^{100}\}$.
Then $n > t > (3c_{10})^{12}$.
Also, if $t > \sqrt{n}\log{n}$, then $G$ has
a vertex whose neighborhood contains at least
$t > \frac{n}{\sqrt{n}}\log{n} > \frac{n}{t}\log^2{n}$ vertices. Since $G$
is triangle-free, this neighborhood forms an independent set, which
contains at least $2^{\frac{n}{t}\log^2{n}} > 2^{\frac{n}{2400t}\log^2{t}}$ subsets, which are also independent.
Thus we may assume that $t \leq \sqrt{n}\log{n}$; in particular,
$G$ satisfies the hypotheses of Theorem \ref{algthm}.
If the algorithm terminates at line \ref{retturan}, then $G$ contains
an independent set of size at least $\frac{n}{2t}\log^2{t}$; so
$G$ contains at least $2^{\frac{n}{2t}\log^2{t}} > 2^{\frac{n}{2400t}{\log^2{t}}}$ independent sets,
and we are done.
Thus we may assume that Algorithm \ref{algorithm} terminates at line \ref{reth}.
Consequently, at each iteration $i$, the algorithm applies
Lemma \ref{thelemma} to pick a sparse graph with $k=\lfloor \frac{1}{2}\frac{n}{100t} \rfloor$ vertices.
The vertices in this graph are chosen from
\[
L_i = \{v \in V(M_i): deg(v) \leq 10t_i\}.
\]
Note that
\[
n_it_i = \sum_{v\in L_i}{deg(v)}+\sum_{v\in V(M_i)-L_i}{deg(v)} \geq \sum_{v\in V(M_i)-L_i}{deg(v)} \geq (n_i-|L_i|)10t_i.
\]
This, together with \ref{vertcond} in Lemma \ref{thelemma}, implies $|L_i| \geq \frac{9}{10}n_i > \frac{9}{10}n_{i-1}/2 \geq \frac{9}{10}n/2^i$.
At least half of the $k$-sets in $L_i$ satify the conditions of Lemma \ref{thelemma},
so the number of choices for $H_i$ is at least
\[
\frac{1}{2}\binom{|L_i|}{k} \geq \frac{1}{2}\binom{.9n/2^i}{k}.
\]
Therefore, the number of choices
for the sequence $H_1,\dots,H_R$ is at least
\begin{align}\label{thct}
\prod_{i=0}^{R-1}{\frac{1}{2}\binom{|L_i|}{k}} \geq
\frac{1}{2^R}\prod_{i=0}^{R-1}{\binom{.9n/2^{i}}{k}}
&\geq \frac{1}{2^R}(\frac{.9n}{k})^{kR}2^{-kR^2/2} \notag \\
&= 2^{kR\log{.9n}-kR\log{k}-kR^2/2-R} \notag \\
&= 2^{kR(\log{n}-\log{k})-kR^2/2+kR\log{.9}-R} \notag \\
&> 2^{kR(\log{n}-\log{k})-\frac{kR}{2}\frac{\log{t}}{2}-kR-R} \notag \\
&> 2^{kR(\log{n}-\log{k})-\frac{kR}{4}\log{t}-2kR}.
\end{align}
Recall that Algorithm \ref{algorithm} obtains an independent set $I$ of size
$\lceil \frac{48}{50}kR \rceil$ from the graph $H=H_1 \cup \dots \cup H_R$. For a fixed
$I$, the number of graphs $H$ that yield $I$ is at most the number of
possibilities for $H-I$. This is at most $\binom{n}{|V(H-I)|}$,
which is at most
\begin{align}\label{thict}
\binom{n}{\frac{2}{50}kR} \leq (\frac{50ne}{2kR})^{\frac{2}{50}kR}
&= 2^{\frac{2}{50}kR\log{n} + \frac{2}{50}kR\log{50e} - \frac{2}{50}kR\log{2kR} } \notag \\
&< 2^{\frac{2}{50}kR(\log{n}-\log{k}) + \frac{2}{50}kR\log{50e}} \notag \\
&< 2^{\frac{2}{50}kR(\log{n}-\log{k}) + kR}.
\end{align}
For a fixed $H$, the number of partitions $H_1\cup\dots\cup H_R = H$ is
at most the number of partitions of $kR$ elements into $R$ sets of size $k$,
which is less than
\begin{equation}\label{tpct}
\binom{kR}{k}^R \leq (Re)^{kR} = 2^{kR\log{Re}} < 2^{kR\log{R} +2kR} < 2^{kR\frac{1}{4}R +2kR} \leq 2^{kR\frac{1}{8}\log{t} +2kR}.
\end{equation}
Since each $H$ yields an independent set, the total number of independent
sets that can be returned at line \ref{reth} of the algorithm is at least
\begin{equation*}
\frac{\text{\# of ways to obtain $H$}}
{(\text{\# of $H$ that yield a fixed $I$})(\text{\# of partitions that yield $H$})}.
\end{equation*}
Since $\frac{n}{268t} \leq k \leq \frac{n}{200t}$ and $R > (\log{t})/3$,
\eqref{thct}, \eqref{thict}, and \eqref{tpct} imply that this is at least
\begin{align*}
2^{\frac{48}{50}kR(\log{n}-\log{k}) - \frac{5}{8}kR\log{t}-5kR}
& \geq 2^{\frac{48}{50}kR\log{200t} - \frac{5}{8}kR\log{t}-5kR} \\
&> 2^{\frac{134}{400}kR\log{t}} \\
&> 2^{\frac{134}{1200}k\log^2{t}}\\
&> 2^{\frac{1}{2400}\frac{n}{t}\log^2{t}}.
\end{align*}
\qed
\bibliographystyle{amsplain}
|
\section{Introduction}
The properties of the host
galaxies of the different types of AGN and their environment, up to
several hundred kpc, can give us valuable information on the nature of the general AGN
population, as well as on different properties of each AGN subtype.
In addition, the availability nowadays of large automatically
constructed galaxy catalogues, like the SDSS, can provide
the necessary statistical significance for these type of analyses.
However, great caution should be taken when interpreting
results based on large databases, since the larger the sample size the
less control usually one has on the spectral and other details of the individual galaxy entries.
It could then be difficult to address important questions, such
as : Do the Unification paradigm explains all cases of type 1 and type 2 AGN?
What is the true connection between galaxy interactions, star formation and
nuclear activity? What is the lifetime of these phenomena? How do LINERs
fit in the general picture and can all be considered AGN? Do
evolutionary trends affect the AGN phenomenology?
Nowadays, it is widely accepted that the accretion of
material into a massive black hole (MBH), located at the galactic
center, is responsible for the detected excess emission (radiation not
emitted by stellar photospheres) in the AGN's spectra and such black
holes do exist in all elliptical galaxies and spiral galaxy bulges
(Kormendy and Richstone 1995; Magorrian et al. 1998), including our
own (e.g. Melia \& Falcke 2001). However, we still lack a complete
understanding of the various aspects of activity, for example the
triggering mechanism and the feeding of the black hole, the physical
properties of the
accretion disk and the obscuring torus predicted by the unified
scheme (Antonucci et al. 1993), the origin of jets in radio loud
objects, the connection with star formation and
the role of the AGN feedback. Even the exact mechanism that
produces the observed IR, X-ray, and gamma-ray emission, is still only partially
understood (e.g. Le{\'o}n-Tavares et al. 2011). The unification model itself, although successful in many
cases, has not been able to fully explain all the AGN phenomenology (among others, the role of
interactions on induced activity; Koulouridis et al. 2006a,b and
references therein).
Despite observational difficulties and limitations, there have been
many attempts, based on different diagnostics, to investigate the
possible triggering mechanisms of nuclear activity. Most agree that
the accretion of material into a MBH (Lynden-Bell 1969) is the
mechanism responsible for the emission, but it is still necessary to
understand the feeding mechanism of the black hole. It is known and
widely accepted that interactions between galaxies can force gas and
molecular clouds towards the galactic center, where they become
compressed and produce starburst events.(e.g. Li et al. 2008;
Ellison et al. 2008; Ideue et al. 2012). Many also believe that the
same mechanism could give birth to an active nucleus
(e.g. Umemura 1998; Kawakatu et al. 2006; Ellison et al. 2011;
Silverman et al. 2011, Villforth et al. 2012). Despite the fact that
the exact mechanism is still unknown, in the local universe a minimum
accretion rate of $\sim 10^{-6\pm1} M_\odot$/yr is needed in order to fuel
the black hole (Ho 2008). At such low accretion rates, compared to the host galaxy, nuclear activity is probably relatively weak and most of the spectral signatures of the AGN are "buried". Theoretically the feeding of the black hole can only be achieved by means of a non axisymmetric perturbation which induces mass inflow. Such
perturbations can be provided by interactions and the result of the
inflow is the feeding of the black hole and the activation of the AGN
phase, maybe $\sim \rm 50-250\; Myr$ after the initial interaction
took place (see below). An interaction certainly predicts such a time
delay, since after the material has piled up around the inner Linblad
resonance, enhancing star formation, it can be channeled towards the
nucleus by loosing significant amounts of angular momentum, a process
which is not instantaneous.
Indeed, post starburst stellar populations have been
observed around AGN (Dultzin-Hacyan \& Benitez 1994; Maiolino \& Rieke 1995;
Nelson \& Whittle 1996; Hunt et al. 1997; Maiolino et al. 1997; Cid Fernandes,
Storchi-Bergmann \& Schmitt 1998; Boisson et al. 2000, 2004;
Cid Fernandes et al. 2001, 2004, 2005) and in close proximity to the
core ($\sim$50 pc). This fact implies the continuity of
these two states and a delay of 50-250 Myr between the onset of the
starburst and the feeding of the AGN (e.g., M{\"u}ller S{\'a}nchez et
al. 2008; Wild, Heckman, \& Charlot 2010; Davies et al. 2012), which
may reach the peak of its activity
after $\sim$500 Myr (Kaviraj et al. 2011). Ballantyne, Everett \&
Murray (2006) studying the Cosmic XRay Background (CXRB) concluded
that Seyfert galaxies (dominating in the production of the CXRB) are
likely fueled by minor mergers or interactions that can trigger a
circumnuclear star formation event, but that there may be a
significant delay between the interaction and the ignition of the
nucleus. Davies et al. (2007), analyzing star formation in the nuclei
of nine Seyfert galaxies found recent, but no longer active,
starbursts which occurred 10 - 300 Myr ago. Further support for an
interaction-activity relation was recently provided by HI observations
of Tang et al. (2008), who found that 94\% of the Seyfert galaxies in
their sample were disturbed in contrast to their control sample (where
only 19\% were disturbed), but see also Georgakakis et al. (2009) and
Cisternas et al. (2011) in the AEGIS and cosmos surveys, respectively.
This paper is the third in a series of 3-dimensional studies of the
environment of active galaxies (Koulouridis et al. 2006a,b),
extending previous 2D analyses (Dultzin et al. 1999, Krongold et
al. 2002) in an effort to shed more light to the starburst/AGN
connection and to the evolutionary scenario, triggered by interactions, proposed in our previous
papers. It is a follow-up spectroscopic pilot study aiming at investigating the possible
effects of interactions on the neighbours of our Seyfert
galaxies and understanding the conditions necessary for the different types of
activity.
In \S 2 we will discuss our galaxy samples and we will present our observations
and data reduction.
The spectroscopic analysis and classification of the galaxies, basic host galaxy properties, results from STARLIGHT stellar population synthesis code and the analysis
of the available X-ray
observations are presented in section \S 3. Finally, in section \S 4 we
will interpret our results and draw our conclusions. All distances are calculated taking into account the local velocity field (which includes the effects of the following structures : Virgo, Great Attractor and Shapley) for the standard $\rm\Lambda$CDM cosmology ($\Omega_m$=0.27, $\Omega_\Lambda$=0.73). Throughout our paper we use $H_{\circ}=100 h$ km s$^{-1}$ Mpc$^{-1}$, following our previous study on the same samples.
\section{Data}
\subsection{Sample Definition and Previous Results}
The samples of active galaxies were initially compiled from the
catalogue of Lipovetskyj, Neisvestnyj \& Neisvetnaya (1987), which
itself is a compilation of all Seyfert galaxies known at the time from
various surveys and in various frequencies (optical, X-ray, radio,
infrared). It includes all extended objects and several starlike
objects with absolute magnitudes lower than -24. Available multifrequency data
are listed, including: coordinates,
redshifts, Seyfert type (and sub-type), UBVR-photoelectric
magnitudes, morphological types, fluxes in $\rm H\beta$ and [OIII]5007, JHKLN fluxes,
far-IR (IRAS) fluxes,
radio fluxes at 6 and 11 cm, monochromatic X-Ray fluxes in 0.3 - 3.5 and 2
- 10 keV. All data can be found online at the vizier database
(http://vizier.cfa.harvard.edu/viz-bin/VizieR?-source=VII/173). About
half of the listed Seyfert galaxies can also be found in the IRAS
catalogue.
Dultzin-Hacyan et al. (1999) selected from the catalogue two volume
limited and complete samples, consisting of 72 Sy1 and 60 Sy2, to
study their projected circumgalactic environment. In Koulouridis et al. (2006a)
we used practically the same samples in order to verify their results, using in
addition redshift data from the CFA2 and SSRS surveys and our own
deeper spectroscopic observations.
Well selected control samples
(same redshift, diameter and morphology distributions) were used for
the comparison in both studies.
Using the CfA2 and SSRS redshift catalogues, and our own
deeper low-resolution spectroscopic
observations (reaching to $m_B\sim 18.5$), we searched for neighbours
within a projected distance $R\leq 100$ h$^{-1}$ kpc and a radial
velocity separation $\delta u\leq 600$km/sec and we found that:
\begin{itemize}
\item The Sy1 galaxies and their control sample show a similar (consistent within
1$\sigma$ Poisson uncertainty) fraction of objects having at least one close
neighbour.
\item There is a significantly higher fraction of Sy2 galaxies
having a near neighbour, especially within $D\leq 75$ h$^{-1}$ kpc,
with respect to both their control sample and the Sy1 galaxies.
\item The large-scale environment of Sy1 galaxies (D = 1 $h^{-1}$ Mpc
and $\delta u\leq 1000$km/sec) is denser than that
of Sy2 galaxies, although consistent with their respective control samples.
\item Using deeper spectroscopic observations of the neighbors for a random subsample
of 22 Sy1 and 22 Sy2 galaxies we found that the differences between
the close environment of Sy1 and Sy2's persists even when going to
fainter neighbours, correspond to a magnitude similar to that of the
Large Magellanic Cloud.
\end{itemize}
For the purposes of the present study we obtained new
medium-resolution spectroscopy, in order to resolve the $H\alpha$ and [NII]
lines - unresolved in our original low-resolution spectra, of
all the neighbours around the aforementioned subsamples
of the 22 Sy1 \& 22 Sy2, respectively. In Table 1 and 2
we present the names, celestial coordinates, $O_{MAPS}$ magnitudes
\footnote{$O$
(blue) POSS I plate magnitudes of the Minnesota Automated Plate
Scanner (MAPS) system. We use $O_{MAPS}$ magnitudes because Zwicky magnitudes were not
available for the fainter neighbours, and we needed a homogeneous
magnitude system for all our objects.} and redshifts of the
Sy1 and Sy2 galaxies which have at least one close neighbour (within
$\delta u <600$km/sec). The full samples are presented in detail in Koulouridis et al. (2006).
Note that we have kept the original neighbours enumeration of
the previous papers (for example, in table 2, NGC1358 has only
neighbour 2, since neighbour 1 had $\delta u > 600$km/sec).
\subsection{Spectroscopic Observations}
We have obtained medium-resolution spectroscopic
data of all the neighbouring galaxies in our samples in order to classify
them according to their optical emission lines (\S 2.3).
Optical spectra were taken with the Boller \& Chivens spectrograph
mounted on the 2.1m telescope at the Observatorio Astron\'omico
Nacional in San Pedro M\'artir (OAN-SPM). Observations were carried out during
photometric conditions. All spectra were obtained with a 2$\farcs$5 slit.
The typical wavelength range
was 4000-8000 \AA\ and the spectral resolution R=8\AA\ .
Spectrophotometric standard stars were observed every night.
The data reduction was carried out with the IRAF\footnote{IRAF is
distributed by National Optical Astronomy Observatories
operated by the Association of Universities for Research in
Astronomy, Inc. under cooperative agreement with the National
Science Foundation.} package following a standard procedure. Spectra
were bias-subtracted and corrected with dome
flat-field frames. Arc-lamp (CuHeNeAr) exposures were used for
wavelength calibration. All spectra can be found in Appendix A.
\subsection{Analysis and Classification Method}
In this section we present results of our spectroscopic observations
of all the neighbours with $D\le100\;h^{-1}$ kpc and $m_{O_{MAPS}}\mincir
18.5$ for the samples of Sy1 and Sy2 galaxies. We have also
used SDSS spectra when available.
Our aim was to measure six emission lines: H$\beta \; \lambda
4861$, H$\alpha \; \lambda 6563$, [NII] $\lambda 6583$, [OIII]
$\lambda 5007$, [SII] $\lambda 6716$ and [SII] $\lambda 6731$,
in order to
classify our galaxies, using the Baldwin, Phillips \& Terlevich (1981,
hereafter BPT) and Veilleux \& Osterbrock (1987) diagrams.
For the cases that it was not possible to measure the H$\beta$ and [OIII] emission lines,
we use the more approximate classification by Stasi{\'n}ska et al. (2006).
Based on the above, we adopted the following classification scheme:
\begin{itemize}
\item absorption line galaxies (ALG), i.e. galaxies with no emission lines.
\item galaxies with emission lines (ELG), meaning that they exhibit
nuclear or/and recent star forming activity.
\end{itemize}
Flux ratios for the emission lines mentioned above have been measured
after subtracting the host galaxy contamination from each
spectrum. We disentangle the spectral contribution of the host galaxy
from the observed spectra by using the stellar population synthesis
code \texttt{STARLIGHT}\footnote{http://starlight.ufsc.br/}.
Spectra processing and fits were carried in the same fashion as
described in section 3.1 of Le\'on-Tavares et al. (2011). For a
detailed description of the \texttt{STARLIGHT } code and its
scientific results, we refer to the papers of the SEAGal
collaboration (Cid-Fernandes et al. 2005, Mateus et al. 2006; Asari et
al. 2007; Cid-Fernandes et al. 2007). We only note that we have calculated the 1$\sigma$
standard deviation of the flux as follows (Tresse et al. 1999):
\[\sigma=\sigma_cd\sqrt{2N_{pix}+EW/d}\,\,\,\,\,(1)\]
where $\sigma_c$ is the standard deviation of the continuum about the
emission line, $d$ is the spectral dispersion in \AA\ per pixel and
$N_{pix}$ is the base-width of the emission line in pixels. In our case the parameter $d \sim 4$\AA/pix, while for the SDSS spectra is $\sim 1.1$\AA/pix for the H$\beta$ area and $\sim 1.5$\AA/pix for the H$\alpha$ area.
To the above we have added in quadrature the errors of the Gaussian fitting of the emission lines. We should note here that in some cases the B telluric band is very close to the [SII] doublet (see for example NGC 1019-N2 on the left of the doublet or UGC 7064-N1B on the right of the doublet) introducing a further uncertainty on the calculation of the flux. In all such cases we have simultaneously fitted the telluric absorption and the emission lines in order to have a better measure of the [SII] doublet's flux. Although we do not have an exact evaluation of the uncertainty due to the above spectral feature, we presume (at least for the cases that the B telluric band is close to the doublet) that the reported error is underestimated. \footnote{We should also note that the standard deviation of the continuum about the [SII] doublet was calculated after the subtraction of the B telluric band.}
Although it is possible to
distinguish between a Star Forming Nucleus (SFN) galaxy\footnote{We choose to call SFN all
galaxies with prominent emission lines that do not show AGN activity.} and an AGN using only the
[NII]/H$\alpha$ ratio, we cannot distinguish between a low ionization
(LINER) and a high ionization (Seyfert) AGN galaxy. We have also
measured [OI] ($\lambda=6300$) when possible, as an extra
indicator of AGN activity. However, the weakness of the line in most cases did
not allow further use of it in a separate BPT diagram.
\begin{figure}
\centering
\resizebox{8cm}{8cm}{\includegraphics{Seyferts_final.ps}}
\resizebox{8cm}{8cm}{\includegraphics{Seyferts_sii.ps}}
\caption{
(a) BPT diagram, and (b) Veilleux \& Osterbrock classification diagram of
the neighbours of Sy1 and Sy2 galaxies. The neighbours of Sy2 and Sy1 galaxies are
indicated by black triangles and red squares, respectively. For clarity, only the errorbars for those
galaxies with the largest uncertainties in the [OIII]/H$\beta$ ratio are presented in the diagrams.}
\end{figure}
In Fig.1a we plot the line ratios log([OIII]/H$\beta$) versus log([NII]/H$\alpha$)
(BPT diagram) for those neighbours of Seyfert galaxies for which we
have available the four necessary emission lines
\footnote{We have excluded one merger neighbour
(UGC7064-N1) since its two nuclei are in an advanced merging state and
their properties are most probably independent of any interaction
which may have with the central active galaxy}. We also
plot the Kauffmann et al. (2003a) separation line between SFN and AGN
galaxies, given by:
$$\log([{\rm OIII}]/\rm H\beta)=\frac{0.61}{(\log([{\rm
NII}]/H\alpha)-0.05}+1.3\;,$$
and the corresponding one of Kewley et al. (2001):
$$\log([{\rm OIII}]/\rm H\beta)=\frac{0.61}{(\log([{\rm NII}]/\rm H\alpha)-0.47}+1.19
\;.$$
We also plot in Fig.1b the line ratios log([OIII]$/\rm H\beta$) vs
log([SII]$/\rm H\alpha$). Qualitatively, the same results as those
presented in Fig.1a are repeated
here as well. The dividing line is given by Kewley at
al. (2006). However, we do not have the respective line of Kauffmann
et al. (2003a), as
it is not available in the literature, and thus we cannot separate
pure star forming galaxies from
composite objects. Since, as we have already discussed, the measurement of the [SII] doublet is probably contaminated by absorption of the B telluric band, we will draw our results based on the [NII] forbidden line.
We can now classify our objects in the following categories:
\begin{itemize}
\item SFN: all the objects which are found below the line of Kaufmann et al.
\item AGN: the objects which are found above the line of Kewley et al.
\item TO (transition object): the ones that are found
between the two lines and exhibit characteristics of both nuclear activity and recent star formation.
\end{itemize}
We do not attempt to divide the star forming galaxies into more
subcategories since such a categorization appears to be highly subjective
and depends on the applied methodology (e.g. Knapen \& James, 2009).
Note that for one of our objects (ESO 545-G013-N1) the $\rm H\beta$ and [OIII] $(\lambda5007)$ lines were not observed and therefore we classified it using the more approximate method of Stasi{\'n}ska et al. (2006), which is based solely on the NII/H$\alpha$ ratio.
in order to evaluate this method, we applied it to all our galaxies and found a consistency with the BPT classification in all cases but one (see Table 3).
Further classification of the Seyfert galaxies in type 1 and type 2 was
obtained by direct visual examination of the spectra from the broadening of the
emission lines. No broad lines were discovered in the spectrum of the two neighbours classified as AGN
and therefore they should be considered as type 2.
In Tables 1 and 2 we list, for all neighbours, their line ratios and the two different
classifications.
We also measured the equivalent width of the H$\alpha$ emission
line, in order to use it as an extra indicator of the galaxy's star forming history, in addition with the STARLIGHT code's results. The minimum Equivalent Width, defined as the integrated local continuum rms noise normalized to the level of the local continuum, at a 5$\sigma$ confidence level, is found to be EW$_{min}\sim 1$\AA.
We should note here that in a small number of cases the [OIII] and H$\beta$
lines were detected only after the subtraction of the continuum. We
calculated the 1$\sigma$ standard deviation of the EW as follows
(Tresse et al. 1999):
\[\sigma_{EW}=\frac{EW}{F}\sigma_cd\sqrt{2N_{pix}+EW/d+(EW/d)^2/N_{pix}}\,\,\,\,\,(2)\]
where $\sigma_c$ is the standard deviation of the continuum about the
emission line, $d$ is the spectral dispersion in \AA\ per pixel,
$N_{pix}$ is the base-width of the emission line in pixels and F the
flux of the emission line.
\section{Results and analysis.}
\subsection{Activity of the neighbours.}
In this section we discuss in more detail the results of our
spectroscopy and classification. We have excluded the merging neighbour of UGC 7064, since
the properties of its two nuclei are more affected by their mutual
interaction rather than
by their neighbouring Seyfert.
We can draw our first results for
each sample separately inspecting Table 1 and 2.
From the analyzed 15
neighbours of Sy1 only 4 are ALGs, while 8 of them are
SFNs, 2 are classified as TOs and one is classified as AGN. Similar results
hold for the neighbours of Sy2 galaxies. 4 out of 13 neighbours do not
present emission lines, 6 are SFNs and 3 are TOs.
Therefore, at least 70\% of the neighbours, within $100 \; h^{-1}$ kpc, of both type of
Seyfert galaxies have emission lines. We should note
here that Ho et al. (1997), studying a magnitude limited sample of
galaxies ($B_T\leq$12.5), came up with a similar high percentage of
activity (86\%). However, the results of our sample of faint
neighbours cannot be directly compared with those of Ho et al. due to
the brighter magnitude limit of the latter.
We can extract one of the most interesting results of our analysis by
examining Fig.1, i.e., that the neighbours of Sy2 galaxies have
systematically larger values
of [OIII]$/\rm H\beta$ than the neighbours of Sy1 galaxies. Using a
Kolmogorov-Smirnov two-sample test for the [OIII]$/\rm H\beta$ ratio
we find that the null hypothesis that the samples are drawn from the
same parent population is rejected at a 99.9\% level. Especially for
those galaxies that exhibit only star-formation, the ratio
[OIII]/H$\alpha$ is mainly related to their ionization level. This
fact could be an indication of a more recent starburst event in the
neighbours of Sy2 galaxies than of Sy1's, caused possibly by the
interaction with a neighbouring galaxy, or an effect of the galaxy
downsizing i.e. more massive galaxies have formed their stellar
populations earlier than less massive ones (Asari et al. (2007) argue that the location of galaxies on the BPT diagram is considered to be a result of downsizing).
Should the downsizing explanation be true, the ionization level can be
considered as an indicator of metallicity, which is closely related to
the stellar mass. Thus, galaxies having lower values of
[OIII]/H$\beta$ would be more massive and would have higher
metallicities, indicative of an older average age of the stellar
population. In Table 1 and 2 we can see a weak trend of the mean
stellar metallicity ($<Z>$) values (extracted from STARLIGHT) for the
Sy2 SFN neighbours being lower with respect to that of the Sy1's and although
stellar masses cannot be directly derived from our data, most low
metallicity SFNs are also faint and small in size (Table 3). However,
no trend can be found by comparing the average age of the stellar
populations, and given the small number of galaxies these results remain rather
inconclusive.
The Equivalent Width of the $\rm H\alpha$ emission line is also a good
indicator of the star formation history, since it represents the ratio
of present to past star formation, i.e. during a starburst event young
massive stars strengthen the emission lines and enhance their EW, but
as time passes the strength of the emission line fades, the continuum
rises again and the value of the EW declines. The highest values of
the EW($\rm H\alpha$) can be found in the spectra of the Star Forming
neighbours of our Sy2 sample, while on the other hand some of the lowest values
can be found in the respective Sy1's neighbours spectra.
A more direct way to to explore the possibility that the differences
of the ionization level is due to the age of the interaction of the
central active galaxies with its neighbour, is by the determination of the age of the
most recent peak of star formation with the "STARLIGHT" code. As it
was expected however, most of the star forming galaxies present a
recent event within the last 20 Myr, a necessary fact in order to
detect strong emission lines and we can not detect any significant
differences between Sy1 and Sy2 SFN neighbours. On the other hand, an
interesting result is the fact that six out of seven Sy2's non-SFN
(ALG, AGN or TO) companions present a recent star formation peak $<$30
Myr, while six out seven Sy1's corresponding neighbours are "quiet"
for more than 100 Myr. The above fact may indicate that indeed the Sy1
galaxies have interacted with their neighbour earlier than the Sy2s.
Summarizing our main results of this section:
\begin{itemize}
\item More than 70\% of the neighbours of the two AGN samples exhibit optical
emission lines, indicating recent star formation and/or nuclear activity.
\item Around 30\% of the neighbors of Sy1 and Sy2 galaxies show the presence
of AGN activity, mainly in the form of TOs.
\item The neighbours of Sy2s are systematically more ionized than the
neighbours of Sy1s and their EW(H$\alpha$) values tend also to be higher.
\item Most of the non-SFN neighbours of Sy2 galaxies show a recent
starburst event ($<$30 Myr), while the corresponding age for most of
the Sy1's neighbours is $>$100 Myr.
\item the previous two results indicate differences in the star
formation history of the neighbours of different types of AGN as
well as in the age of the most recent interaction.
\end{itemize}
Finally we should note how close to a
composite state are the neighbours of active galaxies, in agreement
with Kewley at al. (2006a) who showed
that the star forming members of close pairs, lie closer to the
classification line than the star forming field galaxies.
We suggest that galaxies between the curves of Kauffmann
et al. (2003) and Kewley et al. (2001) possibly migrate from a pure
star forming phase to
a pure AGN phase. This suggestion is of great importance to the
formulation of a possible
evolutionary scenario and will also be discussed further in \S4.
\subsection{Magnitude and distance analysis}
Since we have already applied a homogeneous magnitude system to our samples,
we can now study whether there is a correlation between the activity
of an interacting pair of galaxies and their magnitudes.
The activity-magnitude comparison is performed by examining the
absolute magnitude difference between the neighbour
and the central active galaxy ($\Delta M$),
with small values ($\Delta M <1.5$) indicating a stronger pair interactions (we tag these
pairs as equally bright).
A further parameter that can be used is the absolute magnitude of
the neighbour, indicating its size. On average, absolute
magnitude and size are correlated in small redshift intervals (as it is in our
case) and therefore we can safely presume that a faint galaxy
is also small in size and a bright one is large. The latter has been also optically
inspected for our galaxies to further confirm the correlation (see
also maps of Fig.3), while the median absolute magnitude $M=-17.49$ is
considered to be the separating limit between bright and faint
companions. In addition we also examine the isophotal diameters at
25.0 B-mag arcsec$^{-2}$) from the Third Reference Catalogue of bright
galaxies (RC3) to compare with the absolute magnitudes, by considering
any neighbour with $D/D_{AGN}<1/2$ as being small. In two cases,
because of lack of RC3 data, their near-infrared isophotal diameters
(at 20.0 K-mag arcsec$^{-2}$) from the "Two Micron All Sky Survey Team
2MASS Extended Objects" (2MASS) catalogue) were used for the
comparison. We should note that only in the case of NGC 1241 the
diameter criterion is not in agreement with the absolute magnitude
criterion (marginally) and by inspecting also the SDSS image we
concluded that the neighbour is indeed small.
Finally, radial separation can also be considered as a crucial factor
of the strength of the interaction.
In Table 3 we list all the above mentioned values plus three
respective indices than take values between 0 and 1. With 1 we denote a
value that is in favor of the interaction, with 0 the opposite. In
more detail, if the radial separation $R$ is less than 50$h^{-1}$kpc
the respective index $I_D$ is 1 and the same is true for bright
neighbours and equally bright pairs, since all these factors may
affect positively the interactions between two galaxies. The sum of
the three indices is also listed in Table 3. Obviously the strength of
the interaction of a neighbour with the sum of the three indices equal
to 0 (i.e. small and faraway
neighbour of a large AGN) would be significantly different from one
with a sum equal to 3 (i.e. large and close galaxy of a comparable sized AGN).
It therefore becomes evident that:
\begin{itemize}
\item All faint neighbours and all neighbours of a non-equally
bright pair of galaxies are preferentially absorption line or purely
SFN.
\item All neighbours which host an AGN or are transition objects (TO), fall in
the bright category and are neighbours of an equally bright pair.
\item All neighbours with interaction indices sim$\leq 1$ are purely star forming galaxies.
\item All ALGs, AGN and TO galaxies have interaction indices sum$\geq 2$ (except NGC863-N1).
\end{itemize}
From our results we can infer that when a faint/small galaxy comes in interaction with another galaxy,
the encounter induces at most a starburst but no AGN activity in the small galaxy; however it can trigger a bright AGN in the larger one. This could be due to the absence in small galaxies
of a massive black hole (Wang, Kauffmann 2007; Volonteri et al. 2008).
If this assertion is correct, only galaxies which experience a major close
interaction or merger can exhibit AGN
activity and this could be the reason why AGN hosts are more frequently found in
early type galaxies (e.g. Marquez \& Moles 1994; Moles, Marquez, \&
Perez 1995; Ho et al. 1997; Knapen et al. 2000; Wake et
al. 2004). This can also account for the large fraction of star forming
galaxies among our samples of neighbours.
To cover all aspects of this issue, we should mention here that Galaz
et al. (2011) showed that the fraction of low surface brightness
galaxies hosting an AGN is significantly lower than the corresponding fraction
of high surface brightness galaxies, independently of the mass.
So the deficiency of AGN in faint galaxies seems to be due to an intrinsic
inability of these galaxies to host or to feed a massive black hole.
Our results indicate that the interaction of a bright galaxy
especially in an equally bright pair results in an AGN or an ALG.
Finding some massive galaxies, members of an equally bright
interacting pair, without emission lines implies either a non-eventful
interaction or a delay of the outcome of the interaction. On the other hand, weak star formation
or low luminosity nuclear activity may not be detectable by optical
spectroscopy, although it could possibly be detected in X-rays. Such
an analysis is presented below.
\subsection{The XMM-Newton observations}
We explore here, using the {\it XMM-Newton} public
archive, whether the neighbours show X-ray
activity. We find that 13 target fields have been
observed by {\it XMM-Newton}. However, some of them are very bright
and have been observed in partial window mode, rendering the
observations in center of the Field-of-View unusable (NGC5548, NGC863,
1H1142-178, NGC7469).
The list of the remaining observations (13 neighbours and 9 central
Seyfert galaxies) is shown in Table 4, in which
we present X-ray fluxes for the detections as well as upper limits for
the non-detected sources. The fluxes have been taken from
the 2XMM catalogue (Watson et al. 2009).
The fluxes refer to the total 0.2-12 keV band
for the PN detector or the combined MOS detectors in the case
where PN fluxes are not available and are estimated using a
photon index of $\Gamma=1.7$ and an average Galactic column density of
$\rm N_H = 3\times 10^{20}$ $\rm cm^{-2}$. Luminosities were
estimated using the same spectral parameters.
In the same table we quote the 2XMM hardness ratios,
derived from the 1-2 keV and 2-4.5 keV bands (hardness ratio-3
according to the 2XMM catalogue notation).
The upper limits, derived using the {\sl FLIX} software, are
estimated following the method of Carrera et al. (2007).
This provides upper-limits to the X-ray flux at a given
point in the sky covered by {\it XMM-Newton} pointings.
The radius used for deriving the upper limit was 20 or 30 arcsec
depending on the presence of contaminating nearby sources.
In Fig.2 we present the X-ray to optical flux diagram
$f_X-f_B$ (e.g. Stocke et al. 1991).
\begin{figure}
\resizebox{8cm}{8cm}{\includegraphics{fxfo.eps}}
\caption{The X-ray (0.2-12 keV) to optical (B-band) flux
diagram for both the central active galaxy targets (solid circles) and the neighbors
(open circles). The triangles (upper limits) denote the neighbors
with no X-ray detection. The upper, lower solid line and the dash
line correspond to $f_X/f_B=+1, -1, -2 $ respectively. The only neighbour (open circle) that lies in the AGN regime is NGC 7682-N1}
\end{figure}
This diagram provides an idea on whether a galaxy may host an active nucleus.
This is because AGN have enhanced X-ray emission for a given optical magnitude
relative to ALG galaxies. The space usually populated by AGN is shown
between the continuous lines. The central Seyfert galaxies are shown
as filled points, but since X-ray flux has not been corrected for
X-ray absorption,
a number of absorbed AGN galaxies lie between the lower continuous line and the dashed line, while the
heavily absorbed Sy2 NGC 7743 (Akylas \& Georgantopoulos 2009), lie far below the dashed line.
One neighbour which lies in the AGN regime (NGC 7682-N1) can be clearly
seen. This has been classified as a TO galaxy in the
optical spectroscopic analysis and is one of the three
neighbours (for which XMM-Newton observations are available, see Table 4)
having an active nucleus based on optical spectroscopy.
Additional information on the nature of our sources can be extracted from the
hardness ratios.
Two sources NGC 526-N2 and NGC 1358-N2 have hardness ratios suggesting an
absorption of $\rm N_H \approx 10^{22}$ $\rm cm^{-2}$,
consistent with the presence of a moderately obscured active
nucleus. Both these galaxies present no optical emission lines and thus are
classified as ALG, based on their optical spectra. In other words, the
lack of optical emission lines from the
nucleus of these objects could be a result of obscuration and indeed this seems to be the case, since the detection limit of the EW of emission lines is low enough. In
addition, we should mention here that all galaxies among those which fall in fields
observed by XMM-Newton, classified as ALG through
optical spectroscopy, present X-ray emission.
In contrast, all SF galaxies except one
in the X-ray subsample do not show an X-ray detection.
We should note here that unobscured low accretion rate Sy2 objects
and/or low luminosity AGN, where the Narrow Line Region (NLR) cannot
be detected by means of optical spectroscopy, or even X-ray binaries
may account for the X-ray detection of unobscured ALG
galaxies. However, emission from X-ray binaries is not detected in the
spectra of the SFNs rendering this interpretation less plausible. This
analysis therefore implies that the total fraction of neighbours of
AGN that show recent star formation or AGN, based on optical
spectroscopy or X-ray observations, is at least $80\%$ and possibly
quite higher. This matter will be fully addressed in future work.
\section{Discussion \& Conclusions}
In this paper we investigate the close environment ($\leq
100 \; h^{-1}$kpc) of a local sample ($z<0.034$) of
AGN. In particular we explore the spectroscopic, photometric and X-ray properties of 30 neighbouring galaxies
around 10 Sy1 and 13 Sy2 galaxies.
Based on optical spectroscopy, in our current study we have found that
the large majority of these neighbours show some activity, mostly recent
star-formation (emission line spectrum) but AGN as well. In addition,
our X-ray analysis of a subsample of neighbours with public XMM-Newton
observations showed that the neighbours which are classified as ALG
based on optical spectroscopy might have a low luminosity active core,
since all of them are X-ray detected, while two out of five appear to
have a moderately obscured active nucleus. The X-ray detections could
be due to X-ray binaries, but we argue that this is less probable
since the pure star forming neighbours do not show any X-ray emission
down to the flux limit of the available observations. From both
optical spectroscopy and X-ray observations, it becomes clear that the
fraction of AGN's neighbours which exhibit recent star formation
and/or nuclear activity, within 100 $h^{-1}$ Mpc, is $>80\%$ and
possibly higher.
Furthermore, the close neighbours of Sy1 galaxies, especially the
SFNs, are less ionized and have lower values of EW(H$\alpha$) with
respect to those of Sy2 and thus seem to be a different, more evolved
population than those of Sy2s. Other discovered trends of metallicity,
host galaxy size and age of the most recent starburst event indicate
possible physical differences between the neighbours of Sy1 \& Sy2
galaxies as well, a fact that may link AGN activity with interactions.
Indeed, over the past two decades there have been several studies
which supported that the idea of an evolutionary
sequence from starburst to Seyfert galaxies (e.g. Storchi-Bergmann et
al. 2001, see also introduction). Furthermore, there are also studies
that separate type I from type II objects (e.g. Hunt et al. 1997;
Maiolino et al. 1997, Gu et al. 2001) implying that recent
star-formation is only present in type II objects (see also Coldwell
et al. 2009). Based on the number and proximity of close ($\mincir
60-100\;h^{-1}$kpc) neighbours, around different types of active (Sy1,
Sy2 and BIRG) galaxies (e.g. Dultzin-Hacyan et al. 1999; Krongold et
al. 2002; Koulouridis et al. 2006a,b), a very interesting evolutionary
sequence has been suggested, starting with a close interaction that
triggers the formation of a nuclear starburst, subsequently evolving
to a type 2 Seyfert, and finally to a Sy1. Recent observational results by Villarroel et al.(2012) and Kollatschny et
al.(2012) also seem to support this scheme.
This sequence is likely independent of luminosity, as similar trends have been proposed for
LINERs (Krongold et al. 2003) and ULIRGs and Quasars (Fiore et
al. 2008 and references therein). The above findings were also supported by numerical simulations (Hopkins et al. 2008) which outlined such an evolutionary scheme for
merging galaxies. The proposed activity evolution can explain
the excess of starbursts and type 2 AGN in interacting systems, as
well as the lack of type 1 AGN in compact groups of galaxies
(Mart{\'{\i}}nez et al. 2008) and galaxy pairs (e.g. Gonzalez et
al. 2008).
Since the physical properties of the neighbours should be reflected on
the state of the central active galaxy,
we argue that our results may be in the same direction as those of our
previous papers (Koulouridis et al 2006a,b), supporting an
evolutionary sequence of galaxy activity, driven by interactions, the
main path of which follows the sequence of induced star formation, Sy2
and finally Sy1 phase. A time delay should exist between the pure star
forming and AGN phase (see discussion in the introduction), where
active nucleus and circumnuclear starburst coexist. In this initial
phase, the nucleus is heavily obscured by the still star forming
molecular clouds and it can be observed as a transition stage of
composite Sy2-starburst objects. We should note here that according to
Ballantyne, Everett \& Murray (2006) a non-evolving torus cannot
provide the AGN obscuration over all cosmic time and that extra
obscuration by star formation is needed.
The most probable manner for the AGN to dominate is to eliminate the
starburst, possibly by the AGN outflows or by radiation pressure. We
point out that a great theoretical success of the starburst/AGN
connection is the quenching of the induced star formation by the AGN
feedback, which can explain the formation of red and dead elliptical
galaxies (e.g. Springel et al. 2005a; Di Matteo et al. 2005; Khalatyan
et al. 2008). This can be achieved by outflows from the core which
have enough energy to dissipate the material around it and thus
suffocate star formation (e.g. Krongold et al. 2007, 2009; Blustin et
al. 2008; Hopkins \& Elvis 2010; Novak, Ostriker \& Ciotti 2011;
Cano-D{\'{\i}}az et al. 2012; Zubovas \& King 2012). Recent
observational studies and simulations have shown that AGN's ionized
outflows may carry enough energy to cease star formation in the host
galaxy rapidly, in less than 1 Gyr (see for example Kaviraj et al. 2011). As
the starburst fades (see relevant discussion and references in the introduction), the Seyfert 2
state starts dominating, to be followed at the end by a totally
unobscured Sy1 state, plausibly $\sim 1$ Gyr after the initial
interaction (see Krongold et al. 2002). More details about the
co-evolution of the torus and the AGN are given by Liu \& Zhang (2011),
supporting our evolutionary scheme. We should note here that recent
observations (Hasinger et al. 2008; Treister et al. 2010) verified a
significant increase of the type 2 AGN fraction with redshift, a fact
which is in agreement with our evolutionary scheme.
Alternatively, there is a possibility that the SFN neighbours of Sy1
galaxies are systematically more
massive with respect to those of Sy2 and that their older stellar
population is due to downsizing, i.e. more massive galaxies have
evolved earlier, while less massive ones exhibit more recent star
formation and thus younger stellar population. However, there is no
obvious explanation on why more massive galaxies should be located
preferentially near Sy1 galaxies and not Sy2. The combination of both
downsizing and the interaction driven sequence, as presented
previously, can also be at work.
We stress that the suggested evolutionary scenario does not
invalidate completely the unification scheme. It implies that the orientation of the torus can
determine the AGN phenomenology only at specific phases of the evolutionary sequence.
In particular, this probably occurs when the obscuring
molecular clouds form the torus (possibly when the AGN activity
reaches its peak $\sim$0.5 Gyr after the initial interactions (Kaviraj
et al. 2011)) and before being completely swept away (possibly
after 1 Gyr (Krongold et al. 2002)).
From our point of view, in an ever evolving universe an evolutionary
scheme, is more probable than the original
unification paradigm which proposes a rather static view of AGN. Of
course, orientation could and
should also play a role between the obscured Sy2 and Sy1 phase, when
the relaxing obscuring material forms a toroidal structure.
There are still many unresolved issues and caveats concerning
these suggestions, since the evolutionary sequence is not unique and should also
depend on the geometry, the density and other factors of the obscuring and
the accreting material, as well as on the mass of the host galaxy and its
black hole. Furthermore, the sample presented in this pilot study is
rather small and the results should be considered as indicative and
should be confirmed by analysis of larger samples.
\subsection*{Acknowledgments}
EK thanks the IUNAM and INAOE, were a major part of this work was
completed, for their warm hospitality. We also thank OAGH and OAN-SPM staff for excellent assistance and
technical support at the telescopes.
VC acknowledges funding by CONACyT research grants 54480
and 15149 (M\'exico).
MP acknowledges funding by the Mexican Government research grant
No. CONACyT 49878-F and DD
support from grant PAPIIT IN111610 from DGAPA, UNAM.
YK acknowledges support from CONACyT 168519 grant and UNAM-DGAPA PAPIIT IN103712 grant.
This research has made use of the USNO-B catalog (Monet et al. 2003)
and the MAPS Catalog of POSS I (Cabanela et al. 2003) supported by the
University of Minnesota (the APS databases can be accessed at
http://aps.umn.edu/).
The STARLIGHT project is supported by the Brazilian agencies CNPq,
CAPES and FAPESP and by the FranceBrazil CAPES/Cofecub programme.
Funding for the SDSS and SDSS-II has been provided by the Alfred P. Sloan
Foundation, the Participating Institutions, the National Science Foundation,
the U.S. Department of Energy, the National Aeronautics and Space
Administration, the Japanese Monbukagakusho, the Max Planck Society, and the
Higher Education Funding Council for England. The SDSS Web Site is
http://www.sdss.org/.
Finally, we would like to thank the anonymous referee for his or hers comments and suggestions that helped to significantly improve our paper.
|
\section{Introduction}
The calculation via dispersion relations of the ratio of the real part to the imaginary
part of the forward spin-non-flip amplitude, $\rho(s,t) $, does not agree with the data
until one gets to high energies, and it misses all the interesting intermediate-energy
structures.
On the theory side, the situation is very complex and uncertain. Analyticity showed that
one could not do without a real part, while
polarization data proved that it was not possible to ignore spin complications, as
the real part of the spin-non-flip amplitude has a zero, around which the contribution
of the spin-flip amplitude, which decreases quite slowly with energy, cannot be ignored.
On the experimental side, the situation is not entirely clear cut either
\cite{Rev-LHC}, and one of the difficulties is due to the
lack of experimental data at high energies and small momentum transfer.
In this talk, we consider in great detail the situation concerning $\rho(s,t)$.
The model we propose takes into account all known features of the near-forward
proton-proton and proton-antiproton data, {\it i.e.} different slopes for the
spin-non-flip and the spin-flip amplitudes, the value of total cross sections and of
$\rho(s,t)$, the relative phase of the Coulomb and hadron amplitudes and the
form factors of the nucleons.
\section{Impact of the Coulomb-hadron phase}
Let us first compare different approximations
for the Coulomb-hadron interference used in fits to the experimental
$p\bar{p}$-scattering data \cite{Disser-data}. First, we use
the simple West-Yennie form of the relative phase \cite{wy}.
This leads to values for $\rho(s,0)$ shown in the second column of Table 1.
The results show the distribution of the values of $\rho(s)$ extracted
from the experiments. In two cases, they lie slightly above
$\rho_{exp}$ (at $p_L = 4.066, \ 5.603, \ 5.94 \ $ GeV/c); in three cases
they lie considerably higher than $ \ \rho_{exp}\ $ (at $p_L = 5.72, \ 6.23 \ $
GeV/c) and in one case they lie below (at $p_L = 3.7 \ $GeV/c).
If we take the slightly more complicated phase proposed by Cahn \cite{can},
the results are almost the same (see the third column of Table 1).
Finally, if we use the expression derived by one of us \cite{selmp1,PRD-Sum},
taking into account the two-photon amplitude and using a dipole form factor,
the fit gives different values for $\rho(s)$ (see the last column of Table 1):
the results lie above $\rho_{exp}$ for all the considered energies, so that
the difference with the predictions of the dispersion analysis gets worse, as shown in Fig. 1.
\label{sec:figures}
\begin{figure}
\begin{center}
\includegraphics[width=0.45\textwidth] {fig.eps}
\end{center}
\caption{$\rho(s,0)$ - the ratio of the real part to the imaginary part of the
elastic scattering amplitude for proton-antiproton scattering at low
energies. The curve shows the dispersion relation description for
$p \bar{p}$ scattering \cite{Kroll}, and the stars are the result of our analysis.
} \label{Fig_2}
\end{figure}
\begin{table*}
\caption{The dependence of $\rho(s,0)$ on the model used for the Coulomb-hadron phase in proton-antiproton scattering. N is the number of data points.} \label{Table-3}
\vspace{.1cm}
\begin{center}
\begin{tabular}{|c|c|c||c|c|c|} \hline
& & & & & \\
\large{ $p_{L}(GeV/c)$ }& \large{N} &\large{ $\rho_{exper.}$ }\cite{Disser-data}& \large{
$\rho$}(${phase \cite{wy}})$ &
\large{$\rho$}($ {phase \cite{can} }$)& \large{$\rho$}(${phase \cite{selmp1,PRD-Sum}})$ \\
& & & & & \\
\hline
& & & & & \\
3.702 & 34 & $+0.018 \pm 0.03$ & $+0.0077 \pm 0.02$ & $+0.0078 \pm 0.08$ & $+0.028
\pm 0.08$\\
4.066 & 34 & $-0.015 \pm 0.03$ & $+0.0377 \pm 0.02$ & $+0.0378 \pm 0.08$ & $+0.0324
\pm 0.08$\\
5.603 & 215& $-0.047 \pm 0.03$ & $+0.035 \pm 0.02$ & $+0.036 \pm 0.08$ & $-0.0017
\pm 0.08$\\
5.724 & 115& $-0.051 \pm 0.03$ & $+0.0139 \pm 0.02$ & $+0.014 \pm 0.08$ & $-0.0088
\pm 0.08$\\
5.941 & 140& $-0.063 \pm 0.03$ & $-0.0003 \pm 0.02$ & $-0.004 \pm 0.08$ & $-0.0055
\pm 0.08$\\
6.234 & 34 & $-0.06 \pm 0.03$ & $+0.0162 \pm 0.02$ & $+0.0162 \pm 0.08$ & $-0.0216
\pm 0.08$ \\
& & & & & \\
\hline
\end{tabular}
\end{center}
\end{table*}
\section{Impact of the spin-flip amplitude}
In most analyses, one assumes that
the imaginary and real parts of the spin-non-flip
amplitude have an exponential behaviour with the same $t$ slope, and that the
imaginary and real parts of the spin-flip amplitudes, without the
kinematic factor $\sqrt{|t|}$, are proportional to the corresponding spin-non-flip parts of the amplitude, with a proportionality constant independent of $s$.
In \cite{slope-MPL99} it was shown that if the slope of the spin-flip amplitude is bigger than that for spin non-flip, $B_{sf}=2B_{nf}$,
the contribution of the spin-flip amplitude
can be felt in the differential
cross sections of elastic hadron scattering at small $|t|$.
As it is not possible to calculate the hadronic amplitudes
from first principles,
we have to resort to some assumptions about their $s$ and $t$ dependencies \cite{PS-EPA02,CPS-EPA04}.
Here, we use this simple model for the spin-flip
amplitude and study its impact on the determination of $\rho(s,t)$.
We take
the spin-non-flip and spin-flip amplitudes in the simplest exponential form
\begin{eqnarray}
F^{h}_{nf}&=& h_{nf} \ [i+\rho(s,0)] \ e^{B_{nf} t/2};\\
F^{h}_{sf}&=& \sqrt{-t}/m_{p} \ h_{sf} \ [i+\rho(s,0)] \ e^{B^{sf} t/2},
\end{eqnarray}
with $B_{sf}=2B_{nf}$.
The differential cross section in this case will be
\begin{eqnarray}
\frac{d\sigma}{dt} =
2 \pi \ [|F_{nf}|^{2} + 2 |F_{sf}|^{2}
], \label{dsdt-sf}
\end{eqnarray}
where the amplitudes $F_{nf}$ and $F_{sf}$ will include the
corresponding electromagnetic parts and the Coulomb-hadron
phase factors as mentioned previously.
The results of our new fits of the proton-antiproton experimental data
for $ p_L$ in $[3.7,6.2] \ $GeV/c are presented in Table 2.
The changes of $\chi^2$ after the inclusion
of the spin-flip amplitude are measured by the coefficient
\begin{eqnarray}
R_{\chi} \ = \ \frac{\chi^2_{ \ without \ sf.}
- \chi^2_{ \ with \ sf.} }{\chi^2_{ \ without \ sf.}}.
\label{Rchi}
\end{eqnarray}
We again obtain values of $\rho$ close to zero and prevalently positive.
Once again, as seen from Fig. 1, the results do not agree with the prediction by the dispersion analysis \cite{Kroll}.
\begin{table*}
\caption{Spin dependence of proton-antiproton elastic scattering}
\label{Table-3}
\vspace{.1cm}
\begin{center}
\begin{tabular}{|c|c|c||c|c|c|} \hline
& & & & & \\
\large{ $p_{L(GeV/c)}$ } & \large{ N} &\large{ $\rho_{exp.}$ } &\large{$R_{\chi}$}
&\large{ $\rho_{model}$} &\large{ $h_{sf}$, GeV } \\
& & & & & \\ \hline
& & & & & \\
3.702 & 34 & $+0.018 \pm 0.03$ & $8 \%$ & $+0.057 \pm 0.02$ & $ 49.8 \pm 1.4$\\
4.066 & 34 & $-0.015 \pm 0.03$ & $25 \%$ & $+0.052 \pm 0.009$ & $ 48.9 \pm 0.7$\\
5.603 & 215& $-0.047 \pm 0.03$ & $3.5 \%$ & $+0.014 \pm 0.005$ & $ 35.6 \pm 4.$\\
5.724 & 115& $-0.051 \pm 0.03$ & $6.5 \%$ & $+0.023 \pm 0.004$ & $ 38.2 \pm 4.5$\\
5.941 & 140& $-0.063 \pm 0.03$ & $4.5 \%$ & $+0.007 \pm 0.003$ & $ 43.2 \pm 0.4$\\
& & & & & \\
\hline
\end{tabular}
\end{center}
\end{table*}
\section{Conclusion }
The present analysis, which includes the contributions of Coulomb interference and spin effects,
shows a contradiction between the extracted value of $\rho(s,0)$
and the predictions from the analysis based on dispersion relations.
If such a situation is confirmed by future new data from the LHC experiments, it could
reveal new
effects such as , for example, a fundamental length of the order of 1 TeV.
It is likely, however, that the theoretical analysis can be further developed, to
include
additional corrections connected with possible oscillations in the scattering
amplitude and
with the $t$-dependence of the spin-flip scattering amplitude.
We hope that
the forward experiments at NICA will also give valuable information for the improvement of our theoretical understanding of this question.
|
\section{Introduction}
Neutron stars (NSs) can be powered by different sources of energy:
rotation, accretion, residual heat and magnetic field. Up to now,
more than twenty X-ray pulsating sources are identified as
isolated, slowly rotating ($P\sim 2-10~$s) neutron stars, whose
large X-ray luminosity $L_x\sim 10^{33}-10^{35}~$erg/s cannot
usually be explained in terms of rotational energy, at variance
with rotation-powered pulsars (RPPs). Historically, they are
classified as Anomalous X-ray Pulsars (AXPs) or Soft Gamma
Repeaters (SGRs), and they are recognized as magnetars. In these objects, MF decay and evolution is the dominant energy source and crustal deformations produced by the huge internal magnetic stresses are the likely cause of bursts and episodic giant flares. In most cases, the spin-down inferred external magnetic field turns out to be super-strong. Recently, increasing evidence gathered in favor of a unified scenario for isolated neutron stars, their different observed manifestations being mainly due to a different topology and strength of the internal MF, and to evolution (\cite{pons11}).
The typical soft X-ray ($< 10~$keV) spectrum of magnetars is given by a thermal component ($kT\sim 0.3-0.5~$keV), plus a hard tail phenomenologically described by a power law with a photon index $\Gamma\sim 3-4$ \cite{mereghetti08}. This hard tail is believed to be produced by resonant compton (up)scattering (RCS) of photons emitted by the star surface onto mildly relativistic particles flowing in magnetosphere. In order to take into account for this effect, several codes have been developed in recent years \cite{fernandez07,ntz08a,ntz08b,fernandez11}. The two basic inputs are the external MF and a prescription for the particle space and velocity distribution. With these ingredients, and with a given energy distribution for the incoming surface photons (the seed spectrum), the MC simulations of RCS provide the emerging, reprocessed spectrum, which must be compared with observational data.
Magnetars also exhibit a high-energy ($\sim 20$--200 keV) power-law tail, the origin of which is still under debate. One of the possible mechanisms is again RCS, whose efficiency at high energy is very sensitive to MF and plasma velocity. In this paper we discuss the effect of the geometry of the MF, which is obtained from new calculations of magnetostatic, force-free equilibria.
\section{Force-free twisted magnetosphere}\label{sec_forcefree}
The shape of the internal MF of NSs is largely unknown, but it is thought to be complex, with toroidal and poloidal components of similar strength, and with significant deviations from a dipolar geometry. In the crust, a large toroidal field is a necessary ingredient to explain the observed thermal and bursting properties of magnetars, even in the case in which they show a low value of the inferred external dipolar field \cite{turolla11}. The complex internal topology must be smoothly matched to an external solution, so that the internal MF geometry at the star surface has a strong effect on the external MF. In this framework, the RPP phenomenolgy is expected to be better explained by the presence of a NS with a simple, nearly potential, dipolar field, while magnetars activity is compatible with the presence of a complex external field, as a result of the transfer of magnetic helicity from the internal to the external field. A non-potential MF needs supporting currents, and charges can be either extracted from the surface or produced by one-photon pair creation in the strong MF \cite{beloborodov07}.
In the inner magnetospheric region MF lines corotate with the star, with footprints connected with the crustal field, which evolves on long timescales ($10^3-10^5$ years). Furthermore, resistive processes in the magnetosphere act on a typical timescale of years \cite{beloborodov09}, much longer than the typical response of the tenuous plasma, whose Alfv\'en velocity is near to the velocity of light. For these reasons, a reasonable approximation is to build stationary, force-free magnetospheric solutions, assuming that the footprints are anchored in the crust and neglecting resistivity. Then, the ideal MHD force-free equation is a simple balance between electromagnetic forces, provided that inertial, pressure, and gravity terms are negligible
\begin{equation}\label{eq_electromagnetic_forces}
\rho_e \vec{E}+\frac{1}{c}\vec{J}\times\vec{B}=0~,
\end{equation}
where we have introduced the rotationally induced electric field $\vec{E}=-\vec{v}_{rot}\times \vec{B}$, the related charge density $\rho_e$ and the current density $\vec{J}=c\vec{\nabla}\times\vec{B}/4\pi$.
The electric field gives corrections of order $(\rho \Omega /c)^2$. Neglecting such contributions is safe if we consider the region closest to a magnetar surface ($r\lesssim 100 r_\star$), and remembering that the light cylinder for slow rotators lies typically at a few thousands stellar radii ($r_l=c/\Omega$ and magnetar periods are in the 1--10~s range). Therefore, condition (\ref{eq_electromagnetic_forces}) reduces simply to $\vec{J}\parallel \vec{B}$: the particles can flow only along MF lines. Even if the pulsar equation simplifies in this limit, it still contains an arbitrary function, which basically describes the toroidal field as a function of the poloidal one (see \cite{vigano11} for the mathematical formulation). In the literature, only the simplest (semi-)analytical solutions have been derived and later employed in applications. In
particular, in the context of magnetars, the self-similar twisted dipole family of solutions is the most popular choice. Proposed for solar corona studies \cite{low90}, and later extended to the magnetar scenario \cite{tlk02}, in self-similar models the radial dependence of the MF components is described by a simple power law, $B_i\propto r^{-(p+2)}$. These solutions are semi-analytical and described by only one parameter, the radial index $p$ (see \cite{pavan09} for globally twisted multipolar fields). However, globally twisted solutions necessary evolve in time and magnetic configurations in which the twist affects only a limited bundle of field lines are likely required by observations \cite{beloborodov09}.
As an alternative to self-similar models, in \cite{vigano11} we presented a numerical method to build general solutions. The force-free solutions are obtained by iteration starting from an arbitrary, non-force-free poloidal plus toroidal field. By the introduction of a fictitious velocity field proportional to the Lorentz force, the code is designed to dissipate only the part of the current not parallel to the MF. The boundary conditions are an arbitrary, fixed radial MF at the surface, $B_r(r_\star,\theta)$, and the requirement of a continuous matching with a potential dipole at large distances $r_{out}\ge 100 r_\star$. This method is able to produce general configurations, whose features may be radically different from the self-similar solutions. In this work, we try to reproduce some $j$-bundle like structure, aiming at providing input for MC simulations \cite{ntz08b}. In particular, we focus on two models, whose MF and current distribution are shown in Fig. \ref{fig:model12}. The first one is obtained fixing a dipolar form for radial MF, while the second one has a concentration of field lines near the south pole. To get these profiles, we have chosen the $\phi$ component of the vector potential as:
\begin{eqnarray}\label{aphi_models}
&& \mbox{model 1:} \qquad A_\phi(\theta,r_\star)=B_0\frac{\sin\theta}{2} \\
&& \mbox{model 2:} \qquad A_\phi(\theta,r_\star)=B_0\frac{\sin\theta}{2}\left[1+4e^{-\frac{(\theta-0.9\pi)}{0.2^2}}\right]
\end{eqnarray}
from which we recover $B_r=(\vec{\nabla}\times A_\phi\hat{\phi})_r$ ($B_0$ is the value of $B_r$ at north pole). Both models have high helicity, but it is not homogeneously distributed like in self-similar models. Model 1 is characterized by an equatorial asymmetry, with current and twist concentrated in a closed bundle near the equatorial region and a $j-$bundle near the southern semi-axis. Model 2 is more extreme: its helicity is $\sim$ 25 times larger than for model 1. Both are supposed to be more realistic than the self-similar models, since currents are more concentrated near the axis.
\begin{figure}
\centering
\includegraphics[width=.23\textwidth]{./images/blog_scatt_model1bb.eps}
\includegraphics[width=.23\textwidth]{./images/jlog_model1bb.eps}
\includegraphics[width=.23\textwidth]{./images/blog_scatt_model2bb.eps}
\includegraphics[width=.23\textwidth]{./images/jlog_model2bb.eps}
\caption{Numerical model 1 (left) and model 2 (right). First panel:
$|\vec B|$ (grey logarithmic scale) with scattering surfaces for photons of 1, 3 and $5~$keV (here $\beta=0.5$, $\mu=0.5$; see text); second panel: current intensity
$|\vec{J}|/c$ (grey linear scale), in units of $10^{14}~$G$/r_\star$.} \label{fig:model12}
\end{figure}
\section{Resonant Compton scattering}
The MC code we employ is presented in
\cite{ntz08a,ntz08b} and we refer to these works for further
details. Here we just summarize some basic concepts. It is assumed
that currents flow in a non-potential magnetosphere and that the
electron density is large enough for the medium to be optically
thick to resonant scattering: photons produced by the cooling star
surface can be then up-scattered by RCS, populating the hard tail
of the spectrum. In the stellar frame, the resonant energy
corresponds to the Doppler-shifted gyration frequency of the
electron
\begin{equation}
\omega_D(B,\beta)=\frac{1}{\gamma(1-\beta\mu)}\frac{eB}{m_ec}
\end{equation}
where $B=|\vec{B}|$ is the local intensity of MF, $c\beta$ the
electron velocity, $\gamma=1/\sqrt{1-\beta^2}$ the Lorentz
factor, and $\mu$ the cosine of the angle between the photon
direction and the electron momentum. The non-resonant
contributions are negligible, since the resonant cross-section is
up to five orders of magnitude larger than the Thomson
cross-section. Figs. \ref{fig:model12} show the scattering
surfaces in the two models, for photons of 1, 3 and $5~$keV,
assuming $\beta=0.5$ and $\mu=0.5$. The surfaces are far from
being spherically symmetric. In model 2, which has the strongest
helicity, the scattering surfaces of soft X-ray photons lie tens
of stellar radii away from the surface, especially near the axis.
The scattering optical depth for a given seed photon energy depends on the local particle density $n_e$, the velocity of the scatterer and the local value of the MF. Since photons can be pushed to energies of some hundreds keVs, non-conservative scattering and the fully relativistic QED cross sections are used \cite{ntz08b}. The velocity distribution is assumed to be a one-dimensional (along the field line), relativistic Maxwellian distribution with a plasma temperature $T_e$ and a bulk velocity $\beta_{bulk}$, as described in \cite{ntz08a}. The strong simplification in this approach is that the velocity distribution does not depend on position, which reduces the microphysical inputs to two parameters: the temperature of the plasma $T_e$ and $\beta_{bulk}$. Only recently, a more consistent treatment of the charges space and velocity distribution was attempted \cite{beloborodov11}. These new results show that the velocity distribution is non-trivial and strongly depends on position, which opens a new range of possibilities, so far unexplored.
Given a seed spectrum for the photons emitted from the surface
(assumed as a blackbody with temperature $T_\star$), the MC code
follows the photon propagation through the scattering
magnetosphere. When a photon enters in a parameter region where no
more resonant scatterings are possible, the photon escapes to
infinity, where its energy and direction are stored. The sky at
infinity is divided into patches, so that viewing effects can be
accounted for (e.g. if line-of-sight is along the angles
$\theta_s$, $\phi_s$, only photons collected in the patch which
contains that pair of angles are analyzed). The angle-averaged
spectrum is obtained simply by averaging over all patches.
\section{Results and discussion}
In previous works \cite{zane09}, self-similar magnetospheric
models were used to obtain synthetic spectra that were
satisfactorily fitted to X-ray observations, giving typical values
of $\beta_{bulk}\in[0.1,0.7]$ and $\Delta\phi\in[0.4,2.0]$. In
this paper we explore the dependence of the spectrum on the
magnetosphere model by comparing results obtained with
self-similar models and with our new numerical solutions for
force-free magnetospheres. We present results comparing models
with a magnetic field intensity at the pole of $B_p=10^{14}$~G.
The seed spectrum is assumed to be a $kT_\star=0.4~$keV Planckian
isotropic distribution for both ordinary and extraordinary photons. In Figs. \ref{fig:beta_bfinal1} and
\ref{fig:beta09} we plot the angle-averaged spectra, i.e., the
distribution of all photons escaped to infinity, independently on
the final direction. Fig. \ref{fig:beta_bfinal1} shows the effects
produced by changes in the bulk velocity, keeping fixed the MF
configuration (model 1), and the electron temperature ($kT_e=20~$keV). The main fact is the the spectrum in the region $E>10~$keV
becomes harder as $\beta_{bulk}$ is increased, while it is
depleted of photons of energies in the $1-10~$keV range. A similar,
but less pronounced effect, is obtained by increasing the electron
temperature \cite{ntz08b}.
In Fig. \ref{fig:beta09} we compare spectra obtained with fixed
values of $kT_e=20~$keV and $\beta_{bulk}=0.9$, varying only the
MF topology. The three lines correspond to model 1 (solid), model
2 (dashes) and a self-similar model (dash-dotted line) with
$\Delta\phi=1.36$, which approximately has the same total helicity
of model 1. We have chosen a high value of $\beta_{bulk}$ to show
a case where the effects are larger, but our conclusions do not
change qualitatively for lower values of $\beta_{bulk}$. The
high-energy cutoff produced by electron recoil is also evident. We
see that the major differences arising from changes in MF topology
appear in the hard tail of the spectra, which is clearly harder
for the self-similar model. However, the thermal part of the spectrum is more depleted,
especially for model 2.
\begin{figure}
\begin{minipage}[t]{3in}
\centering
\includegraphics[width=.95\textwidth]{./images/beta_bfinal1bb.eps}
\caption{Angle-averaged spectra for $kT_e=20~$keV, model 1, changing $\beta_{bulk}$.}
\label{fig:beta_bfinal1}
\end{minipage}
\hspace{0.5cm}
\begin{minipage}[t]{3in}
\centering
\includegraphics[width=.95\textwidth]{./images/beta09bb.eps}
\caption{Angle-averaged spectra for $kT_e=20~$keV, $\beta_{bulk}=0.9$, changing MF model.}
\label{fig:beta09}
\end{minipage}
\end{figure}
However, in the angle-averaged spectrum, the differences due to
the MF topology are partially smeared out, and the comparison
between synthetic spectra as seen from different viewing angles
(first three panles in Fig. \ref{fig:dir}) is more interesting. The comptonization
degree reflects the inhomogeneous particle density distribution
and therefore the particular current distribution produces
important differences between different viewing angles. Since
neutron stars rotate, the study of pulse profiles and
phase-resolved spectra can trace the geometric features of the
scattering region (tens of stellar radii). In the lower right panel we show the light curves for the three MF models, in bands 0.5--10~keV and 20--200~keV, for an oblique rotator with a certain line of sight ($\xi=\chi=30^\circ$ in \cite{ntz08a}).
For self-similar models (lower left panel), the thermal part of the spectrum shows a smooth dependence with the viewing angle, which in turn translates into relatively regular light curves. In models 1 and 2, differences in the spectra induced by viewing angles are more pronounced. In these cases, the spectrum is much
more irregular and asymmetries are larger. In particular, model 1 (upper left) shows a softer spectrum when seen from northern latitudes, with important
spectral differences. The pulsed fraction of model 1 is very high in the hard X-ray band, while it is comparable with the self-similar model in the soft range. Model 2 (upper right) has a more symmetric distribution of currents, and the comptonization degree at different angles depends on the energy band in a non trivial way. This results in large pulsed fraction for both energy ranges and in notable differences between their pulse profiles.
\begin{figure}
\centering
\includegraphics[width=.47\textwidth]{./images/dir_model1_beta09bb.eps}
\includegraphics[width=.47\textwidth]{./images/dir_model2_beta09bb.eps}\\
\medskip
\includegraphics[width=.47\textwidth]{./images/dir_twist_beta09bb.eps}
\includegraphics[width=.47\textwidth]{./images/fig_lc_3030.eps}
\caption{Synthetic spectra for model 1 (upper left panel), model 2 (upper right), self-similar model (lower left) with the parameters indicated in the figures. Legends indicate the four viewing angles here considered: aligned with north or south pole, just above and just below the equator. The seed blackbody is shown for comparison (solid line). Fourth panel shows the pulse profiles obtained for model 1 (solid line), model 2 (dotted), self-similar (dashed), in the energy ranges 0.5--10 keV (thin lines) and 20--200 keV (thick lines), with $\xi=\chi=30^\circ$ (see text).}
\label{fig:dir}
\end{figure}
The high energy tail, seen at different colatitudes, can vary by one order of magnitude or more. Investigating the geometry can help to recognize the different components seen in the hard tails via pulse phase spectroscopy. On the other hand, the macrophysical approach used in this work to obtain the MF topology should be accompanied by the corresponding microphysical description to determine the velocity distribution of particles, here simply taken as a model parameter. A fully consistent description of both MF geometry and the spatial and velocity distribution of particles is needed to advance our interpretation of magnetar spectra.
\section*{References}
|
\section{Introduction}
John Bell \cite{BELL} has shown that local realistic models are impossible for quantum mechanics of two qubits. After some years researchers started to
ask questions about the Bell theorem for more complicated systems (compare the classic review in which such problems are not covered, \cite{CLAUSER}). First results concerning possible extensions to entangled states of pairs of $d$-state systems,
with $d \geq 3$, arrived in early eighties \cite{MERMIN}.
The early research was confined to Stern-Gerlach
type measurements (which are parametrized by only two numbers, defining the direction of the measurement axis). Some authors speculated that correspondence principle suggests that non-classicality should diminish with growing $d$.
However, on one hand in the meantime we witnessed a surprise in the form of the GHZ theorem \cite{GHZ}: for three or more qubits the conflict
between local realism and quantum mechanics is much sharper than for two. Thus, with the increase of the dimensionality of the full system, non-classicality of quantum correlations can grow.
On the other hand, Peres and Gisin \cite{PERES} considered dichotomic observables applied to maximally entangled pairs of qu$d$its, and showed that violations of the CHSH inequalities robustly survive in the limit of $d \to \infty$.
The concept of multiport interferometers, first discussed in the context of quantum entanglement by Klyshko \cite{KLYSHKO}, gave a hope for operational realizations of unitary transformations much richer than those linking measurement Stern-Gerlach bases for higher spins.
Proposals of Bell experiments with the multiports were presented in \cite{CONTROL, multiport}. Multiport interferometers were shown to be capable to reproduce all finite dimensional unitary
transformations \cite{RECK}, i.e., one can have access to the full U($d$) group. This lead to the discovery that two maximally entangled qu$d$its violate local realism more strongly than qubits, and that this violation grows with $d$, see \cite{KASZLIKOWSKI}. The strength of the non-classicality was measured via a ``white'' noise resistance. One takes
a family of mixed states $
\varrho= v \varrho_{state} + (1-v)\varrho_{noise},
$
where $\varrho_{state}$ represents the state under consideration, and $\varrho_{noise}=\frac{1}{d^2}\openone$ is the maximally mixed state. The lower is the threshold value $v=v_{crit}$ beyond which state does not violate local realism in a given Bell experiment, the bigger is the noise admixture $1 -v_{crit}$, which is needed to erase the non-classicality of the state in the experiment.
Note that $v_{crit}$ is a natural generalization of the ``factor of violation'' of a Bell inequality. E.g. for the CHSH inequality, if $\hat B$ is the Bell operator, its local realistic bound can be violated maximally by a factor of $v_{crit}^{-1}$. The same can be said about CGLMP inequalities \cite{CGLMP}. The results were obtained via a numerical analysis, which is a prototype of the one which will be presented. The values were confirmed by Collins et al. \cite{CGLMP} who derived a set of tight correlation Bell (CGLMP) inequalities specific for qu$d$it measurements. This tool helped to discover a strange property of two qu$d$it states. Maximally entangled states do not violate the CGLMP inequalities for qutrits maximally. The Schmidt decompositions of the optimal states does have all amplitudes of the same modulus --- one of them is smaller \cite{ACIN}. The results were generalized further by Chen at al. \cite{CHEN}. With a new version of a qu$d$it Bell inequality, the optimality of non-maximally entangled states was shown in arbitrary dimension, \cite{ZOHREN}. Note however, that in the papers a specific set of local observables was used, defined by a set of local phase shifts (in the Schmidt decomposition basis for the state), followed by an ``unbiased'' multiport beamsplitters (class M1 in Fig. 1), and finally detectors. Thus, non-classicality was not fully mapped.
Here we report an exhaustive numerical analysis of the two-qutrit Bell experiment involving two settings per observer. The observables are absolutely general, all possible ones are studied (that is the unitary transformations defining the measurement bases are forming the U(3) group). Such an analysis allows us to map the strength of violations of local realism (as measured by noise resistance) for all possible pure two-qutrit states. In this way we discover that Schmidt rank-2 states give rise to highest violations. This phenomenon continues if we increase the number of settings for the observers to three and four. The investigations were also extended to higher dimensional systems, for which we tested a few specially chosen states. The phenomenon seems to persist, and is even more pronounced. As a by-product of our investigations we have found for two qutrits a relatively simple set of unitary transformations (defined by just three phases), which gives a map of violations of local realism which diverges from the one for the full U(3) group defined measurements by maximally just 1.5\%, and for a broad class of states leads to the same level of non-classicality. We have also studied some interesting mixed states, which contain bound entanglement.
\section{Description of the method}
Let us move on to the quantum processes that we analyze, and the numerical approach. We begin with the two-qutrit case.
In our numerical analysis (called {\ttfamily steam-roller}) we consider a class of pure states of two qutrits, which contains all possible Schmidt decompositions:
\begin{equation}
|\psi(\alpha, \beta)\rangle = \cos \alpha |00\rangle + \sin \alpha (\cos \beta |11\rangle + \sin \beta |22\rangle).
\label{state-jg}
\end{equation}
Two spatially separated observers perform measurements of $m$ alternative local noncommuting trichotomic observables: $A_1, A_2, ... A_m$ for Alice and $B_1, B_2, ... B_m$ for Bob. We assume that they measure observables defined by a set of phase-shifters and (see Fig. \ref{Ufig-jg}):
\begin{itemize}
\item (M1) one unbiased three-port beam-splitter (tritter, for its properties, see\cite{multiport}),
\item (M2) two unbiased three-port beam-splitters,
\item (M3) three unbiased three-port beam-splitters,
\end{itemize}
and in the ultimate case
\begin{itemize}
\item (U(3)) any three-dimensional unitary transformation belongs to the U(3) group (for parametrization see e.g. \cite{SUN}).
\end{itemize}
\begin{figure}[!ht]
\includegraphics[width=0.35\textwidth]{U.eps}
\caption{\label{Ufig-jg} Measurement devices represented by unitary transformations M1, M2 and M3. The inputs are on the right hand side. Boxes represent unbiased symmetric three-port beamsplitters \cite{CONTROL,multiport}.}
\end{figure}
A $d$ input and output unbiased multiport performs a unitary transformation described by the Fourier matrix $U_{kl}={d}^{-{1}/{2}}e^{i{(2kl\pi}/{d})}, $ where $k,l=0,1...,d-1.$
The unitary transformations M1 and M2, obviously cannot reproduce the full U(3) group. We have checked that this is also the case for M3 transformations.
By saying that an experiment is local realistic we understand that it has a local realistic model for the assumed set of settings.
In order to obtain the value of the critical visibility $v_{crit}$ to allow such models for any observables from a given class,
we follow the procedure of the kind first used in Refs. \cite{GRUCA, KASZLIKOWSKI} (but with a much more advanced implementation and structure). The method is described in the Appendix.
\section{Results}
We have computed a map of strength of violation of local realism by pure two qutrit states. There one can find a range of the $\alpha$ and $\beta$ parameters of the state (\ref{state-jg}), for which the critical visibility is below the lowest known critical visibility ($0.6861$) in the case of the CGLMP inequality, for the asymmetric state $|\psi_{asym}\rangle$ found in ref. \cite{ACIN} (see Fig. \ref{violmap}). The lowest critical visibility is 0.6821 and surprisingly corresponds to the Schmidt rank-2 states $|\psi_{sym}^{rank-2}\rangle$: $|\psi(90^{\circ}, 45^{\circ})\rangle = (|11\rangle + |22\rangle)/\sqrt{2}$, or $|\psi(45^{\circ}, 0^{\circ})\rangle = (|00\rangle + |11\rangle)/\sqrt{2}$ and $|\psi(45^{\circ}, 90^{\circ})\rangle = (|00\rangle + |22\rangle)/\sqrt{2}$ (outside the map, but trivially related with the previous one). We can observe this phenomenon by using the most general local three-dimensional observables (U(3)) and transformations M3, while for less general observables (e.g. M1, M2), it is unobservable.
Despite the fact that the model noise well describes e.g. dark counts, one may criticize ``white'' noise resistance as a measure of non-classicality, see e.g. \cite{ADGL}. Thus, we have studied also other types of noise.
For a product noise admixture, of the form $\rho_A \otimes \rho_B$, where $\rho_{A(B)}$ is a reduced density matrix of the state (\ref{state-jg}) of the system A(B), the critical visibility for the state $|\psi_{sym}^{rank-2}\rangle$ is equal to 0.7071, whereas for the states $|\psi_{asym}\rangle$, $|\psi_{sym}\rangle$ is the same as in the case of white noise admixture, respectively 0.6962, 0.6861. For a dephasing noise, $\cos^2{\alpha} |00\rangle\langle00| + \sin^2{\alpha}( \cos^2{\beta} |11\rangle\langle11| + \sin^2{\beta} |22\rangle\langle22|)$, we observe violations of local realism for any state $|\psi(\alpha,\beta)\rangle$ ($\alpha>0$) and any amount of the noise ($v<1$). Such a feature is well known for qubits, however for qutrits thus far it was only a conjecture.
\begin{figure}[t]
\includegraphics[width=0.45\textwidth]{MAP-1-1.eps}
\caption{\label{violmap} The map of critical visibilities $v_{crit}$, that is noise resistance of non-classical correlations for the state $|\psi(\alpha, \beta)\rangle = \cos \alpha |00\rangle + \sin \alpha (\cos \beta |11\rangle + \sin \beta |22\rangle)$ for the case in which both observers use any observables. If $v>v_{crit}$, there does not exist any local realistic model describing quantum probabilities of experimental events.}
\end{figure}
\begin{table*}
\begin{tabular}{c c c c c c c c} \hline \hline
$d$ & $|\psi_{sym}^{rank-d}\rangle$ & $|\psi_{asym}^{rank-d}\rangle$ &
$|\psi_{sym}^{rank-4}\rangle$ & $|\psi_{asym}^{rank-4}\rangle$ &
$|\psi_{sym}^{rank-3}\rangle$ & $|\psi_{asym}^{rank-3}\rangle$ &
$|\psi_{sym}^{rank-2}\rangle$ \\ \hline
3 & 0.6962 & 0.6861 & - & - & - & - & 0.6821 \\
4 & 0.6906 & 0.6728 & - & - & 0.6824 & 0.6725 & 0.6442 \\
5 & 0.6871 & 0.6632 & 0.6819 & 0.6637 & 0.6584 & 0.6485& 0.6071 \\
\hline \hline
\end{tabular}
\caption{\label{min} The critical visibilities for special states.
The symbol $|\psi_{sym}^{rank-k}\rangle$ for $k=d$ stands for a maximally entangled state of two qu$d$its, that is with symmetric (all equal) Schmidt coefficients. If $k<d$ it is a state of two qu$d$its resembling a maximally entangled symmetric state for $k$-dimensional systems. The states $|\psi_{asym}^{rank-k}\rangle$ represent in the case $k=d$ the one which maximally violates the CGLMP inequality for qu$d$its, whereas for $k<d$, they are two qu$d$it states resembling the one optimal for the CGLMP inequality for a $k$ dimensional problem. The observers perform unrestricted measurements determined locally by U($d$).}
\end{table*}
We can also observe such phenomena in the 4-dimensional case. For the most general local 4-dimensional observables defined by the U(4) group, the critical visibility obtained for a symmetric Schmidt rank-2 state (e.g. $(|00\rangle + |11\rangle)/\sqrt{2}$) is equal to $0.6442$, whereas the critical visibility necessary for violation the CGLMP inequality is equal to $0.6728$ and $0.6906$ for the asymmetric \cite{CHEN} and symmetric \cite{CGLMP} states respectively (see Tab. \ref{min}).
The critical values for rank-2 states, Tab. \ref{min}, agree with those reported in \cite{ADGL}, where such numbers were found employing an {\em ad hoc} approach, aimed at showing conceptual problems related with the noise resistance as measure of non-classicality. Here we establish that these are the lowest possible critical visibilities for two-qutrit states and give a strong numerically supported conjecture that it is so for higher-dimensional systems. One could claim that a rank-2 state is effectively a two qubit state. However, as it is clear from our graphs, in the case of qutrits there is a continuous family of (strongly asymmetric) rank-3 states, which violate local realism stronger (in terms of noise resistance) than the optimal state for violation of the CGLMP inequality.
We compared our computer code, which is equivalent to a full set of Bell inequalities for the given problem, with the CGLMP inequality. For two qutrit states of the form $|\psi(\alpha, 45^{\circ})\rangle$,
if we restrict observables to M1 type, the numerical method gives the same\footnote{All comparisons are accurate to four decimal places.} critical visibility as the CGLMP inequality \cite{CGLMP}. Note that such observables were used in earlier works, refs \cite{KASZLIKOWSKI,CGLMP,ADGL,ACIN,CHEN,ZOHREN}.
In the case of observables of M2 kind, for $\alpha > 73^{\circ}$ the critical visibility obtained with {\ttfamily steam-roller} is lower than predicted by CGLMP inequality (see Fig. \ref{CGLMP-fig}a). For observables parametrized by M3 (and full U(3)) the advantage of the numerical method is even better visible and occurs already for $\alpha > 70^{\circ}$ (see Fig. \ref{CGLMP-fig}b).
\begin{figure}[!h]
\includegraphics[width=0.40\textwidth]{sr-cglmp-5.eps}
\caption{\label{CGLMP-fig} The dotted lines represent the critical visibility, which is necessary to violate the CGLMP inequality by states $|\psi(\alpha, 45^{\circ})\rangle$. The solid lines correspond to the critical visibility obtained with {\ttfamily steam-roller}. M2 (left graph) and M3 (right graph) transformations are considered here. The points at the curves with an undefined tangent indicate that different Bell inequalities are maximally violated by states to the left and right of the point.}
\end{figure}
For the two-qutrit case, there are some ranges of the parameter $\alpha$ in $|\psi(\alpha, 45^{\circ})\rangle$, for which:
\begin{itemize}
\item the observable generated by M2 gives a better critical visibility than the ones obtained for the observable M1 ($\alpha < 49^{\circ}$ and $\alpha > 73^{\circ}$).
\item the observable for M3 gives a better critical visibility than the ones obtained for the observable M2 ($\alpha < 54^{\circ}$ and $\alpha > 70^{\circ}$).
\end{itemize}
All of these cases are presented in Fig. \ref{2Dplot}. The critical visibilities obtained for the M3 and U(3) observables are the same for the $|\psi(\alpha, 45^{\circ})\rangle$ state. For other values of $\beta$ the highest difference $\delta v_{crit}^{\rm U(3)-M3}(\%)$ between critical visibilities obtained for these observables is less than 1\%.
If one increases the number of settings, up to four on each side, this does not reduce the critical visibility for any of the types of observables.
\begin{figure}[!ht]
\includegraphics[width=0.45\textwidth]{sr-sym-2.eps}
\caption{\label{2Dplot} The critical visibilities for states $|\psi(\alpha, 45^{\circ})\rangle$. Each line corresponds to a different type of the unitary transformations (M1, M2, and M3). Up to four settings per observer were used, the relations did not change.}
\end{figure}
\subsection{Almost perfect two-qutrit Bell interferometer}
Let us consider the M3 interferometer shown in Fig. \ref{Ufig-jg} with $\phi_1=\phi_5=\phi_6=0$. The differences between critical visibilities obtained for such transformation and U(3) are shown in Fig. \ref{violmapprec}a for the state $|\psi(\alpha, \beta)\rangle$ and in Fig. \ref{violmapprec}b for the state $|\psi(\alpha, 45^{\circ})\rangle$. They are less than 1.5\% for any $\alpha$ and $\beta$, and in many cases the same (including specific points related to the symmetric $|\psi_{sym}\rangle$ and asymmtric $|\psi_{asym}\rangle$ state). Thus if one is interested in situations in which violations of local realism play an essential role this relatively simple device might be optimal.
\begin{figure}[t]
\includegraphics[width=0.40\textwidth]{MAP-2-1.eps}
\caption{\label{violmapprec} a) Differences between critical visibilities obtained for the U(3) and M3 (with $\phi_1=\phi_5=\phi_6=0$) transformations for the state $|\psi(\alpha, \beta)\rangle$. Peak points are lower than 1.5\%; b) Differences between critical visibilities obtained for the U(3) and M3 (with $\phi_1=\phi_5=\phi_6=0$) transformations for the states $|\psi(\alpha, 45^{\circ})\rangle$.}
\end{figure}
\subsection{Question of violation of local realism by bound entangled states}
As we have at our disposal a tool which allows to test violation of local realism for any local observables, we also analyzed some classes of two-particle bound entangled states, namely:
the three-dimensional Bennett state \cite{BENNETT}, the three-dimensional Horodecki state \cite{HORODECKIBOUND} and its generalization to the four- and five-dimensional states \cite{CHRUSCINSKI}, the four-dimensional Pankowski -- Horodecki state \cite{PANKOWSKI}. We established that for U(3), and in the last case U(4) transformations (that is all possible local measurements with two settings per observer) one does not observe violations of local realism.
\section{Final remarks}
We have charted unknown territories in the map of non-classicality of two qutrit states, and made some excursions to higher dimensional problems. The obtained results were difficult to anticipate. Our method is quite flexible and we plan to move on to the yet uncharted multi qudit states. Note, that the computer code, after special processing of experimental data is able to decide whether the data allow a local realistic model or not. Such applications will be found in forthcoming manuscripts. The general theory of the code itself, and particular solutions, will be presented in ref. \cite{GONDZIO}.
{\em Acknowledgments.---} The work is supported by EU program QESSENCE
(Contract No.248095) and MNiSW (NCN) Grant no. N202 208538.
WL acknowledges financial support from European Social
Fund as a part of the project ``Educators for the elite
- integrated training program for PhD students, postdocs
and professors as academic teachers at University of
Gda\'nsk'' within the framework of Human Capital Operational
Programme, Action 4.1.1, Improving the quality
of educational offer of tertiary education institutions. JG akcnowledges DS IFTiA UG.
|
\section{Introduction}
In the last few years a great attention has been devoted to the
theory of Sobolev spaces $W^{1,q}$ on metric measure spaces
$(X,{\sf d},{\mbox{\boldmath$m$}})$, see for instance \cite{Heinonen07} and
\cite{Hajlasz-Koskela}
for an overview on this subject. These definitions of Sobolev spaces
usually come with a weak definition of modulus of gradient, in
particular the notion of $q$-upper gradient has been introduced in
\cite{Koskela-MacManus} and used in \cite{Shanmugalingam00} for a
Sobolev space theory. Also, in \cite{Shanmugalingam00} the notion of
minimal $q$-upper gradient has been proved to be equivalent to the
notion of relaxed upper gradient arising in Cheeger's paper
\cite{Cheeger00}.
In this paper we consider a notion of gradient $|\nabla f|_{*,q}$
stronger than the one of \cite{Cheeger00}, because in the
approximation procedure we use Lipschitz functions and their slopes
as upper gradients, and a notion of $q$-weak upper gradient $|\nabla
f|_{w,q}$ weaker than the one of \cite{Shanmugalingam00}, and prove
their equivalence. As a consequence all four notions of gradient
turn out to be equivalent. A byproduct of our equivalence result is
the following density in energy of Lipschitz functions: if $f\in
L^q(X,{\mbox{\boldmath$m$}})$ has a $q$-weak upper gradient $|\nabla f|_{w,q}$ in
$L^q(X,{\mbox{\boldmath$m$}})$, then there exist Lipschitz functions $f_n$ convergent
to $f$ in $L^q(X,{\mbox{\boldmath$m$}})$ satisfying (here $|\nabla f_n|$ is the slope
of $f_n$)
\begin{equation}\label{densitylip}
\lim_{n\to\infty}\int_X\bigl||\nabla f_n|-|\nabla
f|_{w,q}\bigr|^q\,\d{\mbox{\boldmath$m$}}=0.
\end{equation}
Notice that we can use Mazur's lemma to improve this convergence to
strong convergence in $W^{1,q}(X,{\sf d},{\mbox{\boldmath$m$}})$, as soon as this space
is reflexive; this happens for instance in the context of the spaces
with Riemannian Ricci bounds from below considered in
\cite{Ambrosio-Gigli-Savare11bis}, with $q=2$.
We emphasize that our density result does not depend on doubling and
Poincar\'e assumptions on the metric measure structure; as it is
well known (see Theorem~4.14 and Theorem~4.24 in \cite{Cheeger00}),
these assumptions ensure the density in Sobolev norm of Lipschitz
functions, even in the Lusin sense (i.e. the Lipschitz approximating
functions $f_n$ coincide with $f$ on larger and larger sets). On the
other hand, the density in energy \eqref{densitylip} suffices for
many purposes, for instance the extension by approximation, from
Lipschitz to Sobolev functions, of functional inequalities like the
Poincar\'e or Sobolev inequality. For instance, our result can be used to
show that if $(X,{\sf d})$ is complete and separable and ${\mbox{\boldmath$m$}}$ is a Borel measure finite on
bounded sets, then the Poincar\'e inequality
$$
\int_{B_r(x)}|f(y)-f_{B_r(x)}|\,\d{\mbox{\boldmath$m$}}(y)\leq Cr\int_{B_{\lambda r}(x)}|\nabla f|(y)\,d{\mbox{\boldmath$m$}}(y)
$$
holds for all $f:X\to\mathbb{R}$ Lipschitz on bounded sets if and only if it holds in the form
$$
\int_{B_r(x)}|f(y)-f_{B_r(x)}|\,\d{\mbox{\boldmath$m$}}(y)\leq Cr\int_{B_{\lambda r}(x)}g(y)\,d{\mbox{\boldmath$m$}}(y)
$$
for all pairs $(f,g)$ with $f$ Borel and $g$
upper gradient of $f$.
This equivalence was proven in \cite{Heinonen-Koskela99} for proper,
quasiconvex and doubling metric measure spaces, while in
\cite{Koskela_removable} (choosing $X=\mathbb{R}^n\setminus E$ for suitable
compact sets $E$) it is proven that completeness of the space can't
be dropped.
The new notions of gradient, as well as their equivalence, have been
proved in \cite{Ambrosio-Gigli-Savare11} in the case $q=2$, see
Corollary~6.3 therein. Here we extend the result to general
exponents $q\in (1,\infty)$ and we give a presentation more focussed
on the equivalence problem. While the traditional proof of density
of Lipschitz functions relies on Poincar\'e inequality, maximal
functions and covering arguments to construct the ``optimal''
approximating Lipschitz functions $f_n$, our proof is more indirect
and provides the approximating functions using the $L^2$-gradient
flow of $\mathbb{C}_q(f):=q^{-1}\int_X|\nabla f|_{*,q}^q\,\d{\mbox{\boldmath$m$}}$ and the
analysis of the dissipation rate along this flow of a suitable
``entropy'' $\int\Phi_q(f)\,\d{\mbox{\boldmath$m$}}$ (in the case $q=2$, $\Phi(z)=z\log z$). This way we prove that $|\nabla
f|_{w,q}=|\nabla f|_{*,q}$ ${\mbox{\boldmath$m$}}$-a.e., and then \eqref{densitylip}
follows by a general property of the minimal $q$-relaxed slope
$|\nabla f|_{*,q}$, see Proposition~\ref{prop:easy}.
The paper is organized as follows. In Section~2 we recall some
preliminary facts on absolutely continuous curves and gradient
flows. We also introduce the $p$-th Wasserstein distance and the
so-called superposition principle, that allows to pass from an
``Eulerian'' formulation (i.e. in terms of a curve of measures or a
curve of probability densities) to a ``Lagrangian'' one, namely a
probability measure in the space of absolutely continuous paths;
this will be the only tool from optimal transportation theory used
in the paper.\\* In Section~3 we study the pointwise properties of
the Hopf-Lax semigroup
$$
Q_tf(x):=\inf_{y\in X}f(y)+\frac{{\sf d}^p(x,y)}{pt^{p-1}}.
$$
In comparison with Section~3 of \cite{Ambrosio-Gigli-Savare11},
dealing with the case $p=2$, we consider for the sake of simplicity
only locally compact spaces and finite distances, but the
proofs can be modified to deal with more general cases, see also
Section~8. The results of this section overlap with those of the
forthcoming paper \cite{Gozlan} by Gozlan, Roberto and Samson, where
the HL semigroup is used in connection with the proof of transport
entropy inequalities.\\* In Section 4 we introduce the four
definitions of gradients we will be dealing with, namely:
\begin{itemize}
\item[(1)]
the Cheeger gradient $|\nabla f|_{C,q}$ of \cite{Cheeger00} arising from the relaxation
of upper gradients;
\item[(2)]
the minimal relaxed slope $|\nabla f|_{*,q}$ of
\cite{Ambrosio-Gigli-Savare11} arising from the relaxation of the
slope of Lipschitz functions;
\item[(3)]
the minimal $q$-upper gradient
$|\nabla f|_{S,q}$ of \cite{Koskela-MacManus,Shanmugalingam00},
based on the validity of the upper gradient property out of a ${\rm
Mod}_q$-null set of curves;
\item[(4)] the minimal $q$-weak upper gradient of \cite{Ambrosio-Gigli-Savare11},
based on the validity of the upper gradient property out of a
$q$-null set of curves.
\end{itemize}
While presenting these definitions we will point out natural
relations between them, that lead to the chain of inequalities
$$
|\nabla f|_{w,q}\leq|\nabla f|_{S,q}\leq|\nabla f|_{C,q}\leq |\nabla
f|_{*,q}\qquad\text{${\mbox{\boldmath$m$}}$-a.e. in $X$,}
$$
with the concepts of \cite{Ambrosio-Gigli-Savare11} at the extreme
sides.
Section~5 contains some basically well known properties of weak
gradients, namely chain rules and stability under weak convergence.
Section~6 contains the basic facts we shall need on the gradient flow of the lower
semicontinuous functional $\mathbb{C}_q$ we need, in particular the entropy
dissipation rate
$$
\frac{\d}{\d t}\int_X\Phi(f_t)\,\d{\mbox{\boldmath$m$}}=-\int_X\Phi''(f_t)|\nabla
f_t|^q_{*,q}\,\d{\mbox{\boldmath$m$}}
$$
along this gradient flow.
In Section~7 we prove the equivalence of gradients. Starting from a
function $f$ with $|\nabla f|_{w,q}\in L^q(X,{\mbox{\boldmath$m$}})$ we approximate it
by the gradient flow of $f_t$ of $\mathbb{C}_q$ starting from $f$ and we use
the weak upper gradient property to get
$$
\limsup_{t\downarrow 0}\frac1t\int_0^t\int_X\frac{|\nabla
f_s|_{*,q}^q}{f_s^{p-1}}\,d{\mbox{\boldmath$m$}}\d s\leq\int_X\frac{|\nabla
f|_{w,q}^q}{f^{p-1}}\,d{\mbox{\boldmath$m$}}
$$
where $p=q/(q-1)$ is the dual exponent of $q$. Using the stability
properties of the relaxed gradients we eventually get $|\nabla
f|_{*,q}\leq|\nabla f|_{*,w}$ ${\mbox{\boldmath$m$}}$-a.e. in $X$.
Finally, Section~8 discusses some potential extensions of the
results of this paper: we indicate how spaces which are not locally
compact and measures that are locally finite can be achieved. Other
extensions require probably a separate investigation, as the case of
Orlicz spaces and the limiting case $q=1$, corresponding to
$W^{1,1}$ and $BV$ spaces. In this latter case the lack of
reflexivity of $L^1(X,{\mbox{\boldmath$m$}})$ poses problems even in the definition of
the minimal gradients and we discuss this very briefly.
\smallskip
\noindent {\bf Acknowledgement.} The authors acknowledge the support
of the ERC ADG GeMeThNES. The authors also thank P.Koskela for
useful comments during the preparation of the paper.
\section{Preliminary notions}\label{sec:preliminary}
In this section we introduce some notation and recall a few basic
facts on absolutely continuous functions, gradient flows of convex
functionals and optimal transportation, see also
\cite{Ambrosio-Gigli-Savare08}, \cite{Villani09} as general
references.
\subsection{Absolutely continuous curves and slopes}
Let $(X,{\sf d})$ be a metric space, $J\subset\mathbb{R}$ a closed interval and
$J\ni t\mapsto x_t\in X$. We say that $(x_t)$ is \emph{absolutely
continuous} if
$$
{\sf d}(x_s,x_t)\leq\int_s^tg(r)\,\d r\qquad\forall s,\,t\in J,\,\,s<t
$$
for some $g\in L^1(J)$. It turns out that, if $(x_t)$ is absolutely
continuous, there is a minimal function $g$ with this property,
called \emph{metric speed}, denoted by $|\dot{x}_t|$ and given for
a.e. $t\in J$ by
$$
|\dot{x}_t|=\lim_{s\to t}\frac{{\sf d}(x_s,x_t)}{|s-t|}.
$$
See \cite[Theorem~1.1.2]{Ambrosio-Gigli-Savare08} for the simple
proof.
We will denote by $C([0,1],X)$ the space of continuous curves from
$[0,1]$ to $(X,{\sf d})$ endowed with the $\sup$ norm. The set
$AC^p([0,1],X)\subset C([0,1],X)$ consists of all absolutely
continuous curves $\gamma$ such that $\int_0^1|\dot\gamma_t|^p\,\d
t<\infty$: it is the countable union of the sets $\{\gamma:\
\int_0^1|\dot\gamma_t|^p\,\d t\leq n\}$, which are easily seen to be
closed if $p>1$. Thus $AC^p([0,1],X)$ is a Borel subset of
$C([0,1],X)$. The \emph{evaluation maps} $\e_t:C([0,1],X)\to X$ are
defined by
\[
\e_t(\gamma):=\gamma_t,
\]
and are clearly continuous.
Given $f:X\to\mathbb{R}$, we define \emph{slope} (also called local
Lipschitz constant) by
$$
|\nabla f|(x):=\varlimsup_{y\to x}\frac{|f(y)-f(x)|}{{\sf d}(y,x)}.
$$
For $f,\,g:X\to\mathbb{R}$ Lipschitz it clearly holds
\begin{subequations}
\begin{align}
\label{eq:subadd}
|\nabla(\alpha f+\beta g)|&\leq|\alpha||\nabla f|+|\beta||\nabla g|\qquad\forall \alpha,\beta\in\mathbb{R},\\
\label{eq:leibn} |\nabla (fg)|&\leq |f||\nabla g|+|g||\nabla f|.
\end{align}
\end{subequations}
{We shall also need the following calculus lemma.}
\begin{lemma}\label{lem:Fibonacci}
{
Let $f:(0,1)\to\mathbb{R}$, $q\in [1,\infty]$, $g\in L^q(0,1)$ nonnegative be satisfying
$$
|f(s)-f(t)|\leq\bigl|\int_s^t g(r)\,\d r\bigr|\qquad\text{for $\Leb{2}$-a.e. $(s,t)\in (0,1)^2$.}
$$
Then $f\in W^{1,q}(0,1)$ and $|f'|\leq g$ a.e. in $(0,1)$.}
\end{lemma}
\begin{proof}
Let $N\subset (0,1)^2$ be the $\Leb{2}$-negligible subset where the above
inequality fails. Choosing $s\in (0,1)$, whose existence is ensured by Fubini's theorem, such that
$(s,t)\notin N$ for a.e. $t\in (0,1)$, we obtain that $f\in L^\infty(0,1)$.
Since the set $\{(t,h)\in (0,1)^2:\ (t,t+h)\in N\cap (0,1)^2\}$
is $\Leb{2}$-negligible as well, we can apply Fubini's theorem to obtain that for a.e. $h$ it holds
$(t,t+h)\notin N$ for a.e. $t\in (0,1)$. Let $h_i\downarrow 0$ with this property and use
the identities
$$
\int_0^1f(t)\frac{\phi(t+h)-\phi(t)}{h}\,\d t=-\int_0^1\frac{f(t-h)-f(t)}{-h}\phi(t)\,\d t
$$
with $\phi\in C^1_c(0,1)$ and $h=h_i$ sufficiently small to get
$$
\biggl|\int_0^1f(t)\phi'(t)\,\d t\biggr|\leq\int_0^1g(t)|\phi(t)|\,\d t.
$$
It follows that the distributional derivative of $f$ is a signed
measure $\eta$ with finite total variation which satisfies
\begin{displaymath}
-\int_0^1f\phi'\,\d t=\int_0^1 \phi\,\d\eta,\qquad
\Bigl|\int_0^1 \phi\,\d\eta\Bigr|\le \int_0^1g|\phi|\,\d
t\quad\text{for every }\phi\in C^1_c(0,1);
\end{displaymath}
therefore $\eta$ is absolutely continuous with respect to the Lebesgue
measure with $|\eta|\le g\Leb 1$.
This gives the $W^{1,1}(0,1)$ regularity and, at the same time, the inequality
$|f'|\leq g$ a.e. in $(0,1)$. The case $q>1$ immediately follows
by applying this inequality when $g\in L^q(0,1)$.
\end{proof}
\subsection{Gradient flows of convex functionals}
Let $H$ be an Hilbert space, $\Psi:H\to\mathbb{R}\cup\{+\infty\}$ convex and
lower semicontinuous and $D(\Psi)=\{\Psi<\infty\}$ its finiteness
domain. Recall that a gradient flow $x:(0,\infty)\to H$ of $\Psi$ is
a locally absolutely continuous map with values in $D(\Psi)$
satisfying
$$
-\frac{\d}{{\d t}}x_t\in\partial^-\Psi(x_t)\qquad\text{for a.e. $t\in
(0,\infty)$.}
$$
Here $\partial^-\Psi(x)$ is the subdifferential of $\Psi$, defined
at any $x\in D(\Psi)$ by
$$
\partial^-\Psi(x):=\left\{p\in H^*:\ \Psi(y)\geq\Psi(x)+\langle
p,y-x\rangle\,\,\forall y\in H\right\}.
$$
We shall use the fact that for all $x_0\in\overline{D(\Psi)}$ there
exists a unique gradient flow $x_t$ of $\Psi$ starting from $x_0$,
i.e. $x_t\to x_0$ as $t\downarrow 0$, and that $t\mapsto\Psi(x_t)$
is nonincreasing and locally absolutely continuous in $(0,\infty)$.
In addition, this unique solution exhibits a regularizing effect,
namely $-\tfrac{\d}{\d t}x_t$ is for a.e. $t\in (0,\infty)$ the
element of minimal norm in
$\partial^-\Psi(x_t)$.
\subsection{The space $(\Probabilities X,W_p)$ and the superposition principle}
Let $(X,{\sf d})$ be a complete and separable metric space and $p\in
[1,\infty)$. We use the notation $\Probabilities X$ for the set of all Borel
probability measures on $X$. Given $\mu,\,\nu\in\Probabilities X$, we define
the Wasserstein (extended) distance $W_p(\mu,\nu)\in [0,\infty]$
between them as
\[
W_p^p(\mu,\nu):=\min\int {\sf d}^p(x,y)\,\d{\mbox{\boldmath$\gamma$}}(x,y).
\]
Here the minimization is made in the class $\Gamma(\mu,\nu)$ of all
probability measures ${\mbox{\boldmath$\gamma$}}$ on $X\times X$ such that
$\pi^1_\#{\mbox{\boldmath$\gamma$}}=\mu$ and $\pi^2_\#{\mbox{\boldmath$\gamma$}}=\nu$, where
$\pi^i:X\times X\to X$, $i=1,\,2$, are the coordinate projections
and $f_\#:\Probabilities{Y}\to\Probabilities{Z}$ is the push-forward operator induced
by a Borel map $f:Y\to Z$.
An equivalent definition of $W_p$ comes from the dual formulation of
the transport problem. In the case when $(X,{\sf d})$ has finite
diameter the dual formulation takes the simplified form
\begin{equation}
\label{eq:dualitabase} \frac1pW_p^p(\mu,\nu)=\sup_{\psi\in{\rm
Lip}(X)}\int\psi\, \d\mu+\int \psi^c\,\d\nu,
\end{equation}
where the $c$-transform $\psi^c$ is defined by
\[
\psi^c(y):=\inf_{x\in X}\frac{{\sf d}^p(x,y)}p-\psi(x).
\]
We will need the following result, proved in \cite{Lisini07}: it
shows how to associate to an absolutely continuous curve $\mu_t$
w.r.t. $W_p$ a plan ${\mbox{\boldmath$\pi$}}\in\Probabilities{C([0,1],X)}$ representing the
curve itself (see also \cite[Theorem~8.2.1]{Ambrosio-Gigli-Savare08}
for the Euclidean case).
\begin{proposition}[Superposition principle]\label{prop:lisini}
Let $(X,{\sf d})$ be a complete and separable metric space with ${\sf d}$
bounded, $p\in (1,\infty)$ and let $\mu_t\in AC^p\bigl([0,T];(\Probabilities X,W_p)\bigr)$.
Then there exists ${\mbox{\boldmath$\pi$}}\in\Probabilities{C([0,1],X)}$,
concentrated on $AC^p([0,1],X)$, such that $(\e_t)_\sharp{\mbox{\boldmath$\pi$}}=\mu_t$
for any $t\in[0,T]$ and
\begin{equation}\label{eq:Lisini}
\int|\dot\gamma_t|^p\,\d{\mbox{\boldmath$\pi$}}(\gamma)=|\dot\mu_t|^p\qquad \text{for
a.e. $t\in [0,T]$.}
\end{equation}
\end{proposition}
\section{Hopf-Lax formula and Hamilton-Jacobi equation}\label{sec:hopflax}
Aim of this section is to study the properties of the Hopf-Lax
formula in a metric setting and its relations with the
Hamilton-Jacobi equation. Here we assume for simplicity that
$(X,{\sf d})$ is a compact metric space, see Section~\ref{sextensions}
for a more general discussion. Notice that there is no reference
measure ${\mbox{\boldmath$m$}}$ here. We fix a power $p\in (1,\infty)$ and denote by
$q$ the dual exponent.
Let $f:X\to\mathbb{R}$ be a Lipschitz function. For $t>0$ define
\begin{equation}\label{eq:Nicola1}
F(t,x,y):=f(y)+\frac{{\sf d}^p(x,y)}{pt^{p-1}},
\end{equation}
and the function $Q_tf:X\to\mathbb{R}$ by
\begin{equation}\label{eq:Nicola2}
Q_tf(x):=\inf_{y\in X}F(t,x,y)=\min_{y\in X}F(t,x,y).
\end{equation}
Also, we introduce the functions $D^+,\,D^-:X\times(0,\infty)\to\mathbb{R}$
as
\begin{equation}\label{eq:defdpm}
\begin{split}
D^+(x,t)&:=\max\, {\sf d}(x,y),\\
D^-(x,t)&:=\min\, {\sf d}(x,y),\\
\end{split}
\end{equation}
where, in both cases, the $y$'s vary among all minima of
$F(t,x,\cdot)$. We also set $Q_0f=f$ and $D^\pm(x,0)=0$. Arguing as
in \cite[Lemma~3.1.2]{Ambrosio-Gigli-Savare08} it is easy to check
that the map $[0,\infty)\ni(t,x)\mapsto Q_tf(x)$ is continuous.
Furthermore, the fact that $f$ is Lipschitz easily yields
\begin{equation}
\label{eq:boundD} D^-(x,t)\leq D^+(x,t)\leq t(p\Lip(f))^{1/(p-1)}.
\end{equation}
\begin{proposition}[Monotonicity of $D^\pm$]\label{prop:dmon}
For all $x\in X$ it holds
\begin{equation}\label{eq:basic_mono}
D^+(x,t)\leq D^-(x,s)\qquad 0\leq t< s.
\end{equation}
As a consequence, $D^+(x,\cdot)$ and $D^-(x,\cdot)$ are both
nondecreasing, and they coincide with at most countably many
exceptions in $[0,\infty)$.
\end{proposition}
\begin{proof}
Fix $x\in X$. For $t=0$ there is nothing to prove. Now pick $0<t<s$
and choose $x_t$ and $x_s$ minimizers of $F(t,x,\cdot)$ and
$F(s,x,\cdot)$ respectively, such that ${\sf d}(x,x_t)=D^+(x,t)$ and $
{\sf d}(x,x_s)=D^-(x,s)$. The minimality of $x_t,\,x_s$ gives
\[
\begin{split}
f(x_t)+\frac{{\sf d}^p(x_t,x)}{pt^{p-1}}&\leq f(x_s)+\frac{{\sf d}^p(x_s,x)}{pt^{p-1}}\\
f(x_s)+\frac{{\sf d}^p(x_s,x)}{ps^{p-1}}&\leq
f(x_t)+\frac{{\sf d}^p(x_t,x)}{ps^{p-1}}.
\end{split}
\]
Adding up and using the fact that $\tfrac1t\geq\tfrac 1s$ we deduce
\[
D^+(x,t)={\sf d}(x_t,x)\leq {\sf d}(x_s,x)= D^-(x,s),
\]
which is \eqref{eq:basic_mono}. Combining this with the inequality
$D^-\leq D^+$ we immediately obtain that both functions are
nonincreasing. At a point of right continuity of $D^-(x,\cdot)$ we
get
$$
D^+(x,t)\leq\inf_{s>t}D^-(x,s)=D^-(x,t).
$$
This implies that the two functions coincide out of a countable set.
\end{proof}
Next, we examine the semicontinuity properties of $D^\pm$. These
properties imply that points $(x,t)$ where the equality
$D^+(x,t)=D^-(x,t)$ occurs are continuity points for both $D^+$ and
$D^-$.
\begin{proposition}[Semicontinuity of $D^\pm$]
$D^+$ is upper semicontinuous and $D^-$ is lower semicontinuous in
$X\times [0,\infty)$.
\end{proposition}
\begin{proof}
We prove lower semicontinuity of $D^-$, the proof of upper
semicontinuity of $D^+$ being similar. Let $(x_i,t_i)$ be any
sequence converging to $(x,t)$ such that the limit of $D^-(x_i,t_i)$
exists and assume that $t>0$ (the case $t=0$ is trivial). For every
$i$, let $(y_i)$ be a minimum of $F(t_i,x_i,\cdot)$ for which
${\sf d}(y_i,x_i)=D^-(x_i,t_i)$, so that
\[
f(y_i)+\frac{{\sf d}^p(y_i,x_i)}{pt_i^{p-1}}=Q_{t_i}f(x_i).
\]
The continuity of $(x,t)\mapsto Q_tf(x)$ gives that
$\lim_iQ_{t_i}f(x_i)=Q_tf(x)$, thus
\[
\lim_{i\to\infty} f(y_i)+\frac{{\sf d}^p(y_i,x)}{pt^{p-1}}=Q_tf(x),
\]
that is: $i\mapsto y_i$ is a minimizing sequence for $F(t,x,\cdot)$.
Since $(X,{\sf d})$ is compact, possibly passing to a subsequence, not
relabeled, we may assume that $(y_i)$ converges to some $y$ as
$i\to\infty$. Therefore
\[
D^-(x,t)\leq
{\sf d}(x,y)=\lim_{i\to\infty}{\sf d}(x_i,y_i)=\lim_{i\to\infty}D^-(x_i,t_i).
\]
\end{proof}
\begin{proposition}[Time derivative of
$Q_tf$]\label{prop:timederivative} The map $t\mapsto Q_tf$ is
Lipschitz from $[0,\infty)$ to $C(X)$ and, for all $x\in X$, it
satisfies:
\begin{equation}\label{eq:Dini1}
\frac{\d}{\d
t}Q_tf(x)=-\frac{1}{q}\bigl[\frac{D^{\pm}(x,t)}{t}\bigr]^p,
\end{equation}
for any $t>0$, with at most countably many exceptions.
\end{proposition}
\begin{proof}
Let $t<s$ and $x_t$, $x_s$ be minima of $F(t,x,\cdot)$
and $F(s,x,\cdot)$. We have
\[
\begin{split}
Q_sf(x)-Q_tf(x)&\leq F(s,x,x_t)-F(t,x,x_t)=\frac{{\sf d}^p(x,x_t)}{p}\frac{t^{p-1}-s^{p-1}}{t^{p-1}s^{p-1}},\\
Q_sf(x)-Q_tf(x)&\geq
F(s,x,x_s)-F(t,x,x_s)=\frac{{\sf d}^p(x,x_s)}{p}\frac{t^{p-1}-s^{p-1}}{t^{p-1}s^{p-1}},
\end{split}
\]
which gives that $t\mapsto Q_tf(x)$ is Lipschitz in $(\eps,\infty)$
for any $\eps>0$ uniformly with respect to $x\in X$. Also, dividing
by $(s-t)$ and taking Proposition~\ref{prop:dmon} into account, we
get \eqref{eq:Dini1}. Now notice that from \eqref{eq:boundD} we get
that $q|\frac{\d}{\d t}Q_tf(x)|\leq p^q[\Lip(f)]^q$ for any $x\in X$
and a.e. $t$, which, together with the pointwise convergence of
$Q_tf$ to $f$ as $t\downarrow 0$, yields that $t\mapsto Q_tf\in
C(X)$ is Lipschitz in $[0,\infty)$.
\end{proof}
In the next proposition we bound from above the slope of $Q_tf$ at $x$ with
$|D^+(x,t)/t|^{p-1}$; actually we shall prove a more precise statement, in connection with \S\ref{sec:improveslope},
which involves the \emph{asymptotic Lipschitz constant}
\begin{equation}\label{eq:asymlip}
{\rm Lip}_a(f,x):=\inf_{r>0}{\rm Lip}\bigl(f,B_r(x)\bigr)=\lim_{r\downarrow 0}{\rm Lip}\bigl(f,B_r(x)\bigr).
\end{equation}
Notice that ${\rm Lip}(f)\geq {\rm Lip}_a(f,x)\geq |\nabla f|^*(x)$, where $|\nabla f|^*$ is the upper semicontinuous envelope
of the slope of $f$. The second inequality is easily seen to be an equality in length spaces.
\begin{proposition}[Bound on the asymptotic Lipschitz constant of $Q_tf$]\label{prop:slopesqt}
For $(x,t)\in X\times (0,\infty)$ it holds:
\begin{equation}
\label{eq:hjbss} {\rm Lip}_a(Q_tf,x)\leq
\bigl[\frac{D^+(x,t)}t\bigr]^{p-1}.
\end{equation}
\end{proposition}
In particular ${\rm Lip}(Q_t(f))\leq p{\rm Lip}(f)$.
\begin{proof} Fix $y,\,z\in X$, $t\in (0,\infty)$ and a minimizer $\bar y$ for
$F(t,y,\cdot)$. Since it holds
\[
\begin{split}
Q_tf(z)-Q_tf(y)&\leq F(t,z,\bar y)-F(t,y,\bar y)=
f(\bar y)+\frac{{\sf d}^p(z,\bar y)}{pt^{p-1}}-f(\bar y)-\frac{{\sf d}^p(y,\bar y)}{pt^{p-1}}\\
&\leq
\frac{({\sf d}(z,y)+{\sf d}(y,\bar y))^p}{pt^{p-1}}-\frac{{\sf d}^p(x_i,y_i)}{pt^{p-1}}
\\
&\leq \frac{{\sf d}(z,y)}{t^{p-1}}\bigl({\sf d}(z,y)+D^+(y,t)\bigr)^{p-1},
\end{split}
\]
so that dividing by ${\sf d}(z,y)$ and inverting the roles of $y$ and $z$ gives
$$
{\rm Lip}\bigl(Q_tf,B_r(x)\bigr)\leq t^{1-p}\bigl(\sup_{y\in B_r(x)}D^+(y,t)\bigr)^{p-1}.
$$
Letting $r\downarrow 0$ and using the upper
semicontinuity of $D^+$ we get \eqref{eq:hjbss}.
Finally, the bound on the Lipschitz constant of $Q_tf$ follows
directly from \eqref{eq:boundD} and \eqref{eq:hjbss}.
\end{proof}
\begin{theorem}[Subsolution of HJ]\label{thm:subsol}
For every $x\in X$ it holds
\begin{equation}\label{eq:hjbsus}
\frac{\d}{\d t}Q_tf(x)+\frac{1}{q}|\nabla Q_tf|^q(x)\leq 0
\end{equation}
for every $t\in (0,\infty)$, with at most countably many exceptions.
\end{theorem}
\begin{proof}
The claim is a direct consequence of
Propositions~\ref{prop:timederivative} and
Proposition~\ref{prop:slopesqt}.
\end{proof}
Notice also that \eqref{eq:hjbss} allows to write the HJ sub solution property
in a stronger form using the asymptotic Lipschitz constant ${\rm Lip}_a(Q_tf,\cdot)$ in place of
$|\nabla Q_t f|$, namely for all $x\in X$ it holds
\begin{equation}\label{eq:hjbsusbis}
\frac{\d}{\d t}Q_tf(x)+\frac{1}{q}({\rm Lip}_a(Q_tf,x))^q\leq 0
\end{equation}
for every $t\in (0,\infty)$, with at most countably many exceptions.
We just proved that in an arbitrary metric space the Hopf-Lax
formula produces subsolutions of the Hamilton-Jacobi equations. In
geodesic spaces this result can be improved to get solutions. Since
we shall not need the result, we just state it (the proof is
analogous to \cite[Proposition~3.6]{Ambrosio-Gigli-Savare11}).
\begin{theorem}[Supersolution of HJ]\label{thm:supersol}
Assume that $(X,{\sf d})$ is a geodesic space. Then equality holds in
\eqref{eq:hjbss}. In particular, for all $x\in X$ it holds
\[
\frac{\d}{\d t}Q_tf(x)+\frac{1}{q}|\nabla Q_tf|^q(x)=0
\]
for every $t\in(0,\infty)$, with at most countably many exceptions.
\end{theorem}
\section{Weak gradients}\label{sec:weakgra}
Let $(X,{\sf d})$ be a complete and separable metric space and let
${\mbox{\boldmath$m$}}$ be a nonnegative $\sigma$-finite Borel measure in $X$. In this
section we introduce and compare four notions of weak gradients, the
gradient $|\nabla f|_{C,q}$ introduced in \cite{Cheeger00}, the
gradient $|\nabla f|_{S,q}$ introduced in \cite{Koskela-MacManus}
and further studied in \cite{Shanmugalingam00} and the gradients
$\relgradq fq$ and $\weakgradq fq$ whose definition can be obtained
adapting to general power functions the approach of
\cite{Ambrosio-Gigli-Savare11}. We will also see that
\begin{equation}\label{allinequalities}
|\nabla f|_{w,q}\leq|\nabla f|_{S,q}\leq|\nabla f|_{C,q}\leq |\nabla
f|_{*,q}\qquad\text{${\mbox{\boldmath$m$}}$-a.e. in $X$.}
\end{equation}
We shall prove in Section~\ref{sequivalence} that actually all
inequalities are equalities, by proving equality of the two extreme
sides. As in the previous section, we shall denote by $p$ the dual
exponent of $q$.
\subsection{Upper gradients}
Following \cite{Heinonen-Koskela98}, we say that a Borel function
$g$ is an upper gradient of a Borel function $f:X\to\mathbb{R}$ if the
inequality
\begin{equation}\label{eq:uppergradient}
\biggl|\int_{\partial\gamma}f\biggr|\leq\int_\gamma g
\end{equation}
holds for all absolutely continuous curves $\gamma:[0,1]\to X$. Here
$\int_{\partial\gamma} f=f(\gamma_1)-f(\gamma_0)$, while
$\int_\gamma g=\int_0^1g(\gamma_s)|\dot\gamma_s|\,\d s$.
It is well-known and easy to check that the slope is an upper
gradient, for locally Lipschitz functions.
\subsection{Cheeger's gradient $|\nabla f|_{C,q}$}
The following definition is taken from \cite{Cheeger00}, where weak
gradients are defined from upper gradients via a relaxation
procedure.
\begin{definition}[$q$-relaxed upper gradient]\label{def:cheeger00} We say that
$g\in L^q(X,{\mbox{\boldmath$m$}})$ is a $q$-relaxed upper gradient of $f\in
L^q(X,{\mbox{\boldmath$m$}})$ if there exist $\tilde{g}\in L^q(X,{\mbox{\boldmath$m$}})$, functions
$f_n\in L^q(X,{\mbox{\boldmath$m$}})$ and upper gradient $g_n$ of $f_n$ such that:
\begin{itemize}
\item[(a)] $f_n\to f$ in $L^q(X,{\mbox{\boldmath$m$}})$ and $g_n$ weakly converge to
$\tilde{g}$ in $L^q(X,{\mbox{\boldmath$m$}})$;
\item[(b)] $\tilde{g}\leq g$ ${\mbox{\boldmath$m$}}$-a.e. in $X$.
\end{itemize}
We say that $g$ is a minimal $q$-relaxed upper gradient of $f$ if
its $L^q(X,{\mbox{\boldmath$m$}})$ norm is minimal among $q$-relaxed upper gradients.
We shall denote by $|\nabla f|_{C,q}$ the minimal $q$-relaxed upper
gradient.
\end{definition}
\subsection{Minimal $q$-relaxed slope $|\nabla f|_{*,q}$}
The second definition of weak gradient we shall consider is a
variant of the previous one and arises by relaxing the integral of
the $q$-th power of the slope of Lipschitz functions. In comparison
with Definition~\ref{def:cheeger00}, we are considering only
Lipschitz approximating functions and we are taking their slopes as
upper gradients. In the spirit of the Sobolev space theory, it
should be considered as an ``$H$ definition'', since an
approximation with Lipschitz functions is involved.
\begin{definition}[Relaxed slope]\label{def:genuppergrad} We say that $g\in L^q(X,{\mbox{\boldmath$m$}})$ is a
$q$-relaxed slope of $f\in L^q(X,{\mbox{\boldmath$m$}})$ if there exist $\tilde{g}\in
L^q(X,{\mbox{\boldmath$m$}})$ and Lipschitz functions $f_n\in L^q(X,{\mbox{\boldmath$m$}})$ such that:
\begin{itemize}
\item[(a)] $f_n\to f$ in $L^q(X,{\mbox{\boldmath$m$}})$ and $|\nabla f_n|$ weakly converge to
$\tilde{g}$ in $L^q(X,{\mbox{\boldmath$m$}})$;
\item[(b)] $\tilde{g}\leq g$ ${\mbox{\boldmath$m$}}$-a.e. in $X$.
\end{itemize}
We say that $g$ is the minimal $q$-relaxed slope of $f$ if its
$L^q(X,{\mbox{\boldmath$m$}})$ norm is minimal among $q$-relaxed slopes. We shall
denote by $\relgradq fq$ the minimal $q$-relaxed slope.
\end{definition}
By this definition and the sequential compactness of weak
topologies, any $L^q$ limit of Lipschitz functions $f_n$ with
$\int|\nabla f_n|^q\,\d{\mbox{\boldmath$m$}}$ uniformly bounded has a $q$-relaxed
slope. On the other hand, using Mazur's lemma (see
\cite[Lemma~4.3]{Ambrosio-Gigli-Savare11} for details), the
definition of $q$-relaxed slope would be unchanged if the weak
convergence of $|\nabla f_n|$ in (a) were replaced by the condition
$|\nabla f_n|\leq g_n$ and $g_n\to\tilde{g}$ strongly in
$L^q(X,{\mbox{\boldmath$m$}})$. This alternative characterization of $q$-relaxed
slopes is suitable for diagonal arguments and proves, together with
\eqref{eq:subadd}, that the collection of $q$-relaxed slopes is a
closed convex set, possibly empty. Hence, thanks to the uniform
convexity of $L^q(X,{\mbox{\boldmath$m$}})$, the definition of $\relgradq fq$ is well
posed. Also, arguing as in \cite{Ambrosio-Gigli-Savare11} and using once more the uniform
convexity of $L^q(X,{\mbox{\boldmath$m$}})$, it is not difficult to show the following result:
\begin{proposition}\label{prop:easy}
If $f\in L^q(X,{\mbox{\boldmath$m$}})$ has a $q$-relaxed slope then there exist
Lipschitz functions $f_n$ satisfying
\begin{equation}\label{densitylip1}
\lim_{n\to\infty}\int_X|f_n-f|^q\,\d{\mbox{\boldmath$m$}}+\int_X\bigl||\nabla
f_n|-|\nabla f|_{*,q}\bigr|^q\,\d{\mbox{\boldmath$m$}}=0.
\end{equation}
\end{proposition}
Since the slope is an upper gradient for Lipschitz functions it
turns out that any $q$-relaxed slope is a $q$-relaxed upper
gradient, hence
\begin{equation}\label{allinequalities1}
|\nabla f|_{C,q}\leq |\nabla f|_{*,q}\qquad\text{${\mbox{\boldmath$m$}}$-a.e. in $X$}
\end{equation}
whenever $f$ has a $q$-relaxed slope.
\begin{remark}\label{rem:whyq}
{\rm Notice that in principle the integrability of $f$ could be
decoupled from the integrability of the gradient, because no global
Poincar\'e inequality can be expected at this level of generality.
Indeed, to increase the symmetry with the next two gradients, one
might even consider the convergence ${\mbox{\boldmath$m$}}$-a.e. of the approximating
functions, removing any integrability assumption. We have left the
convergence in $L^q$ because this presentation is more consistent
with the usual presentations of Sobolev spaces, and the definitions
given in \cite{Cheeger00} and \cite{Ambrosio-Gigli-Savare11}. Using
locality and a truncation argument, the definitions can be extended
to more general classes of functions, see
\eqref{eq:extendedrelaxed}.}\fr
\end{remark}
\subsection{$q$-upper gradients and $|\nabla f|_{S,q}$}
Here we recall a weak definition of upper gradient, taken from
\cite{Koskela-MacManus} and further studied in
\cite{Shanmugalingam00} in connection with the theory of Sobolev
spaces, where we allow for exceptions in \eqref{eq:uppergradient}.
Recall that, for $\Gamma\subset AC([0,1],X)$, the $q$-modulus ${\rm
Mod}_q(\Gamma)$ is defined by (see \cite{Fuglede} for a systematic
analysis of this concept)
\begin{equation}
\label{eq:defmod2} {\rm
Mod}_q(\Gamma):=\inf\Big\{\int_X\rho^q\,\d{\mbox{\boldmath$m$}}: \ \int_\gamma\rho\geq
1\ \ \forall \gamma\in\Gamma\Big\}.
\end{equation}
We say that $\Gamma$ is ${\rm Mod}_q$-negligible if ${\rm
Mod}_q(\Gamma)=0$. Accordingly, we say that a Borel function
$g:X\to[0,\infty]$ is a $q$-upper gradient of $f$ if there exist a
function $\tilde f$ and a ${\rm Mod}_q$-negligible set $\Gamma$ such
that $\tilde{f}=f$ ${\mbox{\boldmath$m$}}$-a.e. in $X$ and
\[
\big|\tilde f(\gamma_0)-\tilde f(\gamma_1)\big|\leq\int_\gamma
g\qquad\forall \gamma\in AC([0,1],X)\setminus\Gamma.
\]
It is not hard to prove that the collection of all $q$-upper
gradients of $f$ is convex and closed, so that we can call minimal
$q$-upper gradient, and denote by $|\nabla f|_{S,q}$, the element
with minimal $L^q(X,{\mbox{\boldmath$m$}})$ norm. Furthermore, the inequality
\begin{equation}\label{allinequalities2}
|\nabla f|_{S,q}\leq|\nabla f|_{C,q}\qquad\text{${\mbox{\boldmath$m$}}$-a.e. in $X$}
\end{equation}
(namely, the fact that all $q$-relaxed upper gradients are $q$-upper
gradients) follows by a stability property of $q$-upper gradients
very similar to the one stated in Theorem~\ref{thm:stabweak} below
for $q$-weak upper gradients, see
\cite[Lemma~4.11]{Shanmugalingam00}. Finally, an observation due to
Fuglede (see Remark~\ref{rem:Fuglede} below) shows that any
$q$-upper gradient can be strongly approximated in $L^q(X,{\mbox{\boldmath$m$}})$ by
upper gradients. This has been used in \cite{Shanmugalingam00} to
show that the equality $|\nabla f|_{S,q}=|\nabla f|_{C,q}$
${\mbox{\boldmath$m$}}$-a.e. in $X$ holds.
\begin{remark}[Fuglede]\label{rem:Fuglede}{\rm
If ${\rm Mod}_q(\Gamma)=0$ and $\eps>0$, then we can find $\rho\in
L^q(X,{\mbox{\boldmath$m$}})$ with $\|\rho\|_q<\eps$ and $\int_\gamma\rho=\infty$ for
all $\gamma\in\Gamma$. Indeed, if we choose functions $\rho_n\in
L^q(X,{\mbox{\boldmath$m$}})$ with $\|\rho_n\|_q<1/n$ and $\int_\gamma\rho_n\geq 1$
for all $\gamma\in\Gamma$, the function
$$
\rho:=\sum_{n\geq 1} \frac\delta{n}\rho_n
$$
has the required property for $\delta=\delta(\eps)>0$ small
enough.}\fr
\end{remark}
\subsection{$q$-weak upper gradients and $|\nabla f|_{w,q}$}
Recall that the evaluation maps ${\mathrm e}_t:C([0,1],X)\to X$ are defined
by ${\mathrm e}_t(\gamma):=\gamma_t$. We also introduce the restriction
maps ${\rm restr}_t^s: C([0,1],X)\to C([0,1],X)$, $0\le t\le s\le
1$, given by
\begin{equation}
{\rm restr}_t^s(\gamma)_r:=\gamma_{(1-r)t+rs},\label{eq:93}
\end{equation}
so that ${\rm restr}_t^s$ ``stretches'' the restriction of the curve
to $[s,t]$ to the whole of $[0,1]$.
Our definition of $q$-weak upper gradient still allows for
exceptions in \eqref{eq:uppergradient}, but with a different notion
of exceptional set, see also Remark~\ref{rem:comparenullsets} below.
\begin{definition}[Test plans and negligible sets of curves]\label{def:testplans}
We say that a probability measure ${\mbox{\boldmath$\pi$}}\in\Probabilities{C([0,1],X)}$ is a
$p$-\emph{test plan} if ${\mbox{\boldmath$\pi$}}$ is concentrated on $AC^p([0,1],X)$,
$\iint_0^1|\dot\gamma_t|^p\d t\,\d{\mbox{\boldmath$\pi$}}<\infty$ and there exists a
constant $C({\mbox{\boldmath$\pi$}})$ such that
\begin{equation}
(\e_t)_\#{\mbox{\boldmath$\pi$}} \leq C({\mbox{\boldmath$\pi$}}){\mbox{\boldmath$m$}}\qquad\forall t\in[0,1].
\label{eq:1}
\end{equation}
A Borel set $A\subset C([0,1],X)$ is said to be
$q$-\emph{negligible} if ${\mbox{\boldmath$\pi$}}(A)=0$ for any $p$-test plan ${\mbox{\boldmath$\pi$}}$. A
property which holds for every $\gamma\in C([0,1],X)$, except
possibly a $q$-negligible set, is said to hold for $q$-almost every
curve.
\end{definition}
Observe that, by definition, $C([0,1],X)\setminus AC^p([0,1],X)$ is
$q$-negligible, so the notion starts to be meaningful when we look
at subsets $A$ of $AC^p([0,1],X)$.
\begin{remark}
\label{re:easy}
\upshape
An easy consequence of condition \eqref{eq:1} is that if two
${\mbox{\boldmath$m$}}$-measurable functions $f,\,g:X\to\mathbb{R}$ coincide up to a
${\mbox{\boldmath$m$}}$-negligible set and $\mathcal T$ is an at most countable subset
of $[0,1]$, then the functions
$f\circ \gamma$ and $g\circ \gamma$ coincide in $\mathcal T$
for $q$-almost every curve
$\gamma$.
Moreover, choosing an arbitrary $p$-test plan ${\mbox{\boldmath$\pi$}}$ and applying Fubini's
Theorem to the product measure $\Leb 1\times {\mbox{\boldmath$\pi$}}$
in $(0,1)\times C([0,1];X)$ we also obtain that
$f\circ\gamma=g\circ\gamma$ $\Leb 1$-a.e.\ in $(0,1)$ for
${\mbox{\boldmath$\pi$}}$-a.e.\ curve $\gamma$; since ${\mbox{\boldmath$\pi$}}$ is arbitrary, the same
property holds for $q$-a.e.\ $\gamma$.
\end{remark}
Coupled with the definition of
$q$-negligible set of curves, there are the definitions of $q$-weak upper gradient and
of functions which are Sobolev along $q$-a.e. curve.
\begin{definition}[$q$-weak upper gradients]
A Borel function
$g:X\to[0,\infty]$ is a $q$-weak upper gradient of $f:X\to \mathbb{R}$ if
\begin{equation}
\label{eq:inweak} \left|\int_{\partial\gamma}f\right|\leq
\int_\gamma g\qquad\text{for $q$-a.e. $\gamma$.}
\end{equation}
\end{definition}
\begin{definition}[Sobolev functions along $q$-a.e. curve]
A function $f:X\to\mathbb{R}$ is Sobolev along $q$-a.e. curve if for
$q$-a.e. curve $\gamma$ the function $f\circ\gamma$ coincides a.e.
in $[0,1]$ and in $\{0,1\}$ with an absolutely continuous map
$f_\gamma:[0,1]\to\mathbb{R}$.
\end{definition}
By Remark \ref{re:easy} applied to $\mathcal T:=\{0,1\}$, \eqref{eq:inweak} does not depend on
the particular representative of $f$ in the class of ${\mbox{\boldmath$m$}}$-measurable
function coinciding with $f$ up to a ${\mbox{\boldmath$m$}}$-negligible set.
The same Remark also shows that the property of being Sobolev along
$q$-q.e.\ curve $\gamma$ is independent of the representative in the
class of ${\mbox{\boldmath$m$}}$-measurable functions coinciding with $f$ ${\mbox{\boldmath$m$}}$-a.e.\
in $X$.
In the next remark, using Lemma~\ref{lem:Fibonacci}, we prove that the existence of a $q$-weak
upper gradient $g$ such that $\int_\gamma g<\infty$ for
$q$-a.e.\ $\gamma$ (in particular if $g\in L^q(X,{\mbox{\boldmath$m$}})$) implies Sobolev regularity along $q$-a.e.\ curve. Notice that
only recently we realized that the validity of this implication, compare with the definitions
given in \cite{Ambrosio-Gigli-Savare11}, only apparently stronger.
\begin{remark}[Equivalence with the axiomatization in \cite{Ambrosio-Gigli-Savare11}]
\label{re:restr}{\rm Notice that if ${\mbox{\boldmath$\pi$}}$ is a $p$-test plan, so is $({\rm
restr}_t^s)_\sharp{\mbox{\boldmath$\pi$}}$. Hence if $g$ is a $q$-weak upper gradient
of $f$ such that $\int_\gamma g<\infty$ for
$q$-a.e.\ $\gamma$, then for every $t<s$ in $[0,1]$ it holds
\[
|f(\gamma_s)-f(\gamma_t)|\leq \int_t^s g(\gamma_r)|\dot\gamma_r|\,\d
r \qquad\text{for $q$-a.e. $\gamma$.}
\]
Let ${\mbox{\boldmath$\pi$}}$ be a $p$-test plan: by Fubini's theorem applied
to the product measure $\Leb2\times{\mbox{\boldmath$\pi$}}$ in $(0,1)^2\times
C([0,1];X)$, it follows that for ${\mbox{\boldmath$\pi$}}$-a.e. $\gamma$ the function
$f$ satisfies
\[
|f(\gamma_s)-f(\gamma_t)|\leq \Bigl|\int_t^s g(\gamma_r)|\dot\gamma_r|\,\d
r \Bigr|\qquad\text{for $\Leb{2}$-a.e. $(t,s)\in (0,1)^2$.}
\]
An analogous argument shows that
\begin{equation}
\label{eq:2}
\left\{
\begin{aligned}
\textstyle |f(\gamma_s)-f(\gamma_0)|&\textstyle
\leq \int_0^s
g(\gamma_r)|\dot\gamma_r|\,\d r\\
\textstyle |f(\gamma_1)-f(\gamma_s)|&\textstyle \leq \int_s^1
g(\gamma_r)|\dot\gamma_r|\,\d r
\end{aligned}\right.
\qquad\text{for $\Leb{1}$-a.e. $s\in (0,1)$.}
\end{equation}
Since $g\circ \gamma|\dot \gamma|\in L^1(0,1)$ for
${\mbox{\boldmath$\pi$}}$-a.e.\ $\gamma$,
by Lemma~\ref{lem:Fibonacci} it follows that $f\circ\gamma\in W^{1,1}(0,1)$
for ${\mbox{\boldmath$\pi$}}$-a.e. $\gamma$, and
\begin{equation}\label{eq:pointwisewug}
\biggl|\frac{\d}{{\d t}}(f\circ\gamma)\biggr|\leq
g\circ\gamma|\dot\gamma|\quad\text{a.e. in $(0,1)$, for
${\mbox{\boldmath$\pi$}}$-a.e. $\gamma$.}
\end{equation}
Since ${\mbox{\boldmath$\pi$}}$ is arbitrary, we conclude that $f\circ\gamma\in
W^{1,1}(0,1)$ for $q$-a.e.\ $\gamma$, and therefore it admits an
absolutely continuous representative $f_\gamma$; moreover,
by \eqref{eq:2}, it is immediate to check that $f(\gamma(t))=f_\gamma(t)$ for $t\in \{0,1\}$ and $q$-a.e.\ $\gamma$.
\fr }
\end{remark}
Using the same argument given in the previous remark it is
immediate to show that if $f$ is Sobolev along $q$-a.e. curve it holds
\begin{equation}\label{eq:locweak}
\text{$g_i$, $i=1,2$ $q$-weak upper gradients of $f$}\quad\Longrightarrow\quad \text{$\min\{g_1,g_2\}$ $q$-weak upper gradient of $f$.}
\end{equation}
Using this stability property we can recover, again, a distinguished
minimal object.
\begin{definition}[Minimal $q$-weak upper gradient]
Let $f:X\to\mathbb{R}$ be Sobolev along $q$-a.e. curve.
The minimal $q$-weak upper gradient $\weakgradq fq$ of $f$
is the $q$-weak upper gradient characterized, up to
${\mbox{\boldmath$m$}}$-negligible sets, by the property
\begin{equation}\label{eq:defweakgrad}
\weakgradq fq\leq g\qquad\text{${\mbox{\boldmath$m$}}$-a.e. in $X$, for every $q$-weak upper
gradient $g$ of $f$.}
\end{equation}
\end{definition}
Uniqueness of the minimal weak upper gradient is obvious. For
existence, since ${\mbox{\boldmath$m$}}$ is $\sigma$-finite we can find a Borel and
${\mbox{\boldmath$m$}}$-integrable function $\theta:X\to (0,\infty)$ and $\weakgradq
fq :=\inf_n g_n$, where $g_n$ are $q$-weak upper gradients which
provide a minimizing sequence in
$$
\inf\left\{\int_X \theta\, {\rm tan}^{-1}g\,\d{\mbox{\boldmath$m$}}:\ \text{$g$ is a
$q$-weak upper gradient of $f$}\right\}.
$$
We immediately see, thanks to \eqref{eq:locweak}, that we can assume
with no loss of generality that $g_{n+1}\leq g_n$. Hence, by
monotone convergence, the function $\weakgradq fq$ is a $q$-weak
upper gradient of $f$ and $\int_X \theta\,{\rm tan}^{-1}g\,\d{\mbox{\boldmath$m$}}$ is
minimal at $g=\weakgrad fq$. This minimality, in conjunction with
\eqref{eq:locweak}, gives \eqref{eq:defweakgrad}.
\begin{remark}\label{rem:comparenullsets}{\rm
Observe that for a Borel set $\Gamma\subset C([0,1],X)$ and a test
plan ${\mbox{\boldmath$\pi$}}$, integrating on $\Gamma$ w.r.t. ${\mbox{\boldmath$\pi$}}$ the inequality
$\int_\gamma \rho\geq 1$ and then minimizing over $\rho$, we get
$$
{\mbox{\boldmath$\pi$}}(\Gamma)\leq (C({\mbox{\boldmath$\pi$}}))^{1/q}\bigl({\rm
Mod}_q(\Gamma)\bigr)^{1/q}\biggl(\iint_0^1|\dot\gamma|^p\,\d
s\,\d{\mbox{\boldmath$\pi$}}(\gamma)\biggr)^{1/p},
$$
which shows that any ${\rm Mod}_q$-negligible set of curves is also
$q$-negligible according to Definition~\ref{def:testplans}. This
immediately gives that any $q$-upper gradient is a $q$-weak upper
gradient, so that
\begin{equation}\label{allinequalities3}
|\nabla f|_{w,q}\leq|\nabla f|_{S,q}\qquad\text{${\mbox{\boldmath$m$}}$-a.e. in $X$.}
\end{equation}
}\fr
\end{remark}
Notice that the combination of \eqref{allinequalities1},
\eqref{allinequalities2} and \eqref{allinequalities3} gives
\eqref{allinequalities}.
\section{Some properties of weak gradients}
In order to close the chain of inequalities in
\eqref{allinequalities} we need some properties of the weak
gradients introduced in the previous section. The following locality
lemma follows by the same arguments in \cite{Cheeger00} or adapting
to the case $q\neq 2$ the proof in
\cite[Lemma~4.4]{Ambrosio-Gigli-Savare11}.
\begin{lemma}[Pointwise minimality of $\relgradq fq$]\label{le:local}
Let $g_1,\,g_2$ be two $q$-relaxed slopes of $f$. Then
$\min\{g_1,g_2\}$ is a $q$-relaxed slope as well. In particular, not
only the $L^q$ norm of $\relgradq fq$ is minimal, but also
$\relgradq fq\leq g$ ${\mbox{\boldmath$m$}}$-a.e. in $X$ for any relaxed slope $g$ of
$f$.
\end{lemma}
The previous pointwise minimality property immediately yields
\begin{equation}
\label{eq:facile} \relgradq fq\leq |\nabla f|\qquad\text{${\mbox{\boldmath$m$}}$-a.e.
in $X$}
\end{equation}
for any Lipschitz function $f:X\to\mathbb{R}$.
Also the proof of locality and chain rule is quite standard, see
\cite{Cheeger00} and \cite[Proposition~4.8]{Ambrosio-Gigli-Savare11}
for the case $q=2$ (the same proof works in the general case).
\begin{proposition}[Locality and chain rule]\label{prop:chain}
If $f\in L^q(X,{\mbox{\boldmath$m$}})$ has a $q$-relaxed slope, the following
properties hold.
\begin{itemize}
\item[(a)] $\relgradq hq=\relgradq fq$ ${\mbox{\boldmath$m$}}$-a.e. in $\{h=f\}$
whenever $f$ has a $q$-relaxed slope.
\item[(b)] $\relgradq {\phi(f)}q\leq |\phi'(f)|\relgradq fq$ for any $C^1$ and Lipschitz function
$\phi$ on an interval containing the image of $f$. Equality holds if
$\phi$ is nondecreasing.
\end{itemize}
\end{proposition}
Next we consider the stability of $q$-weak upper gradients (as we
said, similar properties hold for $q$-upper gradients, see
\cite[Lemma~4.11]{Shanmugalingam00} but we shall not need them).
\begin{theorem}[Stability w.r.t. ${\mbox{\boldmath$m$}}$-a.e. convergence]\label{thm:stabweak}
Assume that $f_n$ are ${\mbox{\boldmath$m$}}$-measurable, Sobolev along $q$-a.e. curve
and that $g_n\in L^q(X,{\mbox{\boldmath$m$}})$ are $q$-weak upper gradients of $f_n$. Assume
furthermore that $f_n(x)\to f(x)\in\mathbb{R}$ for ${\mbox{\boldmath$m$}}$-a.e. $x\in X$ and
that $(g_n)$ weakly converges to $g$ in $L^q(X,{\mbox{\boldmath$m$}})$. Then $g$ is a
$q$-weak upper gradient of $f$.
\end{theorem}
\begin{proof}
Fix a $p$-test plan ${\mbox{\boldmath$\pi$}}$ and $\theta\in L^1(X,{\mbox{\boldmath$m$}})$ strictly
positive (its existence is ensured by the $\sigma$-finiteness
assumption on ${\mbox{\boldmath$m$}}$). By Mazur's theorem we can find convex
combinations
$$
h_n:=\sum_{i=N_h+1}^{N_{h+1}}\alpha_ig_i\qquad\text{with
$\alpha_i\geq 0$, $\sum_{i=N_h+1}^{N_{h+1}}\alpha_i=1$,
$N_h\to\infty$}
$$
converging strongly to $g$ in $L^q(X,{\mbox{\boldmath$m$}})$. Denoting by $\tilde f_n$
the corresponding convex combinations of $f_n$, $h_n$ are weak upper
gradients of $\tilde f_n$ and still $\tilde f_n\to f$ ${\mbox{\boldmath$m$}}$-a.e. in
$X$.
Since for every nonnegative Borel function $\varphi:X\to [0,\infty]$
it holds (with $C=C({\mbox{\boldmath$\pi$}})$)
\begin{align}
\notag\int\Big(\int_{\gamma}\varphi\Big)\,\d{\mbox{\boldmath$\pi$}}&=
\int\Big(\int_0^1 \varphi(\gamma_t)|\dot
\gamma_t|\,\d t\Big)\,\d{\mbox{\boldmath$\pi$}}
\le
\int\Big(\int_0^1\varphi^q(\gamma_t)\,\d
t\Big)^{1/q}
\Big(\int_0^1 |\dot
\gamma_t|^p\,\d t\Big)^{1/p}\,\d{\mbox{\boldmath$\pi$}}
\\&\notag
\le
\Big(\int_0^1 \int\varphi^q\,\d({\mathrm e}_t)_\sharp{\mbox{\boldmath$\pi$}}\,\d
t\Big)^{1/q}
\Big(\iint_0^1|\dot\gamma_t|^p\,\d t\,\d{\mbox{\boldmath$\pi$}}\Big)^{1/p}
\\ &\le \Big(C\int\varphi^q\,\d{\mbox{\boldmath$m$}}\Big)^{1/q} \Big(\iint_0^1|\dot\gamma_t|^p\,\d t\,\d{\mbox{\boldmath$\pi$}}\Big)^{1/p},
\label{eq:21}
\end{align}
we obtain, for $\bar C:=C^{1/q}\Big(\iint_0^1|\dot\gamma_t|^p\,\d
t\,\d{\mbox{\boldmath$\pi$}}\Big)^{1/p}$
\begin{align*}
\int&\biggl(\int_{\gamma}|h_n-g|+\min\{|\tilde{f}_n-f|,\theta\}\biggr)\,\d{\mbox{\boldmath$\pi$}}\leq
\bar C\Big(\|h_n-g\|_q+ \|\min\{|\tilde{f}_n-f|,\theta\}\|_q\Big)
\to 0.
\end{align*}
By a diagonal argument we can find a subsequence $n(k)$ such that
$$\int_\gamma|h_{n(k)}-g|+\min\{|\tilde{f}_{n(k)}-f|,\theta\}\to 0$$ as
$k\to\infty$ for ${\mbox{\boldmath$\pi$}}$-a.e. $\gamma$. Since $\tilde{f}_n$ converge
${\mbox{\boldmath$m$}}$-a.e. to $f$ and the marginals of ${\mbox{\boldmath$\pi$}}$ are absolutely
continuous w.r.t. ${\mbox{\boldmath$m$}}$ we have also that for ${\mbox{\boldmath$\pi$}}$-a.e. $\gamma$
it holds $\tilde{f}_n(\gamma_0)\to f(\gamma_0)$ and
$\tilde{f}_n(\gamma_1)\to f(\gamma_1)$.
If we fix a curve $\gamma$ satisfying these convergence properties,
since $(\tilde{f}_{n(k)})_\gamma$ are equi-absolutely continuous
(being their derivatives bounded by
$h_{n(k)}\circ\gamma|\dot\gamma|$) and a further subsequence of
$\tilde{f}_{n(k)}$ converges a.e. in $[0,1]$ and in $\{0,1\}$ to
$f(\gamma_s)$, we can pass to the limit to obtain an absolutely
continuous function $f_\gamma$ equal to $f(\gamma_s)$ a.e. in
$[0,1]$ and in $\{0,1\}$ with derivative bounded by
$g(\gamma_s)|\dot\gamma_s|$. Since ${\mbox{\boldmath$\pi$}}$ is arbitrary we conclude
that $f$ is Sobolev along $q$-a.e. curve and that $h$ is a weak
upper gradient of $f$.
\end{proof}
It is natural to ask whether $r$-upper gradients really depend on $r$ or not.
A natural conjecture is the following: let $r\in (1,\infty)$ and $f:X\to\mathbb{R}$ Borel. Assume that ${\mbox{\boldmath$m$}}$ is a
finite measure and that $f$ has a $r$-upper gradient in
$L^r(X,{\mbox{\boldmath$m$}})$. Then, for all $q\in (1,r]$, $f$ has a $q$-upper
gradient and $|\nabla f|_{S,q}=|\nabla f|_{S,r}$ ${\mbox{\boldmath$m$}}$-a.e. in $X$.
Notice however that the ``converse'' implication, namely
\begin{equation}\label{eq:koskela}
\qquad\text{$f$ has a $q$-upper gradient in
$L^r(X,{\mbox{\boldmath$m$}})$}\,\,\Rightarrow\,\,\text{$f$ has a $r$-upper gradient
in $L^r(X,{\mbox{\boldmath$m$}})$}
\end{equation}
for $1<q<r<\infty$ does not hold in general. A counterexample has
been shown to us by P.Koskela: consider the set $X$ equal to the
union of the first and third quadrant in $\mathbb{R}^2$, and take as
function $f$ the characteristic function of the first quadrant.
Since the collection of all curves passing from the first to the
third quadrant is ${\rm Mod}_2$-negligible (just take, for $\alpha\in (0,1)$,
the family of curves $\rho_\alpha(x)=\alpha|x|^{\alpha-1}$, and let $\alpha\downarrow 0$)
it follows that $f$ has a $2$-upper gradient equal to $0$. On the other hand, $f$ is discontinuous along the pencil of
curves $\gamma_\theta(t):=(2t-1)(\cos\theta,\sin\theta)$ indexed by
$\theta\in [0,\pi/2]$, and since this family of curves is not ${\rm
Mod}_r$-negligible for $r>2$ it follows that \eqref{eq:koskela} fails for $f$.
In order to show that the family of curves is not ${\rm Mod}_r$-negligible for $r>2$,
suffices to notice that $\int_{\gamma_\theta}g\geq 1$ implies
$$
\frac{1}{2}\leq\biggl(\int_0^1 g^r(\gamma_\theta(t))|2t-1|\,\d t\biggr)^{1/r}
\biggl(\int_0^1|2t-1|^{-r'/r}\,{\d t}\biggr)^{1/r'}.
$$
Since $r>2$ implies $r'/r<1$, integrating both sides in $[0,\pi/2]$ gives
a lower bound on the $L^r$ norm of $g$ with a positive constant $c(r)$.
In the presence of doubling and a $(1,q)$-Poincar\'e inequality,
\eqref{eq:koskela} holds, following the Lipschitz approximation
argument in Theorem~4.14 and Theorem~4.24 of \cite{Cheeger00} (we
shall not need this fact in the sequel).
\section{Cheeger's functional and its gradient flow}
In this section we assume that $(X,{\sf d})$ is complete and separable
and that ${\mbox{\boldmath$m$}}$ is a finite Borel measure. As in the previous
sections, $q\in (1,\infty)$ and $p$ is the dual exponent. In order
to apply the theory of gradient flows of convex functionals in
Hilbert spaces, when $q>2$ we need to extend $\relgradq fq$ also to
functions in $L^2(X,{\mbox{\boldmath$m$}})$ (because Definition~\ref{def:genuppergrad}
was given for $L^q(X,{\mbox{\boldmath$m$}})$ functions). To this aim, we denote
$f_N:=\max\{-N,\min\{f,N\}\}$ and set
\begin{equation}\label{eq:mathcalC}
\mathcal C:=\left\{f:X\to\mathbb{R}:\ \text{$f_N$ has a $q$-relaxed slope
for all $N\in\mathbb{N}$}\right\}.
\end{equation}
Accordingly, for all $f\in\mathcal C$ we set
\begin{equation}\label{eq:extendedrelaxed}
\relgradq fq:=\relgradq {f_N}q\qquad\text{${\mbox{\boldmath$m$}}$-a.e. in $\{|f|<N\}$}
\end{equation}
for all $N\in\mathbb{N}$. We can use the locality property in
Proposition~\ref{prop:chain}(a) to show that this definition is well
posed, up to ${\mbox{\boldmath$m$}}$-negligible sets, and consistent with the previous
one. Furthermore, locality and chain rules still apply, so we shall
not use a distinguished notation for the new gradient.
Although we work with a stronger definition of weak gradient,
compared to $|\nabla f|_{C,q}$, we call Cheeger's $q$-functional the
energy on $L^2(X,{\mbox{\boldmath$m$}})$ defined by
\begin{equation}\label{def:Cheeger}
\mathbb{C}_q(f):=\frac{1}{q}\int_X |\nabla f|_{*,q}^q \,\d{\mbox{\boldmath$m$}},
\end{equation}
set to $+\infty$ if $f\in L^2(X,{\mbox{\boldmath$m$}})\setminus\mathcal C$.
\begin{theorem} \label{thm:cheeger} Cheeger's $q$-functional
$\mathbb{C}_q$ is convex and lower semicontinuous in $L^2(X,{\mbox{\boldmath$m$}})$.
\end{theorem}
\begin{proof} The proof of convexity is elementary, so we focus on
lower semicontinuity. Let $(f_n)$ be convergent to $f$ in
$L^2(X,{\mbox{\boldmath$m$}})$ and we can assume, possibly extracting a subsequence
and with no loss of generality, that $\mathbb{C}_q(f_n)$ converges to a
finite limit.
Assume first that all $f_n$ have $q$-relaxed slope, so that that
$\relgradq {f_n}q$ is uniformly bounded in $L^q(X,{\mbox{\boldmath$m$}})$. Let
$f_{n(k)}$ be a subsequence such that $\relgradq {f_{n(k)}}q$ weakly
converges to $g$ in $L^q(X,{\mbox{\boldmath$m$}})$. Then $g$ is a $q$-relaxed slope of
$f$ and
$$
\mathbb{C}_q(f)\leq\frac1q\int_X|g|^q\,\d{\mbox{\boldmath$m$}}\leq\liminf_{k\to\infty}\frac1q
\int_X|\nabla f_{n(k)}|^q_{*,q}\,\d{\mbox{\boldmath$m$}}
=\liminf_{n\to\infty}\mathbb{C}_q(f_n).
$$
In the general case when $f_n\in{\mathcal C}$ we consider the
functions $f^N_n:=\max\{-N,\min\{f,N\}\}$ to conclude from the
inequality $|\nabla f^N_n|_{*,q}\leq|\nabla f_n|_{*,q}$ that
$f^N:=\max\{-N,\min\{f,N\}\}$ has $q$-relaxed slope for any $N\in\mathbb{N}$
and
$$
\int_X|\nabla f^N|_{*,q}^q\,\d{\mbox{\boldmath$m$}}\leq \liminf_{n\to\infty}
\int_X|\nabla f^N_n|_{*,q}^q\,\d{\mbox{\boldmath$m$}}\leq\liminf_{n\to\infty}
\int_X|\nabla f_n|_{*,q}^q\,\d{\mbox{\boldmath$m$}}.
$$
Passing to the limit as $N\to\infty$, the conclusion follows by
monotone convergence.
\end{proof}
\begin{remark}\label{rem:basiclsc} {\rm More generally, the same argument proves the
$L^2(X,{\mbox{\boldmath$m$}})$-lower semicontinuity of the functional
$$
f\mapsto\int_X \frac{|\nabla f|_{*,q}^q}{|f|^\alpha}\,\d{\mbox{\boldmath$m$}}
$$
in $\mathcal C$, for any $\alpha>0$. Indeed, locality and chain rule
allow the reduction to nonnegative functions $f_n$ and we can use
the truncation argument of Theorem~\ref{thm:cheeger} to reduce
ourselves to functions with values in an interval $[c,C]$ with
$0<c\leq C<\infty$. In this class, we can again use the chain rule
to prove the identity
$$
\int_X|\nabla f^\beta|^q_{*,q}\,\d{\mbox{\boldmath$m$}}= |\beta|^q\int_X\frac{|\nabla
f|_{*,q}^q}{|f|^\alpha}\,\d{\mbox{\boldmath$m$}}
$$
with $\beta:=1-\alpha/q$ to obtain the result when $\alpha\neq q$.
If $\alpha=q$ we use a logarithmic transformation.
}\fr
\end{remark}
Since the finiteness domain of $\mathbb{C}_q$ is dense in $L^2(X,{\mbox{\boldmath$m$}})$ (it
includes bounded Lipschitz functions), the Hilbertian theory of
gradient flows (see for instance \cite{Brezis73},
\cite{Ambrosio-Gigli-Savare08}) can be applied to Cheeger's
functional \eqref{def:Cheeger} to provide, for all $f_0\in
L^2(X,{\mbox{\boldmath$m$}})$, a locally absolutely continuous map $t\mapsto f_t$ from
$(0,\infty)$ to $L^2(X,{\mbox{\boldmath$m$}})$, with $f_t\to f_0$ as $t\downarrow 0$,
whose derivative satisfies
\begin{equation}\label{eq:ODE}
\frac{d}{dt}f_t\in -\partial^-\mathbb{C}_q(f_t)\qquad\text{for a.e. $t\in
(0,\infty)$.}
\end{equation}
Having in mind the regularizing effect of gradient flows, namely the
selection of elements with minimal $L^2(X,{\mbox{\boldmath$m$}})$ norm in
$\partial^-\mathbb{C}_q$, the following definition is natural.
\begin{definition}[$q$-Laplacian]\label{def:delta}
The $q$-Laplacian $\Delta_q f$ of $f\in L^2(X,{\mbox{\boldmath$m$}})$ is defined for
those $f$ such that $\partial^-\mathbb{C}_q(f)\neq\emptyset$. For those $f$,
$-\Delta_q f$ is the element of minimal $L^2(X,{\mbox{\boldmath$m$}})$ norm in
$\partial^-\mathbb{C}_q(f)$. The domain of $\Delta_q$ will be denoted by
$D(\Delta_q)$.
\end{definition}
\begin{remark}[Potential lack of linearity]\label{re:laplnonlin}{\rm
It should be observed that, even in the case $q=2$, in general the
Laplacian is \emph{not} a linear operator. Still, the trivial
implication
\[
v\in\partial^- \mathbb{C}_q(f)\qquad\Longrightarrow\qquad \lambda^{q-1}
v\in\partial^- \mathbb{C}_q(\lambda f),\quad\forall \lambda\in\mathbb{R},
\]
ensures that the $q$-Laplacian (and so the gradient flow of $\mathbb{C}_q$)
is $(q-1)$-homogenous. }\fr
\end{remark}
We can now write
$$
\frac{\d}{\d t}f_t=\Delta_q f_t
$$
for gradient flows $f_t$ of $\mathbb{C}_q$, the derivative being understood
in $L^2(X,{\mbox{\boldmath$m$}})$, in accordance with the classical case.
\begin{proposition}[Integration by parts]
\label{prop:deltaineq} For all $f\in D(\Delta_q)$, $g\in D(\mathbb{C}_q)$ it
holds
\begin{equation}
\label{eq:delta1} -\int_X g\Delta_q f\,\d{\mbox{\boldmath$m$}}\leq \int_X \relgradq
gq|\nabla f|_{*,q}^{q-1}\,\d{\mbox{\boldmath$m$}}.
\end{equation}
Equality holds if $g=\phi(f)$ with $\phi\in C^1(\mathbb{R})$ with bounded
derivative on the image of $f$.
\end{proposition}
\begin{proof}
Since $-\Delta_q f\in\partial^-\mathbb{C}_q(f)$ it holds
\[
\mathbb{C}_q(f)-\int_X \eps g\Delta_q f\,\d{\mbox{\boldmath$m$}}\leq \mathbb{C}_q(f+\eps
g),\qquad\forall g\in L^q(X,{\mbox{\boldmath$m$}}),\,\,\eps\in\mathbb{R}.
\]
For $\eps>0$, $\relgradq fq+\eps \relgradq gq$ is a $q$-relaxed
slope of $f+\eps g$ (possibly not minimal) whenever $f$ and $g$ have
$q$-relaxed slope. By truncation, it is immediate to obtain from
this fact that $f,\,g\in\mathcal C$ implies $f+\eps g\in\mathcal C$
and
$$
\relgradq {(f+\eps g)}q \leq\relgradq fq+\eps \relgradq
gq\qquad\text{${\mbox{\boldmath$m$}}$-a.e. in $X$.}
$$
Thus it holds $q\mathbb{C}_q(f+\eps g)\leq\int_X(\relgradq fq+\eps\relgradq
gq)^q\,\d{\mbox{\boldmath$m$}}$ and therefore
\[
-\int_X\eps g\Delta_q f\,\d{\mbox{\boldmath$m$}}\leq \frac1q\int_X(\relgradq
fq+\eps\relgradq gq)^q-|\nabla f|_{*,q}^q\,\d{\mbox{\boldmath$m$}}=\eps\int_X\relgradq
gq|\nabla f|^{q-1}_{*,q}\,\d{\mbox{\boldmath$m$}}+o(\eps).
\]
Dividing by $\eps$ and letting $\eps\downarrow 0$ we get
\eqref{eq:delta1}.
For the second statement we recall that $\relgradq {(f+\eps
\phi(f))}q=(1+\eps \phi'(f))\relgradq fq$ for $|\eps|$ small enough.
Hence
\[
\mathbb{C}_q(f+\eps \phi(f))-\mathbb{C}_q(f)= \frac{1}{q}\int_X|\nabla
f|_{*,q}^q\bigl((1+\eps \phi'(f))^q-1\bigr)\,\d{\mbox{\boldmath$m$}}=\eps\int_X|\nabla
f|_{*,q}^q \phi'(f)\,\d{\mbox{\boldmath$m$}}+o(\eps),
\]
which implies that for any $v\in \partial^-\mathbb{C}_q(f)$ it holds
$\int_Xv \phi(f)\,\d{\mbox{\boldmath$m$}}=\int_X|\nabla f|_{*,q}^q\phi'(f)\,\d{\mbox{\boldmath$m$}}$,
and gives the thesis with $v=-\Delta_q f$.
\end{proof}
\begin{proposition}[Some properties of the gradient flow of $\mathbb{C}_q$]\label{prop:basecal}
Let $f_0\in L^2(X,{\mbox{\boldmath$m$}})$ and let $(f_t)$ be the gradient flow of
$\mathbb{C}_q$ starting from $f_0$. Then the following properties hold.\\*
\noindent (Mass preservation) $\int f_t\,\d{\mbox{\boldmath$m$}}=\int f_0\,\d{\mbox{\boldmath$m$}}$ for
any $t\geq 0$.\\* \noindent (Maximum principle) If $f_0\leq C$
(resp. $f_0\geq c$) ${\mbox{\boldmath$m$}}$-a.e. in $X$, then $f_t\leq C$ (resp
$f_t\geq c$) ${\mbox{\boldmath$m$}}$-a.e. in $X$ for any $t\geq 0$.\\* (Energy
dissipation) Suppose $0<c\leq f_0\leq C<\infty$ ${\mbox{\boldmath$m$}}$-a.e. in $X$
and $\Phi\in C^2([c,C])$. Then $t\mapsto\int\Phi(f_t)\,\d{\mbox{\boldmath$m$}}$ is
locally absolutely continuous in $(0,\infty)$ and it holds
\[
\frac{\d}{{\d t}}\int \Phi(f_t)\,\d{\mbox{\boldmath$m$}}=-\int\Phi''(f_t)|\nabla
f_t|_{*,q}^q\,\d{\mbox{\boldmath$m$}}\qquad\text{for a.e. $t\in (0,\infty)$.}
\]
\end{proposition}
\begin{proof} (Mass preservation) Just notice that from \eqref{eq:delta1} we get
\[
\left|\frac{\d}{{\d t}}\int f_t\,\d{\mbox{\boldmath$m$}}\right|=\left|\int
\mathbf{1}\cdot\Delta_q f_t\,\d{\mbox{\boldmath$m$}}\right|\leq\int\relgradq{\mathbf
1}q{|\nabla f_t|_{*,q}^q}\,\d{\mbox{\boldmath$m$}}=0\quad\text{for a.e. $t>0$},
\]
where $\mathbf 1$ is the function identically equal to 1, which has
minimal $q$-relaxed slope equal to 0 by \eqref{eq:facile}.\\*
(Maximum principle) Fix $f\in L^2(X,{\mbox{\boldmath$m$}})$, $\tau>0$ and, according
to the so-called implicit Euler scheme, let $f^\tau$ be the unique
minimizer of
\[
g\qquad\mapsto\qquad \mathbb{C}_q(g)+\frac{1}{2\tau}\int_X|g-f|^2\,\d{\mbox{\boldmath$m$}}.
\]
Assume that $f\leq C$. We claim that in this case $f^\tau\leq C$ as
well. Indeed, if this is not the case we can consider the competitor
$g:=\min\{f^\tau,C\}$ in the above minimization problem. By locality
we get $\mathbb{C}(g)\leq\mathbb{C}(f^\tau)$ and the $L^2$ distance of $f$ and $g$
is strictly smaller than the one of $f$ and $f^\tau$ as soon as
${\mbox{\boldmath$m$}}(\{f^\tau>C\})>0$, which is a contradiction. Starting from
$f_0$, iterating this procedure, and using the fact that the
implicit Euler scheme converges as $\tau\downarrow 0$ (see
\cite{Brezis73}, \cite{Ambrosio-Gigli-Savare08} for details) to the
gradient flow we get the conclusion.\\* (Energy dissipation) Since
$t\mapsto f_t\in L^2(X,{\mbox{\boldmath$m$}})$ is locally absolutely continuous and,
by the maximum principle, $f_t$ take their values in $[c,C]$
${\mbox{\boldmath$m$}}$-a.e., from the fact that $\Phi$ is Lipschitz in $[c,C]$ we get
the claimed absolute continuity statement. Now notice that we have
$\tfrac{\d}{\d t}\int \Phi(f_t) \,\d{\mbox{\boldmath$m$}}=\int \Phi'(f_t)\Delta_q
f_t\,\d{\mbox{\boldmath$m$}}$ for a.e. $t>0$. Since $\Phi'$ belongs to $C^1([c,C])$,
from \eqref{eq:delta1} with $g=\Phi'(f_t)$ we get the conclusion.
\end{proof}
\section{Equivalence of gradients}\label{sequivalence}
In this section we prove the equivalence of weak gradients. We
assume that $(X,{\sf d})$ is compact (this assumption is used to be
able to apply the results of Section~\ref{sec:hopflax} and in
Lemma~\ref{le:kuwada}, to apply \eqref{eq:dualitabase}) and that
${\mbox{\boldmath$m$}}$ is a finite Borel measure, so that the $L^2$-gradient flow of
$\mathbb{C}_q$ can be used.
We start with the following proposition, which relates energy
dissipation to a (sharp) combination of $q$-weak gradients and
metric dissipation in $W_p$.
\begin{proposition}\label{prop:boundweak}
Let $\mu_t=f_t{\mbox{\boldmath$m$}}$ be a curve in $AC^p([0,1],(\Probabilities X,W_p))$. Assume
that for some $0<c<C<\infty$ it holds $c\leq f_t\leq C$ ${\mbox{\boldmath$m$}}$-a.e.
in $X$ for any $t\in[0,1]$, and that $f_0$ is Sobolev along $q$-a.e.
curve with $\weakgradq{f_0}q\in L^q(X,{\mbox{\boldmath$m$}})$. Then for all $\Phi\in
C^2([c,C])$ convex it holds
\[
\int \Phi(f_0)\,\d{\mbox{\boldmath$m$}}-\int\Phi(f_t)\,\d{\mbox{\boldmath$m$}}\leq
\frac1q\iint_0^t\bigl(\Phi''(f_0)|\nabla f_0|_{w,q}\bigr)^qf_s\,\d
s\,\d{\mbox{\boldmath$m$}}+\frac1p\int_0^t|\dot\mu_s|^p\,\d s\qquad\forall t>0.
\]
\end{proposition}
\begin{proof} Let ${\mbox{\boldmath$\pi$}}\in\Probabilities{C([0,1],X)}$ be a plan associated to the curve
$(\mu_t)$ as in Proposition~\ref{prop:lisini}. The assumption
$f_t\leq C$ ${\mbox{\boldmath$m$}}$-a.e. and the fact that
$\iint_0^1|\dot\gamma_t|^p\,\d
t\,\d{\mbox{\boldmath$\pi$}}(\gamma)=\int|\dot\mu_t|^p\,\d t<\infty$ guarantee that
${\mbox{\boldmath$\pi$}}$ is a $p$-test plan. Now notice that it holds
$\weakgradq{\Phi'(f_0)}q=\Phi''(f_0)\weakgradq{f_0}q$ (it follows
easily from the characterization \eqref{eq:pointwisewug}), thus we
get
\[
\begin{split}
\int \Phi(f_0)-\int\Phi(f_t)\,\d{\mbox{\boldmath$m$}}&\leq
\int \Phi'(f_0)(f_0-f_t)\,\d{\mbox{\boldmath$m$}}=\int \Phi'(f_0)\circ \e_0-\Phi'(f_0)\circ \e_t\,\d{\mbox{\boldmath$\pi$}}\\
&\leq\iint_0^t\Phi''(f_0(\gamma_s))\weakgradq{f_0}q(\gamma_s)|\dot\gamma_s|\,\d s\,\d{\mbox{\boldmath$\pi$}}(\gamma)\\
&\leq\frac1q\iint_0^t\bigl(\Phi''(f_0(\gamma_s))|\nabla
f_0|_{w,q}(\gamma_s)\bigr)^q\,\d s\,\d{\mbox{\boldmath$\pi$}}(\gamma)
+\frac1p\iint_0^t|\dot\gamma_s|^p\,\d s\,\d{\mbox{\boldmath$\pi$}}(\gamma)\\
&=\frac1q\iint_0^t\bigl(\Phi''(f_0)|\nabla f_0|_{w,q}\bigr)^qf_s\,\d
s\,\d{\mbox{\boldmath$m$}}+\frac1p\int_0^t|\dot\mu_s|^p\,\d s.
\end{split}
\]
\end{proof}
The key argument to achieve the identification is the following
lemma which gives a sharp bound on the $W_p$-speed of the
$L^2$-gradient flow of $\mathbb{C}_q$. This lemma has been introduced in
\cite{Kuwada10} and then used in
\cite{GigliKuwadaOhta10,Ambrosio-Gigli-Savare11} to study the heat
flow on metric measure spaces.
\begin{lemma}[Kuwada's lemma]\label{le:kuwada}
Let $f_0\in L^q(X,{\mbox{\boldmath$m$}})$ and let $(f_t)$ be the gradient flow of
$\mathbb{C}_q$ starting from $f_0$. Assume that for some $0<c<C<\infty$ it
holds $c\leq f_0\leq C$ ${\mbox{\boldmath$m$}}$-a.e. in $X$, and that $\int
f_0\,\d{\mbox{\boldmath$m$}}=1$. Then the curve $t\mapsto \mu_t:=f_t{\mbox{\boldmath$m$}}\in\Probabilities X$ is
absolutely continuous w.r.t. $W_p$ and it holds
\[
|\dot\mu_t|^p\leq\int\frac{|\nabla f_t|_{*,q}^q}{f_t^{p-1}}\,\d
{\mbox{\boldmath$m$}}\qquad\text{for a.e. $t\in (0,\infty)$.}
\]
\end{lemma}
\begin{proof}
We start from the duality formula \eqref{eq:dualitabase} (written
with $\varphi=-\psi$)
\begin{equation}\label{eq:dualityQ}
\frac{W_p^p(\mu,\nu)}p=\sup_{\varphi\in{\rm Lip}(X)}\int_X
Q_1\varphi\, d\nu-\int_X\varphi\,d\mu.
\end{equation}
where $Q_t\varphi$ is defined in \eqref{eq:Nicola1} and
\eqref{eq:Nicola2}, so that $Q_1\varphi=\psi^c$. Fix $\varphi\in{\rm
Lip}(X)$ and recall (Proposition~\ref{prop:timederivative}) that the
map $t\mapsto Q_t\varphi$ is Lipschitz with values in $C(X)$, in
particular also as a $L^2(X,{\mbox{\boldmath$m$}})$-valued map.
Fix also $0\leq t<s$, set $\ell=(s-t)$ and recall that since $(f_t)$
is a gradient flow of $\mathbb{C}_q$ in $L^2(X,{\mbox{\boldmath$m$}})$, the map $[0,\ell]\ni
\tau\mapsto f_{t+\tau}$ is absolutely continuous with values in
$L^2(X,{\mbox{\boldmath$m$}})$. Therefore, since both factors are uniformly bounded,
the map $[0,\ell]\ni\tau\mapsto Q_{\frac\tau\ell}\varphi f_{t+\tau}$
is absolutely continuous with values in $L^2(X,{\mbox{\boldmath$m$}})$. In addition,
the equality
\[
\frac{Q_{\frac{\tau+h}\ell}\varphi
f_{t+\tau+h}-Q_{\frac{\tau}\ell}\varphi
f_{t+\tau}}{h}=f_{t+\tau}\frac{Q_{\frac{\tau+h}\ell}-Q_{\frac\tau\ell}\varphi
}{h}+Q_{\frac{\tau+h}\ell}\varphi\frac{ f_{t+\tau+h}-
f_{t+\tau}}{h},
\]
together with the uniform continuity of $(x,\tau)\mapsto
Q_{\frac\tau\ell}\varphi(x)$ shows that the derivative of
$\tau\mapsto Q_{\frac\tau\ell}\varphi f_{t+\tau}$ can be computed
via the Leibniz rule.
We have:
\begin{equation}
\label{eq:step1}
\begin{split}
\int_X Q_1\varphi\,\d\mu_s-\int_X\varphi \,\d\mu_t&
=\int Q_1\varphi f_{t+\ell}\,\d{\mbox{\boldmath$m$}}-\int_X\varphi f_t\,\d{\mbox{\boldmath$m$}}
=\int_X\int_0^\ell\frac{\d}{\d\tau}\big(Q_{\frac\tau\ell}\varphi f_{t+\tau}\big)d\tau \,\d{\mbox{\boldmath$m$}}\\
&\leq\int_X\int_0^\ell -\frac{|\nabla Q_{\frac\tau\ell}\varphi
|^q}{q\ell}f_{t+\tau}+
Q_{\frac\tau\ell}\varphi \Delta_q f_{t+\tau}\,\d\tau \,\d{\mbox{\boldmath$m$}},\\
\end{split}
\end{equation}
having used Theorem~\ref{thm:subsol}.
Observe that by inequalities \eqref{eq:delta1} and \eqref{eq:facile}
we have
\begin{equation}
\label{eq:sarannouguali}
\begin{split}
\int_X Q_{\frac\tau\ell}\varphi \Delta_q f_{t+\tau} \,\d{\mbox{\boldmath$m$}}& \leq
\int_X\relgradq{Q_{\frac\tau\ell}\varphi}q|\nabla
f_{t+\tau}|_{*,q}^{q-1}\,\d{\mbox{\boldmath$m$}}\leq \int_X|\nabla
Q_{\frac\tau\ell}\varphi||\nabla f_{t+\tau}|_{*,q}^{q-1} \,\d{\mbox{\boldmath$m$}}\\
&\leq \frac1{q\ell}\int_X|\nabla Q_{\frac\tau\ell}\varphi
|^qf_{t+\tau}d{\mbox{\boldmath$m$}}+\frac{\ell^{p-1}} p\int_X\frac{|\nabla
f_{t+\tau}|_{*,q}^q}{f_{t+\tau}^{p-1}}\,\d{\mbox{\boldmath$m$}}.
\end{split}
\end{equation}
Plugging this inequality in \eqref{eq:step1}, we obtain
\[
\int_X Q_1\varphi \,\d\mu_s-\int_X\varphi \,\d\mu_t\leq
\frac{\ell^{p-1}} p\int_0^\ell\int_X\frac{|\nabla
f_{t+\tau}|_{*,q}^q}{f_{t+\tau}^{p-1}}\,\d{\mbox{\boldmath$m$}}.
\]
This latter bound does not depend on $\varphi$, so from
\eqref{eq:dualityQ} we deduce
\[
W_p^p(\mu_t,\mu_s)\leq \ell^{p-1}\int_0^\ell\int_X\frac{|\nabla
f_{t+\tau}|_{*,q}^q}{f^{p-1}_{t+\tau}}\,\d{\mbox{\boldmath$m$}}.
\]
At Lebesgue points of $r\mapsto\int_X|\nabla
f_r|_{*,q}^q/f_r^{p-1}\,\d{\mbox{\boldmath$m$}}$ where the metric speed exists we
obtain the stated pointwise bound on the metric speed.
\end{proof}
The following result provides equivalence between weak and relaxed
gradients. Recall that the set $\mathcal C$ was defined in
\eqref{eq:mathcalC}.
\begin{theorem}\label{thm:graduguali}
Let $f:X\to\mathbb{R}$ Borel. Assume that $f$ is Sobolev along $q$-a.e.
curve and that $\weakgradq fq\in L^q(X,{\mbox{\boldmath$m$}})$. Then $f\in \mathcal C$
and $\relgradq fq=\weakgradq fq$ ${\mbox{\boldmath$m$}}$-a.e. in $X$.
\end{theorem}
\begin{proof}
Up to a truncation argument and addition of a constant, we can
assume that $0<c\leq f\leq C<\infty$ ${\mbox{\boldmath$m$}}$-a.e. for some $0<c\leq
C<\infty$. Let $(g_t)$ be the $L^2$-gradient flow of $\mathbb{C}_q$ starting
from $g_0:=f$ and let us choose $\Phi\in C^2([c,C])$ in such a way
that $\Phi''(z)=z^{1-p}$ in $[c,C]$. Recall that $c\leq g_t\leq C$
${\mbox{\boldmath$m$}}$-a.e. in $X$ and that from Proposition~\ref{prop:basecal} we
have
\begin{equation}\label{eq:Amerio}
\int\Phi(g_0)\,\d{\mbox{\boldmath$m$}}-\int\Phi(g_t)\,\d{\mbox{\boldmath$m$}}=\int_0^t\int_X\Phi''(g_s)|\nabla
g_s|_{*,q}^q\d{\mbox{\boldmath$m$}}\,\d s\qquad\forall t\in [0,\infty).
\end{equation}
In particular this gives that $\int_0^\infty\int_X\Phi''(g_s)|\nabla
g_s|_{*,q}^q\,\d{\mbox{\boldmath$m$}}\,\d s$ is finite. Setting $\mu_t=g_t{\mbox{\boldmath$m$}}$,
Lemma~\ref{le:kuwada} and the lower bound on $g_t$ give that
$\mu_t\in AC^p\bigl((0,\infty),(\Probabilities X,W_p)\bigr)$, so that
Proposition~\ref{prop:boundweak} and Lemma~\ref{le:kuwada} yield
\[
\int \Phi(g_0)\,\d{\mbox{\boldmath$m$}}-\int
\Phi(g_t)\,\d{\mbox{\boldmath$m$}}\leq\frac1q\int_0^t\int_X\bigl(\Phi''(g_0)|\nabla
g_0|_{w,q}\bigr)^q g_s\,\d{\mbox{\boldmath$m$}}\,\d s+\frac1p\int_0^t\int_X\frac{|\nabla
g_s|_{*,q}^q}{g_s^{p-1}}\,\d{\mbox{\boldmath$m$}}\,\d s.
\]
Hence, comparing this last expression with \eqref{eq:Amerio}, our
choice of $\Phi$ gives
\[
\frac1q\iint_0^t\,\frac{|\nabla g_s|_{*,q}^q}{g_s^{p-1}}\d
s\,\d{\mbox{\boldmath$m$}}\leq\int_0^t\int_X\frac1q \bigl(\frac{|\nabla
g_0|_{w,q}}{g_0^{p-1}}\bigr)^q g_s\,\d{\mbox{\boldmath$m$}}\,\d s.
\]
Now, the bound $f\geq c>0$ ensures $\Phi''(g_0)|\nabla g_0|_{*,q}\in
L^q(X,{\mbox{\boldmath$m$}})$. In addition, the
maximum principle together with the convergence of $g_s$ to $g_0$ in
$L^2(X,{\mbox{\boldmath$m$}})$ as $s\downarrow 0$ grants that the convergence is also
weak$^*$ in $L^\infty(X,{\mbox{\boldmath$m$}})$, therefore
\[
\limsup_{t\downarrow 0}\frac{1}{t}\iint_0^t\,\frac{|\nabla
g_s|_{*,q}^q}{g_s^{p-1}}\d s\,\d{\mbox{\boldmath$m$}}\leq\int_X \frac{|\nabla
g_0|_{w,q}^q}{g_0^{q(p-1)}}g_0\d{\mbox{\boldmath$m$}}=\int_X \frac{|\nabla
g_0|_{w,q}^q}{g_0^{p-1}}\,\d{\mbox{\boldmath$m$}}.
\]
The lower semicontinuity property stated in
Remark~\ref{rem:basiclsc} with $\alpha=p-1$ then gives
\[
\int_X \frac{|\nabla g_0|_{*,q}^q}{g_0^{p-1}}\,\d{\mbox{\boldmath$m$}}\leq \int_X
\frac{|\nabla g_0|_{w,q}^q}{g_0^{p-1}}\,\d{\mbox{\boldmath$m$}}.
\]
This, together with the inequality $\weakgradq {g_0}q\leq\relgradq
{g_0}q$ ${\mbox{\boldmath$m$}}$-a.e. in $X$, gives the conclusion.
\end{proof}
In particular, taking into account \eqref{allinequalities}, we
obtain the following equivalence result. We state it for
$L^q(X,{\mbox{\boldmath$m$}})$ functions because in the definition of $q$-relaxed
upper gradient and $q$-relaxed slope this integrability assumption
is made (see also Remark~\ref{rem:whyq}), while no integrability is
made in the other two definitions. It is also clear that if we
extend the ``relaxed'' definitions of gradient by truncation, as in
\eqref{eq:extendedrelaxed}, then equivalence goes beyond
$L^q(X,{\mbox{\boldmath$m$}})$ functions.
\begin{theorem}[Equivalence of weak gradients]
\label{thm:gradugualibis} Let $f\in L^q(X,{\mbox{\boldmath$m$}})$. Then the
following four properties are equivalent:
\begin{itemize}
\item[(i)] $f$ has a $q$-relaxed upper gradient;
\item[(ii)] $f$ has a $q$-relaxed slope;
\item[(iii)] $f$ has a $q$-upper gradient in $L^q(X,{\mbox{\boldmath$m$}})$;
\item[(iv)] $f$ has a $q$-weak upper gradient in $L^q(X,{\mbox{\boldmath$m$}})$.
\end{itemize}
In addition, the minimal $q$-relaxed upper gradient, the minimal
$q$-relaxed slope, the minimal $q$-upper gradient and the minimal
$q$-weak upper gradient coincide ${\mbox{\boldmath$m$}}$-a.e. in $X$.
\end{theorem}
\begin{proof} If either of the four properties holds for some gradient $g$,
then \eqref{allinequalities} gives that $f$ is Sobolev along
$q$-a.e. curve and $|\nabla f|_{w,q}\leq g$ ${\mbox{\boldmath$m$}}$-a.e. in $X$. Then,
Theorem~\ref{thm:graduguali} yields $|\nabla f|_{*,q}\leq g$
${\mbox{\boldmath$m$}}$-a.e. in $X$ and we can invoke \eqref{allinequalities} again to
obtain that all four properties hold and the corresponding weak
gradients are equal.
\end{proof}
\section{Further comments and extensions}\label{sextensions}
In this section we point out how our main results, namely
Theorem~\ref{thm:graduguali} and Theorem~\ref{thm:gradugualibis} can
be extended to more general metric measure spaces. Recall that, in
the previous section, we derived them under the assumptions that
$(X,{\sf d})$ is a compact metric space and that ${\mbox{\boldmath$m$}}$ is a finite
measure.
\subsection{The role of the compactness assumption in Section~\ref{sec:hopflax}}
The compactness assumption is not really needed, and suffices to
assume that $(X,{\sf d})$ is a complete metric space. The only
difference appears at the level of the definition of $D^\pm(x,t)$,
since in this case existence of minimizers is not ensured and one
has to work with minimizing sequences. This results in longer
proofs, but the arguments remain essentially the same, see
\cite{Ambrosio-Gigli-Savare11} for a detailed proof in the case
$p=q=2$. Thanks to this remark, the proof of the equivalence results
immediately extends to complete and separable metric measure spaces
with $(X,{\sf d},{\mbox{\boldmath$m$}})$ with ${\sf d}$ bounded and ${\mbox{\boldmath$m$}}$ finite.
Also, it is worthwhile to remark that all results (except of course
the Lipschitz bounds on $Q_tf$ and the continuity of $t\mapsto Q_tf$
from $[0,\infty)$ to $C(X)$) of Section~\ref{sec:hopflax} remain
valid for lower semicontinuous functions $f:X\to\mathbb{R}\cup\{+\infty\}$
satisfying
$$
f(x)\geq -C\bigl(1+{\sf d}^r(x,\bar x)\bigr)\qquad\forall x\in X
$$
for suitable $\bar x\in X$, $C\geq 0$, $r\in [0,p)$.
\subsection{Locally finite metric measure spaces}
We say that a metric measure space $(X,{\sf d},{\mbox{\boldmath$m$}})$ is locally finite
if $(X,{\sf d})$ is complete and separable and any $x\in{\rm
supp\,}{\mbox{\boldmath$m$}}$ has a neighbourhood $U$ with finite ${\mbox{\boldmath$m$}}$-measure.
For any locally finite metric measure space it is not difficult to
find (choosing for instance as $U$ balls with ${\mbox{\boldmath$m$}}$-negligible
boundary) a nondecreasing sequence of open sets $A_h$ whose union
covers ${\mbox{\boldmath$m$}}$-almost all of $X$ and whose boundaries $\partial A_h$
are ${\mbox{\boldmath$m$}}$-negligible. Then, setting $X_h=\overline{A_h}$, we can
apply the equivalence results in all metric measure spaces
$(X_h,{\sf d},{\mbox{\boldmath$m$}})$ to obtain the equivalence in $(X,{\sf d},{\mbox{\boldmath$m$}})$. This
is due to the fact that the minimal $q$-weak upper gradient
satisfies this local-to-global property (see
\cite[Theorem~4.20]{Ambrosio-Gigli-Savare11bis} for a proof in the
case $p=q=2$):
\begin{equation}\label{eq:locglob1}
|\nabla f|_{X,w,q}=|\nabla f|_{X_h,w,q}\qquad\text{${\mbox{\boldmath$m$}}$-a.e. in
$X_h$.}
\end{equation}
An analogous property holds for the larger gradient, namely the minimal
$q$-relaxed slope (arguing as in \cite[Lemma 4.11]{Ambrosio-Gigli-Savare11}):
\begin{equation}\label{eq:locglob2}
|\nabla f|_{X,*,q}=|\nabla f|_{X_h,*,q}\qquad\text{${\mbox{\boldmath$m$}}$-a.e. in
$X_h$.}
\end{equation}
Combining \eqref{eq:locglob1} and \eqref{eq:locglob2} gives
the identification result for all gradients and all locally finite metric measure spaces.
\subsection{An enforcement of the density result}\label{sec:improveslope}
In Theorem~\ref{thm:graduguali} we proved that
if $f:X\to\mathbb{R}$ is Borel and $f$ is Sobolev along $q$-a.e. curve
and $\weakgradq fq\in L^q(X,{\mbox{\boldmath$m$}})$, then there exist Lipschitz functions
$f_n$ convergent to $f$ ${\mbox{\boldmath$m$}}$-a.e. in $X$ and satisfying
\begin{equation}\label{eq:lavuolenicola}
|\nabla f_n|\to \weakgradq fq\qquad\text{in $L^q(X,{\mbox{\boldmath$m$}})$.}
\end{equation}
This follows by a diagonal argument, thanks to the fact that all truncations $f_N$ of $f$ satisfy $\mathbb{C}_q(f_N)\leq\tfrac1q
\int_X\weakgradq fq^q\,\d{\mbox{\boldmath$m$}}$. It is worthwhile to notice that \eqref{eq:lavuolenicola}
can be improved asking the existence of Lipschitz functions $f_n$ such that
${\rm Lip}_a(f_n,\cdot)\to\weakgradq fq$ in $L^q(X,{\mbox{\boldmath$m$}})$, where ${\rm Lip}_a(f,\cdot)$ is the asymptotic Lipschitz constant
defined in \eqref{eq:asymlip}: the key observation is
that, as noticed in \eqref{eq:hjbsusbis}, the Hamilton-Jacobi subsolution property holds with the new, and larger, pseudo gradient
${\rm Lip}_a(g,\cdot)$. Starting from this observation, and using the convexity inequality
$$
{\rm Lip}_a\bigl((1-\chi)f+\chi g\bigr)\leq\bigl(1-\chi(x)\bigr){\rm Lip}_a(f,x)+\chi(x){\rm Lip}_a(g,x)+{\rm Lip}(\chi)|f(x)-g(x)|
$$
for $\chi:X\to [0,1]$ Lipschitz and $f,\,g:X\to\mathbb{R}$ Lipschitz,
one can build Cheeger's energy by minimizing the integrals of ${\rm Lip}_a(f_n,\cdot)$ instead of the integral of $|\nabla g|$,
still getting a convex and lower semicontinuous functional and a corresponding relaxed gradient.
Then, \eqref{eq:hjbsusbis} provides Kuwada's Lemma~\ref{le:kuwada} for the new Cheeger energy and the proof of Theorem~\ref{thm:graduguali}
can repeated word by word.
\subsection{Orlicz-Wasserstein spaces}
Another potential extension, that we shall not develop here, is for
general Lagrangians-Hamiltonians: one can consider the functions
$$
Q_tf(x):=\inf_{y\in X} f(y)+tL\bigl(\frac{{\sf d}(y,x)}{t}\bigr)
$$
and prove that $\tfrac{\d}{\d t}Q_tf+H(\nabla Q_tf)\leq 0$ with
$H=L^*$. This way, also gradients in Orlicz spaces as $LLogL$ could
be considered. On the other hand, the Orlicz-Wasserstein distances
$$
W_L(\mu,\nu):=\inf\left\{\lambda>0:\
\inf_{\sppi\in\Gamma(\mu,\nu)}\int
L\bigl(\frac{{\sf d}(x,y)}{\lambda}\bigr)\,\d{\mbox{\boldmath$\pi$}}\leq 1\right\}
$$
have not been considered much so far (except in \cite{Sturm-Orlicz}
and more implicitly in \cite{FigalliGangbo,Villani09}) and the
extension of Lisini's superposition theorem to this class of
distances is not known, although expected to be true. These
extensions might be particularly interesting to deal with the
limiting case $q\downarrow 1$, where the Wasserstein exponent $p$
goes to $\infty$ (for instance $LlogL$ integrability of gradients
corresponds to exponential integrability of metric derivative on
curves) .
\subsection{$W^{1,1}$ and $BV$ spaces}
In this subsection we discuss the limiting case $q=1$, $p=\infty$
and assume for the sake of simplicity that $(X,{\sf d})$ is locally
compact and separable. Following the approach in \cite{Miranda03}, for
any open set $A\subset X$ we can define
$$
|Df|(A):=\inf\left\{\liminf_{h\to\infty}\int_A|\nabla f_h|\,\d{\mbox{\boldmath$m$}}:\
f_h\in {\rm Lip}_{\rm loc}(A),\,\,f_h\to f\,\,\text{in $L^1_{\rm
loc}(A)$}\right\}.
$$
It is possible to show that, whenever $|Df|(X)<\infty$, the set
function $A\mapsto|Df|(A)$ is the restriction to open sets of $X$ of
a finite Borel measure, that we still denote by $|Df|$. In the case
when $|Df|$ is abolutely continuous with respect to ${\mbox{\boldmath$m$}}$,
corresponding to the Sobolev space $W^{1,1}$ we may define $|\nabla
f|_{*,1}$ as the density of $|Df|$ w.r.t. ${\mbox{\boldmath$m$}}$.
This approach corresponds to $1$-relaxed slopes. Coming to $1$-weak
upper gradients, it is natural to consider $\infty$-test plans as
probability measures ${\mbox{\boldmath$\pi$}}$ concentrated on Lipschitz curves and to
define exceptional sets of curves using this class of test plans.
Then the class of functions which are $BV$ along $1$-almost every
curve can be defined. It is not hard to show that if
$|Df|(X)<\infty$ and ${\mbox{\boldmath$\pi$}}$ is a $\infty$-test plan such that
$(\e_t)_\#{\mbox{\boldmath$\pi$}}\leq C({\mbox{\boldmath$\pi$}}){\mbox{\boldmath$m$}}$ for all $t\in [0,1]$ then the following
inequality between measures in $X$ holds:
$$
\int \gamma_\sharp|D(f\circ\gamma)|\,d{\mbox{\boldmath$\pi$}}(\gamma)\leq C({\mbox{\boldmath$\pi$}})\|{\rm
Lip}(\gamma)\|_{L^\infty(\sppi)}|Df|,
$$
where $|D(f\circ\gamma)|$ is the total variation measure of the map
$f\circ\gamma:[0,1]\to\mathbb{R}$. This provides one connection between
$1$-weak upper gradients and $1$-relaxed slopes, while in \cite{Ambrosio-DiMarino12}
the arguments of this paper are adapted to show that the supremum of
$$
\frac{1}{C({\mbox{\boldmath$\pi$}})\|{\rm
Lip}(\gamma)\|_{L^\infty(\sppi)}}\int \gamma_\sharp|D(f\circ\gamma)|\,d{\mbox{\boldmath$\pi$}}(\gamma)
$$
in the lattice of measures coincides with $|Df|$.
\def$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'${$'$}
|
\section{Introduction}
The motion of a satellite around a planet, which revolves on a circular orbit around the Sun, can be modelled by the Hill's model of the three-body problem. Starting from the circular restricted three body problem (CRTBP), the Hill's approximation is achieved by translating the origin of the rotating reference frame to the planet and the unit of length is scaled by the factor $\mu^{1/3}$, where $\mu$ is the mass parameter of CRTBP. Then we let $\mu\rightarrow 0$. The present study is restricted to the planar motion. Under these assumptions we obtain the Hill's equations (Szebehely, 1967)
\begin{equation} \label{EqCHillEqs}
\ddot{\xi}-2\dot{\eta}=3\xi-\xi\,\rho^{-3},\quad \ddot{\eta}+2\dot{\xi}=-\eta\,\rho^{-3}
\end{equation}
where $\rho=(\xi^{2}+\eta^{2})^{1/2}$. These equations admit an integral of motion, the well known Jacobi integral
\begin{equation} \label{EqJacobi}
C_H=3\xi^2+2\rho^{-1}-(\dot \xi^2+\dot \eta^2),
\end{equation}
and are invariant under the symmetries
$$
\mathbf{\Sigma}:(t,\xi,\eta)\rightarrow (-t, \xi,-\eta)\quad \textnormal{and} \quad \mathbf{\Sigma'}:(t,\xi,\eta)\rightarrow (-t, -\xi,\eta).
$$
Although the circular Hill problem (CH) has only two degrees of freedom and is, moreover, autonomous, conservative and parameter free, it is not integrable (Meletlidou et al., 2001; Morales-Ruiz et al., 2005) and shows rich dynamics, which are depicted by computing Poincar\'{e} surfaces of section (H\'enon, 1970; Chauvineau and Mignard, 1991).
\begin{figure}
\centering
\includegraphics[width=9cm]{CHfamilies.eps}
\caption {Families of periodic orbits of the Circular Hill problem. Thick (blue) or thin (red) curves indicate stable or unstable families, respectively. Grey regions indicate forbidden areas.}
\label{FigCHfams}
\end{figure}
A detailed study of the CH problem has been given by H\'{e}non (1969,1970). He found the main families of periodic orbits, explored the phase space by using Poincar\'{e} sections and computed the width of the stability regions. As we know, the families of periodic orbits consist the backbone of the phase space. The most important of them are presented in Fig. \ref{FigCHfams} (see also H\'{e}non, 2003; Batkhin and Batkhina, 2009). The domain $\xi>0$ corresponds to prograde and $\xi<0$ to retrograde motion. The two equilibrium (Lagrange) points $L_1$ and $L_2$ are found for $C_H^L=3^{4/3}$ at the locations $\xi=\pm 3^{-1/3}$, $\eta=0$. The distance $R_H=3^{-1/3}$ defines the Hill radius. We can see from Fig. \ref{FigCHfams} that families of periodic orbits of retrograde satellites extend to distances much larger than the Hill radius and, interestingly, some of them are stable.
The Hill's approximation can be applied in the same way to the spatial problem (H\'enon, 1974). Such a model has been used in space mission orbit design (Villac, 2003, 2008). A single averaged model for the spatial Hill's problem has been also studied, providing particular solutions (see Vashkov'yak and Teslenko (2008) and references therein). Hill models have also been derived for the case of binary satellites or asteroids (H\'enon and Petit, 1986; Chauvineau and Mignard, 1990). In these models the center of mass of the binary moves along a circular orbit around the Sun. A model where the center of the binary moves in an elliptic orbit has been given by Moons et al. (1988). This model uses a different formalism and aims to the study of the changes of orbital elements due to close encounters. An extensive list of references on the dynamics of satellite motion is given in Waldvogel (1999). The general three body problem may be also used for studying satellite or binary motion (Hadjidemetriou and Voyatzis, 2011).
The elliptic Hill (EH) model can be derived by considering the same Hill's approximation assumptions used in the CH model, but letting the planet move on an elliptic orbit around the Sun. Such a model has been introduced by Ichtiaroglou (1980, 1981) who also computed a few families of periodic orbits. A further study (Ichtiaroglou and Voyatzis, 1990) showed that all periodic orbits found were strongly unstable. It is obvious that the EH is a more appropriate model than the CH for studying the dynamics of a small satellite around a planet (or an asteroid) with eccentric motion. The aim of the present work is to study the dynamics of the EH problem in more detail, focusing on the main qualitative features of the phase space and orbit evolution.
In the next section we describe briefly the EH model and discuss the existence and continuation of periodic orbits. In section \ref{SecFamilies} we present the results of the computations of families of periodic orbits, giving particular attention to the stable ones. In section 4 we explore the phase space of the model by computing various maps of stability and study the effect of the planetary eccentricity to the stability of the orbits. Our conclusions are given in section 5.
\section{The Elliptic Hill model}
Considering two primaries of masses $m_0$ (the Sun) and $m_1$ (the planet) that revolve on the plane $Oxy$ in a Keplerian ellipse around their center of mass $O$, the motion of the planet along the rotating axis $Ox$ with angular momentum $P_\theta$ (see Hadjidemetriou 1975) is described by the equation
\begin{equation} \label{EqERequP}
\ddot x_1 -Q^2 x_1^{-3}+(1-\mu)^3 x_1^{-2}=0,
\end{equation}
where $\mu=m_1/(m_0+m_1)$ and $Q=(1-\mu)P_\theta/\mu$. By assuming the initial conditions $x_1(0)=x_{10}$ and $\dot\theta(0)=1$, where $\dot\theta$ is the angular velocity, we get $Q=x_{10}^2$.
The motion of a massless body, which moves on the same plane $Oxy$, is described by the equations
\begin{equation} \label{EqERequB}
\begin{array}{l}
\ddot x=2Qx_1^{-2}\dot{y} + Q^2 x_1^{-4} x - 2Q x_1^{-3}\dot{x}_1 y+\mu (x_1-x) r_1^{-3}-\left (\mu x_1+(1-\mu)x \right ) r_2^{-3} \\
\ddot y=-2Qx_1^{-2}\dot{x} + Q^2 x_1^{-4} y + 2Q x_1^{-3}\dot{x}_1 x-\mu y r_1^{-3}-(1-\mu) y r_2^{-3}, \end{array}
\end{equation}
where $r_1^2=(x-x_1)^2+y^2$ and $r_2^2=(x+\frac{\mu}{1-\mu} x_1)^2+y^2$.
Equations (\ref{EqERequP}) and (\ref{EqERequB}) are the equations of the {\em restricted elliptic three body problem} in the rotating frame.
We apply the Hill's transformation
\begin{equation} \label{EqHillTrans}
x=x_1+\mu^{1/3}\xi,\quad y=\mu^{1/3}\eta
\end{equation}
to equations (\ref{EqERequP}) and (\ref{EqERequB}) and, by letting $\mu\rightarrow 0$, we get (Ichtiaroglou, 1980)
\begin{equation} \label{EqEHequx1}
\ddot x_1 -x_{10}^4 x_1^{-3}+x_1^{-2}=0
\end{equation}
and
\begin{equation} \label{EqEHequ}
\begin{array}{l}
\ddot \xi=2 x_{10}^2 x_1^{-2}\dot{\eta} + \left ( 2 x_1^{-3}+x_{10}^4 x_1^{-4}-\rho^{-3} \right ) \xi - 2 x_{10}^2 x_1^{-3}\dot{x}_1 \eta\\
\ddot \eta=-2 x_{10}^2 x_1^{-2}\dot{\xi} + \left (- x_1^{-3}+x_{10}^4 x_1^{-4}-\rho^{-3} \right )\eta + 2 x_{10}^2 x_1^{-3}\dot{x}_1 \xi
\end{array}
\end{equation}
with $\rho^2=\xi^2+\eta^2$.
The equation (\ref{EqEHequx1}) describes Keplerian motion with eccentricity, semi-major axis and period given by
\begin{equation}
e_p=x_{10}^3-1,\quad a=\frac{(1+e_p)^{1/3}}{(1-e_p)},\quad T=2 \pi \sqrt{\frac{(1+e_p)}{(1-e_p)^3}},
\end{equation}
respectively. For bounded motion, $x_{10}$ should be restricted in such a way that $-1<e_p<1$. For $e_p>0$ ($e_p<0$) the planet is at periapsis (apoapsis) at $t=0$ and it has $\dot{x}_1(0)=0$. For $e_p=0$ ($x_1=x_{10}=1$) we obtain the equations (\ref{EqCHillEqs}) of the CH model.
Fixing the value of the planetary eccentricity $e_p\neq 0$, the equations (\ref{EqEHequ}), which describe the motion of the massless satellite, constitute a periodic non-autonomous system of period $T$. The system (\ref{EqEHequ}) possesses only the symmetry $\mathbf{\Sigma}$ and an orbit of initial conditions
$$
\xi(0)=\xi_0, \quad \eta(0)=\dot{\xi}(0)=0, \quad \dot{\eta}(0)=\dot{\eta}_0
$$
is a {\em symmetric periodic orbit} of period $T'$ {\em if} it satisfies the periodicity conditions
$$
\dot{\xi}(T'/2;\xi_0,\dot{\eta}_0)=0,\quad \eta(T'/2;\xi_0,\dot{\eta}_0)=0.
$$
The period $T'$ has to be an integer multiple of the planetary period $T$, i.e. $T'=\kappa T$, $\kappa=1,2,..$.
By varying the planetary eccentricity $e_p$, a symmetric periodic orbit continues to exist, so that we get a monoparametric family (Ichtiaroglou, 1981). All periodic orbits that belong to such a family are isolated in phase space and have a period $T'$ which depends on $e_p$. For $e_p=0$, a family of the EH problem crosses a family of the CH problem. The crossing point must be a periodic orbit of the CH problem with period $T_c=T'/\lambda$, where $\lambda=1,2,..$ is the {\em multiplicity} of the periodic orbit. Since for $e_p=0$ it is $T=2\pi$, we conclude that the periodic orbits of the circular problem with period $T_c=2\kappa \pi/\lambda$ can be considered as ``bifurcation'' points for the EH problem. Therefore these orbits are continued for $e_p\neq 0$ with multiplicity $\lambda$ and period $T'$=$\lambda T_c$=$2\kappa \pi$. The ratio $\kappa/\lambda$ is the {\em resonance} of the periodic orbit.
Obviously, the number of bifurcation points on each family of the CH problem is infinite. In this study we restrict our attention to families that are of relatively small multiplicity and bifurcate mainly from stable periodic orbits. The linear stability of the orbits is determined by the two conjugate pairs of eigenvalues of the monodromy matrix of the system (\ref{EqEHequ}). An orbit is {\em stable} (s) only when all eigenvalues lie on the unit circle. If one or two pairs of eigenvalues lie on the real axis we have {\em single} (u) or {\em double} (uu) instability, respectively. Finally when all eigenvalues are outside the unit circle (but they are not real) we have {\em complex instability} (cu). The different kinds of stability can be determined by computing stability indices as in Broucke (1969) (see, also, Ichtiaroglou and Voyatzis, 1990). A necessary (but not sufficient) condition for a family of the EH problem to emanate (from $e_p=0$) with a stable branch is the bifurcation point to be a stable periodic orbit. Certainly, the stability type can change along the family.
\begin{figure}[htb]
\centering
$\begin{array}{ccc}
\includegraphics[width=5.5cm]{CHBifurc1.eps} & \qquad &
\includegraphics[width=5.5cm]{CHBifurc2.eps}\\
\textnormal{(a)} & \qquad & \textnormal{(b)}
\end{array} $
\caption{Bifurcation points for continuation from the CH problem to the EH problem {\bf a} from retrograde families {\bf b} from prograde families of periodic orbits. The corresponding resonances are indicated.}
\label{FigBif}
\end{figure}
In Fig. \ref{FigBif} we present the bifurcation points that are studied in the present work. The resonance $\kappa/\lambda$ is indicated for each point. Most bifurcation points of Fig. \ref{FigBif} belong to the family $f$ of retrograde orbits (left panel) and to the stable parts of families $g$, $g'$ of prograde orbits (right panel). The bifurcation points of families $g3$, $Hg$ and $Hm$ play a special role, which will be discussed in the following section. We note that, as the Jacobi constant increases, the main families $f$ and $g$ tend to $\xi_0=0$ and the Hill approximation degenerates into the two-body problem. In this case the bifurcation points to the EH model are of larger and larger multiplicity as $\xi_0\rightarrow 0$.
\section{Families of periodic orbits} \label{SecFamilies}
A monoparametric family of the EH problem, which emanates from a $\kappa/\lambda$ resonant bifurcation point, will be denoted as $F_e^{\kappa/\lambda}$, where $F$ indicates the name of the family of the CH model where the bifurcation point belongs to. For the circular problem we adopt the naming introduced by H\'{e}non. The families $a_e^{1/1}$, ${g'}_e^{1/1}$ and $f_e^{1/2}$ have been computed by Ichtiaroglou (1981) and are unstable (Ichtiaroglou and Voyatzis, 1990).
Each orbit of an EH family is represented by a point in the three dimensional space $\Pi_3=\{e_p,\xi_0,\dot{\eta}_0\}$ and a family forms a characteristic curve in this space. In the following we will present, for convenience, the families in the projection plane $(\xi_0,e_p)$. Thick (blue) or thin (red) characteristic curves indicate stable or unstable orbits, respectively. The periodic orbits have been computed by the method of differential corrections and satisfy the periodicity condition with an accuracy of $10^{-12}$. We always start from $e_p=0$, where the period and the initial conditions are known from the CH model.
\subsection{Families $f_e^{\kappa/\lambda}$ - retrograte orbits}
\begin{figure}[tb]
\centering
\includegraphics[width=9cm]{EHfamF1.eps}
\caption {Families $f_e^{1/\lambda}$ of periodic orbits of the EH problem. Thick (blue) or thin (red) curves indicate stable or unstable orbits, respectively.}
\label{Ffamf1}
\end{figure}
\subsubsection{The case $\kappa=1$}
We have computed the families up to multiplicity $\lambda=9$ and present them in Fig. \ref{Ffamf1}. The families $f_e^{1/2}$ and $f_e^{1/3}$ are unstable and extend outside of the Hill radius. The families with $6\leq \lambda \leq 9$ start (at $e_p=0$) as unstable but they posses a stable segment for moderate eccentricity values in both cases of the initial planetary position, namely for periapsis ($e_p>0$) and apoapsis ($e_p<0$).
\begin{figure}[htb]
\centering
$\begin{array}{ccc}
\includegraphics[width=5.5cm]{EHfamF14con.eps} & \qquad &
\includegraphics[width=5.5cm]{EHfamF15con.eps}\\
\textnormal{(a)} & \qquad & \textnormal{(b)}
\end{array} $
\caption{ {\bf a} The continuation of family $f_e^{1/4}$ and its connection with families $Hg_e^{1/1}$. {\bf b} The continuation of family $f_e^{1/5}$ and its connection with family $Hm_e^{1/1}$. A closed path is formed. The symbol (uu) indicates parts of the families with {\em double unstable} orbits.}
\label{Ffamf1plus}
\end{figure}
Families $f_e^{1/4}$ and $f_e^{1/5}$ in Fig. \ref{Ffamf1} seem to terminate abruptly at some critical points at $e_p=\pm 0.215$ and $e_p=\pm 0.185$, respectively. However, as it is shown in Fig. \ref{Ffamf1plus}, at these critical points, denoted as $SN$, saddle-node bifurcations take place. In these bifurcations the $1/1$ resonant families $Hg_e^{1/1\pm}$ and $Hm_e^{1/1}$, which bifurcate from the CH families $Hg$ of multiplicity $\lambda=4$ and $Hm$ of multiplicity $\lambda=5$, respectively, are involved. Particularly, in Fig. \ref{Ffamf1plus}a we obtain that the upper branch ($e_p>0$) of the family $f_e^{1/4}$ and the $Hg_e^{1/1-}$ emanate from the point $SN_1$, while the lower branch ($e_p<0$) of the family $f_e^{1/4}$ and the $Hg_e^{1/1+}$ emanate from $SN_2$. In Fig. \ref{Ffamf1plus}b we see that the families $f_e^{1/5}$ and $Hm_e^{1/1}$ emanate from both critical points $SN_3$ and $SN_4$, appearing for the eccentricity values $e_p=\pm 0.185$. The two families connect smoothly and form a closed characterisric curve.
\
\begin{figure}[htb]
\centering
$\begin{array}{ccc}
\includegraphics[width=5.5cm]{EHfamF2.eps} & \qquad &
\includegraphics[width=5.5cm]{EHfamF23con.eps}\\
\textnormal{(a)} & \qquad & \textnormal{(b)}
\end{array} $
\caption{ {\bf a} Families $f_e^{2/\lambda}$ of periodic orbits of the EH problem. {\bf b} The continuation of family $f_e^{2/3}$ and its connection with family $g3_e^{2/1}$. Stability is indicated as (s) for stable (u) for unstable (uu) for double unstable and (cu) for complex unstable orbits.}
\label{Ffamf2}
\end{figure}
\subsubsection{The case $\kappa=2$}
The computed families are presented in Fig. \ref{Ffamf2}a. The families with $\lambda\geq 5$ are unstable for the periapsis case but they start as stable when the planet is initially at apoapsis. The stable segment extends up to high eccentricity values in all cases. In contrast, the family $f_e^{2/3}$ is unstable for $e_p<0$ and starts as stable for $e_p>0$. Then it seems to terminate but, as in the case of the family $f_e^{1/5}$, it connects with the family $g3_e^{2/1}$. The overall characteristic curve is given in Fig. \ref{Ffamf2}b. It forms a closed path, along which we find all possible types of linear stability. We note that the stable part is located outside the Hill sphere.
\begin{figure}[tb]
\centering
\includegraphics[width=7cm]{EHfamF3.eps}
\caption {Families $f_e^{3/\lambda}$ of periodic orbits of the EH problem. Thick (blue) or thin (red) curves indicate stable or unstable orbits, respectively.}
\label{Ffamf3}
\end{figure}
\subsubsection{The case $\kappa=3$}
We computed the families $f_e^{3/4}$, $f_e^{3/5}$ and $f_e^{3/7}$ presented in Fig. \ref{Ffamf3}. All the corresponding bifurcation points are located outside the Hill sphere. The family $f_e^{3/4}$ is unstable for small and moderate values of the planetary eccentricity, but it posses stable segments for $|e_p|>0.5$. We note that for $e_p<0$ the family becomes stable when it enters the Hill sphere. On the other hand, the family $f_e^{3/5}$ is stable approximately in the interval $-0.25<e_p<0.25$. Now, the family becomes unstable when it enters the Hill sphere for $e_p<0$. The family $f_e^{3/7}$ is unstable in all the explored domain.
\subsection{Families $g_e^{\kappa/\lambda}$ and ${g'}_e^{\kappa/\lambda}$ - prograte orbits}
In this case all bifurcation points and families are located inside the Hill sphere.
\begin{figure}[htb]
\centering
$\begin{array}{ccc}
\includegraphics[width=5.5cm]{EHfamG.eps} & \qquad &
\includegraphics[width=5.5cm]{EHfamGT.eps}\\
\textnormal{(a)} & \qquad & \textnormal{(b)}
\end{array} $
\caption{ {\bf a} Families $g_e^{\kappa/\lambda}$ {\bf b} Families ${g'}_e^{\kappa/\lambda}$. Thick (blue) or thin (red) solid curves indicate stable or simply unstable orbits, respectively, and dashed curves correspond to complex unstable orbits. }
\label{FfamG}
\end{figure}
\subsubsection{Families $g_e^{\kappa/\lambda}$}
We computed the families $g_e^{1/\lambda}$ with $5\leq\lambda\leq 9$ and the families $g_e^{2/\lambda}$ with $\lambda=11$, 13 and 15. They are presented in Fig. \ref{FfamG}a. For $e_p>0$ all families are unstable. For $e_p<0$ the families $g_e^{1/\lambda}$ are also unstable but the families $g_e^{2/\lambda}$ start us stable and become unstable when the planetary eccentricity exceeds a critical value, which is different for each case.
\subsubsection{Families ${g'}_e^{\kappa/\lambda}$}
The CH family $g'$ has two branches (see Fig. \ref{FigCHfams}) and each resonant bifurcation point appears twice, one in each branch; we use the notation ``$+$'' and ``$-$'' to distinguish between them. In this case the convergence of our numerical approach becomes delicate. The families we were able to compute are presented in Fig. \ref{FfamG}b. Also in this case we see that segments of stable orbits exist for low eccentricities.
\section{Phase space numerical exploration}
In this section we study the evolution of generally non-periodic orbits, associated with particular grids of initial conditions and various values of the planetary eccentricity, $e_p$. The main orbit classification aims at distinguishing the orbits between bounded and escaping ones. We set the escape criterion to $\rho>15\,R_H$. Additionally, along the numerical integration of the orbit, we compute a Fast Lyapunov Indicator (FLI, Froeschl\'e and Lega, 2000), in order to classify bounded orbits as regular or chaotic. The FLI, as defined by Voyatzis (2008), is given by the relation
\begin{equation}
FLI=\sup_{t\leq t_{max}} \frac{1}{t}|\delta(t)|,
\label{EqFLIndic}
\end{equation}
where $\delta(t)$ is the deviation vector, whose evolution is given by the solution of the variational equations. In Fig. \ref{FigFLIevol} we present the behaviour of the FLI for some typical trajectories. Case (1) corresponds to a regular (quasi-periodic) orbit, while case (4) corresponds to an irregular orbit that escapes at some $t<t_{max}$. Cases (2) and (3) are rather rare. Case (2) corresponds to a chaotic orbit which is sticky and does not escape for $t<t_{max}$. Case (3) corresponds to a trajectory that lies very close to an unstable periodic orbit. It is a typical case for all unstable periodic orbits presented in the previous section with relatively small planetary eccentricity ($e_p\lessapprox 0.2$). When the orbit follows the stable (unstable) manifold of the periodic orbit, the FLI decreases (increases) (see also Skokos et al., 2007). The overall (average) FLI evolution shows a linear increase in logarithmic scale, which is an indication of the existing instability.
\begin{figure}[tb]
\centering
\includegraphics[width=7cm]{fli_evol.eps}
\caption {The evolution of the FLI (without including the supremum) for some typical cases. Case (1) corresponds to a regular orbit, case (2) to a chaotic orbit and case (4) to a chaotic orbit that escapes for $t<t_{max}$. Case (3) refers to an unstable periodic orbit (see the text).}
\label{FigFLIevol}
\end{figure}
We have computed stability maps corresponding to a $N\times M$ grid of initial conditions. For each orbit numerical integration is performed up to $t_{max}=30000$ time units or $FLI<FLI_{max}$(=$10^{10}$) or until the escape condition is fulfilled. Strongly chaotic orbits escape in relatively short time intervals. We observed that even if an orbit reaches the maximum FLI value, it shows a weakly chaotic evolution, at least for $t<t_{max}$.
\begin{figure}[tb]
\centering
\includegraphics[width=9cm]{grid_ecc0.eps}
\caption {H\'enon stability map for the CH model ($e_p=0$). The orbits of the grid are classified according to the color bar on the right of the map. The main families of periodic orbits are also presented with solid (stable) and dashed (unstable) curves. The two horizontal lines $R_H$ determine the boundaries of the Hill's sphere. For $\xi_0=0$ we get collision orbits (c.o.) while the grey regions (f.a.) indicate forbitten regions.}
\label{FigGridHenon0}
\end{figure}
\begin{figure}
\centering
$\begin{array}{ccc}
\includegraphics[width=5.5cm]{grid_ecc48p.eps} & \qquad & \includegraphics[width=5.5cm]{grid_ecc48a.eps}\\
\textnormal{(a)} & \qquad & \textnormal{(b)} \\
\includegraphics[width=5.5cm]{grid_ecc20p.eps} & \qquad & \includegraphics[width=5.5cm]{grid_ecc20a.eps}\\
\textnormal{(c)} & \qquad & \textnormal{(d)} \\
\includegraphics[width=5.5cm]{grid_ecc50p.eps} & \qquad & \includegraphics[width=5.5cm]{grid_ecc50a.eps}\\
\textnormal{(e)} & \qquad & \textnormal{(f)}
\end{array} $
\caption{H\'enon stability maps of grid size $150\times 100$ for the indicated planetary eccentricity values. Panels on the left correspond to the {\em periapsis} case and panels on the right to the {\em apoapsis} case. The color map is as in Fig. \ref{FigGridHenon0}.}
\label{FigGridHenonEH}
\end{figure}
\subsection{H\'enon stability maps}
In the circular Hill model, which has two degrees of freedom, the method of Poincar\'e section can depict clearly the topology of phase space and the qualitative characteristics of the trajectories (H\'enon, 1970; Chauvineau and Mignard, 1991). For $C_H>3^{4/3}$ the zero velocity curves define closed regions on the $O\xi\eta$ plane and all orbits, either regular or chaotic, are bounded. For $C_H<3^{4/3}$ the majority of chaotic orbits are escape orbits. H\'enon (1970) determined the borders of bounded motion by considering the plane $C_H-\xi_0$ of initial conditions, with $\eta_0=0$, and $\dot \eta_0>0$, which are defined by the Jacobi integral (\ref{EqJacobi}). This diagram of bounded motion, which is called {\em H\'enon diagram}, has been used also by Shen and Tremaine (2008) in order to obtain bounded satellite motion by using numerical integrations of a 6-body model, which includes the Sun, the four giant planets and the satellite as a massless body. The H\'enon diagrams showed that regions of bounded motion are located near the main families ($f$, $g$ and $g'$) of stable periodic orbits.
Following our methodology, the H\'enon map (or diagram) for the CH model ($e_p=0$) is given in Fig. \ref{FigGridHenon0}. Empty (white) regions correspond to escape orbits. The regular orbits (dark regions) are distributed around the stable families of periodic orbits and this is in agreement with the results of H\'enon (1970). The reader can observe the tangle of stability around the stable family $Hn$, and this fact indicates the efficiency of the FLI method to provide detailed stability maps. Chaotic orbits (say orbits with $FLI\gtrapprox10^6$) are almost absent in the map, since the majority of chaotic orbits escape during the numerical integration.
Using the same initial conditions as those of Fig. \ref{FigGridHenon0}, we set $e_p\neq 0$ and repeat the computations. Note that positive or negative values of $e_p$ indicate that the planet is initially at periapsis ($p$-case) or apoapsis ($a$-case), respectively. Also, we remark that the points of the characteristic curves $f$, $g$ and $g'$ do not correspond to periodic orbits in this case.
In Figs. \ref{FigGridHenonEH}a and \ref{FigGridHenonEH}b we present the stability maps for the $p$ and $a$ case, respectively, and for $e_p=0.048$ (Jupiter's eccentricity). We can observe that the width of the stable region of retrograde motion (i.e. around the family $f$) has shrunk but not significantly. The tail of this region ($C_H<0$) seems now to be located below the characteristic curve $f$ for the $p$-case and above it for the $a$-case. A rather significant effect of the non-zero planetary eccentricity value is observed in case of prograde orbits and mainly those located close to the $g'$ family, where chaotic orbits appear. This is more clear in the $a$-case, where now the ordered orbits are obtained for higher values of $C_H$, compared to those of the circular case.
By increasing the planetary eccentricity to $e_p=0.2$ (periapsis case), we see from the corresponding map of Fig. \ref{FigGridHenonEH}c that the area of the stability regions remains practically unchanged for retrograde as well as prograde orbits. However, for the apoapsis case ($e_p=-0.2$) the stability region of retrograde orbits has shrunk significantly, while the stability region of prograde orbits has disappeared completely from the map (actually stability can be found for higher values of $C_H$). Finally in Figs. \ref{FigGridHenonEH}e and \ref{FigGridHenonEH}f the case of a relatively high eccentricity value $(|e_p|=0.5)$ is shown. It is remarkable that in the $p$-case we can still find regions of stable orbits for both retrograde and prograde cases. But retrograde stable orbits are found almost only inside the Hill's sphere. In contrast, the $a$-case shows an extensive reduction of the area of stable orbits. The only significant region of stable (retrograde) orbits left is located outside the Hill sphere.
\subsection{Stability maps along CH families}
The H\'enon maps reveal that, in the case of the EH model, regions of stable satellite orbits are located around the stable families of the CH model. The effect of the planetary eccentricity to the periodic orbits of a given family, say $A$, of the circular model can be studied by constructing stability maps on the plane $\xi_0 - e_p$, where for each value $\xi_0$ we complete the set of initial conditions with $\eta_0=\dot\xi_0=0$ and $\dot\eta_0$ has the value that corresponds to the particular periodic orbit of the given family $A$. Thus, the initial conditions of the line $e_p=0$ correspond exactly to the family $A$ of periodic orbits.
The stability map along the family $f$ is shown in Fig. \ref{FigGridCHfamf}. Inside the Hill's sphere ($-3^{-1/3}<\xi_0<0$) the stable orbits dominate and exist up to relatively high eccentricity values. Outside the Hill's sphere, we obtain a region of stability around the family $f$ ($e_p=0$). Two of the gaps in this region are associated with the crossing of the unstable family $g3$ by the family $f$ and can be also seen in the H\'enon maps of Figs. \ref{FigGridHenon0} and \ref{FigGridHenonEH}a,b. This region is restricted to low (absolute) eccentricity values, because we have seen that, as the eccentricity increases, the stable regions located outside the Hill's sphere move away from the characteristic curve $f$.
The stability map along the family $g$ is shown in Fig. \ref{FigGridCHfamg}a. The point B indicates the pitchfork bifurcation of the family $g'$ from the family $g$ (H\'enon, 1969). No stable orbits are found around the unstable part of the family (on the right of the bifurcation point B). Around the stable part of the family, regions of ordered motion dominate, but the asymmetry between the periapsis and apoapsis cases is clearly seen. A similar asymmetry is also seen in the map along the family $g'$, shown in Fig. \ref{FigGridCHfamg}b. So we conclude that the {\em periapsis} case gives significantly more stable prograde orbits, compared to the {\em apoapsis} case, in agreement with the results obtained from the H\'enon maps.
\begin{figure}[tb]
\centering
\includegraphics[width=8cm]{grid_f.eps}
\caption {Stability map ($150\times 200$ grid) along the stable family $f$ of retrograde orbits. The color map is as in Fig. \ref{FigGridHenon0}.}
\label{FigGridCHfamf}
\end{figure}
\begin{figure}[tb]
\centering
$\begin{array}{ccc}
\includegraphics[width=5.5cm]{grid_g.eps} & \qquad & \includegraphics[width=5.5cm]{grid_gt.eps}\\
\textnormal{(a)} & \qquad & \textnormal{(b)} \\
\end{array} $
\caption{Stability maps ($150\times 200$ grid) along {\bf a.} family $g$ {\bf b.} family $g'$. The color map is as in Fig. \ref{FigGridHenon0}.}
\label{FigGridCHfamg}
\end{figure}
\subsection{Stability maps along EH families}
In the previous cases we associated regions of stable motion of the EH model with the families of periodic orbits of the CH model. Additionally, new regions of stability are expected to exist around stable periodic orbits of the EH model. We construct stability maps centered along a given family $F$ of periodic orbits and defined by grids of initial conditions on the plane $e_p$-$\Delta\xi_0$. Each point of such a grid corresponds to the initial conditions $\bar{\xi}_0+\Delta\xi_0$, $\eta_0=\dot\xi_0=0$ and $\bar{\dot\eta}_0$, where ($\bar{\xi}_0$, $\bar{\dot\eta}_0$) are the initial conditions of the periodic orbit of family $F$ at the planetary eccentricity $e_p$, which has a period $T'$. Thus, the family $F$ is located on the axis $\Delta\xi_0=0$. For these maps we have set $t_{max}=20000 T'$ and $FLI_{max}=10^{20}$.
In Figs. \ref{FigGridEHfam}a,b we show the stability maps for the families $f_e^{1/6}$ and $f_e^{2/5}$, respectively. Family $f_e^{1/6}$ is stable in the segments $-0.35<e_p<-0.23$ and $0.23<e_p<0.35$. The stability map shows the existence of two islands of stability (I1 and I2) around these segments. For $|e_p|>0.35$, where the family is unstable, strongly irregular orbits exist, which escape after a few periods. In the interval $-0.23<e_p<0.23$, where the family is also unstable, the existing instability gives a very thin chaotic zone in phase space, which is not clearly seen due to the resolution of the map. The wide stability regions (R1 and R2) that exist close to the unstable zone are associated with the stable family $f$ of the circular problem, as we have mentioned above.
A similar dynamical situation is seen in the map of the family $f_e^{2/5}$. This family is stable for $-0.56<e_p<0$ and close to this segment we can observe the existence of regimes (I1-I3) with stable orbits. For $e_p>0$ the family is unstable, but for small eccentricities the instability zone is very thin. In this instability zone the orbits are weakly chaotic and their FLI evolution is like that of Fig. \ref{FigFLIevol} (case 3). The neighbouring stability regions R1 and R2 extend up to $e_p\approx 0.25$; for larger planetary eccentricities the orbits escape.
\begin{figure}[tb]
\centering
$\begin{array}{ccc}
\includegraphics[width=5.5cm]{grid_ef16.eps} & \qquad & \includegraphics[width=5.5cm]{grid_ef25.eps}\\
\textnormal{(a)} & \qquad & \textnormal{(b)} \\
\includegraphics[width=4cm]{palletegeo2.eps} & & \\
\textnormal{(c)} & &
\end{array} $
\caption{Stability maps of grid size $150\times 100$ along {\bf a.} family $f_e^{1/6}$ {\bf b.} family $f_e^{2/5}$. The color map is indicated in panel {\bf c} .}
\label{FigGridEHfam}
\end{figure}
\section{Conclusions}
The elliptic Hill's model is an extension of the famous circular Hill's model. It is obtained by applying the classical Hill's transformation to the equations of the elliptic restricted three body problem in a rotating frame. In this paper we studied the main qualitative features of its dynamics by computing (i) families of periodic orbits and their stability and (ii) stability maps for the determination of regions with regular orbits.
The families of periodic orbits in the EH model bifurcate from the families of the CH model and continue with parameter the planetary eccentricity, $e_p$. We can define an infinite set of bifurcation points on the CH families ($e_p=0$) by increasing the multiplicity of the periodic orbits. We have determined a large set of bifurcation points that correspond mainly to stable periodic orbits, since we are interested in locating stable families. For each bifurcation point we computed the family of periodic orbits by varying the planetary eccentricity, $e_p$. We classified the families according to their resonance and discussed their structure. We found that most of the families extend up to high absolute $e_p$ values for both the periapsis and the apoapsis case. However, in some cases the EH families reach a maximum value of $|e_p|$, where a saddle-node bifurcation takes place. Then a new EH family is generated, which also meet a bifurcation point on a CH family. Nevertheless, there may exist EH families that do not originate from critical points of the circular problem, but such families cannot be computed by the approximation followed in this paper.
Most of the families are found to start as unstable, but in many cases family segments exist having stable periodic orbits even for relatively high absolute eccentricity values ($e_p>0.5$). Generally, continuing the families for $e_p>0$ and $e_p<0$ we find that at least one case gives simply unstable orbits. An exception is the family $f_e^{3/5}$, which in both cases starts as stable. Family segments of double and complex instability also exist.
In order to determine the phase space regions of regular orbits, we computed maps of dynamical stability defined on various planes. The majority of initial conditions corresponds to orbits that are irregular-chaotic and finally escape. However, the H\'enon stability maps showed that around the stable families $f$ and $g$ of the CH model there exist wide regions of regular motion. As the planetary eccentricity increases, the stability regions are affected significantly only when the planet is at apoapsis for $t=0$. Prograde orbits are affected more than retrograde ones.
The stable periodic orbits of the EH model are located at the ``centers'' of phase space regions with regular orbits. Instead, unstable periodic orbits are associated with chaotic orbits in their vicinity. For large absolute eccentricity values such chaotic orbits escape after a few planetary revolutions. For $|e_p|\lessapprox 0.2$ the chaotic regions around the unstable periodic orbits are thin and are surrounded by regions of regular orbits. In these regions the orbits are trapped and do not escape, at least for relatively long time spans.
\vspace{0.5cm}
\noindent{\bf Acknowledgements} The authors would like to thank the anonymous reviewers for their comments and fruitful suggestions to improve the
quality of the paper.
|
\section{Introduction}
\label{sec:Introduction}
To protect digital content from unauthorized redistribution, distributors embed watermarks in the content such that, if a customer distributes his copy of the content, the distributor can see this copy, extract the watermark and see which user it belongs to. By embedding a unique watermark for each different user, the distributor can always determine from the detected watermark which of the customers is guilty. However, several users could cooperate to form a coalition, and compare their differently watermarked copies to look for the watermark. Assuming that the original data is the same for all users, the differences they detect are differences in their watermarks. The colluders can then distort this watermark, and distribute a copy which matches all their copies on the positions where they detected no differences, and has some possibly non-deterministic output on the detected watermark positions. Since the watermark does not match any user's watermark exactly, finding the guilty users is non-trivial.
In this paper we focus on the problem of constructing efficient collusion-resistant schemes for tracing pirates, which involves finding a way to choose watermark symbols for each user (the traitor tracing code) and a way to trace a detected copy back to the guilty users (an accusation algorithm). In particular, we will focus on the application of such schemes in the dynamic setting, where the pirate output is detected in real-time, before the next watermark symbols are embedded in consecutive segments of the content. We will show that by building upon the (static) Tardos scheme \cite{tardos03}, we can construct efficient and flexible dynamic traitor tracing schemes. The number of watermark symbols needed in our schemes is a significant improvement compared to the scheme of Tassa \cite{tassa05}, and our schemes can be easily adjusted when the model is slightly different from the standard dynamic traitor tracing model \cite{berkman01,fiat01,roelse11,tassa05}.
\subsection{Model}
\label{sub:Introduction-Model}
Let us first formally describe the mathematical model for the problem discussed in this paper. First, some entity called the distributor controls the database of watermarks and distributes the content. The recipients, each receiving a watermarked copy of the content, are referred to as users. We write $U = \{1, \ldots, n\}$ for the set of all users, and we commonly use the symbol $j$ for indexing these users. For the watermarks, we refer to the sequence of watermarking symbols assigned to a user $j$ by the vector $\vec{X}_j$, which is also called a codeword. We write $\ell$ for the total number of watermark symbols in a codeword, so that each codeword $\vec{X}_j$ has length $\ell$, and we commonly use the symbol $i$ to index the watermark positions. We write $\mathcal{X}$ for the algorithm used to generate the codewords $\vec{X}_j$. In this paper we only focus on watermark symbols from a binary alphabet, so that $(\vec{X}_j)_i \in \{0,1\}$ for all $i,j$. A common way to represent the traitor tracing code is by putting all codewords $\vec{X}_j$ as rows in a matrix $X$, so that $X_{j,i} = (\vec{X}_j)_i$ is the symbol on position $i$ of user $j$.
After assigning a codeword to each user, the codewords are embedded in the data as watermarks. The watermarked copies are sent to the users, and some of the users (called the pirates or colluders) collude to create a pirate copy. The pirates form a subset $C \subseteq U$, and we use $c = |C|$ for the number of pirates in the coalition. The pirate copy has some distorted watermark, denoted by $\vec{y}$. We assume that if on some position $i$ all pirates see the same symbol, they output this symbol. This assumption is known in the literature as the marking assumption. On other positions we assume pirates simply choose one of the two symbols to output. This choice of pirate symbols can be formalized by denoting a pirate strategy by a (probabilistic) function $\rho$, which maps a code matrix $X$ (or the part of the matrix visible to them) to a forgery $\vec{y}$. After the coalition generates a pirate copy, we assume the distributor detects it and uses some accusation algorithm $\sigma$ to map the forgery $\vec{y}$ to some subset $\sigma(\vec{y}) = \hat{C} \subseteq U$ of accused users. These users are then disconnected from the system. Ideally $\hat{C} = C$, but this may not always be achievable.
\paragraph{Static schemes.} We distinguish between two types of schemes. In static schemes, the process ends after one run of the above algorithm with a fixed codelength $\ell$, and the set $\hat{C}$ is the final set of accused users. So the complete codewords are generated and distributed, the pirates generate and distribute a pirate copy, and the distributor detects this output and calculates the set of accused users. In this case an elementary result is that one can never have any certainty of catching all pirates. After all, the coalition could decide to sacrifice one of its members, so that $\vec{y} = \vec{X}_j$ for some $j \in C$. Then it is impossible to distinguish between other pirates $j' \in C \setminus \{j\}$ and innocent users $j' \in U \setminus C$. However, static schemes do exist that achieve catching at least one guilty user and not accusing any innocent users with high probability. The original Tardos scheme \cite{tardos03} belongs to this class of schemes.
\paragraph{Dynamic schemes.} The other type of scheme is the class of dynamic schemes, where the process of sending out symbols, detecting pirate output and running an accusation algorithm is repeated multiple times. In this case, if a user is caught, he is immediately cut off from the system and can no longer access the content. These dynamic scenarios for example apply to live broadcasts, such as pay-tv. The distributor broadcasts the content, while the pirates directly output a pirate copy of the content. The distributor then listens in on this pirate broadcast, extracts the watermarks, and uses this information for the choice of watermarks for the next segment of the content. We assume that the pirates always try to keep their broadcast running, so that if one of the pirates is disconnected, the other pirates will take over. Ideally one demands that the set of accused users always matches the exact coalition, i.e.\ $\hat{C} = C$, and with dynamic schemes we can also achieve this with high probability, as we will see later. The new schemes we present in this paper belong to this class of schemes.
As mentioned earlier, we call static schemes successful if with high probability, at least one guilty user is caught, and no innocent users are accused. With dynamic schemes one can catch all pirates, so we only call such schemes successful if with high probability, all pirates are caught and no innocent users are accused. This leads to the following definitions of soundness and static/dynamic completeness.
\begin{definition}[Soundness and completeness] \label{def:Secureness}
Let $(\mathcal{X}, \sigma)$ be a traitor tracing scheme, let $c_0 \geq 2$ and let $\epsilon_1,\epsilon_2 \in (0,1)$. Then this scheme is called $\epsilon_1$-sound, if for all coalitions $C \subseteq U$ and pirate strategies $\rho$, the probability of accusing one or more innocent users is bounded from above by
\begin{align*}
P(\hat{C} \not\subseteq C) \leq \epsilon_1.
\end{align*}
A static traitor tracing scheme $(\mathcal{X}, \sigma)$ is called static $(\epsilon_2, c_0)$-complete, if for all coalitions $C \subseteq U$ of size at most $c_0$ and for all pirate strategies $\rho$, the probability of not catching \textit{any} pirates is bounded from above by
\begin{align*}
P(C \cap \hat{C} = \emptyset) \leq \epsilon_2.
\end{align*}
Finally, a dynamic traitor tracing scheme $(\mathcal{X}, \sigma)$ is called dynamic $(\epsilon_2, c_0)$-complete, if for all coalitions $C \subseteq U$ of size at most $c_0$ and for all pirate strategies $\rho$, the probability of not catching \textit{all} pirates is bounded from above by
\begin{align*}
P(C \not\subseteq \hat{C}) \leq \epsilon_2.
\end{align*}
\end{definition}
Note that we distinguish between $c$, the \textit{actual} collusion size, and $c_0$, the \textit{estimated} collusion size used by the distributor to build the traitor tracing scheme. Since $c$ is usually unknown, the distributor has to make a guess $c_0 \approx c$, which has to be sufficiently large to guarantee security, and sufficiently small to guarantee efficiency.
In the following sections we will omit the $c_0$ in the completeness property if the parameter is implicit. Similarly, when $\epsilon_1$ or $\epsilon_2$ is implicit, we simply call a scheme sound or complete. As we will see later, in the schemes discussed in this paper, $\epsilon_1/n$ and $\epsilon_2$ are closely related. We will use the notation $\eta = \ln(\epsilon_2)/\ln(\epsilon_1/n)$ to denote the log ratio of these error probabilities. In most practical scenarios we have $\epsilon_1/n < \epsilon_2$, so usually $\eta \in (0,1)$.
\subsection{Related work}
\label{sub:Introduction-RelatedWork}
The schemes in this paper all build upon the Tardos scheme \cite{tardos03}, introduced in 2003. This is an efficient static traitor tracing scheme, and it was the first scheme to achieve $\epsilon_1$-soundness and $(\epsilon_2, c_0)$-completeness with a codelength of $\ell = O(c_0^2 \ln(n/\epsilon_1))$. In the same paper it was proved that this order codelength is asymptotically optimal for large $c$. The original Tardos scheme had a codelength of $\ell = 100 c_0^2 \ln(n/\epsilon_1)$, and several improvements of the Tardos scheme have been suggested to reduce the constant before the $c_0^2 \ln(n/\epsilon_1)$. We mention two in particular: the improved analysis done by Blayer and Tassa \cite{blayer08}; and the introduction of a symmetric score function by \v{S}kori\'{c} et al. \cite{skoric08}. Laarhoven and De Weger combined these improvements \cite{laarhoven11} to get even shorter codelength constants. For $c_0 \geq 2$ and $\eta \leq 1$, this construction gives codelengths of $\ell < 24 c_0^2 \ln(n/\epsilon_1)$, with the constant further decreasing as $c_0$ increases or $\eta$ decreases. For asymptotically large $c_0$, this construction leads to codelengths satisfying $\ell = [\frac{\pi^2}{2} + O(c_0^{-1/3})] c_0^2 \ln(n/\epsilon_1)$. The symmetric Tardos scheme and its properties are discussed in Section~\ref{sec:Preliminaries}.
For the dynamic setting, we mention four papers. In 2001, Fiat and Tassa~\cite{fiat01} constructed a deterministic scheme, i.e., a scheme with $\epsilon_1 = \epsilon_2 = 0$. The number of symbols needed to catch pirates in that scheme is only $\ell = O(c \log n)$, but the alphabet size required is $q = 2c + 1$. In the same year, Berkman et al.~\cite{berkman01} proposed several deterministic schemes using a smaller alphabet of size $q = c + 1$, with codelengths ranging from $O(c^3\log_2(n))$ to $O(c^2 + c\log_2(n))$. In 2005, Tassa~\cite{tassa05} combined the dynamic scheme of Fiat and Tassa~\cite{fiat01} with the static scheme of Boneh and Shaw~\cite{boneh98}, to get a dynamic scheme using a binary alphabet, with a codelength of $\ell = O(c^4 \log_2(n) \ln(c/\epsilon_1))$. In the same paper it was suggested that using the Tardos scheme instead of the scheme of Boneh and Shaw as a building block may decrease the codelength by a factor $c$, thus possibly giving a codelength of $\ell = O(c^3 \log_2(n) \ln(c/\epsilon_1))$. In 2011, Roelse~\cite{roelse11} presented another deterministic scheme. As in the generalization of the scheme of Fiat and Tassa presented by Berkman et al.~\cite{berkman01}, in the scheme of Roelse the alphabet size equals $kc + 1$ with $k \geq 2$ and for a fixed value of $k$, the codelength is $O(c \log n)$. Moreover, the real-time computational cost and the bandwidth usage are logarithmic in $n$, instead of linear in $n$ as in the scheme of Fiat and Tassa and its generalization of Berkman et al.~\cite{berkman01}.
\subsection{Contributions and outline}
\label{sub:Introduction-Contributions}
First we show that the static Tardos scheme can be extended to a dynamic traitor tracing scheme in an efficient way, allowing us to catch the whole coalition instead of at least one colluder. This dynamic scheme has a codelength of $\ell = O(c c_0 \ln(n/\epsilon_1))$, where the constants only slightly increase compared to the constants of Laarhoven and De Weger~\cite{laarhoven11}. The adjustments do not influence the method of generating codewords, so these can still be generated in advance.
To avoid the loss of efficiency caused by having to choose a value $c_0$, we then show how to create a $c_0$-independent ``universal" dynamic scheme that does not require a sharp estimate of $c$ as input. The property that the codewords can be generated in advance is left unchanged, while the scheme also has several advantages with respect to flexibility, detailed in Section~\ref{sec:Discussion}. The codelength of this scheme is also $\ell = O(c^2 \ln(n/\epsilon_1))$, thus improving upon the results of Tassa~\cite{tassa05} by roughly a factor $O(c^2)$ and upon the suggested improvement of Tassa by a factor $O(c)$.
The paper is organized as follows. First, we recall the construction of the static symmetric Tardos scheme and its properties in Section~\ref{sec:Preliminaries}. This scheme and its results will be used as the foundation for the dynamic Tardos scheme, which we present in Section~\ref{sec:DynamicTardos}. Then, in Section~\ref{sec:SemiDynamicTardos} we present a modification of the dynamic Tardos scheme when the setting is not fully dynamic. In Section~\ref{sec:UniversalTardos} we then present the universal Tardos scheme, which is an extension of the dynamic Tardos scheme that does not require a sharp bound on $c$ as input. In Section~\ref{sec:Discussion}, we discuss the results and argue that our schemes have several advantages with respect to flexibility as well. Finally, in Section~\ref{sec:OpenProblems} we list some open problems raised by our work.
This paper is mainly based on results from the first author's Master's thesis \cite{msc11}.
\section{Preliminaries: The Tardos scheme}
\label{sec:Preliminaries}
The results in the next sections all build upon results from the (static) symmetric Tardos scheme, so we first discuss this scheme here. Since the codeword generation of the schemes discussed in this paper all use (a variant of) the arcsine distribution, we also explicitly mention this distribution below.
\subsection{Arcsine distribution}
\label{sub:Preliminaries-Arcsine}
The standard arcsine distribution function $F(p)$ on $[0,1]$, and its associated probability density function $f(p)$, are given by:
\begin{align}
F(p) = \frac{2}{\pi} \arcsin(\sqrt{p}), \quad f(p) = \frac{1}{\pi \sqrt{p (1 - p)}}. \label{dist2}
\end{align}
This distribution function will be used in Section~\ref{sec:UniversalTardos}. In Sections~\ref{sec:Preliminaries}, \ref{sec:DynamicTardos} and \ref{sec:SemiDynamicTardos} we will use a variant of this distribution function, where the values of $p$ cannot be arbitrarily close to $0$ and $1$, as this generally leads to a high probability of accusing innocent users. Tardos~\cite{tardos03} therefore used the arcsine distribution with a certain small cutoff parameter $\delta > 0$, such that $p$ is always between $\delta$ and $1 - \delta$. By scaling $F$ and $f$ appropriately on this interval, this leads to the following distribution functions $F_{\delta}$ and associated probability density functions $f_{\delta}$:
\begin{align}
F_{\delta}(p) &= \frac{2 \arcsin(\sqrt{p}) - 2 \arcsin(\sqrt{\delta})}{\pi - 4 \arcsin(\sqrt{\delta})}, \label{dist1} \\
f_{\delta}(p) &= \frac{1}{(\pi - 4 \arcsin(\sqrt{\delta}))\sqrt{p (1 - p)}}.
\end{align}
Note that taking $\delta = 0$ (i.e., using no cutoff) leads to $F_0(p) \equiv F(p)$.
\subsection{Construction}
\label{sub:Preliminaries-Construction}
The Tardos scheme, with parameters $d_{\ell}, d_z, d_{\delta}$ as used by Blayer and Tassa \cite{blayer08} and Laarhoven and De Weger \cite{laarhoven11}, and with the symmetric score function introduced by \v{S}kori\'{c} et al.~\cite{skoric08}, is described below.
\begin{enumerate}
\item \textbf{Initialization phase}
\begin{enumerate}
\item Take the codelength as $\ell = d_{\ell} c_0^2 \ln(n/\epsilon_1)$.
\item Take the threshold as $Z = d_z c_0 \ln(n/\epsilon_1)$.
\item Take the cutoff parameter as $\delta = 1/(d_{\delta} c_0^{4/3})$. \footnote{Previously \cite{blayer08, skoric08, tardos03, laarhoven11}, it was common to parametrize the offset $\delta$ as $\delta = 1/(d_{\delta} c_0)$. However, Laarhoven and De Weger \cite{laarhoven11} showed that to get an optimal codelength, $\delta$ should scale as $c_0^{-4/3}$ rather than $c_0^{-1}$. Therefore we now use $\delta = 1/(d_{\delta} c_0^{4/3})$, with $d_{\delta}$ converging to a non-zero constant for asymptotically large $c_0$.}
\end{enumerate}
\item \textbf{Codeword generation} \\
For each position $1 \leq i \leq \ell$:
\begin{enumerate}
\item Select $p_i \in [\delta, 1 - \delta]$ from the distribution function $F_{\delta}(p)$ defined in \eqref{dist1}.
\item For each user $j \in U$, generate the $i$th entry of the codeword of user $j$ according to $P(X_{j,i} = 1) = p_i$ and $P(X_{j,i} = 0) = 1 - p_i$.
\end{enumerate}
\item \textbf{Distribution of codewords}\\
Send to each user $j \in U$ their codeword $\vec{X}_j = (X_{j,1}, \ldots, X_{j,\ell})$, embedded as a watermark in the content.
\item \textbf{Detection of pirate output} \\
Detect the pirate output, and extract the watermark $\vec{y} = (y_1, \ldots, y_{\ell})$.
\item \textbf{Accusation phase} \\
For each user $j \in U$:
\begin{enumerate}
\item For each position $1 \leq i \leq \ell$, calculate the user's score $S_{j,i}$ for this position according to:
\begin{align}
S_{j,i} = \begin{cases}
+\sqrt{(1 - p_i)/p_i} & \text{if $X_{j,i} = 1, y_i = 1$}, \\
-\sqrt{(1 - p_i)/p_i} & \text{if $X_{j,i} = 1, y_i = 0$}, \\
-\sqrt{p_i/(1 - p_i)} & \text{if $X_{j,i} = 0, y_i = 1$}, \\
+\sqrt{p_i/(1 - p_i)} & \text{if $X_{j,i} = 0, y_i = 0$}.
\end{cases} \label{scoresym}
\end{align}
\item Calculate the user's total score $S_j(\ell) = \sum_{i = 1}^{\ell} S_{j,i}$.
\item User $j$ is accused (i.e.\ $j \in \hat{C}$) iff $S_j(\ell) > Z$.
\end{enumerate}
\end{enumerate}
\subsection{Soundness}
\label{sub:Preliminaries-Soundness}
For the above construction, one can prove soundness and static completeness, provided the constants $d_{\ell}, d_z, d_{\delta}$ satisfy certain requirements. For soundness, Laarhoven and De Weger~\cite{laarhoven11} proved the following lemma. Here $h(x) = (e^x - 1 - x)/x^2$, which is a strictly increasing function from $(0,\infty)$ to $(\frac{1}{2}, \infty)$.
\begin{lemma} \cite[Lemma 1]{laarhoven11} \label{lem:Soundness}
Let the Tardos scheme be constructed as in Section~\ref{sub:Preliminaries-Construction}. Let $j$ be some arbitrary innocent user, and let $a > 0$. Then
\begin{align*}
E\left(e^{a S_j(\ell) c_0^{-1}}\right) \leq \left(\frac{\epsilon_1}{n}\right)^{-a \lambda_a d_{\ell}},
\end{align*}
where $\lambda_a = a h(a \sqrt{d_{\delta}} c_0^{-1/3})$.
\end{lemma}
Now if the following condition of soundness is satisfied,
\begin{align*}
\exists \ a > 0: \quad a \left(d_z - \lambda_a d_{\ell}\right) \geq 1, \tag{S} \label{eq:S}
\end{align*}
then using the Markov inequality and Lemma~\ref{lem:Soundness} with this $a$, for innocent users $j$ we get
\begin{align*}
P(j \in \hat{C}) & \leq P(S_j(\ell) > Z) = P(e^{a S_j(\ell) c_0^{-1}} > e^{a Z c_0^{-1}}) \\
& \leq \frac{E\left(e^{a S_j(\ell) c_0^{-1}}\right)}{e^{a Z c_0^{-1}}} \leq \left(\frac{\epsilon_1}{n}\right)^{a (d_z - \lambda_a d_{\ell})} \leq \frac{\epsilon_1}{n}.
\end{align*}
So the probability that no innocent user is accused is at least $(1 - \frac{\epsilon_1}{n})^n \geq 1 - \epsilon_1$, as was also shown by Laarhoven and De Weger~\cite[Theorem 3]{laarhoven11}.
\subsection{Static completeness}
\label{sub:Preliminaries-Completeness}
To prove static completeness, Laarhoven and De Weger~\cite{laarhoven11} used the following lemma. Below, and throughout the rest of this paper, $S(\ell) = \sum_{j \in C} S_j(\ell)$ represents the total coalition score, i.e., the sum of the scores of all pirates $j \in C$.
\begin{lemma} \cite[Lemma 2]{laarhoven11} \label{lem:Completeness}
Let the Tardos scheme be constructed as in Section~\ref{sub:Preliminaries-Construction}, and let $b > 0$. Then
\begin{align*}
E\left(e^{-b S(\ell) c_0^{-5/3}}\right) \leq \left(\frac{\epsilon_1}{n}\right)^{b \lambda_b d_{\ell} c_0^{1/3}},
\end{align*}
where $\lambda_b = \frac{2}{\pi} - \frac{4}{d_{\delta} \pi} c_0^{-1/3} - b h(b \sqrt{d_{\delta}}) c_0^{-2/3}$.
\end{lemma}
If the following condition of completeness is satisfied,
\begin{align*}
\exists \ b > 0: \quad b (\lambda_b d_{\ell} - d_z) \geq \eta c_0^{-1/3}, \tag{C} \label{eq:C}
\end{align*}
then using the pigeonhole principle, the Markov inequality and Lemma~\ref{lem:Completeness} with this $b$ we get
\begin{align*}
P(C \cap \hat{C} = \emptyset) & \leq P(S(\ell) < c_0 Z) \leq \frac{E\left(e^{-b S(\ell) c_0^{-5/3}}\right)}{e^{-b Z c_0^{-2/3}}} \\
& \leq \left(\frac{\epsilon_1}{n}\right)^{b (\lambda_b d_{\ell} - d_z) c_0^{1/3}} \leq \left(\frac{\epsilon_1}{n}\right)^{\eta} = \epsilon_2.
\end{align*}
So static completeness follows from Lemma~\ref{lem:Completeness} and condition~\eqref{eq:C}, as was also shown by Laarhoven and De Weger~\cite[Theorem 4]{laarhoven11}.
\subsection{Codelengths}
\label{sub:Preliminaries-Optimization}
Blayer and Tassa~\cite{blayer08}, and subsequently Laarhoven and De Weger~\cite{laarhoven11}, gave a detailed analysis to go from requirements \eqref{eq:S} and \eqref{eq:C} to the optimal set of parameters that satisfies the constraints and minimizes $d_{\ell}$. Recall that $\ell = d_{\ell} c_0^2 \ln(n/\epsilon_1)$, so a smaller $d_{\ell}$ gives shorter codelengths, whereas the parameters $d_z$ and $d_{\delta}$ affect only $Z$ and $\delta$, which have no influence on the efficiency of the scheme. In the end, the following result was obtained.
\begin{lemma} \cite[Theorem 6]{laarhoven11} \label{lem:FirstOrder}
Let $\gamma = \left(\frac{2}{3\pi} \right)^{2/3} \approx 0.36$. The asymptotically optimal value for $d_{\ell}$ is
\begin{align*}
d_{\ell} &= \frac{\pi^2}{2} + O(c_0^{-1/3}),
\end{align*}
the associated values for $d_z$ and $d_{\delta}$ are
\begin{align*}
d_z = \pi + O(c_0^{-1/3}), & \quad d_{\delta} = \frac{4}{\gamma} - O\left(\frac{\eta}{\ln c_0}\right),
\end{align*}
and the corresponding values for $a,b,\lambda_a, \lambda_b$ are
\begin{align*}
a = \frac{2}{\pi} - O(c_0^{-1/3}), & \quad b = \frac{\ln c_0}{9 \pi \gamma} - O\left(\ln\left(\frac{\ln c_0}{\eta}\right)\right), \\
\lambda_a = \frac{1}{\pi} + O(c_0^{-1/3}), & \quad \lambda_b = \frac{2}{\pi} - O(c_0^{-1/3}).
\end{align*}
\end{lemma}
A direct consequence of Lemma~\ref{lem:FirstOrder} is the following, which gives the asymptotically optimal scheme parameters for $c_0 \to \infty$.
\begin{corollary} \cite[Corollary 1]{laarhoven11} \label{cor:Asymptotics}
The construction from Section \ref{sub:Preliminaries-Construction} gives an $\epsilon_1$-sound and static $(\epsilon_2,c_0)$-complete scheme with asymptotic scheme parameters
\begin{align*}
\ell \to \frac{\pi^2}{2} c_0^2 \ln(n/\epsilon_1), \quad Z \to \pi c_0 \ln(n/\epsilon_1), \quad \delta \to \frac{\gamma}{4} c_0^{-4/3}.
\end{align*}
\end{corollary}
For further details on the optimal first order constants, see Laarhoven and De Weger~\cite{laarhoven11}.
\subsection{Example}
\label{sub:Preliminaries-Example}
For the next few sections, we will use a running example to compare the codelengths of the several schemes. Let the scheme parameters be given by $c_0 = 25$ pirates, $n = 10^6$ users, and error probabilities $\epsilon_1 = \epsilon_2 = 10^{-3}$. Then $\eta = \frac{1}{3}$, and the optimal values of $d_{\ell}, d_z, d_{\delta}$ can be calculated numerically as
\begin{align*}
d_{\ell} = 8.46, \quad d_z = 4.53, \quad d_{\delta} = 14.36.
\end{align*}
This leads to the scheme parameters
\begin{align*}
\ell = 109\,585, \quad Z = 2345, \quad \delta = 5.09 \cdot 10^{-4}.
\end{align*}
So using these scheme parameters, we know that after $109\,585$ symbols, with probability at least $0.999$ there are no false accusations (regardless of the actual coalition size $c$), and with probability at least $0.999$ at least one pirate is accused if the actual coalition size $c$ does not exceed the bound on the coalition size $c_0 = 25$. In Fig.~\ref{fig:Fig1} we show simulation results for these parameters, with $c = c_0 = 25$. The curves in the figure are the pirate scores $S_j(i)$ for each pirate $j \in C$, while the shaded area is bounded from above by the highest score of an innocent user, and bounded from below by the lowest score of an innocent user in this simulation. In Fig.~\ref{fig:Fig1a} we simulated pirates using the interleaving attack (i.e.\ for each position, they choose a random pirate and output his symbol), and in Fig.~\ref{fig:Fig1b} they used the scapegoat strategy (i.e.\ one pirate, the scapegoat, always outputs his symbol, until he is caught and another pirate is picked as the scapegoat). With the scapegoat strategy, only one pirate is caught, while using the interleaving attack leads to many accused pirates.
\begin{figure}[!t]
\centering
\subfloat[][Interleaving attack]{\includegraphics[width=\columnwidth]{fig1a} \label{fig:Fig1a}} \\
\subfloat[][Scapegoat strategy]{\includegraphics[width=\columnwidth]{fig1b} \label{fig:Fig1b}}
\caption{Simulations of the Tardos scheme, with $c = c_0 = 25$ colluders, $n = 10^6$ users, and error probabilities $\epsilon_1 = \epsilon_2 = 10^{-3}$. The green, shaded area corresponds to the range of innocent user scores, the red lines correspond to pirate scores, and the dashed lines correspond to the threshold $Z$ and codelength $\ell$. In Fig.~\ref{fig:Fig1a} the pirates used the interleaving attack, whereas in Fig.~\ref{fig:Fig1b} they used the scapegoat strategy. In both cases, the total coalition score $S(\ell)$ at time $\ell$ is approximately $72\,000$, but while in the first case the score is evenly divided among the pirates, in the second case one pirate takes all the blame.}
\label{fig:Fig1}
\end{figure}
\section{The dynamic Tardos scheme}
\label{sec:DynamicTardos}
Let us now explain how we create a dynamic scheme from the static Tardos scheme, such that with high probability we catch all colluders, instead of at least one colluder. The change we make is the following. Instead of only comparing the cumulative user scores to $Z$ after $\ell$ symbols, we now compare the scores to $Z$ after every single position $i$. If a user's score exceeds $Z$ at any point in time, he is disconnected immediately and can no longer access the content. His score is then necessarily between $Z$ and $\tilde{Z} := Z + \sqrt{d_{\delta}} c_0^{2/3} > Z + \max_{p_i,X_{j,i},y_i} S_{j,i}$. The other parts of the construction remain the same, except for the values of $d_{\ell}, d_z, d_{\delta}$, which now have to be chosen differently.
\subsection{Construction}
\label{sub:DynamicTardos-Construction}
The scheme again depends on three constants $d_{\ell}, d_z, d_{\delta}$. We will show in Sections~\ref{sub:DynamicTardos-Soundness} and \ref{sub:DynamicTardos-Completeness} that if certain requirements on these constants are satisfied, we can prove soundness and dynamic completeness. Below we say a user is active if he has not yet been disconnected from the scheme. As mentioned before, we assume that the pirates always output some watermarked data, unless all of the pirates are disconnected. In that case, the traitor tracing scheme terminates.
\begin{enumerate}
\item \textbf{Initialization phase}
\begin{enumerate}
\item Take the codelength as $\ell = d_{\ell} c_0^2 \ln(n/\epsilon_1)$.
\item Take the threshold as $Z = d_z c_0 \ln(n/\epsilon_1)$.
\item Take the cutoff parameter as $\delta = 1/(d_{\delta} c_0^{4/3})$.
\item Set initial user scores at $S_j(0) = 0$.
\end{enumerate}
\item \textbf{Codeword generation} \\
For each position $1 \leq i \leq \ell$:
\begin{enumerate}
\item Select $p_i \in [\delta, 1 - \delta]$ from $F_{\delta}(p)$ defined in \eqref{dist1}.
\item Generate $X_{j,i} \in \{0,1\}$ using $P(X_{j,i} = 1) = p_i$.
\end{enumerate}
\item \textbf{Distribution/Detection/Accusation} \\
For each position $1 \leq i \leq \ell$:
\begin{enumerate}
\item Send to each active user $j$ symbol $X_{j,i}$.
\item Detect the pirate output $y_i$. \\
(If there is no pirate output, terminate.)
\item Calculate scores $S_{j,i}$ using \eqref{scoresym}.
\item For active users $j$, set $S_j(i) = S_j(i-1) + S_{j,i}$. \\
(For inactive users $j$, set $S_j(i) = S_j(i-1)$.)
\item Disconnect all active users $j$ with $S_j(i) > Z$.
\end{enumerate}
\end{enumerate}
In the construction above, we separated the codeword generation from the distribution, detection and accusation. These phases can also be merged by generating $p_i$ and $X_{j,i}$ once we need them. However, we present the scheme as above to emphasize the fact that these phases can indeed be executed sequentially instead of simultaneously, and that the codeword generation can thus be done before the traitor tracing process begins.
\subsection{Soundness}
\label{sub:DynamicTardos-Soundness}
For the dynamic Tardos scheme as given above, we can prove the following result regarding soundness.
\begin{theorem} \label{thm:PDS-DT-Soundness}
Consider the dynamic Tardos scheme in Section~\ref{sub:DynamicTardos-Construction}. If the following condition is satisfied,
\begin{align*}
\exists \ a > 0: \quad a (d_z - \lambda_a d_{\ell}) \geq 1 + \frac{\ln(2)}{\ln(n/\epsilon_1)}, \tag{S'} \label{eq:S'}
\end{align*}
then the scheme is $\epsilon_1$-sound.
\end{theorem}
To prove the theorem, we first prove a relative upper bound on the probability that a single innocent user is accused and disconnected. This bound relates the error probability in the dynamic Tardos scheme to the probability that the user score at time $\ell$ is above $Z$. We then use the proof of the original Tardos scheme to get an absolute upper bound on the soundness error probability, and to prove Theorem~\ref{thm:PDS-DT-Soundness}. Since the relative upper bound gives us an extra factor $2$, and since the terms in \eqref{eq:S'} appear as exponents in the proof, we get an additional term $\ln(2)/\ln(n/\epsilon_1)$ compared to \eqref{eq:S}. Note that this term is small for reasonable values of $n$ and $\epsilon_1$, so this only has a small impact on the right hand side of \eqref{eq:S'}, compared to \eqref{eq:S}.
In the following we write $\tilde{S}_j(i) = \sum_{k=1}^i S_{jk}$ for the \textit{extended} user score. If user $j$ is still active at time $i$, then $\tilde{S}_j(i) = S_j(i)$. But whereas $S_j(i)$ does not change anymore once user $j$ is disconnected, the score $\tilde{S}_j(i)$ does change on every position, even if the user has already been disconnected. The score $\tilde{S}_j$ then calculates the user's score as if he had not been disconnected. Similarly, we write $\tilde{S}(i) = \sum_{j \in C} \tilde{S}_j(i)$ for coalitions $C$. Note that if the last pirate is disconnected at position $i_0 < \ell$, then $S_{j,i}$ and $S_j(i)$ are not defined for $i_0 < i \leq \ell$.
\begin{lemma} \label{lem:DT1}
Let $j \in U$ be an arbitrary innocent user, let $C \subseteq U \setminus \{j\}$ be a pirate coalition and let $\rho$ be some pirate strategy employed by this coalition. Then
\begin{align*}
P(j \in \hat{C}) = P(S_j(\ell) > Z) \leq 2 \cdot P\left(\tilde{S}_j(\ell) > Z\right).
\end{align*}
\end{lemma}
\begin{proof}
Let us define events $A$ and $B$ as
\begin{align*}
A &:= \{j \in \hat{C}\} = \{S_j(\ell) > Z\} = \bigcup_{i=1}^{\ell} \{\tilde{S}_j(i) > Z\}, \\
B &:= \{\tilde{S}_j(\ell) > Z\}.
\end{align*}
We trivially have $P(A \mid B) = 1$. For $P(B \mid A)$, note that under the assumption that $A$ holds, the process $\{\tilde{S}_j(i)\}_{i = i_0}^{\infty}$ starting at position $i_0 = \min \{i: S_j(i) > Z\} \leq \ell$ describes a symmetric random walk with no drift. So we then have $P(\tilde{S}_j(\ell) \geq \tilde{S}_j(i_0)) = 1/2$, and since $S_j(i_0) > Z$ it follows that $P(B \mid A) \geq 1/2$. Finally we apply Bayes' Theorem to $A$ and $B$ to get
\begin{align*}
P(A) = \frac{P(A \mid B)}{P(B \mid A)} \cdot P(B) \leq 2 \cdot P(B).
\end{align*}
This completes the proof.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:PDS-DT-Soundness}]
First, we remark that the distribution of $\tilde{S}_j(\ell)$ is the same as the distribution of the scores $S_j(\ell)$ in the original Tardos scheme, for the same parameters $\ell, Z, \delta$. From the Markov inequality, Lemma~\ref{lem:Soundness} and condition~\eqref{eq:S'} it thus follows that
\begin{align*}
P(\tilde{S}_j\left(\ell) > Z\right) \leq \frac{E\left(e^{a \tilde{S}_j(\ell) c_0^{-1}}\right)}{e^{a Z c_0^{-1}}} \leq \left(\frac{\epsilon_1}{n}\right)^{a (d_z - \lambda_a d_{\ell})} \leq \frac{\epsilon_1}{2n}.
\end{align*}
Using Lemma~\ref{lem:DT1} the result follows.
\end{proof}
\subsection{Dynamic completeness}
\label{sub:DynamicTardos-Completeness}
With the dynamic Tardos scheme, we get the following result regarding dynamic completeness. Recall that here we require that \textit{all} pirates are caught, instead of at least one, as was the case in the original Tardos scheme.
\begin{theorem} \label{thm:PDS-DT-Completeness}
Consider the dynamic Tardos scheme in Section~\ref{sub:DynamicTardos-Construction}. If the following condition is satisfied,
\begin{align*}
\exists \ b > 0: b \left(\lambda_b d_{\ell} - d_z\right) \geq \left(\eta + \frac{\ln(2) + b \sqrt{d_{\delta}}}{\ln(n/\epsilon_1)}\right) c_0^{-1/3}, \tag{C'} \label{eq:C'}
\end{align*}
then the scheme is dynamic $(\epsilon_2,c_0)$-complete.
\end{theorem}
Similar to the proof of soundness, we prove dynamic completeness by relating the error probability to the static completeness error probability of the static Tardos scheme described in Section~\ref{sec:Preliminaries}. Then we use the results from the static scheme to complete the proof. We again see a factor $2$ in the relative upper bound in Lemma~\ref{lem:DT-Comp}, which again comes from a random walk argument, and which explains the additional term $\ln(2)/\ln(n/\epsilon_1)$ in~\eqref{eq:C'}. The other term $b \sqrt{d_{\delta}}/\ln(n/\epsilon_1)$ is a consequence of using $\tilde{Z}$ instead of $Z$ in the proofs. Note that these two terms are generally small, compared to the term $\eta$.
\begin{lemma} \label{lem:DT-Comp}
Let $C$ be a coalition of size at most $c_0$, and let $\rho$ be any pirate strategy employed by this coalition. Then
\begin{align*}
P\left(C \not\subseteq \hat{C}\right) \leq 2 \cdot P\left(\tilde{S}(\ell) < c_0\tilde{Z}\right).
\end{align*}
\end{lemma}
\begin{proof}
First we remark that $P(\tilde{S}(\ell) < c_0\tilde{Z} \mid C \not\subseteq \hat{C}) \geq 1/2$. In other words, if not all pirates are caught by the end, the total extended coalition score will be below $c_0 \tilde{Z}$ with probability at least $1/2$. This is because if $C \not\subseteq \hat{C}$, then $S(\ell) < c_0\tilde{Z}$, and since $\tilde{S}(\ell) - S(\ell) = R(\ell)$ is a symmetric, unbiased random walk, with probability at least $1/2$ we have $R(\ell) < 0$ and as a consequence $\tilde{S}(\ell) < c_0\tilde{Z}$. Next, we use the definition of conditional probabilities to get
\begin{align*}
P(C \not\subseteq \hat{C}) \leq 2 \cdot P(C \not\subseteq \hat{C}) \cdot P(\tilde{S}(\ell) < c_0 \tilde{Z} \mid C \not\subseteq \hat{C}) \\ = 2 \cdot P\left(\tilde{S}(\ell) < c_0 \tilde{Z}, C \not\subseteq \hat{C}\right) \leq 2 \cdot P(\tilde{S}(\ell) < c_0\tilde{Z}).
\end{align*}
This proves the result.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:PDS-DT-Completeness}]
First, note that in the dynamic Tardos scheme, the only extra information pirates receive compared to the static Tardos scheme is the fact whether some of them are disconnected. This information is certainly covered by the information contained in the previous values of $p_i$; if pirates receive $p_1, \ldots, p_{i-1}$, then they can calculate their current scores themselves and calculate whether they would have been disconnected or not. Also note that $\tilde{S}(\ell)$ behaves the same as $S(\ell)$ in the static Tardos scheme, where the total coalition score is calculated for all pirates and all positions, regardless of whether they contributed on that position or not. So if we can prove that even in the static Tardos scheme, and even if coalitions get information about the previous values of $p_i$ (for which $y_i$ was already determined), the probability of keeping the coalition score $S(\ell)$ below $c_0\tilde{Z}$ is bounded by $\epsilon_2/2$, then it follows that also $P(\tilde{S}(\ell) < c_0 \tilde{Z}) \leq \epsilon_2/2$.
For the static Tardos scheme, note that the proof method for the completeness property does not rely on the other values of $p_i$ being secret. In fact, $p_i$ and $p_{i'}$ are independent for $i \neq i'$. The only assumption that is used in that proof is that the Marking Assumption applies, which does apply here, and that the \textit{current} value $p_i$ is hidden before $y_i$ is generated. So here we can also use the proof method of the static Tardos scheme. From the Markov inequality, Lemma~\ref{lem:Completeness} and condition~\eqref{eq:C'}, it thus follows that
\begin{align*}
P\left(\tilde{S}(\ell) < c_0\tilde{Z}\right)
& \leq \frac{E\left(e^{-b \tilde{S}(\ell) c_0^{-5/3}}\right)}{e^{-b (Z + \sqrt{d_{\delta}} c_0^{2/3}) c_0^{-2/3}}} \\
& \leq \left(\frac{\epsilon_1}{n}\right)^{b \left(\lambda_b d_{\ell} - d_z - \frac{\sqrt{d_{\delta}}}{\ln(n/\epsilon_1)} c_0^{-1/3}\right) c_0^{1/3}} \\
& \leq \left(\frac{\epsilon_1}{n}\right)^{\eta + \frac{\ln 2}{\ln(n/\epsilon_1)}} = \frac{\epsilon_2}{2}.
\end{align*}
Using Lemma~\ref{lem:DT-Comp} the result then follows.
\end{proof}
\subsection{Codelengths}
\label{sub:DynamicTardos-Optimization}
The requirements \eqref{eq:S'} and \eqref{eq:C'} are only slightly different from requirements \eqref{eq:S} and \eqref{eq:C}. For asymptotically large $c_0$, these differences even disappear, and the optimal asymptotic codelength is the same as in the static Tardos scheme. In Fig.~\ref{fig:Fig2} we show the optimal values of $d_{\ell}$ in the dynamic Tardos scheme for $\eta = 1$ and $\eta = 0.01$. The different curves correspond to different values of $n/\epsilon_1$, ranging from $n/\epsilon_1 = 10^3$ (the highest values of $d_{\ell}$) to $n/\epsilon_1 = 10^{15}$ (the lowest values of $d_{\ell}$).
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{fig2}
\caption{Optimal values of $d_{\ell}$ in the dynamic Tardos scheme. The dotted line corresponds to the asymptotic optimal value $d_{\ell} = \frac{\pi^2}{2} \approx 4.93$. The bold curves show the values of $d_{\ell}$ in the static Tardos scheme for $\eta = 1$ (top) and $\eta = 0.01$ (bottom) respectively. The five curves slightly above each of the bold curves show the optimal values of $d_{\ell}$ in the dynamic Tardos scheme for $n/\epsilon_1 = 10^{3k}$, for $k = 1$ up to $5$. Higher values of $k$ correspond to lower values of $d_{\ell}$.}
\label{fig:Fig2}
\end{figure}
Note that these values of $d_{\ell}$ correspond to the theoretical codelengths such that with probability at least $1 - \epsilon_1$, by time $\ell$ all of the pirates have been disconnected. This does not mean that the last pirate is likely to be caught \textit{exactly at} time $\ell$; this means that he is likely to be caught \textit{before or at} time $\ell$. So in practice the number of symbols needed to disconnect all traitors may very well be below this theoretical codelength $\ell$, and may even decrease compared to the static Tardos scheme.
Furthermore, if the coalition size is not known, then one generally uses a traitor tracing scheme that is resistant against up to $c_0 > c$ colluders. \v{S}kori\'{c} et al.~\cite{skoric08} showed that in the Tardos scheme, the total coalition score $S(i) = \sum_{j \in C} S_j(i)$ always increases linearly in $i$ with approximately the same slope, regardless of the actual coalition size $c$ or the employed pirate strategy $\rho$. More precisely, the score $S(i)$ behaves as $S(i) \approx i \tilde{\mu}$, with $\tilde{\mu} \approx \frac{2}{\pi}$ only slightly depending on the coalition size $c$ and the pirate strategy $\rho$. Since one chooses $\ell$ and $Z$ such that $S(\ell) \approx \ell \tilde{\mu} \approx c_0 Z$, it follows that $S(\frac{c}{c_0} \ell) \approx c Z$. In other words, to catch a coalition of size $c \leq c_0$, the expected number of symbols needed is approximately $\ell = O(\frac{c}{c_0} \ell) = O(c c_0 \ln(n/\epsilon_1))$. So compared to the static Tardos scheme, where the codelength is fixed in advance at $O(c_0^2 \ln(n/\epsilon_1))$, the codelength is reduced by a factor $\frac{c}{c_0}$. In particular, small coalitions of few pirates are generally caught up to $O(c_0)$ times faster, for $c_0 \gg c$.
\subsection{Example}
\label{sub:DynamicTardos-Example}
Let the scheme parameters be the same as in Section~\ref{sub:Preliminaries-Example}, i.e., $c_0 = 25$, $n = 10^6$ and $\epsilon_1 = \epsilon_2 = 10^{-3}$, so that $\eta = \frac{1}{3}$. The optimal values of $d_{\ell}, d_z, d_{\delta}$ satisfying \eqref{eq:S'} and \eqref{eq:C'} can be calculated numerically as
\begin{align*}
d_{\ell} = 9.00, \quad d_z = 4.73, \quad d_{\delta} = 13.44
\end{align*}
This leads to the scheme parameters
\begin{align*}
\ell = 116\,561, \quad Z = 2448, \quad \delta = 1.02 \cdot 10^{-3}
\end{align*}
In Fig.~\ref{fig:Fig3} we show some simulation results for these parameters, with the actual coalition also consisting of $c = c_0 = 25$ colluders. In Fig.~\ref{fig:Fig3a} the pirates used the interleaving attack, and in Fig.~\ref{fig:Fig3b} they used the scapegoat strategy. In both cases, the whole coalition is caught well before we reach $\ell$ symbols.
\begin{figure}[!t]
\centering
\subfloat[][Interleaving attack]{\includegraphics[width=\columnwidth]{fig3a} \label{fig:Fig3a}} \\
\subfloat[][Scapegoat strategy]{\includegraphics[width=\columnwidth]{fig3b} \label{fig:Fig3b}}
\caption{Simulations of the dynamic Tardos scheme, with the same parameters $c$, $c_0$, $n$, $\epsilon_1$ and $\epsilon_2$ as in Fig.~\ref{fig:Fig1}. Users are now disconnected as soon as their scores exceed the threshold $Z$, i.e., as soon as their corresponding score curves cross the bold horizontal line. In both cases, after less than $95\,000$ symbols all pirates have been caught, which is less than the theoretical codelength $\ell = 116\,561$, and less than the codelength of the static Tardos scheme with the same parameters, $\ell = 109\,585$.}
\label{fig:Fig3}
\end{figure}
\section{The weakly dynamic Tardos scheme}
\label{sec:SemiDynamicTardos}
In the dynamic Tardos scheme, we need to disconnect users as soon as their scores exceed the threshold $Z$. In some scenarios this may not be possible. For example, the pirates may transmit each symbol with a delay.
We call a traitor tracing scheme weakly dynamic if $B$ ($B \geq 1$) symbols are distributed during the delay between the original broadcast and the corresponding pirate output. Observe that the dynamic schemes presented in \cite{berkman01, fiat01, roelse11, tassa05} are not weakly dynamic schemes, as these schemes use the value of each symbol to adapt the distribution of the next symbols (i.e. $B = 0$ for these schemes).
In this section we present two weakly dynamic schemes based on the dynamic Tardos scheme.
First, in Section~\ref{sub:SemiDynamicTardos-First} we present a scheme that achieves a codelength of at most $\ell = d_{\ell} c_0^2 \ln(n/\epsilon_1) + B c_0$, where $d_{\ell}$ is the same as in the dynamic Tardos scheme for the same parameters. For small values of $B$, this means that with codelength which is only slightly higher than in the dynamic Tardos scheme, we can also catch all pirates in a weakly dynamic setting. Then, in Section~\ref{sub:SemiDynamicTardos-Second} we present a scheme that achieves a codelength of $\ell = d_{\ell,B} c_0^2 \ln(n/\epsilon_1)$, where $d_{\ell,B}$ increases with $B$. Since a small increase in $d_{\ell}$ can already lead to a big increase in the codelength, the second scheme generally has a larger codelength than the first scheme.
\subsection{First scheme: $\ell = d_{\ell} c_0^2 \ln(n/\epsilon_1) + B c_0$}
\label{sub:SemiDynamicTardos-First}
The first scheme is based on the following modification to the accusation algorithm of the dynamic Tardos scheme. Suppose a user's score exceeds $Z$ after $i_0$ positions. At position $i_0$ we now disconnect this user. Since this user may have contributed to the next $B$ symbols of the pirate output $\vec{y}$, we disregard the following $B$ `contaminated' positions of the watermark, and do not update the scores for positions $i \in \{i_0 + 1, \ldots, i_0 + B\}$. After those positions we continue the traitor tracing process as in the dynamic Tardos scheme, and we repeat the above procedure each time a user's score exceeds $Z$.
With this modification, the traitor tracing process on those positions that were used for calculating scores is identical to the traitor tracing process of the dynamic Tardos scheme. We can therefore use the analysis from Section~\ref{sec:DynamicTardos} and conclude that with at most $d_{\ell} c_0^2 \ln(n/\epsilon_1)$ positions for which we calculate scores, we can catch any coalition of size $c \leq c_0$. Since we disregarded at most $B c_0$ positions, the pirate broadcast will not last longer than $\ell = d_{\ell} c_0^2 \ln(n/\epsilon_1) + B c_0$ positions in total, where $d_{\ell}$, $d_z$ and $d_{\delta}$ are as in the dynamic Tardos scheme for the same parameters. This means that with at most $B c_0$ more symbols than in the dynamic Tardos scheme, we can also catch coalitions in this weakly dynamic traitor tracing setting.
\subsection{Second scheme: $\ell = d_{\ell,B} c_0^2 \ln(n/\epsilon_1)$}
\label{sub:SemiDynamicTardos-Second}
Instead of using $B c_0$ more symbols, we can also try to adjust the analysis of the dynamic Tardos scheme to the weakly dynamic traitor tracing scenario. We can do this by following the proof methods of the dynamic Tardos scheme, and by making one small adjustment. The change we make in the analysis is to use $\tilde{Z}_B := Z + B \sqrt{d_{\delta}} c_0^{2/3} > Z + B \max_p S_{j,i}(p)$ instead of $\tilde{Z} = Z + \sqrt{d_{\delta}} c_0^{2/3}$ as our new upper bound for the scores of users in the proofs. This results in the following, slightly different condition for dynamic completeness:
\begin{align*}
\exists \ b > 0: b (\lambda_b d_{\ell} - d_z) \geq \left(\eta + \frac{\ln(2) + B b \sqrt{d_{\delta}}}{\ln(n/\epsilon_1)}\right) c_0^{-1/3}. \tag{C''} \label{eq:C''}
\end{align*}
If some parameters $d_{\ell,B}$, $d_{z,B}$, $d_{\delta,B}$ satisfy \eqref{eq:S'} and \eqref{eq:C''}, then using these constants as our scheme parameters, we obtain a $\epsilon_1$-sound and dynamic $(\epsilon_2, c_0)$-complete scheme with a codelength of $\ell = d_{\ell,B} c_0^2 \ln(n/\epsilon_1)$. In Fig.~\ref{fig:Fig4} we show the values of $d_{\ell,B}$ for the parameters $n = 10^6$, $\epsilon_1 = \epsilon_2 = 10^{-3}$, and $\eta = \frac{1}{3}$, for several values of $B$. As the value of $B$ increases, the values of $d_{\ell,B}$ increase as well.
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{fig4}
\caption{Optimal values of $d_{\ell,B}$ in the weakly dynamic Tardos scheme from Section~\ref{sub:SemiDynamicTardos-Second}, for the parameters $n = 10^6$, $\epsilon_1 = \epsilon_2 = 10^{-3}$, and $\eta = \frac{1}{3}$. The bold curve corresponds to the values of $d_{\ell}$ in the static Tardos scheme with the same parameters, while the six curves above this curve correspond to the optimal values of $d_{\ell,B}$ for $B = 1, 2, 4, 8, 16, 32$ respectively. The dotted line corresponds to the asymptotic optimal value $d_{\ell} = \frac{\pi^2}{2} \approx 4.93$. For $B = 1$ we get exactly the codelengths of the dynamic Tardos scheme.}
\label{fig:Fig4}
\end{figure}
\subsection{Example}
\label{sub:SemiDynamicTardos-Example}
As before, let the scheme parameters be given by $c_0 = 25$, $n = 10^6$ and $\epsilon_1 = \epsilon_2 = 10^{-3}$, so that $\eta = \frac{1}{3}$, and let us use $B = 8$. With the first proposed scheme, the codelength increases by $B c_0 = 200$ symbols compared to the dynamic Tardos scheme, giving scheme parameters:
\begin{align*}
\ell = 116\,761, \quad Z = 2448, \quad \delta = 1.02 \cdot 10^{-3}.
\end{align*}
Using the second scheme, the optimal values of $d_{\ell,B}, d_{z,B}, d_{\delta,B}$ satisfying \eqref{eq:S'} and \eqref{eq:C''} for $B = 8$ can be calculated numerically as
\begin{align*}
d_{\ell,B} = 10.16, \quad d_{z,B} = 4.94, \quad d_{\delta,B} = 10.07.
\end{align*}
This leads to the scheme parameters
\begin{align*}
\ell = 131\,587, \quad Z = 2561, \quad \delta = 1.36 \cdot 10^{-3}.
\end{align*}
So in this case, using the first scheme leads to the shortest code.
\section{The universal Tardos scheme}
\label{sec:UniversalTardos}
In this section we present a dynamic scheme that does not require a sharp upper bound $c_0$ on $c$ as input to guarantee quick detection of pirates. This means that even if we set $c_0 = n$, coalitions of any size are caught quickly. We use the word ``universal'' to indicate this universality with respect to the coalition size: coalitions of any size can be caught efficiently with this scheme. Note that in the (dynamic) Tardos scheme, we used the distribution function $F_{\delta}$ where $\delta = \delta(c_0) = O(c_0^{-4/3})$ depends on $c_0$. Instead, we will use a distribution function $F$ that can be used for all values of $c$, so that we can use the same codewords to catch coalitions of any size. In particular, we will use the first $\ell^{(c)} = O(c^2 \ln(n/\epsilon_1))$ symbols to catch coalitions of size $c$, for each $c$ between $2$ and $c_0$. We do this in such a way that if a coalition has some unknown size $c$, then after $\ell^{(c)} = O(c^2 \ln(n/\epsilon_1)$ symbols, the probability of not having caught all members of this coalition is at most $\epsilon_2$. Since we do this for each value of $c$, we now only need $O(c^2 \ln(n/\epsilon_1))$ symbols to catch a coalition of a priori unknown size $c$, compared to the $O(c_0^2 \ln(n/\epsilon_1))$ worst-case codelength of the static and dynamic Tardos schemes, and the $O(c c_0 \ln(n/\epsilon_1))$ practical codelength of the dynamic Tardos scheme.
The only drawback of this new codeword generation method is that a completely universal distribution function, which is completely efficient for all values of $c$, does not seem to exist. More precisely, the proof of soundness of the Tardos scheme requires the cutoff parameter $\delta$ to be sufficiently large in terms of $c$, whereas for completeness we need that $\delta$ approaches $0$ as $c \to \infty$. Our solution to this problem is the following. For generating the values of $p_i$, we use the standard arcsine distribution function $F$ from Eq.~\eqref{dist2}, with no cutoffs. Then, for each value of $c$, we simply disregard those values $p_i$ that are not between the corresponding cutoff $\delta^{(c)}$ and $1 - \delta^{(c)}$. The fraction of values of $p_i$ that is disregarded can be estimated as follows:
\begin{align*}
1 - \int_{\delta^{(c)}}^{1 - \delta^{(c)}} \! f(p) \, \mathrm{d}p = \frac{4}{\pi} \arcsin \sqrt{\delta^{(c)}} = \frac{4 c^{-2/3}}{\pi \sqrt{d_{\delta}}} + O(c^{-2}).
\end{align*}
So the fraction of disregarded positions is very small and decreases when $c$ increases.
\subsection{Construction}
\label{sub:UniversalTardos-Construction}
The construction now basically consists of running several dynamic Tardos schemes simultaneously with shared codewords. So scheme parameters and scores now have to be calculated for each of these schemes, i.e., for each of the values of $c$. We introduce counters $t^{(c)}$ to keep track of the number of positions that have not been disregarded. For each $c$, we then run a dynamic Tardos scheme using the same code $X$ until $t^{(c)} = \ell^{(c)}$.
\begin{enumerate}
\item \textbf{Initialization phase} \\
For each $c \in \{2, \ldots, c_0 = n\}$:
\begin{enumerate}
\item Take the codelength as $\ell^{(c)} = d_{\ell}^{(c)} c^2 \ln(n/\epsilon_1^{(c)})$.
\item Take the threshold as $Z^{(c)} = d_z^{(c)} c \ln(n/\epsilon_1^{(c)})$.
\item Take the cutoff parameter as $\delta^{(c)} = 1/(d_{\delta}^{(c)} c^{4/3})$.
\item Initialize the user scores at $S_j^{(c)}(0) = 0$.
\item Initialize the counters $t^{(c)}$ at $t^{(c)}(0) = 0$.
\end{enumerate}
\item \textbf{Codeword generation} \\
For each position $i \geq 1$:
\begin{enumerate}
\item Select $p_i \in [0, 1]$ from $F(p)$ as defined in \eqref{dist2}.
\item Generate $X_{j,i} \in \{0, 1\}$ using $P(X_{j,i} = 1) = p_i$.
\end{enumerate}
\item \textbf{Distribution/Detection/Accusation} \\
For each position $i \geq 1$:
\begin{enumerate}
\item Send to each active user $j$ symbol $X_{j,i}$.
\item Detect the pirate output $y_i$. \\
(If there is no pirate output, terminate.)
\item Calculate scores $S_{j,i}$ using \eqref{scoresym}.
\item For active users $j$ and values $c$ such that $p_i \in [\delta^{(c)}, 1 - \delta^{(c)}]$, set $S_j^{(c)}(i) = S_j^{(c)}(i-1) + S_{j,i}$. \\
(Otherwise set $S_j^{(c)}(i) = S_j^{(c)}(i-1)$.)
\item For values of $c$ such that $p_i \in [\delta^{(c)}, 1 - \delta^{(c)}]$, set $t^{(c)}(i) = t^{(c)}(i-1) + 1$. \\
(Otherwise set $t^{(c)}(i) = t^{(c)}(i-1)$.)
\item Disconnect all active users $j$ with $S_j^{(c)}(i) > Z^{(c)}$ and $t^{(c)}(i) \leq \ell^{(c)}$ for some $c$.
\end{enumerate}
\end{enumerate}
As was already mentioned in Section~\ref{sub:DynamicTardos-Construction}, if desired the codeword generation can be merged with the distribution/detection/accusation phase. This depends on the scenario and the exact implementation of the scheme.
Also note that several variations can be made to the above construction, to deal with specific situations. One could easily replace $c_0 = n$ by a smaller value of $c_0$ to restrict the amount of memory needed, if a sharper upper bound on $c$ is known. And of course, we may also choose to draw values $p_i$ from $F_{\delta^{(c_0)}}$, as values $p_i \in [0,1] \setminus [\delta^{(c_0)}, 1 - \delta^{(c_0)}]$ are disregarded for all $c$.
A less obvious optimization would be to use a geometric progression of values $c$, e.g., $c \in \{2, 4, 8, 16, \ldots, c_0\}$ and maintain the user scores only for this set of coalition sizes, rather than for all values of $c \in \{2, \ldots, c_0\}$. This significantly reduces the space requirement per user from $O(c_0)$ to $O(\log c_0)$. However, if the actual coalition size is, say, $33$, then the coalition may not be caught until we reach $c = 64$. Since the codelength scales quadratically in $c$, this means that the codelength increases by a worst-case factor of $4$. In general, using any geometric progression with geometric factor $r$ possibly loses a factor $r^2$ in the codelengths. We have chosen to give the construction with many scores per user, to show that we then still obtain the same asymptotic codelengths. But the above construction is just one of the many alternatives to catch coalitions of any size efficiently.
\subsection{Soundness}
\label{sub:UniversalTardos-Soundness}
For the universal Tardos scheme we get the following result regarding soundness.
\begin{theorem} \label{thm:PDS-UT-Soundness}
Consider the universal Tardos scheme in Section~\ref{sub:UniversalTardos-Construction}. If \eqref{eq:S'} is satisfied for each set of parameters $d_z^{(c)}, d_{\ell}^{(c)}, d_{\delta}^{(c)}, \epsilon_1^{(c)}$, and if the $\epsilon_1^{(c)}$ satisfy the following requirement:
\begin{align*}
\sum_{c=2}^{c_0} \epsilon_1^{(c)} \leq \epsilon_1, \tag{E} \label{eq:E}
\end{align*}
then the scheme is $\epsilon_1$-sound.
\end{theorem}
\begin{proof}
For each $c \in \{2, \ldots, c_0\}$, let $\hat{C}^{(c)}$ be the set of users that are accused because their scores $S_j^{(c)}$ exceeded $Z^{(c)}$ before $t^{(c)} > \ell^{(c)}$. Then $\hat{C} = \bigcup_{c=2}^{c_0} \hat{C}^{(c)}$. For any $c$, we can apply Theorem~\ref{thm:PDS-DT-Soundness} to the parameters $d_{\ell}^{(c)}, d_z^{(c)}, d_{\delta}^{(c)}, a^{(c)}$ and $\epsilon_1^{(c)}$ so that we know that the probability that $j \in \hat{C}^{(c)}$ for innocent users $j$ is at most $\epsilon_1^{(c)}/n$. So the overall probability that an innocent user is disconnected is bounded from above by
\begin{align*}
P(j \in \hat{C}) \leq \sum_{c=2}^{c_0} P(j \in \hat{C}^{(c)}) \leq \sum_{c=2}^{c_0} \frac{\epsilon_1^{(c)}}{n} \leq \frac{\epsilon_1}{n}.
\end{align*}
This completes the proof.
\end{proof}
Note that one can choose values $\epsilon_1^{(c)}$ satisfying \eqref{eq:E} such that $O(c^2 \ln(n/\epsilon_1^{(c)})) = O(c^2 \ln(n/\epsilon_1))$, e.g., by taking $\epsilon_1^{(c)} = 6\epsilon_1/(\pi^2 c^2)$. If furthermore $c = n^{o(1)}$ is subpolynomial in $n$, then asymptotically $d_{\ell} c^2 \ln(n/\epsilon_1^{(c)}) = d_{\ell} c^2 \ln(n/\epsilon_1) (1 + o(1))$ and we achieve the same asymptotic codelength as in the static and dynamic Tardos schemes.
\subsection{Dynamic completeness}
\label{sub:UniversalTardos-Completeness}
The main advantage of the universal Tardos scheme is that we can now prove dynamic completeness for all values of $c$.
\begin{theorem} \label{thm:PDS-UT-Completeness}
Consider the universal Tardos scheme in Section~\ref{sub:UniversalTardos-Construction}. If \eqref{eq:C'} is satisfied for each set of parameters $d_z^{(c)}, d_{\ell}^{(c)}, d_{\delta}^{(c)}, \epsilon_1^{(c)}, \eta^{(c)}$, where $\eta^{(c)} = \ln(1/\epsilon_2)/\ln(n/\epsilon_1^{(c)})$, then for each $c \in \{2, \ldots, c_0\}$ the scheme is dynamic $(\epsilon_2, c)$-complete.
\end{theorem}
\begin{proof}
This follows directly from applying Theorem~\ref{thm:PDS-DT-Completeness} to $d_{\ell}^{(c)}, d_z^{(c)}, d_{\delta}^{(c)}, a^{(c)}$ and $\epsilon_1^{(c)}$, where $c$ is the actual (unknown) coalition size.
\end{proof}
To prove that the scheme catches a coalition of size $c$, we only argued that the coalition's score $S^{(c)}(i)$ will exceed $c Z^{(c)}$ before we have seen $\ell^{(c)}$ positions $i$ with $p_i \in [\delta^{(c)}, 1 - \delta^{(c)}]$. In reality, the probability of catching the coalition is much larger than this, since for instance with high probability the coalition score $S^{(c-1)}$ will also exceed $Z^{(c-1)}$ before we have seen $\ell^{(c-1)}$ positions with $p_i \in [\delta^{(c-1)}, 1 - \delta^{(c-1)}]$. And if a pirate is disconnected because for some $k$ his score $S_j^{(k)}$ exceeded the threshold $Z^{(k)}$, then we do not have to wait until $S(i) > c\tilde{Z}^{(c)}$ but only until $S(i) > (c-1) Z^{(c)} + Z^{(k)}$. And since $S(i)$ has a constant slope, as soon as a pirate is caught, the other pirates' scores will increase even faster. In practice we therefore also see that we usually need fewer than $\ell^{(c)}$ positions to catch $c$ colluders.
\subsection{Codelengths}
\label{sub:UniversalTardos-Codelengths}
The theoretical results from the previous subsections are not for exactly $\ell^{(c)}$ watermark positions, but for some number of symbols $T^{(c)}$ such that there are $\ell^{(c)}$ positions $i$ between $1$ and $T^{(c)}$ with $p_i \in [\delta^{(c)}, 1 - \delta^{(c)}]$. The difference $T^{(c)} - \ell^{(c)}$ is a random variable, and is distributed according to a negative binomial distribution with parameters $r = \ell^{(c)}$ (the number of successes we are waiting for) and $p = 1 - P(p_i \in [\delta^{(c)}, 1 - \delta^{(c)}]) = \frac{4}{\pi} \arcsin(\sqrt{\delta^{(c)}})$ (the probability of a success). Because the parameter $p = O(c^{-2/3})$ is very small for large $c$, the difference between $T^{(c)}$ and $\ell^{(c)}$ will also be small. More precisely, $T^{(c)}$ has mean $\ell^{(c)}/(1 - p) = \ell^{(c)}(1 + O(c^{-2/3}))$ and variance $\sigma^2 = \ell^{(c)} p / (1 - p)^2 = O(\ell^{(c)} c^{-2/3})$, and the probability that $T^{(c)}$ exceeds its mean by $m > 0$ decreases exponentially in $m$.
Also note that if some upper bound $c_0 \geq c$ is used for constructing the scheme as described earlier, and if the values of $p_i$ are drawn from $F_{\delta^{(c_0)}}$ instead of $F$, then we have $T^{(c_0)} = \ell^{(c_0)}$, as no values of $p_i$ are disregarded for $c = c_0$. So then the maximum codelength is fixed in advance, at the cost of possibly not catching coalitions of size $c > c_0$.
Finally, note that this scheme is constructed in such a way that coalitions of any (small) size can be caught more efficiently. To catch a coalition of size $c$ we now only use $O(c^2 \ln(n/\epsilon_1))$ symbols. This in comparison to the static and dynamic Tardos scheme, where we need $O(c_0^2 \ln(n/\epsilon_1))$ and $O(c c_0 \ln(n/\epsilon_1))$ symbols respectively, where $c_0$ is again some upper bound on the coalition size used to construct the schemes. So while using the dynamic Tardos scheme already reduces the codelength by a factor $\frac{c}{c_0}$, the universal Tardos scheme shaves off another factor $\frac{c}{c_0}$.
\subsection{Example}
\label{sub:UniversalTardos-Example}
As before, let the scheme parameters be given by $n = 10^6$ and $\epsilon_1 = \epsilon_2 = 10^{-3}$. Let us use $\epsilon_1^{(c)} = 6\epsilon_1/(\pi^2 c^2)$, so that $\sum_{c=2}^{c_0 = n} \epsilon_1^{(c)} \leq \epsilon_1$. Let us assume the coalition again has an actual size of $c = 25$. The optimal values of $d_{\ell}^{(25)}, d_z^{(25)}, d_{\delta}^{(25)}$ satisfying \eqref{eq:S'} and \eqref{eq:C'} can be calculated numerically as
\begin{align*}
d_{\ell}^{(25)} = 8.59, \quad d_z^{(25)} = 4.61, \quad d_{\delta}^{(25)} = 13.83.
\end{align*}
This leads to the corresponding scheme parameters
\begin{align*}
\ell^{(25)} = 148\,457, \ Z^{(25)} = 3188, \ \delta^{(25)} = 9.89 \cdot 10^{-4}.
\end{align*}
In Fig.~\ref{fig:Fig5} we show some simulation results for these parameters, where we only show the thresholds $Z^{(2)}, \ldots, Z^{(25)}$. In Fig.~\ref{fig:Fig5a} we simulated pirates using the interleaving attack, and in Fig.~\ref{fig:Fig5b} the pirates used the scapegoat strategy. As one can see, in the universal Tardos scheme the scapegoat strategy is not a good strategy, as the whole coalition is caught very soon. This is because the scapegoat strategy basically divides the coalition in $25$ coalitions of size $1$, and as mentioned before, small coalitions are caught much sooner in the universal Tardos scheme.
\begin{figure}[!t]
\centering
\subfloat[][Interleaving attack]{\includegraphics[width=\columnwidth]{fig5a} \label{fig:Fig5a}} \\
\subfloat[][Scapegoat strategy]{\includegraphics[width=\columnwidth]{fig5b} \label{fig:Fig5b}}
\caption{Simulations of the universal Tardos scheme, with parameters $c$, $c_0$, $n$, $\epsilon_1$, and $\epsilon_2$ as in Figs.~\ref{fig:Fig1} and \ref{fig:Fig3}. The black bars show the thresholds $Z^{(c)}$, for $c = 2, \ldots, 25$. For each pirate $j$ we only show the score $S_j^{(c)}(i)$ that made him get caught. In reality, all users have $25$ slightly different scores.}
\label{fig:Fig5}
\end{figure}
\section{Discussion}
\label{sec:Discussion}
Comparing the universal Tardos scheme to the static Tardos scheme, we see that the main advantages are that (a) we now have certainty about catching the whole coalition (instead of at least one pirate), and (b) we no longer need the coalition size, or a sharp upper bound on the coalition size, as input. We do need to calculate multiple scores per user, namely one for each possible coalition size $c$. But since the only disadvantage of a large $c_0$ is this larger number of scores per user and thus a larger offline space requirement (which may not be a big issue), $c_0$ can easily be much higher than the expected coalition size $c$. This in contrast to the static and dynamic Tardos schemes, where an increase in $c_0$ means an increase in the theoretical and practical codelengths as well.
In Table \ref{tab1} we list some of the differences between the static, dynamic, weakly dynamic and universal Tardos schemes. Here we assume that the upper bound $c_0$ on the number of colluders is the same for each scheme. The actual coalition size is denoted by $c$. The example referred to in the table is the example used throughout this paper, with $c = c_0 = 25$, $n = 10^6$, and $\epsilon_1 = \epsilon_2 = 10^{-3}$. The practical codelengths are based on $1000$ simulations for each scheme, where the pirates used the interleaving attack in all cases. For the weakly dynamic Tardos scheme we used $B = 8$ in our example.
\begin{table*}[!t]
\caption{A comparison of the Tardos schemes discussed in this paper. \label{tab1}}
\centering
\begin{tabular}{lccccc} \hline
& static & dynamic & weakly dynamic & weakly dynamic & universal \\
& (Section~\ref{sec:Preliminaries}) & (Section~\ref{sec:DynamicTardos}) & (Section~\ref{sub:SemiDynamicTardos-First}) & (Section~\ref{sub:SemiDynamicTardos-Second}) & (Section~\ref{sec:UniversalTardos}) \\ \hline
scores per user & $1$ & $1$ & $1$ & $1$ & $c_0 - 1$ \\
density function & $f^{(c_0)}$ & $f^{(c_0)}$ & $f^{(c_0)}$ & $f^{(c_0)}$ & $f$ \\
blocks & $1$ of size $\ell$ & $\ell$ of size $1$ & $\ell/B$ of size $B$ & $\ell/B$ of size $B$ & $\ell$ of size $1$ \\
guilty caught & at least $1$ & all $c$ & all $c$ & all $c$ & all $c$ \\
expected codelength & $O(c_0^2 \ln(n/\epsilon_1))$ & $O(c c_0 \ln(n/\epsilon_1))$ & $O(c c_0 \ln(n/\epsilon_1))$ & $O(c c_0 \ln(n/\epsilon_1))$ & $O(c^2 \ln(n/\epsilon_1))$ \\
asymptotic codelength & $\frac{\pi^2}{2} c^2 \ln(n/\epsilon_1)$ & $\frac{\pi^2}{2} c^2 \ln(n/\epsilon_1)$ & $\frac{\pi^2}{2} c^2 \ln(n/\epsilon_1)$ & $\frac{\pi^2}{2} c^2 \ln(n/\epsilon_1)$ & $\frac{\pi^2}{2} c^2 \ln(n/\epsilon_1)$ \\
example, theoretical codelength & $109\,585$ & $116\,561$ & $116\,761$ & $131\,587$ & $148\,457$ \\
example, practical codelength & $109\,585$ & $92\,000$ & $92\,000$ & $96\,000$ & $89\,000$ \\ \hline
\end{tabular}
\end{table*}
Since our schemes are dynamic traitor tracing schemes, it makes sense to also compare them to other dynamic schemes from the literature. Recall from Section~\ref{sub:Introduction-RelatedWork} that the scheme of Fiat and Tassa~\cite{fiat01}, the schemes of Berkman et al.~\cite{berkman01} and the scheme of Roelse~\cite{roelse11} are deterministic schemes. That is, each of these schemes always catches all pirates and no user is ever falsely accused, which are advantages compared to probabilistic schemes such as our schemes. An additional advantage of these schemes is that they have very short codelengths. On the other hand, it was shown by Fiat and Tassa~\cite{fiat01} that $q \geq c+1$ for any deterministic scheme, so these schemes cannot be used in scenarios in which a small alphabet size is required.
As is the case with our schemes, the dynamic scheme of Tassa~\cite{tassa05} is probabilistic and uses a binary alphabet (i.e., $q = 2$). The codelengths of these schemes can therefore be compared directly. In particular, the codelength of the scheme of Tassa is $\Theta(c^4 \log_2(n) \ln(n/\epsilon_1))$, which is more than a factor $\Theta(c^2)$ larger than the codelengths of our schemes. In fact, to the best of our knowledge our schemes have the shortest order codelengths of all known binary dynamic traitor tracing schemes.
Below we list some other nice properties of the universal Tardos scheme, which are not related to the codelength or the alphabet size. Most of these properties are inherited from the static Tardos scheme.
\paragraph{Codewords of users are independent.} This means that framing a specific innocent user is basically impossible, as the codewords of the pirates and the pirate output are independent of the innocent users' codewords. Also, a new user can be added to the system easily after the codewords of other users have already been generated, since the codewords of other users do not have to be updated.
\paragraph{Codeword positions are independent.} In other words, the scheme does not need the information obtained from the previous pirate output to generate new symbols for each user. Therefore the codewords can even be generated in advance. This also allows us to effectively tackle weakly dynamic traitor tracing scenarios, as described in Section~\ref{sec:SemiDynamicTardos}.
In particular, the total tracing times of the dynamic schemes presented in \cite{berkman01, fiat01, roelse11, tassa05} are bounded from below by the total delay, defined as the codelength of the scheme times the delay of the pirates' transmission. By comparison, the total tracing times of our weakly dynamic schemes only increase marginally if $B$ increases. As a result, for a large delay (i.e. for a large value of $B$), our weakly dynamic schemes have the shortest total tracing times of all known dynamic schemes.
\paragraph{The distribution of watermark symbols is identical for each position.} This property offers new options, like tracing several coalitions simultaneously, using the same traitor tracing code. This also means that multiple watermarks from several broadcasts can be concatenated and viewed as one long watermark from one longer broadcast, allowing one to catch large coalitions with multiple watermarked broadcasts.
\paragraph{The codeword generation and accusation algorithm are computationally and memory-wise efficient.} The schemes do not require any complicated data structures and computations, and the only memory needed during the broadcast is the scores for each user at that time, and the counters $t^{(c)}$. During the broadcast only simple calculations are needed: computing $S_{j,i}$ (which has to be calculated only once), adding $S_{j,i}$ to those scores $S_j^{(c)}$ where $c$ satisfies a certain condition, and comparing the scores $S_j^{(c)}$ to the thresholds $Z^{(c)}$.
\paragraph{Several instances of the scheme can be run simultaneously.} For example, by using parameters $\{\epsilon_1^{(c)}\}$ with $\sum \epsilon_1^{(c)} \leq 0.01$ and $\{\bar{\epsilon}_1^{(c)}\}$ with $\sum \bar{\epsilon}_1^{(c)} \leq 0.05$ for two different instances of the universal Tardos scheme (using the same codewords), a pirate will first cross one of the thresholds associated to $\{\bar{\epsilon}_1\}$, and only later cross one of the thresholds associated to the $\{\epsilon_1\}$. If we use the $\{\epsilon_1\}$ for disconnecting users, then even before a user is disconnected, we can give some sort of statistic to indicate the `suspiciousness' of this user. If a user then does not cross the highest thresholds, one could still decide whether to disconnect him or not. After all, the choice of $\epsilon_1$ may be arbitrary, and a user that almost crosses the thresholds $Z^{(c)}$ is likely to be guilty as well.
\section{Open problems}
\label{sec:OpenProblems}
Let us conclude with mentioning some open problems for future research.
\subsection{A single-score universal Tardos scheme}
\label{sub:OpenProblems-SingleScore}
Although we argued that the universal Tardos scheme has several advantages over other binary schemes, it has a minor drawback: we have to keep multiple scores for each user, namely for each possible coalition size $c$. To address this issue, one could try making small adjustments to the universal Tardos scheme, or start from the dynamic Tardos scheme and build a different, $c_0$-independent traitor tracing scheme. For instance, would it be possible to change the process of generating the $p_i$'s such that no positions are ever disregarded? Then all scores for one user would be the same, and we would only have to keep one score for each user.
\subsection{A continuous universal Tardos scheme}
\label{sub:OpenProblems-Continuous}
Looking at Fig.~\ref{fig:Fig5} suggests that a continuous threshold function $Z(i)$ might also be an option, with $Z$ depending on the position $i$ instead of on the coalition size $c$. However, for the proof of soundness of the universal Tardos scheme, we simply added up the error probabilities for each threshold and showed that this sum is still less than $\epsilon_1$. If we use a continuous function $Z(i)$ and use this same proof method, this would lead to even smaller values of $\epsilon^{(i)}$ and longer codelengths. Still, theoretically it would be interesting to see if such a continuous threshold function can be constructed.
\subsection{A fully dynamic Tardos scheme}
\label{sub:OpenProblems-FullyDynamic}
Most dynamic schemes find their strength in being able to adjust the next codeword symbols to the previous pirate output. In the dynamic Tardos scheme, we do not use this ability at all, and only use the dynamic setting to disconnect users inbetween. It is an open problem whether better results can be obtained with a fully dynamic Tardos scheme, that does use this extra power given to the distributor.
\subsection{A weakly dynamic deterministic scheme}
\label{sub:OpenProblems-WeaklyDynamic}
The deterministic dynamic schemes in \cite{berkman01, fiat01, roelse11, tassa05} are not designed for the weakly dynamic setting, and it is not obvious how to adapt these schemes to this setting. The design and analysis of efficient weakly dynamic deterministic schemes is therefore an open problem.
\subsection{The dynamic traitor tracing capacity}
\label{sub:OpenProblems-Capacity}
On the other hand, it is also very well possible that no fully dynamic Tardos scheme exists that achieves significantly better codelengths. For the static setting, it is known that the order codelength of the Tardos scheme (quadratic in $c_0$, logarithmic in $n$) is optimal. But what about the dynamic setting? What is the optimal order codelength required to catch all colluders? Our results show that the optimal order codelength is at most quadratic in $c$, but this may not be optimal.
\subsection{A $q$-ary dynamic Tardos scheme}
\label{sub:OpenProblems-qary}
In this paper we discussed several probabilistic dynamic schemes, taking the static binary Tardos scheme and the results of Laarhoven and De Weger~\cite{laarhoven11} as starting points. The design and analysis of $q$-ary probabilistic dynamic traitor tracing schemes is still an open problem. A possible approach for solving this problem is to take the $q$-ary Tardos scheme of \v{S}kori\'{c} et al.~\cite{skoric08} as a starting point.
In a recent paper, Laarhoven et al.~\cite{laarhoven12} presented another approach to solve this problem. It was shown that with a divide-and-conquer construction, any binary dynamic traitor tracing scheme can be turned into a $q$-ary dynamic traitor tracing scheme with a codelength that is roughly a factor $q/2$ smaller than the codelength of the underlying binary scheme. Applying this to the constructions described in this paper, this leads to $q$-ary dynamic Tardos schemes with codelengths of the order $\ell_q = O\left(\frac{c^2}{q} \ln \frac{n}{\epsilon_1}\right)$. Moreover, for fixed $q$ and large $c$, this leads to an asymptotic codelength of $\ell_q \to \frac{\pi^2}{q} c^2 \ln \frac{n}{\epsilon_1}$, compared to the $\ell_2 \to \frac{\pi^2}{2} c^2 \ln \frac{n}{\epsilon_1}$ of the binary schemes presented in this paper. For details, see \cite{laarhoven12}.
\section*{Acknowledgments}
The authors are grateful to the anonymous reviewers for their valuable comments.
|
\section{Introduction}
\label{Introduction}
Can the quantum-mechanical interference pattern in a double-slit experiment be described by a suitable time evolution of a classical statistical ensemble? Can this be extended to the quantum-interference of multi-fermion systems? Can the evolution law for the classical probabilities be causal and local in the sense that the probability distribution at time $t+\epsilon$ can be computed from the one at time $t$, and that the evolution of local properties of the distribution is only influenced by the distribution in a local neighborhood? We answer these questions with a clear ``yes''. While a general discussion how ``no-go-theorems'' as Bell's inequalities \cite{Bell,BS} or the Kochen-Specker theorem \cite{KS} can be circumvented can be found in ref. \cite{GR}, we concentrate in this paper on the construction of a classical statistical model that can describe real physical situations.
We will investigate an Ising-type model for discrete occupation numbers or Ising-spins on the sites of a lattice. Intuitively, if a particle is present on a lattice site $x$ the Ising-spin is up, and if no particle is present it is down. Since the associated occupation numbers only take the values one or zero, one sees already the close analogy to occupation numbers of fermionic multi-particle systems. The notion of presence or absence of particles can be extended to more general ground states, as a half-filled state. Now a flip of the Ising-spin at $x$ from down to up corresponds to a change from a situation with no particle at $x$ to one with a particle at $x$. More precisely, we will consider four or eight ``species'' of Ising-spins. These species correspond to the real degrees of freedom for Majorana or Dirac spinors.
The state of the system or the classical statistical ensemble is characterized by the probability distribution for the possible configurations of Ising-spins. For the formulation of dynamics one has to postulate an evolution law which describes how this probability distribution changes in time. For the Ising-spins there is no notion of underlying continuous and deterministic classical dynamics for ``trajectories'' -each Ising-spin can only take two values. A differential evolution equation has therefore to be formulated on the level of probability distributions. This constitutes the fundamental law defining the dynamics. (It is the analogue of the Liouville equation for a statistical ensemble of Newtonian classical particles.) The only basic restrictions on the form of this law are that all probabilities have to remain positive and the sum of all probabilities must equal one for all times $t$. We propose a specific causal and local evolution law. With this evolution law our Ising-type model describes a quantum field theory for Dirac fermions in an arbitrary external electromagnetic field. The Schr\"odinger equation for a quantum particle in a potential follows for one-particle states in the non-relativistic approximation.
With our proposed evolution equation for a classical statistical ensemble of Ising-spins, the Schr\"odinger equation and all associated quantum phenomena as interference, tunneling, the uncertainty relation and discrete energy levels for stationary states, are implemented within classical statistics. We follow here the concept of ``probabilistic realism'' that there is one reality, but the fundamental laws are formulated as probabilistic laws. It seems not entirely excluded that our evolution law for the probability distribution could be produced by cellular automata \cite{TH}. This seems, however, rather difficult, and it is not needed for our purpose.
Time evolution laws for classical statistical ensembles that account for the full dynamics of quantum particles have been found previously for more restricted settings. For a single particle in a potential a suitable evolution equation for the probability density in phase space has been proposed in ref. \cite{CWP}. It leads to the Schr\"odinger equation after ``coarse graining''. This ``quantum evolution equation'' modifies Liouville's equation. In ref. \cite{CWP} the different behavior of classical and quantum particles is traced back to the specific form of the evolution equation. A continuous interpolation between the ``classical'' and ``quantum'' evolution law can describe ``zwitters'' - particles whose properties interpolate continuously between classical and quantum particles. This earlier approach constitutes already a powerful demonstration that there is no difference in principle between classical statistics and quantum statistics since a continuous interpolation between both is possible. It also allows the formulation of consistent theories that are arbitrarily close to quantum mechanics, with a small parameter accounting for deviations that can be restricted by experiment. However, the generalization of this setting to a multi-particle situation is cumbersome. It is much simpler to base a realistic model directly on the analogy between Ising-spins and occupation numbers for fermions, as done in the present paper. Furthermore, the setting of ref. \cite{CWP} does not account explicitly for the discrete particle properties, while in the present paper they can be associated directly to the discrete Ising-spins.
In ref. \cite{CWF}, \cite{CWES}, \cite{CWMSa} the map between Ising-type models and quantum fermions has been developed. In particular, ref. \cite{CWMSa} has already constructed a classical statistical ensemble for massless Majorana spinors in four dimensions. The present paper builds on these findings. It extends this formalism to include a mass term, to formulate Dirac spinors and to add the coupling to external electromagnetic fields. These steps are needed for a description of realistic situations. The discussion of one-particle states can now describe electrons in a potential. For a non-relativistic approximation we can derive the Schr\"odinger equation from the proposed evolution law for the classical statistical ensemble of Ising-spins.
Our Ising-type model is formulated for a lattice of discrete points in space and we use discrete time steps. (The lattice formulation of the present paper differs in some aspects from ref. \cite{CWMSa}.) As long as the number of lattice points (in space) remains finite all mathematical operations are defined unambiguously and the model is fully regularized. The map between classical statistics for the Ising-type model and quantum fermions can be done on this level where no ambiguities are present. At the very end one may perform the continuum limits of vanishing lattice distance and vanishing size for the time steps. We formulate the model in a way such that the unitary time evolution is already realized for discrete time steps and a finite number of lattice points in space. This implies certain restrictions on the precise lattice formulation for which more details will be discussed in a separate paper.
Our paper is organized as follows: In sect. \ref{Probability distribution and wave function} we formulate the classical statistical ensemble for Ising-spins and we introduce the important concept of the ``classical wave function''. Up to signs this is the square root of the probability distribution. Time evolution laws preserving the positivity and normalization of the probability distribution can be most easily formulated in terms of the classical wave function. Since the normalization condition states that the classical wave function is a vector with unit length, we can formulate the normalization preserving evolution equations as rotations of a vector. We also employ the simple map between the classical wave function and a Grassmann wave function. It has been shown in refs. \cite{CWF}, \cite{CWMSa} that a linear time evolution of a Grassmann wave function can be mapped to a Grassmann functional integral and vice versa. This permits us to formulate the evolution law for the classical wave function and probability distribution directly in terms of the action of a Grassmann functional integral. This possibility is of great help for the implementation of the symmetries preserved by the evolution as, for example, the Lorentz symmetry for relativistic particles.
In sect. \ref{Massless Majorana fermions} we present the lattice action for massless Majorana fermions in four dimensions. Since the classical wave function is a real object, it should be an element of a real Grassmann algebra. The continuum limit of the lattice action realizes Lorentz-symmetry. In sect. \ref{Grassmann wave function} we extract the time evolution of the Grassmann wave function from the real Grassmann functional integral defined in sect. \ref{Massless Majorana fermions}. It obtains by ``integrating'' out the Grassmann variables at past (or future) times. Sect. \ref{Timeevolution} shows that the time evolution is unitary. This holds already for discrete time steps and can therefore be extended in a straightforward way to the continuum limit of infinitesimal time steps. Sect. \ref{Timeevolution} establishes the concrete evolution law for the classical wave function and therefore the probability distribution of the classical ensemble of Ising-spins. In sect. \ref{Observables} we briefly recapitulate how expectation values of observables in the classical statistical ensemble can be computed from the Grassmann wave function.
In sect. \ref{Particlestates} we turn to the notion of statistical states with a fixed number of fermions. We concentrate on the one particle state and show that it describes the propagation of a massless Majorana or Weyl fermion. This can be generalized to multi-fermion states. The equivalence between Majorana and Weyl spinors in four dimensions gives a first glance on the emergence of a complex structure which is discussed in sect. \ref{Complex structure}. The presence of a complex structure is characteristic for quantum mechanics where the ``physics of phases'' plays a crucial role. While single Majorana spinors are described by a real four-component quantum wave function, Weyl spinors are formulated in terms of an equivalent complex two-component quantum wave function. For Dirac spinors a different complex structure will be used and we therefore resume in sect. \ref{Complex structure} briefly the general features of complex structures in our setting of a real Grassmann functional integral and a real classical wave function. In sect. \ref{Massive Majorana Fermions} we finally add a mass term. We present some explicit instructive examples for the classical probability distributions that represent the propagation of massive Majorana spinors.
In sect. \ref{Diracfermions} we turn to Dirac spinors. They are simply represented as two Majorana spinors. The electromagnetic gauge transformations rotate the two Majorana spinors into each other. The coupling to an external electromagnetic field involves therefore both types of Majorana spinors. We discuss the complex structure, the usual Weyl representation of Dirac spinors and the symmetries of the model. From the Grassmann functional integral we extract again the evolution equation for the Grassmann wave function. The associated evolution equation for the probability distribution of the classical statistical ensemble for Ising-spins describes the dynamics of an arbitrary number of electrons and positrons in an external electromagnetic field. Sect. \ref{Quantum mechanics for particle in a potential} discusses the one-particle states. The classical eight-component real wave function obeys the Dirac equation in an external electromagnetic field. The complex representation and the non-relativistic limit are straightforward. This results in the Schr\"odinger equation for a particle in a potential. All characteristic features of quantum mechanics, as interference, tunneling or discrete energy levels can therefore be derived from our evolution equation for a classical statistical ensemble of Ising-spins. This is one more clear concrete example that quantum physics can be described by classical statistics! Even though this is not the main purpose of the present paper, we comment how the discrete particle properties in quantum mechanics are related to the discrete Ising-spins, and how the continuity of the wave function simply reflects the continuity of the classical probability distribution. Particle-wave duality appears as a simple consequence of our classical statistical setting. Our conclusions are presented in sect. \ref{Conclusions}.
\section{Probability distribution and wave function}
\label{Probability distribution and wave function}
A classical statistical ensemble is specified by its states $\tau$ and a probability distribution $\{p_\tau\}$, which associates to every $\tau$ a positive probability $p_\tau\geq 0$. The distribution is normalized $\Sigma_\tau p_\tau=1$. Classical observables take in every state $\tau$ a fixed value $A_\tau$, and the mean value obeys $\langle A\rangle=\Sigma_\tau A_\tau p_\tau$.
\bigskip\noindent
{\bf 1. Generalized Ising model}
\medskip
We will consider an Ising-model type system for discrete Ising-spins $s_\gamma(\vec x)=\pm 1$, with $\vec x$ points on a suitable three-dimensional lattice and $\gamma$ denoting different ``species'' of Ising-spins, $\gamma=1\dots N_s$. For $N_s=4$ our system will be equivalent to a quantum field theory for Majorana or Weyl spinors, while for $N_s=8$ we will describe Dirac spinors. The states $\tau$ are sequences or configurations of Ising-spins, $\tau=\{s_\gamma(\vec x)\}$. Instead of Ising-spins, we will actually use occupation numbers or bits $n_\gamma(\vec x)=\big(s_\gamma(\vec x)+1\big)/2$, such that the states $\tau$ describe bit sequences of numbers $n_\gamma(\vec x)=0,1~,~\tau=\big\{n_\gamma(\vec x)\big\}$. We will consider a setting with $L^3/8$ lattice points ($L$ even) such that the number of states is $2^{(N_sL^3/8)}$. The continuum limit $L\to\infty$ is taken at the end.
The sequences or configurations of occupation numbers $\big\{n_\gamma(\vec x)\big\}$ show already a strong analogy to the basis states of a quantum theory for an arbitrary number of fermions, formulated in the occupation number basis for positions. We will exploit this analogy in order to formulate a fundamental ``evolution law'' for the time evolution of the probability distribution $p_\tau(t)$, such that our system describes a relativistic quantum field theory for fermions.
We emphasize that the form of the evolution law is not known a priori - we do not introduce any underlying deterministic theory of ``classical trajectories'' or similar concepts. We note that no continuous time evolution for individual Ising-spins $s_\gamma(x)$ can be formulated, since $s_\gamma(x)$ only admits discrete values $\pm 1$. On the other hand, the time evolution of the probability density may well be continuous. We postulate that the basic dynamics describes the time evolution of probabilities. The corresponding dynamical law must be specified in order to define the model. The only constraint to be imposed a priori is that is respects for all $t$ the positivity and normalization of the classical probability distribution
\begin{equation}\label{N11}
p_\tau(t)\geq 0~,~\sum_\tau p_\tau(t)=1.
\end{equation}
\bigskip\noindent
{\bf 2. Classical wave function}
\medskip
A useful concept for the construction of a consistent time evolution law is the classical wave function \cite{CWP},
\begin{equation}\label{35A}
q_\tau(t)=s_\tau(t)\sqrt{p_\tau(t)}~,~p_\tau(t)=q^2_\tau(t)~,~s_\tau(t)=\pm 1.
\end{equation}
This is a real function which is given by the square roots of the probabilities up to signs $s_\tau(t)$. Consistent evolution laws correspond to rotations of the vector $q_\tau(t)$,
\begin{eqnarray}\label{35B}
q_\tau(t)=\sum_\rho R_{\tau\rho}(t,t')q_\rho(t'),\nonumber\\
\sum_\rho R_{\tau\rho}(t,t')R_{\sigma\rho}(t,t')=\delta_{\tau\sigma}.
\end{eqnarray}
The positivity of $p_\tau=q^2_\tau$ is automatic, and the normalization of the distribution remains preserved provided it was normalized for some initial time $t_{in}$, since the length of a vector or $\Sigma_\tau q^2_\tau=\Sigma_\tau p_\tau$ is preserved by rotations. We will specify a linear evolution for the wave function. In other words, we will consider rotation matrices $R_{\tau\rho}$ that are independent of the wave function.
The classical wave function specifies the probability distribution uniquely. The specification of an evolution law for the wave function therefore defines the dynamics of the classical statistical system completely. In the other direction, different choices of the sign functions $s_\tau(t)$ correspond to a choice of gauge. For all expectation values and correlations that can be computed from the probability distribution $\big\{p_\tau(t)\big\}$ the gauge choice does not matter. Natural gauge choices preserve the continuity and differentiability properties of the wave function by avoiding arbitrary discrete jumps which would be induced by arbitrary jumps in the sign functions \cite{CWP}. In contrast to the discussion of classical mechanics in a Hilbert space by Koopman and von Neumann \cite{Kop}, the classical wave function is real, such that at this step no phases appear as new degrees of freedom beyond the probability distribution.
Quantum wave functions are usually defined in a complex Hilbert space. Indeed, many characteristic quantum features are closely associated to the ``physics of phases''. In our setting we will define a complex quantum wave function by introducing a complex structure in the real space spanned by $\{q_\tau\}$. The $2^{(N_sL^3/8)}$ real components of the vector $\{q_\tau\}$ can then be associated to a complex vector with dimension $2^{(N_sL^3/8)-1}$. One complex structure can be closely related to the equivalence between Majorana and Weyl spinors in four dimensions \cite{CWMS}, \cite{CWMSab} \cite{CWMSa}, but more general complex structures are possible. There is no conceptual difference between the classical and quantum wave function in our setting.
\bigskip\noindent
{\bf 3. Grassmann wave function}
\medskip
We will specify the time evolution of the wave function $\big\{q_\tau(t)\big\}$ within a formalism based on a real Grassmann algebra. This will make the close connection to fermions most transparent \cite{CWF}. The Grassmann formulation relies on the isomorphism between states $\tau$ and the basis elements $g_\tau$ of a Grassmann algebra that can be constructed from the Grassmann variables $\psi_\gamma(x)$. Each basis element $g_\tau$ is a product of Grassmann variables
\begin{equation}\label{N12}
g_\tau=\psi_{\gamma_1}(x_1)\psi_{\gamma_2}(x_2)\dots
\end{equation}
which is ordered in some convenient way. To be specific, we define some linear ordering of the lattice points and place variables with ``smaller $x$'' to the left, and for each $x$ place smaller $\gamma$ to the left. If a Grassmann element $g_\tau$ contains a given variable $\psi_\gamma(x)$ we put the number $n_\gamma(x)$ in the sequence $\tau$ to $0$, while we take $n_\gamma(x)=1$ if the variable $\psi_\gamma(x)$ does not appear in the product \eqref{N12}. This specifies the map between the $2^{(N_sL^3/8)}$ independent basis elements $g_\tau$ of the Grassmann algebra and the states $\tau$.
An arbitrary element $g$ of the Grassmann element can be expanded in terms of the basis elements
\begin{equation}\label{YA}
g=\sum_\tau q_\tau g_\tau.
\end{equation}
A time dependent wave function $\big\{q_\tau(t)\big\}$ can therefore be associated to a time dependent element of the Grassmann algebra $g(t)$, provided the coefficients $q_\tau(t)$ are real and normalized according to $\sum_\tau q^2_\tau(t)=1$. We will formulate the fundamental evolution law as an evolution law for the ``Grassmann wave function'' $g(t)$. For this purpose we will formulate in the next section a Grassmann functional integral for a quantum field theory of massless Majorana or Weyl fermions. It will involve $N_sL^3/8$ Grassmann variables $\psi_\gamma(t,x)$ for every discrete time point $t$. In section \ref{Grassmann wave function} we will present a prescription how the Grassmann wave function $g(t)$ and the fundamental evolution law can be extracted from this functional integral. This will establish a map between a quantum field theory for fermions and Ising type classical statistical ensembles with dynamics specified by an appropriate evolution law.
\section{Massless Majorana fermions}
\label{Massless Majorana fermions}
{\bf 1. Action}
\medskip
In this section we formulate the quantum field theory of a free Majorana spinor for a discrete space lattice on a torus. It will be defined by a Grassmann functional integral based on a real Grassmann algebra, and we take $N_s=4$. Due to the finite number of $(L/2)^3$ space points we have at any given $t$ a finite number $L^3/2$ of Grassmann variables $\psi_\gamma(t,\vec x)$, with four ``species'' $\gamma=1\dots 4$. (No complex conjugate Grassmann variables are defined at this stage.) For the associated classical statistical ensemble the states correspond to sequences of $L^3/2$ occupation numbers $n_\gamma(\vec x)$ that can take values $n_\gamma(\vec x)=0,1$. The total number of classical states equals $2^{L^3/2}$ and remains finite for finite $L$.
We also will take time $t$ on a finite discrete chain. The functional integral will therefore involve a finite number of Grassmann variables. It is fully regularized and all quantities are well defined. The continuum limit $L\to\infty$ of an infinite number of Grassmann variables will only be taken at the end.
We start with the action for the regularized quantum field theory
\begin{equation}\label{N1}
S=\sum^{t_f-\epsilon}_{t=t_{in}}L(t),
\end{equation}
with Lagrangian
\begin{equation}\label{A17}
L(t)=\sum_x\psi_\gamma(t,x)B_\gamma(t+\epsilon,x).
\end{equation}
Here we define
\begin{equation}\label{7A}
B_\gamma(t+\epsilon,x)=\sum_{\{v\}}Y_{\gamma\delta}\big(\{v\}\big)\psi_\delta(t+\epsilon,x_k+v_k\Delta),
\end{equation}
with
\begin{equation}\label{7B}
\sum_{\{v\}}=\prod_j\Big(\sum_{v_j=\pm 1}\Big)
\end{equation}
and
\begin{equation}\label{B17}
Y_{\gamma\delta}\big(\{v\})=\frac18
\Big[1-\sum_k(v_k+w_k\tilde I)T_k-v_1v_2v_3\tilde I\Big]_{\gamma\delta}.
\end{equation}
Thus $B_\gamma(t+\epsilon,x)$ involves a linear combination of Grassmann variables at lattice sites which are diagonal neighbors of $x$, corresponding to corners of a cube with basis length $2\Delta$, with $x$ being at its center. The sums extend over $j,k,l=1\dots 3$ and each corner corresponds to a particular combination of the three signs $v_j=\pm 1$. We define $w_k$ by $w_1=-v_2v_3,w_2=v_1v_3$, $w_3=-v_1v_2$. Eqs. \eqref{A17}, \eqref{B17} and the following imply a summation over repeated indices $\gamma,\delta=1\dots 4$.
We have also introduced the three real symmetric $4\times 4$ matrices $T_k$ as
\begin{equation}\label{18AA}
T_1={0,1\choose 1,0}~,~T_2={~~0,c\choose -c,0}~,~
T_3={1,~~0\choose 0,-1}~,~T^T_k=T_k,
\end{equation}
with
\begin{equation}\label{18AB}
c={~~0,1\choose -1,0}=i\tau_2,
\end{equation}
and $1$ stands for the unit $2\times 2$ matrix. The product of these matrices yields the real antisymmetric $4\times 4$ matrix
\begin{equation}\label{18ABa}
\tilde I=-{c,0\choose~ 0,c}=T_1T_2T_3~,~\tilde I^T=-\tilde I.
\end{equation}
The sum in eq. \eqref{N1} extends over discrete time points $t_n$, with $\int_t=\epsilon\sum_t=\epsilon\sum_n~,~t_{n+1}-t_n=\epsilon,n\in{\mathbbm Z}$, $t_{in}\leq t_n\leq t_f$. The time-continuum limit is taken as $\epsilon\to 0$ for fixed $t_{in},t_f$. Similarly, we sum in eq. \eqref{A17} over points $x$ of a lattice. For this purpose we consider for every given $t$ a cubic lattice with lattice distance $2\Delta$ and $\int_x=8\Delta^3\sum_x$. We take $\epsilon =\Delta$ and place the space-time lattice points on a hypercubic bcc lattice with distance $2\Delta$ between nearest neighbors, which we call the ``fundamental lattice''. For even $t=2n\epsilon$ the space lattice points are even, $x_k=2m'_k\Delta$, with $n,m'_k\in{\mathbbm Z}$. This will be called the even sublattice. The odd sublattice consists of the odd time points $t=(2n+1)\epsilon$ for which the space points are also odd, $x_k=(2m'_k+1)\Delta$. The action \eqref{N1} involves indeed only Grassmann variables for points on the fundamental lattice. For even $t$ eq. \eqref{A17} involves Grassmann variables $\psi_\gamma(t,x)$ living on the even sublattice, while the combination $B_\gamma(t+\epsilon,x)$ lives on the odd sublattice. For odd $t$ the role of the sublattices is exchanged, now with $\psi$ on the odd and $B$ on the even sublattice. This construction eliminates ``lattice doublers'' - a more detailed discussion of the lattice implementation will be given elsewhere.
The action \eqref{N1} is an element of a real Grassmann algebra - all coefficients multiplying the Grassmann variables $\psi_\gamma(t,x)$ are real. Within the Grassmann algebra the operation of transposition amounts to a total reordering of all Grassmann variables. The action \eqref{N1} is antisymmetric under this operation,
\begin{equation}\label{F3}
S^T=-S.
\end{equation}
If we define formally the Minkowski action $S_M=iS$ the latter is hermitean, $S_M=S^\dagger_M$ since $S^*_M=-S_M$.
As an important ingredient for the probabilistic interpretation and time evolution discussed in the next two sections we observe that one can obtain $B_\gamma$ from $\psi_\gamma$ by a rotation
\begin{eqnarray}\label{N8}
&&B_\gamma(x)=\sum_y\bar R_{\gamma\delta}(x,y)\psi_\delta(y),\nonumber\\
&&\sum_y\bar R_{\eta\delta}(z,y)
\bar R_{\gamma\delta}(x,y)=\delta_{\eta\gamma}\delta(z,x).
\end{eqnarray}
In eq. \eqref{N8} $B_\gamma$ and $\psi_\delta$ refer to the same time arguments, but we note that $x$ is not a point on the space-lattice on which the Grassmann variables $\psi_\delta(y)$ live. The quantity $B_\gamma(x)$ is a linear combination of Grassmann variables living on the corners of a cube with center at $x$. Since the number of lattice points $y$ and the number of centers of cubes $x$ is the same, we may nevertheless interpret $\bar R_{\gamma\delta}(x,y)$ as a quadratic matrix. We may formally extend the space to objects living on a cubic lattice with lattice distance $\Delta$ for each $t$. In this extended space $\bar R$ acts as a rotation matrix despite the fact that $x$ and $y$ refer to points on different sublattices of the fundamental lattice. The second equation \eqref{N8} and the following equations should be interpreted in this sense.
In order to show that $\bar R$ is on orthogonal matrix we write it as a product
\begin{equation}\label{A25}
\bar R=\tilde R_1\tilde R_2\tilde R_3,
\end{equation}
with
\begin{equation}\label{B25}
\tilde R_k=D^+_k-T_k D^-_k.
\end{equation}
Here the shift operators
\begin{equation}\label{C25}
D^\pm_k(x,y)=\frac12\big[\delta(x,y-\Delta_k)\pm\delta
(x,y+\Delta_k)\big]
\end{equation}
act as
\begin{equation}\label{D25}
\sum_yD^\pm_k(x,y)\psi(y)=\frac12\big[\psi(x+\Delta_k)\pm\psi(x-\Delta_k)\big].
\end{equation}
They obey
\begin{eqnarray}\label{22A}
\sum_y D^\pm_k(x,y)D^\pm_k(y,z)&=&\\
\frac14\big[\delta(x,z-2\Delta_k)
&+&\delta(x,z+2\Delta_k)\pm 2\delta(x,y)\big],\nonumber
\end{eqnarray}
and
\begin{eqnarray}\label{E25}
&&\sum_y D^+_k(x,y)D_k^-(y,z)=\sum_yD^-_k(x,y)D^+_k(y,z)=\nonumber\\
&&\quad\frac14\big[\delta(x,z-2\Delta_k)-\delta(x,z+2\Delta_k)\big].
\end{eqnarray}
With
\begin{equation}\label{F25}
(D^\pm_k)^T=\pm D^\pm_k~,~\tilde R^T_k=D^+_k+T_kD^-_k,
\end{equation}
one finds indeed $\tilde R^T_k\tilde R_k=1$. Observing that all shift operators $D^\pm_k,D^\pm_l$ mutually commute, one easily verifies that $B$ in eq. \eqref{B17} indeed obeys
\begin{equation}\label{20A}
B=\tilde R_1\tilde R_2\tilde R_3\psi.
\end{equation}
We next discuss the continuum limits in space and time for the action \eqref{N1}-\eqref{B17}. The space continuum limit $\Delta\to 0$ is characterized by
\begin{equation}\label{G25}
(D^+_k\psi)(x)\to\psi(x)~,~(D^-_k\psi)(x)\to\Delta\partial_k\psi(x),
\end{equation}
with $\partial_k=\partial/\partial x_k$ the continuum derivative. In the continuum limit one finds
\begin{equation}\label{H25}
\tilde R_k\to 1-\Delta T_k\partial_k
\end{equation}
and
\begin{equation}\label{I25}
B_\gamma(x)=\psi_\gamma(x)-\Delta\sum_k(T_k)_{\gamma\delta}
\partial_k\psi_\delta(x)+0(\Delta^2).
\end{equation}
We further employ
\begin{equation}\label{B3}
\partial_t\psi(t)=\frac1\epsilon\big[\psi(t+\epsilon)-\psi(t)\big],
\end{equation}
such that the Grassmann property $\psi^2_\gamma(t,x)=0$ results in
\begin{equation}\label{N4}
\psi_\gamma(t,x)\partial_t\psi_\gamma(t,x)=\frac1\epsilon ~ \psi_\gamma(t,x)\psi_\gamma(t+\epsilon,x).
\end{equation}
For $\Delta=\epsilon$ we obtain the continuum relation
\begin{equation}\label{J25}
\psi(t,x)B(t+\epsilon,x)=\epsilon(\psi\partial_t\psi-\sum_k\psi T_k\partial_k\psi).
\end{equation}
The factor of $\epsilon$ combines with $\Sigma_t$ into $\int_t=\epsilon\Sigma_t$. Finally, the factor $8\Delta^3$ in the relation between $\int_x$ and $\sum_x$ is absorbed in the continuum limit by a rescaling of the Grassmann variables by a factor $(2\Delta)^{(-3/2)}$.
\bigskip\noindent
{\bf 2. Lorentz symmetry}
\medskip
We next show that the action \eqref{N1} is Lorentz-symmetric in the continuum limit $\epsilon=\Delta,\Delta\to 0$. In the continuum limit we write the action as
\begin{equation}\label{F1}
S=\int_{t,x}\big\{\psi_\gamma\partial_t\psi_\gamma-\psi_\gamma
(T_k)_{\gamma\delta}
\partial_k\psi_\delta\big\}.
\end{equation}
It involves now four Grassmann functions $\psi_\gamma(t,x),\gamma=1\dots 4,x=(x_1,x_2,x_3)$. The integral extends over three dimensional space and time, with $\partial_t=\partial/\partial t$ and $\partial_k=\partial/\partial x_k$.
The Lorentz invariance of the action \eqref{F1} is most easily established by employing the real matrices
\begin{equation}\label{F7}
\gamma^0=\left(\begin{array}{rcr}
0&,&\tau_1\\-\tau_1&,&0\end{array}\right)~,~
\gamma^k=-\gamma^0 T_k,
\end{equation}
such that
\begin{equation}\label{F8}
S=-\int_{t,x}\bar\psi\gamma^\mu \partial_\mu \psi~,~\bar\psi=\psi^T\gamma^0,
\end{equation}
where $\mu=(0,k)$ and $\partial_0=\partial_t$. We define $\bar\psi_\gamma=\psi_\delta(\gamma^0)_{\delta\gamma}$. (We consider $\psi$ here as a vector with components $\psi_\gamma$ and suppress the vector indices. The Pauli matrices are denoted by $\tau_k$.) The real $4\times 4$ Dirac matrices $\gamma^\mu$ obey the Clifford algebra
\begin{equation}\label{F9}
\{\gamma^\mu,\gamma^\nu\}=2\eta^{\mu\nu},
\end{equation}
with signature of the metric given by $\eta_{\mu\nu}=diag(-1,1,1,1)$. This can be easily verified by using the relations
\begin{eqnarray}\label{F10}
\{T_k,T_l\}=2\delta_{kl}~,~
\{\gamma^0,T_k\}=0.
\end{eqnarray}
Furthermore, one finds
\begin{equation}\label{F11}
(\gamma^0)^T=-\gamma^0~,~(\gamma^k)^T=\gamma^k,
\end{equation}
and the relations
\begin{eqnarray}\label{F12}
[T_k,T_l]&=&2\epsilon_{klm}\tilde I T_m~,~[T_k, \tilde I]=0~,~\tilde I^2=-1,\nonumber\\
\gamma^0\gamma^1\gamma^2\gamma^3&=&\tilde I~,~\{\gamma^0,\tilde I\}=0~,~
\{\gamma^k,\tilde I\}=0.
\end{eqnarray}
Infinitesimal Lorentz-transformations can be written as
\begin{equation}\label{F4}
\delta\psi_\gamma=-\frac12
\epsilon_{\mu\nu}\left(\Sigma^{\mu\nu}\right)_{\gamma\delta}\psi_\delta~,~\epsilon_{\mu\nu}=-\epsilon_{\nu\mu},
\end{equation}
where we have omitted to indicate the associated transformations of the coordinates. The Lorentz generators $\Sigma^{\mu\nu}$ obtain from the Dirac matrices as
\begin{equation}\label{F13}
\Sigma^{\mu\nu}=-\frac14[\gamma^\mu,\gamma^\nu],
\end{equation}
and obey
\begin{equation}\label{F5}
\Sigma^{0k}=-\frac12 T_k~,~\Sigma^{kl}=-\frac12\epsilon^{klm}\tilde I T_m,
\end{equation}
We recognize in eq. \eqref{F8} the standard Lorentz invariant action for free Majorana fermions in a Majorana representation of the Clifford algebra with real $\gamma^\mu$-matrices.
\bigskip\noindent
{\bf 3. Functional integral}
\medskip
The functional integral is defined by the partition function
\begin{equation}\label{N9}
Z=\int{\cal D}\psi\bar g_f\big[\psi(t_f)\big]e^{-S}
g_{in}\big[\psi(t_{in})\big],
\end{equation}
with the functional measure
\begin{equation}\label{N10}
\int{\cal D}\psi=\prod_{t,x}\int \big(d\psi_4(t,x)\dots d\psi_1(t,x)\big).
\end{equation}
The boundary terms $g_{in}$ and $\bar g_f$ only depend on the Grassmann variables $\psi(t_{in})$ and $\psi(t_f)$, respectively. As we will see below, the boundary terms $\bar g_f$ and $g_{in}$ are related to each other, such that the functional integral \eqref{N9} is fully specified by the choice of $g_{in}$.
The basic definition of $Z$ is formulated on the discrete space-time lattice with finite volume. Thus $Z$ is a well defined real number. We will show later that for a suitable normalization of $g_{in}$ one obtains $Z=1$, independently of $L$. The continuum limit $L\to\infty$ for the partition function is therefore trivial.
We will use the Grassmann function integral \eqref{N9} in order to specify the fundamental evolution law for the probability distribution $\big\{p_\tau(t)\}$ of the classical statistical ensemble for Ising-spins. The positivity and normalization of the probabilities holds for an arbitrary choice of $g_{in}$, provided $\bar g_f$ is related to $g_{in}$ appropriately. For every given $g_{in}$ the probability distribution $\{p_\tau(t)\}$ is uniquely computable for all $t$, such that the functional integral \eqref{N9} indeed specifies the time evolution of the probability distribution.
\section{Grassmann wave function from functional integral}
\label{Grassmann wave function}
In this section we compute for the functional integral \eqref{N9} a Grassmann wave function $g(t)$, which is an element of the Grassmann algebra constructed from the Grassmann variables $\psi_\gamma(t,x)$ at given $t$. The central idea is to ``integrate out'' the Grassmann variables at times $t'\neq t$ \cite{14}. The expansion coefficients of $g(t)$ will specify the classical wave function $\big\{q_\tau(t)\big\}$.
\bigskip\noindent
{\bf 1. Integrating out past and future}
\medskip
The functional integral \eqref{N9} involves variables for arbitrary time points $t_n$. In order to construct a wave function $g(t)$ which only refers to a particular time $t$ we have to ``integrate out'' the information referring to other time points $t'\neq t$ \cite{CWMSa,14}. This can be done by decomposing the action \eqref{N1}
\begin{eqnarray}\label{M1}
S&=&S_<+S_>,\nonumber\\
S_<&=&\sum_{t'<t}L(t')~,~S_>=\sum_{t'\geq t}L(t').
\end{eqnarray}
The wave function $g(t)$ obtains now by integrating out all Grassmann variables for $t'<t$
\begin{equation}\label{M2}
g(t)=\int {\cal D}\psi(t'<t)e^{-S_<}g_{in}.
\end{equation}
We observe that $g(t)$ depends only on the Grassmann variables $\psi(t)$. More precisely, it is an element of the Grassmann algebra that can be constructed from the Grassmann variables $\psi_\gamma(t,x)$.
The conjugate wave function is defined as
\begin{equation}\label{M4}
\tilde g(t)=\int {\cal D} \psi(t'>t)\bar g_f e^{-S_>}.
\end{equation}
Again, this is an element of the Grassmann algebra constructed from $\psi(t)$. In terms of $g$ and $\tilde g$ the partition function reads
\begin{equation}\label{M5}
Z=\int{\cal D}\psi(t)\tilde g(t)g(t),
\end{equation}
where the Grassmann integration $\int {\cal D}\psi(t)$ extends now only over the Grassmann variables $\psi_\gamma(t,x)$ for a given time $t$.
\bigskip\noindent
{\bf 2. Classical probabilities and wave function}
\medskip
We next expand $g$ in terms of the basis elements $g_\tau$ of the Grassmann algebra generated by the variables $\psi_\gamma(t)$,
\begin{equation}\label{M3}
g(t)=\sum_\tau q_\tau(t)g_\tau\big[\psi(t)\big].
\end{equation}
We associate the real coefficients $q_\tau(t)$ with the classical wave function, such that the classical probabilities obtain as $p_\tau(t)=q^2_\tau(t)$. This requires for every $t$ the normalization $\sum_\tau q^2_\tau(t)=1$. We will show in the next section that this normalization condition is indeed obeyed, provided it holds for the initial wave function $g(t_{in})=g_{in}\big[\psi_\gamma(t_{in},x)\big]$.
The conjugate basis elements of the Grassmann algebra $\tilde g_\tau$ are defined \cite{CWF} by the relation
\begin{equation}\label{M6}
\tilde g_\tau g_\tau=\prod_x\prod_\gamma\psi_\gamma(x)
\end{equation}
(no sum over $\tau$) and the requirement that no variable $\psi_\gamma(x)$ appears both in $\tilde g_\tau$ and $g_\tau$. They obey
\begin{equation}\label{M6A}
\int{\cal D}\psi(t)\tilde g_\tau\big[\psi(t)\big]g_\rho\big[\psi(t)\big]=\delta_{\tau\rho}.
\end{equation}
Expanding
\begin{equation}\label{M7}
\tilde g(t)=\sum_\tau\tilde q_\tau(t)\tilde g_\tau\big[\psi(t)\big]
\end{equation}
yields
\begin{equation}\label{M8}
Z=\sum_\tau\tilde q_\tau(t)q_\tau(t).
\end{equation}
We will see in the next section that for a suitable choice of $\bar g_f$ the conjugate wave function obeys for all $t$ the relation $\tilde q_\tau(t)=q_\tau(t)$. Together with the normalization condition $\sum_\tau q^2_\tau=1$ this guarantees $Z=1$.
In consequence, we can express the classical probabilities $p_\tau(t)$ directly in terms of the Grassmann functional integral
\begin{equation}\label{M9}
p_\tau(t)=\int {\cal D}\psi\bar g_f{\cal P}_\tau(t)e^{-S}g_{in},
\end{equation}
with ${\cal P}_\tau(t)$ a projection operator
\begin{eqnarray}\label{M10}
{\cal P}_\tau(t)g_\rho\big[\psi(t)\big]
&=&g_\tau\big[\psi(t)\big]\delta_{\tau\rho},\nonumber\\
\int{\cal D}\psi(t)\tilde g_\sigma\big[\psi(t)\big]{\cal P}_\tau(t)g_\rho\big[\psi(t)\big]
&=&\delta_{\tau\sigma}\delta_{\tau\rho}.
\end{eqnarray}
The formal expression of ${\cal P}_\tau$ in terms of the Grassmann variables $\psi(t)$ and derivatives $\partial/\partial\psi(t)$ can be found in ref. \cite{3A}.
The interpretation of the Grassmann functional integral in terms of classical probabilities is based on the map $g(t)\to \big\{ p_\tau(t)\big\}$, which in turn is related to the map $\{ q_\tau\}\to \{p_\tau\}=\{ q^2_\tau\}$. The map $g(t)\leftrightarrow\big\{q_\tau(t)\big\}$ is invertible. A given state or classical statistical ensemble may be specified by the ``initial value'' at some time $t_0, g(t_0)$, or the associated wave function $\big\{q_\tau(t_0)\big\}$ or classical probability distribution $\{p_\tau(t_0)\big\}$. This is equivalent to the specification of $g_{in}=g(t_{in})$ in the functional integral.
\section{Time evolution}
\label{Timeevolution}
In this section we compute the time evolution of the wave function $\big\{q_\tau(t)\big\}$ and the associated probability distribution $\{p_\tau(t)\big\}=\big\{q^2_\tau(t)\big\}$. This will lead to a type of generalized Schr\"odinger equation for the real wave function $\big\{q_\tau(t)\big\}$, as well as an associated evolution equation for the Grassmann wave function $g(t)$. We will use the properties of this evolution equation in order to establish that the norm $\sum_\tau q^2_\tau(t)$ is conserved, such that $\big\{q^2_\tau(t)\big\}$ can indeed be interpreted as a time dependent probability distribution.
Due to the particular form of the action \eqref{N1}, which only involves one type of Grassmann variables (and no conjugate variables as in ref. \cite{CWF}) the discrete formulation of the functional integral is crucial. We will see that $g(t)$ jumps between two neighboring time points, while it is smooth between $g(t+2\epsilon)$ and $g(t)$. We will therefore distinguish between even and odd time points and use the definition \eqref{M3} of the wave function only for even times. (For odd times one may employ a different definition, which will guarantee the smoothness of the time evolution of $\big\{q_\tau(t)\big\}$ for both even and odd time points \cite{CWMSa}.) In the continuum limit, $\epsilon\to 0$, the time evolution of the wave function is described by a continuous rotation of the vector $\big\{q_\tau(t)\big\}$.
As a consequence of its definition \eqref{M3} the Grassmann wave function obeys the time evolution
\begin{equation}\label{P1}
g(t+\epsilon)=\int{\cal D}\psi(t)e^{-L(t)}g(t).
\end{equation}
This determines $\big\{q_\tau(t+\epsilon)\big\}$ in terms of $\big\{q_\tau(t)\big\}$. Thus the action \eqref{N1}-\eqref{B17} specifies the dynamics how the probability distribution $\big\{p_\tau(t)\big\}$ evolves in time. The particular dynamics of a given model is determined by the form of $B_\gamma$ in eq. \eqref{B17}.
\bigskip\noindent
{\bf 1. Unitary time evolution}
\medskip
We want to establish the relation \eqref{35B} which guarantees the preservation of the sum of probabilities $\sum_\tau p_\tau=\sum_\tau q^2_\tau=1$, with $t'\to t,~t\to t+2\epsilon$. For this purpose we employ the property \eqref{N8}. The relation between $q_\tau(t+2\epsilon)$ and $q_\tau(t)$ is computed from eq. \eqref{P1}.
Inserting the specific form \eqref{A17} for $L(t)$ the evolution equation \eqref{P1} reads
\begin{equation}\label{P2}
g(t+\epsilon)=\int{\cal D}\psi(t)\exp\big\{-\sum_x\sum_\gamma\psi_\gamma(t,x)
B_\gamma(t+\epsilon,x)\big\}g(t).
\end{equation}
We may write $g(t)$ in a product form
\begin{equation}\label{P3}
g(t)=\prod_x\prod_\gamma\big[a_\gamma(t,x)+b_\gamma(t,x)\psi_\gamma(t,x)\big],
\end{equation}
with some fixed ordering convention of the factors assumed, e.g. smaller $\gamma$ to the left for given $x$, and some linear ordering of the lattice points, with ``lower'' points to the left. In the product form eq. \eqref{P2} yields
\begin{eqnarray}\label{P4}
&&g(t+\epsilon)=\int{\cal D}\psi(t)\prod_x\prod_\gamma\\
&&\Big\{\big[1-\psi_\gamma(t,x)B_\gamma(t+\epsilon,x)\big]
\big[a_\gamma(t,x)+b_\gamma(t,x)\psi_\gamma(t,x)\big]\Big\}\nonumber\\
&&\hspace{1.3cm}=\prod_x\prod_\gamma\Big\{b_\gamma(t,x)+\eta_\gamma a_\gamma(t,x)B_\gamma(t+\epsilon,x)\Big\}.\nonumber
\end{eqnarray}
Here we use the fact that each individual Grassmann integration $\int d\psi_\gamma(t,x)$ can be performed easily,
\begin{equation}\label{P4A}
\int d\psi(1-\psi\varphi)(a+b\psi)=b-a\varphi,
\end{equation}
and $\eta_\gamma=\pm 1$ results from the anticommuting properties of the Grassmann variables $\varphi$, with
\begin{equation}\label{P5}
\eta_1=\eta_3=1~,~\eta_2=\eta_4=-1.
\end{equation}
As a result, we can write
\begin{equation}\label{P6}
g(t+\epsilon)=\sum_\tau q_\tau(t){\cal C}g_\tau \big[B_\gamma(t+\epsilon,x)\big],
\end{equation}
where ${\cal C}g_\tau$ obtains from $g_\tau$ by the following replacements: (i) for every factor $\psi_\gamma(x)$ in $g_\tau$ one has a factor $1$ in ${\cal C}g_\tau$; (ii) for every pair $(x,\gamma)$ for which no $\psi_\gamma(x)$ is present in $g_\tau$ one inserts a factor $\eta_\gamma\psi_\gamma(x)$ in ${\cal C}g_\tau$.
The ordering of the factors $\eta_\gamma\psi_\gamma(x)$ is the same as the ordering assumed in the product \eqref{P3}. This implies that we can indeed write the action of ${\cal C}$ on the product \eqref{P3} as
\begin{equation}\label{P7}
{\cal C}g(t)=\prod_x\prod_\gamma\big(b_\gamma(t,x)+\eta_\gamma a_\gamma(t,x)\psi_\gamma (t,x)\big).
\end{equation}
The conjugation operator ${\cal C}$ maps every basis element $g_\tau$ into its conjugate element $\tilde g_\tau$ up to a sign $\sigma_\tau=\pm 1$,
\begin{equation}\label{P8}
{\cal C}g_\tau=\sigma_\tau\tilde g_\tau.
\end{equation}
Applying ${\cal C}$ twice on the product \eqref{P3} multiplies each factor by $\eta_\gamma$. The factors $\eta_\gamma$ drop out due to the even number of minus signs, such that ${\cal C}$ is an involution, ${\cal C}^2=1$, or
\begin{equation}\label{P9}
{\cal C}^2 g_\tau=g_\tau.
\end{equation}
The jump between $g$ and ${\cal C}g$ for neighboring time points suggests the use of eq. \eqref{M3} for the definition of the wave function $\big\{q_\tau(t)\big\}$ only for ``even time points'', namely those that obey $t_n=t_{in}+2n\epsilon,n\in{\mathbbm N}$. Repeating the procedure leading to eq. \eqref{P6} one obtains
\begin{equation}\label{58A}
g(t+2\epsilon)=\sum_\tau q_\tau(t)g_\tau\big[A_\gamma(t+2\epsilon,x)\big],
\end{equation}
where $A_\gamma$ is a linear combination of Grassmann variables $\psi_\delta(t+2\epsilon,x)$ which live on the same sublattice of the fundamental lattice as $\psi_\delta(t,x)$. For an arbitrary linear transformation
\begin{equation}\label{58B}
B_\gamma(t+\epsilon, x)=F_{\gamma\delta}(x,y;t+\epsilon)\psi_\delta(t+\epsilon,y)
\end{equation}
with unit Jacobian, $\det F=1$, we can write
\begin{eqnarray}\label{58C}
g(t+2\epsilon)&=&\int{\cal D}\psi(t+\epsilon)\nonumber\\
&& \exp \big\{-\psi_\gamma(t+\epsilon,x)B_\gamma(t+2\epsilon,x)\big\}g(t+\epsilon)\nonumber\\
&=&\int{\cal D}B(t+\epsilon)\\
&&\exp \big\{-B_\gamma(t+\epsilon,x)A_\gamma(t+2\epsilon,x)\big\}g(t+\epsilon),\nonumber
\end{eqnarray}
with
\begin{eqnarray}\label{58D}
A_\gamma(t+2\epsilon,x)&=&(F^{-1})^T_{\gamma\delta}(x,y;t+\epsilon)\nonumber\\
&& \times F_{\delta\eta}(y,z;t+2\epsilon)\psi_\eta
(t+2\epsilon,z).
\end{eqnarray}
Here we sum over repeated indices $\delta,\eta$ and repeated coordinates $y,z$ and write the Grassmann integration as an integration over new variables $B_\gamma(t+\epsilon,x)$. We next use the expression \eqref{P4} for $g(t+\epsilon)$ and perform the integration over $B_\gamma(t+\epsilon,x)$,
\begin{equation}\label{58E}
g(t+2\epsilon)=\prod_{x,\gamma}\eta_\gamma\big(a_\gamma(t,x)+b_\gamma(t,x)A_\gamma(t+2\epsilon,x)\big).
\end{equation}
This replaces in $g(t)$ every factor $\psi_\gamma(t,x)$ by $A_\gamma(t+2\epsilon,x)$ such that eq. \eqref{58D} specifies the expression $A_\gamma(t+2\epsilon,x)$ appearing in eq. \eqref{58A}.
We next employ the important property that $F$ is given by the orthogonal matrix $\bar R$ in eq. \eqref{N8}, which is independent of $t$. Since $(F^{-1})^T=F$ we end with the simple expression
\begin{equation}\label{58F}
A_\gamma(t+2\epsilon,x)=\bar R^2_{\gamma\eta}(x,z)\psi_\eta(t+2\epsilon,z),
\end{equation}
where
\begin{equation}\label{58G}
\bar R^2_{\gamma\eta}(x,z)=\bar R_{\gamma\delta}(x,y)\bar R_{\delta\eta}(y,z)
\end{equation}
defines a transformation between Grassmann variables on the same sublattice. The matrix $\bar R^2$ is, in turn, also orthogonal $(\bar R^2)^T\bar R^2=1$. We conclude that $g(t+2\epsilon)$ can be obtained from $g(t)$ by a simple rotation of the Grassmann variables.
Since $A_\gamma(x)$ is related to $\psi_\gamma(x)$ by a rotation \eqref{58F}, it is straightforward to show that $g_\tau\big[A_\gamma(x)\big]$ is also connected to $g_\tau\big[\psi_\gamma(x)\big]$ by a rotation among the basis elements
\begin{eqnarray}\label{P16}
g_\tau\big[A_\gamma(x)\big]&=&\sum_\rho
g_\rho\big[\psi_\gamma(x)\big]R_{\rho\tau},\nonumber\\
\sum_\rho R_{\tau\rho}R_{\sigma\rho}&=&\delta_{\tau\sigma}.
\end{eqnarray}
One infers
\begin{eqnarray}\label{P17}
g(t+2\epsilon)&=&\sum_{\tau,\rho}q_\tau(t)
g_\rho\big[\psi_\gamma(t+2\epsilon)\big]R_{\rho\tau}\nonumber\\
&=&\sum_\tau q_\tau(t+2\epsilon)g_\tau\big[\psi_\gamma(t+2\epsilon)\big],
\end{eqnarray}
with a rotated wave function
\begin{equation}\label{P18}
q_\tau(t+2\epsilon)=\sum_\rho R_{\tau\rho} q_\rho(t).
\end{equation}
This establishes eq. \eqref{35B}.
Rotations preserve the length of the vector $\{q_\tau\}$ such that $\sum_\tau q^2_\tau(t)$ is independent of $t$. Choosing
$g_{in}=\sum_\tau q_\tau(t_{in})g_\tau\big[\psi(t_{in})\big]$ with $\sum_\tau q^2_\tau(t_{in})=1$ one infers $\sum_\tau q^2_\tau(t)=1$ for all $t$. Therefore $\big\{p_\tau(t)\big\}=\big\{q^2_\tau(t)\big\}$ has indeed for all $t$ the properties of a probability distribution, namely positivity of all $p_\tau$ and the normalization $\sum_\tau p_\tau=1$. In analogy to quantum mechanics we call a time evolution which preserves the norm of $\big\{q_\tau(t)\big\}$ a ``unitary time evolution''. A unitary time evolution is crucial for the probabilistic interpretation of the functional integral \eqref{N9}.
\bigskip\noindent
{\bf 2. Evolution of conjugate wave function}
\medskip
We next want to show the relation (for $t$ even)
\begin{equation}\label{P19}
\tilde g(t-2\epsilon)=\sum_{\tau,\rho}\tilde q_\tau(t)R_{\tau\rho}\tilde g_\rho
\big[\psi_\gamma(t-2\epsilon,x)\big].
\end{equation}
The definition \eqref{M4} of the conjugate wave function $\tilde q$ then implies
\begin{equation}\label{P20}
\tilde q_\tau(t-2\epsilon)=\sum_\rho \tilde q_\rho(t)R_{\rho\tau},
\end{equation}
such that
\begin{equation}\label{P21}
\tilde q_\tau(t)=\sum_\rho R_{\tau\rho}\tilde q_\rho(t-2\epsilon).
\end{equation}
Comparing with eq. \eqref{P18} one infers that $q_\tau(t)$ and $\tilde q_\tau(t)$ obey the same evolution equation. (By an analogous argument eq. \eqref{P21} also holds for $t$ odd.) If $q$ and $\tilde q$ are equal for some particular time $t_0$, they will remain equal for all $t$.
If $\big\{\tilde q_\tau(t_0)\big\}$ equals $\big\{ q_\tau(t_0)\big\}$ for some time $t_0$ we can use $\tilde q_\tau(t)=q_\tau(t)$ for all $t$ and infer from eq. \eqref{M8}
\begin{equation}\label{P28}
Z=\sum_\tau q^2_\tau(t).
\end{equation}
As it should be, $Z$ remains invariant under rotations \eqref{P18} of the vector $\{q_\tau\}$ and is therefore independent of $t$.
In order to show eq. \eqref{P19} we employ the definition \eqref{M4} which implies
\begin{equation}\label{P13}
\tilde g(t-\epsilon)=\int{\cal D}\psi(t)\tilde g(t)e^{-L(t-\epsilon)},
\end{equation}
or
\begin{eqnarray}\label{P22}
&&\tilde g(t-\epsilon)=\int{\cal D}\psi(t)\tilde g(t)\exp
\big\{-\sum_x\psi_\gamma(t-\epsilon,x)B_\gamma(t,x)\big\}\nonumber\\
&&=\int{\cal D}\psi(t)\tilde g(t)\exp
\big\{-\sum_{x,y}\psi_\gamma(t-\epsilon,x)
\bar R_{\gamma\delta}(x,y)\psi_\delta(t,y)\big\}\nonumber\\
&&=\int{\cal D}\psi(t)\tilde g(t)\exp
\big\{-\sum_x\tilde B_\gamma(t-\epsilon,x)\psi_\gamma(t,x)\big\},
\end{eqnarray}
with
\begin{eqnarray}\label{P23}
\tilde B_\gamma(t,x)&=&\sum_y\psi_\delta(t,y)\bar R_{\delta\gamma}(y,x)\nonumber\\
&=&\sum_y(\bar R^{-1})_{\gamma\delta}(x,y)\psi_\delta(t,y).
\end{eqnarray}
Performing the Grassmann integral one finds
\begin{equation}\label{P24}
\tilde g(t-\epsilon)=\sum_\tau\tilde q_\tau(t)\tilde{\cal C}\tilde g_\tau
\big[\tilde B_\gamma(t-\epsilon,x)\big],
\end{equation}
where the map $\tilde {\cal C}$ acts similarly as ${\cal C}$, with $\eta_\gamma$ replaced by $\tilde \eta_\gamma=-\eta_\gamma,
\tilde\eta_1=\tilde\eta_3=-1,\tilde\eta_2=\tilde\eta_4=1$, and $\tilde{\cal C}^2=1$. A similar step yields
\begin{equation}\label{79A}
\tilde g(t-2\epsilon)=\sum_\tau\tilde q_\tau(t)\tilde g_\tau
\big[\tilde A_\gamma(t-2\epsilon,x)\big]
\end{equation}
with
\begin{equation}\label{79B}
\tilde A_\gamma(t-2\epsilon,x)=(\bar R^2)^{-1}_{\gamma\delta}(x,y)\psi_\delta
(t-2\epsilon,y).
\end{equation}
One concludes
\begin{equation}\label{79Bb}
\tilde g_\tau\big[\tilde A_\gamma(t-2\epsilon,x)\big]
=\sum_\rho\tilde g_\rho\big[\psi_\gamma(t-2\epsilon,x)\big]R^{-1}_{\rho\tau}
\end{equation}
and infers eq. \eqref{P19}.
\bigskip\noindent
{\bf 3. Boundary terms}
\medskip
The final point we have to settle in order to establish the normalization $Z=1$ and the expression \eqref{M9} concerns the equality of $\tilde q(t_0)$ and $q(t_0)$ for some arbitrary time $t_0$. This is achieved by a proper choice of the relation between the boundary terms $\bar g_f$ and $g_{in}$ in eq. \eqref{N9}. For this purpose we may imagine that we (formally) solve the evolution equation \eqref{P18} in order to compute $\big\{q_\tau(t_f)\big\}$ in terms of $\big\{q_\tau(t_{in}\big\}$, $g_{in}=\sum_\tau q_\tau(t_{in}) g_\tau\big[\psi(t_{in})\big]$. (We assume that $t_{in}$ and $t_f$ are even.) Using
\begin{eqnarray}\label{P32}
g(t_f)&=&\sum_\tau q_\tau(t_f)g_\tau\big[\psi(t_f)\big], \\
\tilde g(t_f)&=&\sum_\tau\tilde q_\tau(t_f)\tilde g_\tau\big[\psi(t_f)\big]=\bar g_f,\label{P32a}
\end{eqnarray}
it is sufficient to choose $\bar g_f$ such that $\tilde q_\tau(t_f)=q_\tau(t_f)$. Equivalently, we may specify the wave function $\big\{q_\tau(t_0)\big\}=\big\{\tilde q_\tau(t_0)\big\}$ at some arbitrary even time $t_0$ and compute the corresponding $g_{in}$ and $\bar g_f$ by a solution of the evolution equation, using the fact that the rotation \eqref{P18} can be inverted in order to compute $q(t-\epsilon)$ form $q(t)$.
If we would not adopt the choice $\tilde q_\tau(t_f)=q_\tau(t_f)$ the functional integral \eqref{N9} would amount to a transition amplitude, which is another useful notion in quantum field theory. We are interested, however, in a classical probabilistic setting and therefore focus on the choice of $\bar g_f$ that guarantees $Z=1$ for all $t$. In practice, neither $g_{in}$ nor $\bar g_f$ need to be computed explicitly since we only use the Grassmann functional integral for extracting the evolution equation for the Grassmann wave function $g(\tau)$ and the classical wave function $\big \{q_\tau(t)\big\}$.
\bigskip\noindent
{\bf 4. Continuous evolution equation}
\medskip
Finally, we cast the evolution law \eqref{P18} into the form of a differential time evolution equation by taking the limit $\epsilon\to 0$. This results in a generalized Schr\"odinger type equation for the real wave function $\big\{q_\tau(t)\big\}$,
\begin{equation}\label{P35}
\partial_t q_\tau(t)=\sum_\rho K_{\tau\rho} q_\rho(t).
\end{equation}
Since the evolution describes a rotation, the matrix $K$ is antisymmetric
\begin{equation}\label{P36}
K_{\rho\tau}=-K_{\tau\rho}.
\end{equation}
We identify this evolution equation with the generalized Schr\"odinger equation for a quantum wave function for the special case of a real wave function and purely imaginary and hermitean Hamiltonian $H=i\hbar K$.
The time evolution \eqref{P35} translates directly to the probabilities (no summation over $\tau$ here)
\begin{equation}\label{G14}
\partial_t p_\tau=2\sum_\rho K_{\tau\rho}s_\tau s_\rho\sqrt{p_\tau p_\rho}.
\end{equation}
Once the signs $s_\tau(t_0)$ are fixed by some appropriate convention at a given time $t_0$, the signs $s_\tau(t)$ are computable in terms of the probabilities $p_\tau(t_0)$. This follows since for all $t$ the wave function $q_\tau(t)$ is uniquely fixed by $q_\tau(t_0)$ or $p_\tau(t_0)$, and $p_\tau(t)$ is uniquely determined by $q_\tau(t)$. In principle, it is therefore possible to formulate the time evolution law for the probabilities uniquely in terms of the probabilities $p_\tau(t_0)$. However, an expression of $\partial_tp_\tau(t)$ in terms of $p_\tau(t)$ needs to keep track of the sign functions $s_\tau(t)$. In principle, this can be done by an updating procedure. The sign $s_\tau$ can flip only at times $t$ where $p_\tau(t)=0$. If it is flipped or not at these times is decided by the requirements $p_\tau>0,\Sigma_\tau p_\tau=1$ for the following times \cite{CWMSa}.
Instead of such an updating procedure it is much more convenient to use the wave function and the linear evolution law \eqref{P35} for the description of the classical statistical ensembles associated to the action $S$. The basic reason is the condition of unit norm of the probability distribution. This can be quite cumbersome for a general evolution equation for $\{p_\tau\}$, but it is extremely simple on the level of $\{q_\tau\}$ where only the length of a real vector has to be preserved.
\bigskip\noindent
{\bf 5. Grassmann evolution equation}
\medskip
The matrix $K$ can be extracted from the Grassmann evolution equation
\begin{equation}\label{P37}
\partial_t g={\cal K} g,
\end{equation}
according to
\begin{eqnarray}\label{P34}
\partial_tg(t)&=&\frac{1}{2\epsilon}\big[g(t+2\epsilon)-g(t)\big]={\cal K} g(t)=\sum_\tau q_\tau (t){\cal K} g_\tau \nonumber\\
&=&\sum_{\tau,\rho}q_\tau(t)g_\rho\big[\psi_\gamma(x)\big]K_{\rho\tau}=\sum_\tau\partial_t q_\tau(t)g_\tau\big[\psi_\gamma(x)\big].\nonumber\\
\end{eqnarray}
(In eq. \eqref{P34} we use a fixed basis, corresponding to the basis elements $g_\tau$ constructed from $\psi(t)$ for $g(t)$, and from $\psi(t+2\epsilon)$ for $g(t+2\epsilon)$.) For our model of free fermions the Grassmann evolution generator ${\cal K}$ is given by
\begin{equation}\label{P38}
{\cal K}=\sum_x\frac{\partial}{\partial\psi_\gamma(x)}(T_k)_{\gamma\delta}
\partial_k\psi_\delta(x).
\end{equation}
This yields the matrix element in eq. \eqref{P35},
\begin{equation}\label{P44}
K_{\rho\tau}=\int{\cal D}\psi\tilde g_\rho{\cal K} g_\tau.
\end{equation}
In order to proof the relations \eqref{P37}, \eqref{P38} we
infer from eqs. \eqref{58A}, \eqref{58F} the relation
\begin{equation}\label{89-1}
g(t+2\epsilon)-g(t)=\sum_{\tau} q_\tau(t)\big(g_\tau[\bar R^2\psi]-g_\tau[\psi]\big)
\end{equation}
and use the continuum limit $\epsilon=\Delta\to 0$, cf. eqs. \eqref{A25}, \eqref{H25},
\begin{equation}\label{90-1}
\bar R^{2}\psi=(1-2\epsilon\sum_k T_k\partial_k)\psi.
\end{equation}
We then employ the identity (in linear order in $\epsilon$ and summation over $k$ implied)
\begin{eqnarray}\label{91-1}
&&g_\tau\big[(1-2\epsilon T_k\partial_k)\psi\big]\nonumber\\
&&\qquad=\left(1+2\epsilon \sum_x\frac{\partial}{\partial\psi_\gamma(x)}
(T_k)_{\gamma\delta}\partial_k\psi_\delta(x)\right)g_\tau[\psi]\nonumber\\
&&\qquad=(1+2\epsilon{\cal K})g_\tau[\psi]
\end{eqnarray}
in order to establish eqs. \eqref{P37}, \eqref{P38}.
Taking finally the space-continuum limit by rescaling $\psi_\gamma(x)$ and $\partial/\partial\psi_\gamma(x)$ such that
\begin{equation}\label{P45}
\left\{\frac{\partial}{\partial\psi_\gamma(x)}~,~
\psi_\delta(y)\right\}=
\delta_{\gamma\delta}\delta^3(x-y),
\end{equation}
we arrive at the continuum form of the Grassmann evolution equation
\begin{equation}\label{P46}
\partial_tg={\cal K} g~,~{\cal K}=\int_x
\frac{\partial}{\partial\psi_\gamma(x)}(T_k)_{\gamma\delta}
\partial_k\psi_\delta(x).
\end{equation}
This evolution equation will be the basis for the interpretation of the time dependent wave function $\big\{q_\tau(t)\big\}$ and probability distribution $\big\{p_\tau(t)\big\}$ in terms of propagating fermionic particles.
\section{Observables}
\label{Observables}
Classical observables $A$ take a fixed value $A_\tau$ for every classical state $\tau$. In classical statistics the possible outcomes of measurements of $A$ correspond to the spectrum of possible values $A_\tau$. The expectation value of $A$ obeys
\begin{equation}\label{G15}
\langle A\rangle=\sum_\tau p_\tau A_\tau.
\end{equation}
Our description of the system will be based on these classical statistical rules. For example, we may consider the observable measuring the occupation number $N_\gamma(x)$ of the bit $\gamma$ located at $x$. The spectrum of possible outcomes of measurements consists of values $1$ or $0$, depending if a given state $\tau=[n_\gamma(x)]$ has this particular bit occupied or empty.
For the Grassmann basis element $g_\tau$ associated to $\tau$ one finds $N_\gamma(x)=0$ if $g_\tau$ contains a factor $\psi_\gamma(x)$, and $N_\gamma(x)=1$ otherwise. We can associate to this observable a Grassmann operator ${\cal N}_\gamma(x)$ obeying (no summation here)
\begin{equation}\label{G16}
{\cal N}_\gamma(x)g_\tau=\big(N_\gamma(x)\big)_\tau g_\tau~,~{\cal N}_\gamma(x)=
\frac{\partial}{\partial\psi_\gamma(x)}
\psi_\gamma(x).
\end{equation}
Two occupation number operators ${\cal N}_{\gamma_1}(x_1)$ and ${\cal N}_{\gamma_2}(x_2)$ commute.
In general, we may associate to each classical observable $A$ a diagonal quantum operator $\hat A$ acting on the wave function, defined by
\begin{equation}\label{G16a}
(\hat A q)_\tau=A_\tau q_\tau.
\end{equation}
This yields the quantum rule for expectation values
\begin{equation}\label{G17}
\langle A\rangle=\langle q\hat A q\rangle=\sum_{\tau,\rho}q_\tau\hat A_{\tau\rho}q_\rho,
\end{equation}
with $\hat A$ a diagonal operator $\hat A_{\tau\rho}=A_\tau\delta_{\tau\rho}$. In the Grassmann formulation one uses the associated Grassmann operator ${\cal A}$ obeying
\begin{equation}\label{G18}
{\cal A} g_\tau=A_\tau g_\tau,
\end{equation}
such that
\begin{equation}\label{G19}
\langle A\rangle=\int {\cal D}\psi\tilde g{\cal A} g.
\end{equation}
Here $\tilde g$ is conjugate to $g$, i.e. for $g=\sum_\tau q_\tau g_\tau$ one has $\tilde g=\sum_\tau q_\tau\tilde g_\tau$.
In classical statistics the time evolution of the expectation value is induced by the time evolution of the probability distribution
\begin{equation}\label{95A}
\langle A(t)\rangle=\sum_\tau p_\tau(t)A_\tau.
\end{equation}
This corresponds to the Schr\"odinger picture in quantum mechanics
\begin{equation}\label{95B}
\langle A(t)\rangle=\langle q(t)\hat A q(t)\rangle,
\end{equation}
or the corresponding expression in terms of the Grassmann algebra
\begin{equation}\label{95C}
\langle A(t)\rangle=\int {\cal D}\psi\tilde g(t){\cal A}g(t).
\end{equation}
Using $\partial_t q=Kq$ \eqref{P35} and $\partial_t g={\cal K} g$ \eqref{P46} we infer for the time evolution of the expectation value the relations
\begin{equation}\label{95D}
\partial_t\langle A\rangle=\langle q[\hat A,K]q\rangle=\int{\cal D}\psi\tilde g[{\cal A},{\cal K}]g.
\end{equation}
Conserved quantities are represented by Grassmann operators that commute with ${\cal K},[{\cal A},{\cal K}]=0.$
\section{Particle states}
\label{Particlestates}
Our system admits a conserved particle number, corresponding to the Grassmann operator ${\cal N}$,
\begin{equation}\label{H5}
{\cal N}=\int_y\frac{\partial}{\partial\psi_\gamma(y)}\psi_\gamma(y)~,~[{\cal N},{\cal K}]=0.
\end{equation}
The particle number is Lorentz invariant. We can decompose an arbitrary Grassmann element into eigenstates of ${\cal N}$
\begin{equation}\label{H6}
g=\sum_mA_mg_m~,~{\cal N} g_m=mg_m.
\end{equation}
The time evolution does not mix sectors with different particle number $m$, such that the coefficients $A_m$ are time independent
\begin{equation}\label{H7}
\partial_t g=\sum_mA_m\partial_tg_m~,~\partial_tg_m={\cal K} g_m.
\end{equation}
We can restrict our discussion to eigenstates of ${\cal N}$. The range of $m$ is $[0,B]$, with $B=N_sL^3/8=L^3/2$ the number of independent Grassmann variables.
\bigskip\noindent
{\bf 1. Vacuum}
\medskip
Let us consider some static vacuum state $g_0$ with a fixed particle number $m_0$,
\begin{equation}\label{H8}
{\cal K} g_0=0~,~{\cal N} g_0=m_0g_0~,~\int {\cal D}\psi \tilde g_0 g_0=1.
\end{equation}
An example for a possible vacuum state is the totally empty state $g_0=|0\rangle$, with
\begin{eqnarray}\label{47Aa}
|0\rangle=\prod_\alpha\psi_\alpha&=&\prod_x\prod_\gamma\psi_\gamma(x)
=\prod_x(\psi_1\psi_2\psi_3\psi_4),\nonumber\\
{\cal N}|0\rangle&=&0.
\end{eqnarray}
It obeys
\begin{equation}\label{104Aa}
\int{\cal D}\psi|0\rangle=1~,~|\tilde 0\rangle=1.
\end{equation}
For $m_0\neq 0$ we shift the particle number by an additive ``renormalization'' $n=m-m_0$, such that the vacuum corresponds to $n=0$, and $g=A_n g_n$. An eigenstate of ${\cal N}$ with eigenvalue $m=m_0+n$ is called a $n$-particle state if $n$ is positive, and a $n$-hole state for negative $n$.
\bigskip\noindent
{\bf 2. One-particle and one-hole states}
\medskip
We next define creation and and annihilation operators $a^\dagger_\gamma(x),~a_\gamma(x)$ as
\begin{equation}\label{47A}
a^\dagger_\gamma(x)g=\frac{\partial}{\partial\psi_\gamma(x)}g~,~a_\gamma(x)g=
\psi_\gamma(x)g.
\end{equation}
They obey the standard (anti-)commutation relations
\begin{eqnarray}\label{H6a}
\big \{a^\dagger_\gamma(x),~a_\epsilon(y)\big\}&=&\delta_{\gamma\epsilon}\delta(x-y)~,~
{\cal N}=\int_xa^\dagger_\gamma(x)a_\gamma(x),\nonumber\\
~[a^\dagger_\gamma(x),{\cal N}]&=&-a^\dagger_\gamma(x)~,~[a_\gamma(x),{\cal N}]=a_\gamma(x).
\end{eqnarray}
Acting with the creation operator on the vacuum produces one-particle states
\begin{eqnarray}\label{H7a}
&&g_1(t)=\int_xq_\gamma(t,x)a^\dagger_\gamma(x)g_0={\cal G}_1g_0,\nonumber\\
&&({\cal N}-m_0)g_1=g_1.
\end{eqnarray}
If needed, we may multiply ${\cal G}_1$ with an appropriate normalization factor such that the wave function obeys
\begin{equation}\label{57A}
\int_x\sum_\gamma q^2_\gamma(x)=1,
\end{equation}
and $g_1$ has a standard normalization.
Similarly, a one-hole state with $n=-1$ obtains by employing the annihilation operator
\begin{eqnarray}\label{H8a}
g_{-1}(t)&=&\int_x\hat q_\gamma(t,x)a_\gamma(x)g_0\nonumber\\
&=&-\int_x\bar q_\gamma(t,x)(\gamma^0)_{\gamma\delta}a_\delta(x)g_0=
{\cal G}_{-1}g_0,\nonumber\\
&&~~({\cal N}-m_0)g_{-1}=-g_{-1}.
\end{eqnarray}
No one-hole states exist for the vacuum \eqref{47Aa}, but this issue is different if $m_0\neq 0$, as for example for $g_0=1$ where $m_0=B$.
If we transform the one-particle wave function $q_\gamma(t,x)$ infinitesimally according to
\begin{eqnarray}\label{H9}
\delta q_\gamma=-\frac12\epsilon_{\mu\nu}\left (\Sigma^{\mu\nu}\right)_{\gamma\delta}q_\delta
\end{eqnarray}
the operator ${\cal G}_1$ is Lorentz invariant. (We omit here the part resulting from the change of coordinates.) Similarly, $\hat q_\gamma$ transforms as $q_\gamma$ and the corresponding infinitesimal transformation results in an invariant ${\cal G}_{-1}$. If the vacuum $g_0$ is Lorentz invariant, the Lorentz transformed one-particle or one-hole wave functions will obey the same evolution equations as the original wave functions.
The time evolution of the one particle wave function $q_\gamma$ is given by
\begin{equation}\label{H11}
\partial_t g_1={\cal K} g_1=\int_x(\partial_tq\frac{\partial}{\partial\psi})g_0=\int_x
q\left[{\cal K},\frac{\partial}{\partial\psi}\right]g_0.
\end{equation}
and similar for the hole. With
\begin{equation}\label{54A}
\left[{\cal K},\frac{\partial}{\partial\psi_\gamma(x)}\right]=-
\partial_k\frac{\partial}{\partial\psi_\epsilon(x)}
(T_k)_{\epsilon\gamma}
\end{equation}
and
\begin{eqnarray}\label{H12}
\big[{\cal K},\psi_\gamma(x)\big]=-\partial_k\psi_\epsilon(x)(T_k)_{\epsilon\gamma},
\end{eqnarray}
one obtains Dirac equations for real wave functions (with $\partial_0=\partial_t)$
\begin{equation}\label{H13}
\gamma^\mu \partial_\mu q=0~,~\gamma^\mu \partial_\mu \hat q=0.
\end{equation}
We emphasize that these equations follow for arbitrary static states $g_0$ which obey ${\cal K}g_0=0$.
\bigskip\noindent
{\bf 3. Weyl and Majorana spinors}
\medskip
Let us consider $g_0=|0\rangle$ where $a_\gamma(x)|0\rangle=0$ implies that no hole states exist. There are then only particle states and the propagating degrees of freedom correspond to Majorana fermions. In four dimensions Majorana spinors are equivalent to Weyl spinors \cite{CWMS}. Indeed, we may introduce a complex structure by defining a two-component complex spinor
\begin{equation}\label{H14}
\varphi(x)=\left(\begin{array}{c}
\varphi_1(x)\\\varphi_2(x)
\end{array}\right)~,~
\varphi_1=q_1+iq_2~,~\varphi_2=q_3+iq_4.
\end{equation}
The matrices $T_k=\gamma^0\gamma^k$ are compatible with this complex structure. They are translated to the complex Pauli matrices
\begin{equation}\label{H15}
T_k=\gamma^0\gamma^k\to \tau_k,
\end{equation}
while the operation $\varphi\to i\varphi$ corresponds to the matrix multiplication $q\to \tilde I q$. (Cf. ref. \cite{CWMSa} for more details of the map from Majorana to Weyl spinors.) The Dirac equation reads in the complex basis
\begin{equation}\label{H16}
\partial_t\varphi=\tau_k\partial_k\varphi~,~i\partial_t\varphi=-\tau_k{\cal P}_k\varphi~,~{\cal P}_k\varphi=-i\partial_k\varphi,
\end{equation}
and the Lorentz generators are given by
\begin{equation}\label{H17}
\Sigma^{kl}\to-\frac i2\epsilon^{klm}\tau_m~,~
\Sigma^{0k}\to-\frac12\tau_k.
\end{equation}
In contrast, the multiplications with $\gamma^0$ cannot be represented by a multiplication of $\varphi$ with a complex $2\times 2$ matrix, since
\begin{equation}\label{18}
q\to \gamma^0 q~~\widehat{=}~~\varphi\to -\tau_2\varphi^*.
\end{equation}
If we express $\varphi^*$ in terms of the two-component complex vector
\begin{equation}\label{H19}
\chi=E\varphi^*=-i\tau_2\varphi^*=\left(\begin{array}{r}
-q_3+iq_4\\q_1-iq_2\end{array}\right),
\end{equation}
the transformation $q\to\gamma^0 q$ corresponds to
\begin{equation}\label{H20}
\varphi\to-i\chi~,~\chi\to -i\varphi.
\end{equation}
We may now introduce the complex four component vector
\begin{equation}\label{H21}
\Psi_M=\left(\begin{array}{c}
\varphi\\\chi\end{array}\right),
\end{equation}
for which all matrix multiplications $q\to \gamma^\mu q$ can be represented by multiplication with complex matrices of the Clifford algebra,
\begin{eqnarray}\label{H22}
&&\gamma^0=\left(\begin{array}{cc}0,&-i\\-i,&0\end{array}\right)~,~
\gamma^k=
\left(\begin{array}{cc}0,&-i\tau_k\\i\tau_k,&0\end{array}\right),\\
&&\Sigma^{0k}=-\frac12\left(\begin{array}{cc}
\tau_k,&0\\0,&-\tau_k
\end{array}\right)~,~
\Sigma^{kl}=-\frac i2\epsilon^{klm}
\left(\begin{array}{cc}
\tau_m,&0\\0,&\tau_m
\end{array}\right).\nonumber
\end{eqnarray}
We can also define the matrix
\begin{equation}\label{H23}
\bar\gamma=-i\gamma^0\gamma^1\gamma^2\gamma^3=
\left(\begin{array}{cc}1,&0\\0,&-1\end{array}\right),
\end{equation}
for which $\varphi$ and $\chi$ are eigenvectors with eigenvalues $\pm 1$. (Often $\bar\gamma$ is denoted as $\gamma^5$.) The transformation $q\to\tilde I q$ acts on $\psi_M$ as $\psi_M\to i\bar\gamma \psi_M$. Thus the representation \eqref{H23}, $\bar\gamma=-i\tilde I$, is consistent with the complex structure. Since $\chi$ is not independent of $\varphi$ the spinor $\Psi_M$ obeys the Majorana constraint \cite{CWMS}, \cite{CWMSab}.
\begin{equation}\label{H24}
B^{-1}\Psi_M^*=\Psi_M~,~B=B^{-1}=-\gamma^2=
\left(\begin{array}{cccc}
0&0&0&1\\0&0&-1&0\\0&-1&0&0\\1&0&0&0
\end{array}\right),
\end{equation}
where $B\gamma^\mu B^{-1}=(\gamma^\mu)^*$.
The real Dirac matrices $\gamma^\mu_{(M)}$ in the ``Majorana representation'' \eqref{F7} and the Dirac matrices $\gamma^\mu_{(W)}$ in the ``Weyl representation'' \eqref{H22} are related by a similarity transformation,
\begin{equation}\label{67A}
q=\frac{1}{\sqrt{2}}A\Psi_M~,~\gamma^\mu_{(W)}=A^{-1}\gamma^\mu_{(M)}A,
\end{equation}
with a unitary matrix $A$
\begin{equation}\label{67B}
A=\frac{1}{\sqrt{2}}
\left(\begin{array}{cccc}
1,&0,&,0,&1\\-i,&0,&0,&i\\0,&1,&-1,&0\\0,&-i,&-i,&0
\end{array}\right)~,~A^\dagger A=1.
\end{equation}
\bigskip\noindent
{\bf 4. Weyl particles}
\medskip
The propagating degrees of freedom correspond to the solutions of the evolution equation. They are most simply discussed in terms of the complex equation \eqref{H16}. We can perform a Fourier transform
\begin{equation}\label{H26}
\varphi(t,x)=\int_p\tilde \varphi(t,p)e^{ipx}=\int\frac{d^3p}{(2\pi)^3}\tilde\varphi
(t,p)e^{ipx},
\end{equation}
such that the evolution equation becomes diagonal in momentum space,
\begin{equation}\label{H27}
i\partial_t\tilde\varphi(t,p)=-p_k\tau_k\tilde\varphi(t,p).
\end{equation}
The general solution of eq. \eqref{H16} obeys
\begin{equation}\label{H28}
\tilde\varphi(t,p)=\exp(ip_k\tau_kt)\tilde \varphi(p),
\end{equation}
with arbitrary complex two-component vectors $\tilde\varphi(p)$. This is the standard time evolution for a propagating Weyl particle.
\bigskip\noindent
{\bf 5. Multi-fermion states}
\medskip
States with arbitrary $n$ describe systems of $n$ fermions. This is not surprising in view of our translation of the classical statistical ensemble to a Grassmann functional integral. For $g_0=|0\rangle$ there are only $n$-particle states, and no hole states. They can be constructed by applying $n$ creation operators $a^\dagger_\gamma(x)$ on the vacuum. For example, the two-fermion state obeys
\begin{equation}\label{H33}
g_2(t)=\frac{1}{\sqrt{2}}\int_{x,y}q_{\gamma\epsilon}(t,x,y)a^\dagger_\gamma(x)
a^\dagger_\epsilon(y)g_0.
\end{equation}
Due to the anticommutation relation
\begin{equation}\label{H34}
\big\{a^\dagger_\gamma(x)~,~a^\dagger_\epsilon(y)\big\}=
\left\{\frac{\partial}{\partial\psi_\gamma(x)}~,~
\frac{\partial}{\partial\psi_\epsilon(y)}\right\}=0
\end{equation}
the two-particle wave function is antisymmetric, as appropriate for fermions
\begin{equation}\label{H35}
q_{\gamma\epsilon}(t,x,y)=-q_{\epsilon\gamma}(t,y,x).
\end{equation}
The Grassmann elements $g_2$ describe states with two Weyl spinors. A particular class of states can be constructed as products of appropriately normalized one particle states $q^{(1)},q^{(2)}$
\begin{equation}\label{H36}
q_{\gamma\epsilon}(x,y)=q^{(1)}_\gamma(x)q^{(2)}_\epsilon(y)-q^{(1)}_\epsilon(y)
q^{(2)}_\gamma(x).
\end{equation}
General two particle states are superpositions of such states.
\bigskip\noindent
{\bf 6. Symmetries}
\medskip
Besides Lorentz symmetry the action \eqref{F1} is also invariant under global $SO(2)$ rotations
\begin{eqnarray}\label{71A}
\psi'_1&=&\cos\alpha~\psi_1-\sin\alpha~\psi_2~,~\psi'_2
=\sin\alpha~\psi_1+\cos\alpha~\psi_2,\nonumber\\
\psi'_3&=&\cos\alpha~\psi_3-\sin\alpha~\psi_4~,~\psi'_4
=\sin\alpha~\psi_3+\cos\alpha~\psi_4.\nonumber\\
\end{eqnarray}
This is easily seen from the infinitesimal transformation
\begin{equation}\label{71B}
\delta\psi_\gamma=\alpha\tilde I_{\gamma\delta}\psi_\delta
\end{equation}
and the relations $\tilde I^2=-1,~\tilde I^T=-\tilde I,~[\tilde I,T_k]=0$. These rotations carry over to the one-particle and one-hole wave functions $q_\gamma$ and $\hat q_\gamma$.
Correspondingly, the complex two component wave functions $\varphi$ and $\chi$ transform as
\begin{equation}\label{71C}
\varphi'=e^{i\alpha}\varphi~,~\chi'=e^{-i\alpha}\chi.
\end{equation}
The $SO(2)$ rotations are now realized as $U(1)$ phase rotations. If we define for a general complex field $\eta$ the charge $\bar Q$ by the transformation
\begin{equation}\label{71D}
\eta'=e^{i\alpha \bar Q}\eta
\end{equation}
we infer that $\varphi$ carries charge $\bar Q=1$, while $\chi$ has opposite charge $-1$. If $\varphi$ describes degrees of freedom of an electron, $\chi$ describes the corresponding ones for a positron. We note that charge eigenstates exist only in connection with a complex structure. The real wave function $q_\gamma$ can be encoded both in $\varphi$ and $\chi$ and may therefore describe degrees of freedom with opposite charge.
The action \eqref{F1} is further invariant with respect to discrete symmetries. Among them a parity type reflection maps
\begin{equation}\label{85Aa}
\psi(x)\to\bar{\cal P}\psi(x),
\end{equation}
with
\begin{eqnarray}\label{85Ba}
\big (\bar{\cal P}\psi(x)\big)_\gamma&=&(\gamma^0)_{\gamma\delta}\psi_\delta(-x),
\end{eqnarray}
and we note $\bar{\cal P}^2=-1$. We may associate $\bar{\cal P}$ with a CP-transformation for Weyl spinors. The action \eqref{F1} is further invariant under the discrete transformation $\psi\to\tilde I\psi$. It is therefore also invariant under a parity type transformation where $\bar{\cal P}$ is replaced by $\tilde I\bar {\cal P}$.
\section{Complex structure}
\label{Complex structure}
We have formulated the classical statistical description of a quantum field theory for Majorana spinors in terms of a real Grassmann algebra. All quantities in the functional integral \eqref{N9} and the action \eqref{F1} or \eqref{N1}-\eqref{B17} are real. The introduction of complex Weyl spinors in eq. \eqref{H14} or \eqref{H19} reveals the presence of a complex structure in this real formulation. Such complex structures are the basis for the importance of phases in quantum mechanics. The Schr\"odinger equation for a one-particle state can be formulated in terms of a complex wave function, and this generalizes to multi-particle states.
\bigskip\noindent
{\bf 1. Complex Grassmann variables}
\medskip
Complex Grassmann variables may be introduced in analogy to eq. \eqref{H14}
\begin{equation}\label{71LA}
\zeta_1=\frac{1}{\sqrt{2}}(\psi_1+i\psi_2)~,~
\zeta_2=\frac{1}{\sqrt{2}}(\psi_3+i\psi_4).
\end{equation}
Together with the complex conjugate Grassmann variables
\begin{equation}\label{158A}
\zeta^*_1=\frac{1}{\sqrt{2}}(\psi_1-i\psi_2)~,~\zeta^*_2=\frac{1}{\sqrt{2}}
(\psi_3-i\psi_4)
\end{equation}
we have for every $x$ and $t$ four independent Grassmann variables $\zeta_1,\zeta^*_1,\zeta_2,\zeta^*_2$, which replace $\psi_1,\psi_2,\psi_3,\psi_4$.
A general element of a complex Grassmann algebra can be expanded as
\begin{eqnarray}\label{71M}
g_c&=&\sum_{k,l}c_{\alpha_1\dots \alpha_k,\bar\alpha_1\dots\bar\alpha_l}
(x_1\dots x_k,\bar x_1\dots \bar x_l)\nonumber\\
&&\zeta_{\alpha_1}(x_1)\dots\zeta_{\alpha_k}(x_k)\zeta^*_{\bar\alpha_1}(\bar x_1)\dots
\zeta^*_{\bar\alpha_l}(\bar x_l),
\end{eqnarray}
with complex coefficients $c$. For a real Grassmann algebra the coefficients $c$ are restricted by $g^*=g$. From a general complex $g_c$ we can obtain an element of a real Grassmann algebra as
\begin{equation}\label{87A}
g=\frac12(g_c+g^*_c).
\end{equation}
The action \eqref{F1} is an element of a real Grassmann algebra and reads in terms of $\zeta$
\begin{eqnarray}\label{168D}
S&=&\int_{t,x}\big\{\zeta^\dagger(\partial_t-\tau_k\partial_k)\zeta+\zeta^T
(\partial_t-\tau^*_k\partial_k)\zeta^*\big\}\nonumber\\
&=&2\int_{t,x}\zeta^\dagger(\partial_t-\tau_k\partial_k)\zeta.
\end{eqnarray}
Up to the factor $2$, which may be removed by a rescaling of $\zeta$, this is the action for a free Weyl spinor.
On every factor $\zeta$ the action of a matrix multiplication of $\psi$ by $\tilde I$ amounts to a multiplication with $i$. More precisely, we can interpret eq. \eqref{71LA} as a map $\psi\to\zeta[\psi]$ with the property $\zeta[\tilde I\psi]=i\zeta[\psi]$. For the complex conjugate one has $\zeta^*[\tilde I\psi]=-i\zeta^*[\psi]$. For the infinitesimal transformation \eqref{71B} one concludes
\begin{equation}\label{71N}
\delta\big [\zeta_{\alpha_1}(x_1)\dots \zeta^*_{\bar\alpha_l}(\bar x_l)\big ]=
i\alpha\bar Q\zeta_{\alpha_1}(x_1)\dots\zeta^*_{\bar\alpha_l}(\bar x_l),
\end{equation}
where the charge $\bar Q$ counts the number of factors $\zeta$ minus the number of factors $\zeta^*$ for a given term in the expansion \eqref{71M}. In other words, products with $\bar Q_+$ factors $\zeta$ and $\bar Q_-$ factors $\zeta^*$ are charge eigenstates with $\bar Q=\bar Q_+-\bar Q_-$. States with a given $\bar Q$ are degenerate since many different choices of $\bar Q_+,\bar Q_-$ lead to the same $\bar Q$. For a real Grassmann algebra we use eq. \eqref{87A}. Every term in $g_c$ with $\bar Q\neq 0$ is accompanied by a term with opposite charge $-\bar{Q}$ in $g^*_c$. Thus the expansion \eqref{71M} of eq. \eqref{87A} involves ``charge eigenspaces'' with pairs of opposite axial charge.
It will often be convenient to use an extension to a complex Grassmann algebra, where the coefficients $c$ in eq. \eqref{71M} are arbitrary and charge eigenstates belong to the Grassmann algebra also for $\bar Q\neq 0$. This can be mapped at the end to a real Grassmann algebra by eq. \eqref{87A}. The action of $\bar Q_\pm$ on $g_c$ is represented as
\begin{equation}\label{71O}
\bar Q_+=\int_y\zeta_\alpha(y)\frac{\partial}{\partial\zeta_\alpha(y)}~,~
\bar Q_-=\int_y\zeta^*_\alpha(y)\frac{\partial}{\partial\zeta^*_\alpha(y)}.
\end{equation}
The Grassmann derivatives $\partial/\partial \zeta_\alpha$ obey the standard anticommutation relations
\begin{eqnarray}\label{104A}
\left\{\frac{\partial}{\partial\zeta_\alpha(x)},\zeta_\beta(y)\right\}&=&
\delta_{\alpha\beta}\delta(x-y)~,~
\left\{\frac{\partial}{\partial\zeta_\alpha(x)},\zeta^*_\beta(y)\right\}=0.\nonumber\\
\end{eqnarray}
In the complex basis we can write the time evolution equation in the form
\begin{equation}\label{71Q}
i\partial_t g_c={\cal H}g_c
\end{equation}
with
\begin{equation}\label{71R}
{\cal H}=i\int_x\left[\frac{\partial}{\partial\zeta_\alpha}
(\tau_k\partial_k\zeta)_\alpha+\frac{\partial}{\partial\zeta^*_\alpha}
(\tau^*_k\partial_k\zeta^*)_\alpha\right].
\end{equation}
Both $\bar Q_+$ and $\bar Q_-$ commute with ${\cal H}$. The operators creating one hole or one particle states read
\begin{eqnarray}\label{93A}
\hat q_\gamma\psi_\gamma&=&\frac{1}{\sqrt{2}}
(\hat\varphi^*_\alpha\zeta_\alpha+\hat\varphi_\alpha\zeta^*_\alpha),\nonumber\\
q_\gamma\frac{\partial}{\partial\psi_\gamma}&=&\frac{1}{\sqrt{2}}
\left(\varphi_\alpha\frac{\partial}{\partial\zeta_\alpha}+\varphi^*_\alpha
\frac{\partial}{\partial\zeta^*_\alpha}\right).
\end{eqnarray}
The infinitesimal Lorentz transformations of the complex two-component spinors $\zeta,\delta\zeta=-\frac12\epsilon_{mn}\Sigma^{mn}\zeta$, are represented by the complex $2\times 2$ matrices $\Sigma^{mn}$ given by eq. \eqref{H17}. We also observe that $\tilde\zeta=E\zeta=-i\tau_2\zeta$ transforms as $\delta\tilde\zeta=\frac12\epsilon_{mn}(\Sigma^{mn})^T\tilde\zeta$ such that
\begin{equation}\label{93B}
\frac12\tilde\zeta^T\zeta=\zeta_1\zeta_2=\frac12\big[\psi_1\psi_3-\psi_2\psi_4+i
(\psi_1\psi_4+\psi_2\psi_3)\big]
\end{equation}
is a Lorentz scalar. The same holds for $\zeta^*_1\zeta^*_2$ such that the bilinears $\psi_1\psi_3-\psi_2\psi_4$ and $\psi_1\psi_4+\psi_2\psi_3$ are separately Lorentz scalars. Also the product
\begin{equation}\label{93C}
\zeta_1\zeta_2\zeta^*_1\zeta^*_2=\psi_1\psi_2\psi_3\psi_4=
\frac{1}{24}\epsilon^{\alpha\beta\gamma\delta}
\psi_\alpha\psi_\beta\psi_\gamma\psi_\delta
\end{equation}
is a Lorentz scalar. The functional measure obeys
\begin{eqnarray}\label{93D}
&&\int d \psi_4 d\psi_3 d\psi_2 d\psi_1=\prod_\gamma d\psi_\gamma\nonumber\\
&&=\int d\zeta_2 d\zeta_1\int d\zeta^*_2 d\zeta^*_1=
\int d \zeta d\zeta^*.
\end{eqnarray}
\bigskip\noindent
{\bf 2. General complex structure}
\medskip
A real Grassmann algebra will, in general, admit different possible complex structures. A discussion of phases in quantum mechanics needs a specific choice of the complex structure. For example, our discussion of Dirac-spinors in sect. \ref{Diracfermions} will employ a complex structure different from eq. \eqref{71LA}. We therefore briefly discuss the general properties of complex structures.
In a real even-dimensional vector space a complex structure is given by the existence of an involution $K$, together with a map $I$, obeying
\begin{equation}\label{186A}
K^2=1~,~I^2=-1~,~\{K,I\}=0.
\end{equation}
The matrix $K$ has eigenvalues $\pm 1$ and we may denote eigenstates with positive eigenvalues by $v_R$ and those with negative ones by $v_I,Kv_R=v_R, Kv_I=-v_I$. The matrix $I$ is a map between $v_R$ and $v_I$, implying that the number of independent $v_R$ and $v_I$ are equal. We can choose the $v_I$ such that $Iv_R=v_I,Iv_I=-v_R$ and use these properties for defining a map from the real vectors $v$ to complex vectors $c=v_R+iv_I=c(v)$, with the properties $c(Kv)=\big[c(v)\big]^*$, $c(Iv)=ic(v)$. A linear operator or observable $A$ can be represented by multiplication with a complex matrix if $[A,I]=0$.
As an example, we consider the complex structure underlying eq. \eqref{71LA}. The maps \eqref{H14} or \eqref{71LA} act on $v=\{q_\gamma\}$ or $v=\{\psi_\gamma\}$, and the matrices $K,I$ are given by
\begin{equation}\label{186B}
K=\tilde K=\left(\begin{array}{rr}
\tau_3,&0\\0,&\tau_3
\end{array}\right)~,~I=\tilde I.
\end{equation}
Indeed, the map $\zeta[\psi]$ given by eq. \eqref{71LA} obeys
\begin{equation}\label{186C}
\zeta[\tilde K\psi]=\zeta^*[\psi]~,~\zeta[\tilde I\psi]=i\zeta[\psi].
\end{equation}
A transformation $\psi\to A\psi$ is compatible with this complex structure if $A$ obeys
\begin{equation}\label{186D}
[A,\tilde I]=0.
\end{equation}
In this case $A$ can be represented by complex matrix multiplication acting on $\zeta$,
\begin{equation}\label{186E}
\zeta[A\psi]=\tilde A\zeta[\psi].
\end{equation}
The matrices $T_k$ commute with $\tilde I$ and are therefore compatible with the complex structure
\begin{equation}\label{186F}
\zeta[T_k\psi]=\tau_k\zeta[\psi].
\end{equation}
Also $\tilde I$ and $\Sigma^{\mu\nu}$ (cf. eq. \eqref{F5}) are compatible with the complex structure, where the action of $\Sigma^{\mu\nu}$ on $\zeta$ is given by eq. \eqref{H17}. The matrices $A$ obeying eq. \eqref{186D} form a group, since the product of two matrices $A_1A_2$ again commutes with $\tilde I$. This product is represented by complex matrix multiplication of $\tilde A_1$ and $\tilde A_2$, $\zeta[A_1A_2\psi]=\tilde A_1\tilde A_2\zeta[\psi]$. In contrast, the matrices $\gamma^\mu$ anticommute with $\tilde I$, cf. eq. \eqref{F12}, and are therefore not compatible with the complex structure \eqref{186B}. One finds
\begin{equation}\label{186G}
\zeta[\gamma^0\psi]=-\tau_2\zeta^*.
\end{equation}
The complex structure \eqref{186A}, \eqref{186B} for the Grassmann variables can be extended to a complex structure for the real Grassmann algebra by defining a suitable map $g\to g_c$.
\section{Massive Majorana Fermions}
\label{Massive Majorana Fermions}
The action \eqref{F1} describes massless Majorana or Weyl spinors. For a massive Majorana spinor one adds a mass term
\begin{eqnarray}\label{D1a}
S_m&=&\int_{t,x}\bar\psi(m\tilde I-\tilde m)\psi\nonumber\\
&=&\int_{t,x}\psi_\gamma\big[m(\gamma^0\tilde I)_{\gamma\delta}-\tilde m(\gamma^0)_{\gamma\delta}\big]\psi_\delta.
\end{eqnarray}
In the presence of the mass term \eqref{D1a} the action remains Lorentz-invariant, anti-hermitean and real (for real $m,\tilde m$), and the Minkowski action $S_M=iS$ is hermitean. For discrete time steps the first factor $\psi$ is taken at $t$, and the second at $t+\epsilon$. The $U(1)$-rotations \eqref{71A} do not leave $S_m$ invariant. Indeed, the infinitesimal transformation \eqref{71B} results in
\begin{equation}\label{D3}
\tilde M\to \tilde M+\delta\tilde M~,~\delta\tilde M=-\alpha[\tilde I,\tilde M]=
2\alpha\tilde M\tilde I,
\end{equation}
where
\begin{equation}\label{D2}
\tilde M_{\gamma\delta}=m(\gamma^0\tilde I)_{\gamma\delta}-\tilde m(\gamma^0)_{\gamma\delta}.
\end{equation}
The terms $\sim m$ and $\tilde m$ are rotated into each other,
\begin{equation}\label{D4}
\delta m=-2\alpha\tilde m~,~\delta\tilde m=2\alpha m.
\end{equation}
We may use this transformation in order to choose a convention where $\tilde m=0$. With respect to the parity transformation $\bar{\cal P}$ \eqref{85Ba} the term $\sim \tilde m$ is invariant, while $m$ changes sign. On the other hand, $S_m$ is odd with respect to the discrete transformation $\psi\to\tilde I\psi$. Thus, with respect to the parity transformation $\tilde I\bar{\cal P}$ one finds that $m$ is invariant while $\tilde m$ changes sign.
\bigskip\noindent
{\bf 1. Evolution equation}
\medskip
For massive Majorana spinors the Grassmann time evolution \eqref{P46} obeys
\begin{equation}\label{D5}
{\cal K}=\int_x\frac{\partial}{\partial\psi_\gamma(x)}
\big\{(T_k)_{\gamma\delta}\partial_k-\tilde M_{\gamma\delta}\big\}\psi_\delta(x)={\cal K}_0+{\cal K}_m,
\end{equation}
with
\begin{equation}\label{D6}
{\cal K}_m=-\int_x\frac{\partial}{\partial\psi_\gamma(x)}\tilde M_{\gamma\delta}
\psi_\delta(x).
\end{equation}
The particle number ${\cal N}$ remains conserved, while the charge $\bar Q$ is no longer a conserved quantity in the presence of a mass term.
We may consider general operators of the type
\begin{equation}\label{H1}
{\cal B}_{\epsilon\eta}(y)=\frac{\partial}{\partial\psi_\epsilon(y)}~B_{\epsilon\eta}(y)
\psi_\eta(y),
\end{equation}
with $B_{\epsilon\eta}$ depending on $y$ and derivatives with respect to $y$. They obey the commutation relation
\begin{eqnarray}\label{H2}
&&\big[{\cal B}_{\epsilon\eta}(y),{\cal K}_0\big]=\partial_k
\left\{\frac{\partial}{\partial\psi_\gamma}(T_k)_{\gamma\epsilon}B_{\epsilon\eta}
\psi_\eta\right\}\\
&&\quad +\frac{\partial}{\partial\psi_\gamma}\big\{B_{\gamma\epsilon}
(T_k)_{\epsilon\eta}\partial_k\psi_\eta-(T_k)_{\gamma\epsilon}\partial_k(B_{\epsilon\eta}\psi_\eta)\big\},\nonumber
\end{eqnarray}
where all quantities on the r.h.s. depend on $y_k$ and $\partial_k=\partial/\partial y_{k}$. For the mass contribution one finds
\begin{equation}\label{D7}
[{\cal B}_{\epsilon\eta}(y),{\cal K}_m]=\frac{\partial}{\partial\psi_\gamma(y)}
\big[\tilde M,B(y)\big]_{\gamma\delta}\psi_\delta(y).
\end{equation}
The momentum operator \eqref{H16} reads in the real Grassmann algebra
\begin{equation}\label{H4}
{\cal P}_k=-\int_y\frac{\partial}{\partial\psi}\tilde I\partial_k\psi.
\end{equation}
For $\tilde M\neq 0$ it is no longer conserved, while we may define conserved quantities involving an even number of derivatives
\begin{equation}\label{D7A}
{\cal P}^2_k=-\int_y\frac{\partial}{\partial\psi_\gamma}\partial^2_k\psi_\gamma.
\end{equation}
For any static vacuum state $g_0,{\cal K}g_0=0$, the evolution of the one-particle state \eqref{H7a} follows eq. \eqref{H11}
\begin{equation}\label{D8}
\int_x\partial_t q_\gamma(x)\frac{\partial}{\partial\psi_\gamma(x)}g_0=\int_x q_\gamma(x)
\left[{\cal K},\frac{\partial}{\partial\psi_\gamma(x)}\right]g_0.
\end{equation}
With
\begin{equation}\label{D9}
\left[{\cal K}_m,\frac{\partial}{\partial\psi_\gamma(x)}\right]=-
\frac{\partial}{\partial\psi_\delta(x)}\tilde M_{\delta\gamma}=\tilde M_{\gamma\delta}
\frac{\partial}{\partial\psi_\delta(x)}
\end{equation}
one finds the evolution equation
\begin{equation}\label{D10a}
\partial_tq=(T_k\partial_k-\tilde M)q~,~
{\cal D} q=(\gamma^\mu\partial_\mu-M)q=0,
\end{equation}
where
\begin{equation}\label{D11}
M=-\gamma^0\tilde M=m\tilde I-\tilde m.
\end{equation}
For $\tilde m=0$ the relation $\{\gamma^\mu,\tilde I\}=0$ implies
\begin{equation}\label{D12}
{\cal D}^2=\partial^\mu\partial_\mu-m^2.
\end{equation}
By using
\begin{equation}\label{D13}
\big[{\cal K}_m,\psi_\gamma(x)\big]=\tilde M_{\gamma\delta}\psi_\delta(x)
\end{equation}
we find for the one-hole wave function $\hat q$ \eqref{H8a} the same evolution equation as for the one-particle wave function $q$,
\begin{equation}\label{D14}
{\cal D}\hat q=0.
\end{equation}
It is instructive to write the one-particle evolution equation \eqref{D10a} in terms of the two-component complex Weyl spinor $\varphi$ defined by eq. \eqref{H14}
\begin{equation}\label{D15}
\partial_t\varphi=\tau_k\partial_k\varphi-i\tau_2(m-i\tilde m)\varphi^*.
\end{equation}
The fact that $\varphi^*$ appears shows that the multiplication with $\tilde M$ is not compatible with the complex structure \eqref{186B}. In terms of $\chi$ \eqref{H19} and $\psi_M$ \eqref{H21} one finds
\begin{eqnarray}\label{D16}
\partial_t\varphi&=&\tau_k\partial_k\varphi+(m-i\tilde m)\chi,\nonumber\\
\partial_t\chi&=&-\tau_k\partial_k\chi-(m+i\tilde m)\varphi,
\end{eqnarray}
and
\begin{eqnarray}\label{D17}
&&\partial_t\psi_M=\gamma^0_{(W)}(\gamma^k_{(W)}
\partial_k-im\bar\gamma_{(W)}+\tilde m)\psi_M,\nonumber\\
&&(i\gamma^\mu_{(W)}\partial_\mu+m\bar\gamma_{(W)}+i\tilde m)\psi_M=0,
\end{eqnarray}
with $\gamma^\mu_{(W)}$ the Dirac matrices in the Weyl representation \eqref{H22}.
\bigskip\noindent
{\bf 2. General solution for single massive Majorana
~particle}
\medskip
In order to find the general solution of eq. \eqref{D10a} we expand similar to eq. \eqref{H26}. We may consider modes with a fixed momentum $p$. (As familiar in quantum mechanics, these modes are not normalizable for infinite volume and may be considered as limiting cases of normalizable wave packets.) For every given $p\neq 0$ we can define
\begin{equation}\label{H29}
H(p)=\frac{p_k\tau_k}{|p|}~,~H^2(p)=1~,~p_kp_k=|p|^2,
\end{equation}
such that $H(p)$ has eigenvalues $\pm 1$. Decomposing $\tilde\varphi(p)$ according to the eigenvalues of $H(p)$,
\begin{equation}\label{30}
H(p)\tilde\varphi_\pm(p)=\mp\tilde\varphi_\pm(p),
\end{equation}
and using
\begin{eqnarray}\label{D19}
\tilde \chi_\pm(t,p)&=&-i\tau_2\tilde\varphi^*_\pm(t,-p),\nonumber\\
H(p)\tilde\chi_\pm(t,p)&=&\mp\tilde\chi_\pm(t,p),
\end{eqnarray}
one has
\begin{eqnarray}\label{D18}
\varphi(t,x)&=&\int_p\big(\tilde\varphi_+(t,p)e^{ipx}+
\tilde \varphi_-(t,p)e^{ipx}\big),\nonumber\\
\chi(t,x)&=&\int_p\big(\tilde \chi_+(t,p)e^{ipx}+\tilde\chi_-(t,p)e^{ipx}\big).
\end{eqnarray}
This yields, for $\tilde m=0$,
\begin{eqnarray}\label{D20}
\partial_t\tilde\varphi_\pm(p)&=&\mp i|p|\tilde\varphi_\pm(p)+m\tilde\chi_\pm(p),\nonumber\\
\partial_t\tilde\chi_\pm(p)&=&\pm i|p|\tilde\chi_\pm(p)-m\tilde\varphi_\pm(p),
\end{eqnarray}
with general solution
\begin{eqnarray}\label{D21}
\tilde\varphi_+(t,p)&=&b_+(p)e^{-i\omega t}-
\frac{\omega-|p|}{m}\tau_2b^*_+(-p)e^{i\omega t},\nonumber\\
\tilde\chi_+(t,p)&=&-i\tau_2b^*_+(-p)e^{i\omega t}-i
\frac{\omega-|p|}{m}b_+(p)e^{-i\omega t},\nonumber\\
\tilde\varphi_-(t,p)&=&b_-(p)e^{i\omega t}+
\frac{\omega-|p|}{m}\tau_2 b^*_-(-p)e^{-i\omega t},\\
\tilde\chi_-(t,p)&=&-i\tau_2 b^*_-(-p)e^{-i\omega t}+i
\frac{\omega-|p|}{m}b_-(p)e^{i\omega t},\nonumber
\end{eqnarray}
where
\begin{equation}\label{D22}
\omega=\omega(p)=+\sqrt{|p|^2+m^2}.
\end{equation}
\bigskip\noindent
{\bf 3. Classical probabilities for Majorana fermions}
\medskip
The general solution for $\varphi(t,x)$ depends on four complex functions contained in $b_\pm(p)$. From the real and imaginary part of $\varphi(t,x)$ we can infer the four real functions $q_\gamma(x)$ and therefore the classical probability distributions which describe the one-particle states. As an example, we consider the special solution $b_-(p)=0,b_{+,2}(p)=0,b_{+,1}(p)=f(p)$ real, for which
\begin{eqnarray}\label{D23}
\varphi_1(t,x)&=&\int_pf(p)e^{i(px-\omega t)},\nonumber\\
\varphi_2(t,x)&=&\int_p-i\frac{\omega-|p|}{m}f(-p)e^{i(px+\omega t)},
\end{eqnarray}
or
\begin{eqnarray}\label{D24}
q_1(t,x)&=&\int_p f(p)\cos (px-\omega t),\nonumber\\
q_2(t,x)&=&\int_p f(p)\sin (px-\omega t),\nonumber\\
q_3(t,x)&=&-\int_p f(p)\frac{\omega-|p|}{m}\sin(px-\omega t),\nonumber\\
q_4(t,x)&=&-\int_p f(p)\frac{\omega-|p|}{m}\cos(px-\omega t).
\end{eqnarray}
We may further take $f(p)=f\delta(p-\hat p)$ with $\hat p=(0,0,-P),P>0$, such that, with $\omega=\sqrt{P^2+m^2}$,
\begin{eqnarray}\label{D25}
q_1(t,x)&=&f\cos(Px_3+\omega t)~,~q_2(t,x)=-f\sin (Px_3+\omega t),\nonumber\\
q_3(t,x)&=&f\frac{\omega-P}{m}\sin (Px_3+\omega t),\nonumber\\
q_4(t,x)&=&-f\frac{\omega-P}{m}
\cos (Px_3+\omega t).
\end{eqnarray}
For this particular example the normalization $\int_x q^2_\gamma(x)=1$ requires (with $V$ the space-volume)
\begin{equation}\label{D26}
f^2=\frac{1}{2V}\frac{m^2}{m^2+P^2-P\sqrt{P^2+m^2}}.
\end{equation}
For the general solution \eqref{D24} the probabilities $p_\gamma(t,x)=q^2_\gamma(t,x)$ describe the time dependent probability distribution of a classical statistical ensemble which accounts for all properties of a propagating massive single Majorana fermion. If the vacuum is the totally empty state $g_0=|0\rangle$ the states $(\gamma,x)$ correspond to ``Ising states'' $\tau$ where only one bit at $x$ of species $\gamma$ is occupied and all others are empty. (For a different $g_0$ the classical states associated to $(\gamma,x)$ are more complicated.) The probability of a state $(\gamma,x)$ is given by $p_\gamma(x)= q^2_\gamma(x)$, e.g. according to eq. \eqref{D24} with a proper normalization of $f(p)$.
The probability distribution $\big \{p_\gamma(x)\big\}$ properly accounts for all characteristic quantum interference phenomena for modes with different momenta $p$. This extends to the correct description of interference for multi-particle states. (For a concrete example of a two-particle state for massless Majorana spinors cf. ref. \cite{CWMSa}.) The appearance of interference phenomena is not surprising in view of the formulation of the time evolution in terms of a Grassmann functional integral. On a deeper level, it is related to the fact that simple time evolution equations \eqref{35B} are linear in the classical wave function $q_\tau$, but not linear in the classical probabilities $p_\tau$.
\section{Dirac fermions}
\label{Diracfermions}
Dirac spinors can be composed of two different Majorana spinors with equal mass. The action reads
\begin{eqnarray}\label{E1}
S&=&\int_{t,x}\big\{\psi_1(\partial_t-T_k\partial_k+m\gamma^0\tilde I)\psi_1\nonumber\\
&&+\psi_2(\partial_t-T_k\partial_k+m\gamma^0\tilde I)\psi_2\big\},
\end{eqnarray}
with $T_k,\tilde I,\gamma^0$ given by eqs. \eqref{18AA}, \eqref{18ABa}, \eqref{F7}. (For simplicity we take $\tilde m=0$ in eq. \eqref{D1a}.) The functional integral extends now over the Grassmann variables $\psi_{1,\gamma}(t,x)$ and $\psi_{2,\gamma}(x,t)$. For a free theory the action involves separate pieces for $\psi_1$ and $\psi_2$, $S=S_1+S_2$. If the initial state $g(t_0)=g_{in}$ factorizes into two factors, one involving only $\psi_1$ and the other only $\psi_2$, the Grassmann wave function $g(t)$ factorizes for all $t$. There is no need, however, for $g(t_0)$ to factorize and general states $g(t)$ will not be of a factorisable form.
\bigskip\noindent
{\bf 1. Complex structure}
\medskip
We introduce a complex structure by defining the four-component complex Grassmann variables $\psi_D$
\begin{equation}\label{E2}
\psi_D=\psi_1+i\psi_2~,~\psi^*_D=\psi_1-i\psi_2.
\end{equation}
In terms of the general discussion in sect. \ref{Complex structure} the matrices $K$ and $I$ act as
\begin{equation}\label{E3}
K{\psi_1\choose \psi_2}={~~\psi_1\choose -\psi_2}~,~
I{\psi_1\choose\psi_2}={-\psi_2\choose ~~\psi_1},
\end{equation}
such that $\psi_1$ is the ``real part'' of $\psi_D$, and $\psi_2$ the ``imaginary part''. The complex structure employed for Dirac spinors differs from the complex structure \eqref{186B}. We can express the action \eqref{E1} in terms of $\psi_D$ as
\begin{eqnarray}\label{E4}
S&=&\int_{t,x}\psi^\dagger_D(\partial_t-T_k\partial_k+m\gamma^0
\tilde I)\psi_D\nonumber\\
&=&-\int_{t,x}\bar\psi_D(\gamma^\mu\partial_\mu-m\tilde I)\psi_D,
\end{eqnarray}
where we define
\begin{equation}\label{E5}
\bar\psi_D=\psi^\dagger_D\gamma^0.
\end{equation}
The matrices $\gamma^\mu$ and $\tilde I$ appearing in eqs. \eqref{E4}, \eqref{E5} are all real. We may also introduce a purely imaginary matrix
\begin{equation}\label{E6}
\bar\gamma=-i\gamma^0\gamma^1\gamma^2\gamma^3=-i\tilde I~,~\bar\gamma^2=1.
\end{equation}
It is hermitean and anticommutes with all Dirac matrices $\gamma^\mu$,
\begin{equation}\label{E7}
\bar\gamma^\dagger=\bar\gamma~,~\{\bar\gamma,\gamma^\mu\}=0~,~
[\bar\gamma,\Sigma^{\mu\nu}]=0.
\end{equation}
Often $\bar\gamma$ is denoted as $\gamma^5$. On the level of the eight variables $\psi_{1,\gamma}$ and $\psi_{2,\gamma}$ the matrix $\bar\gamma$ is defined as a real $8\times 8$ matrix, with $I_2$ not acting on the index $\gamma$,
\begin{equation}\label{E8}
\bar\gamma=-I_2\otimes\tilde I~,~I_2={0,-1\choose 1,~~0}.
\end{equation}
(The matrix $I$ defining the complex structure is given by $I_2$ according to eq. \eqref{E3}.)
Since $\bar\gamma$ commutes with the Lorentz-generators $\Sigma^{\mu\nu}$ the projections
\begin{equation}\label{E8A}
\psi_L=\frac12(1+\bar\gamma)\psi_D~,~\psi_R=\frac12(1-\bar\gamma)\psi_D
\end{equation}
are representations of the Lorentz group. They transform as the two inequivalent fundamental spinor representations and describe Weyl spinors. In terms of $\psi_{1,\gamma},\psi_{2,\gamma}$ one has
\begin{equation}\label{E8B}
\psi_{L,R}=\frac12
\left(\begin{array}{c}
\psi_{1,1}\mp\psi_{2,2}+i\psi_{2,1}\pm i\psi_{1,2}\\
\psi_{1,2}\pm\psi_{2,1}+i\psi_{2,2}\mp i\psi_{1,1}\\
\psi_{1,3}\mp\psi_{2,4}+i\psi_{2,3}\pm i\psi_{1,4}\\
\psi_{1,4}\pm\psi_{2,3}+i\psi_{2,4}\mp i\psi_{1,3}
\end{array}\right),
\end{equation}
where the upper sign relates to $\psi_L$ and the lower to $\psi_R$. We observe that both $\psi_L$ and $\psi_R$ receive contributions both from $\psi_1$ and $\psi_2$. They differ from the Weyl spinors $\zeta_1,\zeta_2$ that can be formed from $\psi_1$ and $\psi_2$ according to eq. \eqref{71LA}.
Using $\bar\gamma$ we can write the action \eqref{E4} as
\begin{eqnarray}\label{E9}
S&=&-\int_{t,x}\bar\psi_D(\gamma^\mu\partial_\mu- im\bar\gamma)\psi_D=-iS_M,\nonumber\\
S_M&=&-\int_{t,x}\bar\psi_D(i\gamma^\mu\partial_\mu+m\bar\gamma)\psi_D.
\end{eqnarray}
Defining the hermitean conjugation as a map $\psi_D\to \psi^*_D$, accompanied by a complex conjugation of all coefficients in the Grassmann algebra and a total transposition (reordering) of all Grassmann variables, the hermiticity of the Minkowski action, $S^\dagger_M=S_M$, is easily verified.
\bigskip\noindent
{\bf 2. Charge and electromagnetic fields}
\medskip
The action \eqref{E1} is invariant under all symmetry transformations defined in the preceding sections, i.e. those leaving $S_1$ or $S_2$ separately invariant. There are further symmetries related to transformations between $\psi_1$ and $\psi_2$. Rotations between $\psi_1$ and $\psi_2$ can be accounted for by the infinitesimal transformation
\begin{equation}\label{E10}
\delta\tilde\psi=\delta{\psi_1\choose\psi_2}=-\beta I_2\tilde \psi=
{~~\beta\psi_2\choose -\beta\psi_1}.
\end{equation}
The discretization is the same for $\psi_1$ and $\psi_2$. It therefore respects the invariance with respect to the transformation \eqref{E10}. This rotation can be transferred to a phase rotation of the complex Dirac spinor
\begin{equation}\label{E11}
\delta\psi_D=-i\beta\psi_D.
\end{equation}
We associate the transformations \eqref{E10}, \eqref{E11} with the $U(1)$-transformations of electromagnetism. The Dirac spinor carries charge $Q=-1$ and can be identified with the electron. Both components $\psi_L$ and $\psi_R$ \eqref{E8A} transform as Weyl spinors with the same electric charge $Q$.
We may extend the action by considering Dirac spinors in an external electromagnetic field $A_\mu(t,x)$. As usual this is done by replacing the derivatives $\partial_\mu$ in eq. \eqref{E9} by covariant derivatives $D_\mu=\partial_\mu+ieA_\mu$. This adds to the action a piece
\begin{equation}\label{E12}
\Delta S=-ie\int_{t,x}\bar\psi_D\gamma^\mu A_\mu\psi_D.
\end{equation}
In terms of the real Grassmann algebra involving $\psi_1$ and $\psi_2$ the additional piece reads
\begin{equation}\label{E13}
\Delta S=-e\int_{t,x}\big\{\psi_1(A_0-A_kT_k)\psi_2
-\psi_2(A_0-A_kT_k)\psi_1\big\}.
\end{equation}
For discrete time steps the first Grassmann variable is taken at $t$ and the second at $t+\epsilon$. In the continuum limit the two terms in eq. \eqref{E13} are equal. We may consider the fields $A_\mu$ as ``sources'' for fermion bilinears involving $\psi_1$ and $\psi_2$. For nonzero $A_\mu$ the action is no longer a sum of independent pieces $S_1+S_2$ for $\psi_1$ and $\psi_2$, respectively.
\bigskip\noindent
{\bf 3. Weyl representation}
\medskip
One may use the similarity transformation \eqref{67A}
\begin{equation}\label{E14}
\psi_{D,W}=A^{-1}\psi_D~,~
\gamma^\mu_{(W)}=A^{-1}\gamma^\mu_{(M)}A,
\end{equation}
in order to bring the Dirac matrices to the Weyl representation \eqref{H22}. The action \eqref{E9}, \eqref{E12} retains its form, with the replacements $\gamma^\mu\to\gamma^\mu_{(W)},~\psi_D\to\psi_{D,W},\bar\psi_D\to\psi^\dagger_{D,W}\gamma^0_{(W)}=\bar\psi_{D,W}$. In the Weyl representation the matrix $\bar \gamma$ is diagonal, $\bar\gamma =diag(1,1,-1,-1)$. The Dirac spinor $\psi_{D,W}$ can therefore be written in terms of two-component spinors $\psi_L$ and $\psi_R$,
\begin{eqnarray}\label{E15}
\psi_{D,W}&=&{\psi_L\choose \psi_R}~,~\psi_L=(\zeta_1+i\zeta_2)~,~\psi_R=
(\xi_1+i\xi_2),\nonumber\\
\zeta_a&=&\frac{1}{\sqrt{2}}
{\psi_{a,1}+i\psi_{a,2}\choose \psi_{a,3}+i\psi_{a,4}},\nonumber\\
\xi_a&=&\frac{1}{\sqrt{2}}
{-\psi_{a,3}+i\psi_{a,4}\choose ~~\psi_{a,1}-i\psi_{a,2}}=-i\tau_2\zeta^*_a,
\end{eqnarray}
which, in turn, can be written as linear combinations of $\zeta_1$ and $\zeta_2$ or $\xi_1$ and $\xi_2$.
We observe that in this basis the operation of complex conjugation is represented by an involution $\hat K$ different from $K$ in eq. \eqref{E3}. In terms of the eight-component field $\tilde \psi=(\psi_1,\psi_2)$ one has $\hat K=diag (1,-1,1,-1,-1,1,-1,1)$. The matrix $I$ remains $I_2$ as in eq. \eqref{E3}, and $I_2$ and $\hat K$ anticommute
\begin{equation}\label{E16}
\hat K={\tilde K,~~0\choose \hspace{0.15cm}0,-\tilde K}~,~
I_2={0,-1\choose 1,~~0}~,~\{\hat K,I_2\}=0.
\end{equation}
Thus the similarity transformation \eqref{E14} changes the complex structure. On the level of a real representation it can be expressed by real $8\times 8$ matrices $A'$ such that a similarity transformation may, in principle, transform both $K$ and $I$. The matrix $I$ remains unchanged, since $iA\psi=A(i\psi)$ in the complex representation. For $A\neq A^*$, however, the matrix $K$ changes since $(A\psi)^*\neq A\psi^*$.
In the Weyl basis \eqref{E15} we can take $\psi_L$ and $\psi_R$ as two-component complex spinors. In terms of $\psi_L$ and $\psi_R$ the action reads
\begin{eqnarray}\label{E17}
S&=&-\int_{t,x}\big\{\bar\psi_L\gamma^\mu(\partial_\mu+ieA_\mu)\psi_L+\bar\psi_R
\gamma^\mu(\partial_\mu+ieA_\mu)\psi_R\nonumber\\
&&-im(\bar\psi_R\psi_L-\bar\psi_L\psi_R)\big\},
\end{eqnarray}
where we define
\begin{eqnarray}\label{E18}
\bar\psi&=&(\bar\psi_R,\bar\psi_L)=-i(\psi^\dagger_R~,~\psi^\dagger_L)\nonumber\\
\bar\psi_L&=&\bar\psi\left(\frac{1-\bar\gamma}{2}\right)~,~\bar\psi_R=\bar\psi
\left(\frac{1+\bar\gamma}{2}\right).
\end{eqnarray}
For $m=0$ it is invariant under separate ``chiral phase rotations'' of $\psi_L$ and $\psi_R$. The transformation $\delta\psi_L=i\alpha\psi_L,\delta\psi_R=-i\alpha\psi_R$ corresponds to $\delta\zeta_a=i\alpha\delta\zeta_a$ and therefore to a simultaneous transformation of the type \eqref{71B} for $\psi_1$ and $\psi_2$.
\bigskip\noindent
{\bf 4. Left-handed representation}
\medskip
We can also reformulate the action in terms of two left handed Weyl spinors $\psi_L$ and $\psi^c_L$ by introducing
\begin{eqnarray}\label{E19}
\psi^c_L&=&i\tau_2\psi^*_R=\zeta_1-i\zeta_2,\nonumber\\
\psi_R&=&-i\tau_2(\psi^c_L)^*~,~\bar\psi_R=(\psi^c_L)^T\tau_2.
\end{eqnarray}
This yields
\begin{eqnarray}\label{E20}
S&=&\int_{t,x}\Big\{\psi^\dagger_L\big[\partial_t+ieA_0
-\tau_k(\partial_k+ieA_k)\big]\psi_L\nonumber\\
&&+(\psi^c_L)^\dagger\big[\partial_t-ieA_0-\tau_k(\partial_k-ieA_k)\big]\psi^c_L\nonumber\\
&&+im\big[(\psi^c_L)^T\tau_2\psi_L+(\psi^c_L)^\dagger\tau_2\psi^*_L\big]\Big\}.
\end{eqnarray}
The charge of $\psi^c_L$ is opposite to $\psi_L$ - if $\psi_L$ describes a left-handed electron, $\psi^c_L$ accounts for the left-handed positron.
We may group $\psi_L$ and $\psi^c_L$ into a four-component complex spinor
\begin{equation}\label{269A}
\psi_{LL}={\psi_L\choose \psi^c_L}={\zeta_1+i\zeta_2\choose \zeta_1-i\zeta_2}.
\end{equation}
In this representation the matrix $\tilde I$ defined by eq. \eqref{18ABa} is represented by a multiplication with $i$,
\begin{equation}\label{269B}
\psi_{LL}[\tilde I\tilde \psi]=i\psi_{LL},
\end{equation}
while the matrix $I_2$, which acts on $(\zeta_1,\zeta_2)$ or $(\psi_1,\psi_2)$ as defined in eq. \eqref{E8}, is now represented as
\begin{equation}\label{269C}
\psi_{LL}[I_2\tilde\psi]={i,~~0\choose 0,-i}\psi_{LL}.
\end{equation}
We observe that the role of $\tilde I$ and $I_2$ is exchanged if we compare the complex spinors $\psi_{LL}$ and $\psi_{D,W}$,
\begin{eqnarray}\label{269D}
\psi_{D,W}[I_2\tilde\psi]&=&i\psi_{D,W},\\
\psi_{D,W}[\tilde I\tilde\psi]&=&{i,~~0\choose 0,-i}\psi_{D,W}
=i\bar\gamma\psi_{D,W}.\nonumber
\end{eqnarray}
The complex structures associated to $\psi_{LL}$ and $\psi_{D,W}$ are therefore different. Nevertheless, the operation of complex conjugation of $\psi_{LL}$ or $\psi_{D,W}$ corresponds in the real representation $\tilde\psi$ to the same matrix multiplication with $\hat K=diag(\tilde K,-\tilde K)$.
\bigskip\noindent
{\bf 5. Enhanced symmetries for free massless Dirac spinors}
\medskip
For $m=0,A_\mu=0$ the free theory exhibits a symmetry of $SU(2)$-rotations among $\psi_L$ and $\psi^c_L$,
\begin{equation}\label{E21}
\delta\psi_{LL}=i\omega_j\hat \tau_j\psi_{LL},
\end{equation}
with $\omega_3=\beta$ identified with the transformation \eqref{E11} and $\hat\tau_j$ the Pauli matrices acting on the two components of $\psi_{2d}$ (not on the spinor components of $\psi_L$ and $\psi^c_L$). For example, the transformation with $\omega_1$ acts as
\begin{eqnarray}\label{E22}
\delta\psi_L&=&-\omega_1\tau_2\psi^*_R~,~\delta\psi_R=-\omega_1\tau_2\psi^*_L,\nonumber\\
\delta\zeta_1&=&i\omega_1\zeta_1~,~\delta\zeta_2=-i\omega_1\zeta_2.
\end{eqnarray}
This corresponds to a transformation of the type \eqref{71B}, now with opposite directions for $\psi_1$ and $\psi_2$. The situation is similar for the transformation $\sim\omega_2$, where
\begin{eqnarray}\label{E23}
\delta\zeta_1&=&-i\omega_2\zeta_2~,~\delta\zeta_2=-i\omega_2\zeta_1,\nonumber\\
\delta\psi_1&=&-\omega_2\tilde I\psi_2~,~\delta\psi_2=-\omega_2\tilde I\psi_1
\end{eqnarray}
leads again to a transformation among $\psi_{1,\gamma}$ and $\psi_{2,\gamma}$ which involves $\tilde I$.
\newpage\noindent
{\bf 6. Evolution equation for Dirac fermions}
\medskip
For the action $S+\Delta S$ \eqref{E1}, \eqref{E13} the Grassmann evolution equation $\partial_t g={\cal K}g$ involves ${\cal K}={\cal K}_0+{\cal K}_m+\Delta{\cal K}$
\begin{eqnarray}\label{X1}
{\cal K}_0+{\cal K}_m&=&\int_x\sum_{a=1,2}\frac{\partial}{\partial\psi_a(x)}
(T_k\partial_k-m\gamma^0\tilde I)\psi_a(x),\nonumber\\
\Delta{\cal K}&=&e\int_x\left[\frac{\partial}{\partial\psi_1(x)}
(A_0(x)-A_k(x) T_k)\psi_2(x)\right.\nonumber\\
&&\left.-\frac{\partial}{\partial\psi_2(x)}
(A_0(x)-A_k(x)T_k)\psi_1(x)\right].
\end{eqnarray}
Here we work in the Majorana basis \eqref{18AA}, \eqref{18ABa}, \eqref{F7} and we have suppressed the spinor index $\gamma$. The evolution equation
\begin{equation}\label{231}
\partial_t g(t)={\cal K} g(t)=({\cal K}_0+{\cal K}_m+\Delta{\cal K})g(t)
\end{equation}
describes the dynamics for an arbitrary number of charged relativistic fermions (and their antiparticles) in external electromagnetic fields.
The evolution equation for the classical wave function \eqref{P35}, and therefore for the probability distribution, obtains from eq. \eqref{P44}. In summary, we have formulated a time evolution equation for a classical wave function $q_\tau(t)$ which describes an arbitrary number of charged electrons in external electric and magnetic fields. In the next section we will see that its restriction to one-particle states yields the relativistic Dirac equation. As usual, a non-relativistic approximation will yield the familiar Schr\"odinger equation for a particle in a potential or moving in external magnetic fields.
\medskip\noindent
{\bf 7. Classical observables}
Classical observables can be constructed from linear combinations of products of occupation numbers. We combine the index $a=1,2$ for the two Majorana spinors $\psi_a$ with the index $\gamma$ into a common index $\epsilon=(\gamma,a),\epsilon=1\dots 8$. The Grassmann operators corresponding to the occupation numbers $N_\epsilon(x)$ read then
\begin{equation}\label{233A}
{\cal N}_\epsilon(x)=\frac{\partial}{\partial\psi_\epsilon(x)}\psi_\epsilon(x).
\end{equation}
In a formulation with discrete lattice points they obey
\begin{equation}\label{FF1}
{\cal N}^2_\epsilon(x)={\cal N}_\epsilon(x)~,~\big[{\cal N}_\epsilon(x),{\cal N}_\eta(y)\big]=0.
\end{equation}
The basis elements $g_\tau$ are eigenstates of ${\cal N}_\epsilon(x)$,
\begin{equation}\label{FF3}
{\cal N}_\epsilon(x)g_\tau=\big(N_\epsilon(x)\big)_\tau g_\tau,
\end{equation}
with $\big(N_\epsilon(x)\big)_\tau=0$ if $g_\tau$ contains a factor $\psi_\epsilon(x)$, and $\big(N_\epsilon(x)\big)_\tau=1$ if no such factor is present. The expectation value of a product of occupation numbers obeys
\begin{eqnarray}\label{FF4}
&&\langle N_{\epsilon_1}(x_1)N_{\epsilon_2}(x_2)\dots N_{\epsilon_m}(x_m)\rangle\nonumber\\
&&=\int{\cal D}\psi(t)\tilde g(t){\cal N}_{\epsilon_1}(x_1)\dots {\cal N}_{\epsilon_m}(x_m)g(t)\nonumber\\
&&=\sum_\tau p_\tau(t)\big(N_{\epsilon_1}(x_1)\big)_\tau\dots\big(N_{\epsilon_m}(x_m)\big)_\tau,
\end{eqnarray}
corresponding to the standard rule for classical statistics. Obviously, the expectation values of classical observables can be computed from the probability distribution $p_\tau(t)$ and do not involve the signs $s_\tau$ in eq. \eqref{35A}. Classical products of classical observables commute.
In order to compute the time evolution of expectation values of classical observables we define
\begin{equation}\label{FF5}
{\cal M}_{\epsilon\eta}(x)=\frac{\partial}{\partial\psi_\epsilon(x)}\psi_\eta(x),
\end{equation}
and use the relation
\begin{eqnarray}\label{FF6}
\big[{\cal M}_{\epsilon\eta}(x),{\cal K}\big]&=&\sum_\alpha\big\{(T_k)_{\epsilon\alpha}\partial_k\
\frac{\partial}{\partial\psi_\alpha(x)}\psi_\eta(x)\nonumber\\
&&+(T_k)_{\eta\alpha}\frac{\partial}{\partial\psi_\epsilon(x)}\partial_k\psi_\alpha(x)\\
&&+W_{\epsilon\alpha}(x){\cal M}_{\alpha\eta}(x)-{\cal M}_{\epsilon\alpha}(x)W_{\alpha\eta}(x)\big\},\nonumber
\end{eqnarray}
with
\begin{equation}\label{FF7}
W_{\epsilon\eta}(x)=-m(\gamma^0\tilde I)_{\alpha\beta}
\delta_{ab}+e\big(A_0(x)\delta_{\alpha\beta}-A_k(x)(T_k)_{\alpha\beta}\big)\epsilon_{ab},
\end{equation}
and $\partial_k=\partial/\partial x_k,\epsilon_{ab}=-\epsilon_{ba},\epsilon_{12}=1,\eta=(\beta,b),(T_k)_{\epsilon\eta}=(T_k)_{\alpha\beta}\delta_{ab}=(T_k)_{\eta\epsilon}$. We observe that $W$ is antisymmetric, $W_{\epsilon\eta}(x)=-W_{\eta\epsilon}(x)$. A local particle number can be defined as
\begin{equation}\label{FF8}
{\cal N}(x)=\sum_\epsilon{\cal N}(x)=\sum_\epsilon{\cal M}_{\epsilon\epsilon}(x),
\end{equation}
and obeys
\begin{equation}\label{FF9}
\big[{\cal N}(x),{\cal K} \big]=(T_k)_{\eta\alpha}\partial_k{\cal M}_{\alpha\eta}.
\end{equation}
According to eq. \eqref{95D} the time evolution of the classical mean local particle number involves expectation values of ``off-diagonal'' operators ${\cal M}_{\alpha\eta}$
\begin{equation}\label{FF10}
\partial_t\langle N(x)\rangle=(T_k)_{\eta\alpha}\partial_k\langle {\cal M}_{\alpha\eta}\rangle.
\end{equation}
The particle number
\begin{equation}\label{FF11}
{\cal N}=\int_x{\cal N}(x)~,~[{\cal N},{\cal K}]=0,
\end{equation}
is conserved.
\section{Quantum mechanics for particle in a potential}
\label{Quantum mechanics for particle in a potential}
\noindent
{\bf 1. Dirac equation}
We next concentrate on one-particle states, described by Grassmann elements
\begin{equation}\label{X2}
g_1(t)=\int_x\left(q_{1,\gamma}(t,x)
\frac{\partial}{\partial\psi_{1,\gamma}(x)}+q_{2,\gamma}(t,x)
\frac{\partial}{\partial\psi_{2,\gamma}(x)}\right)g_0.
\end{equation}
Here $g_0$ is some arbitrary static ``vacuum state''. For ${\cal K}g_0=0$ we find for the one-particle wave function the Dirac equation
\begin{eqnarray}\label{X3}
&&(\gamma^\mu\partial_\mu -m\tilde I) q_1=eA_\mu\gamma^\mu q_2,\nonumber\\
&&(\gamma^\mu\partial_\mu -m\tilde I) q_2=-e A_\mu\gamma^\mu q_1.
\end{eqnarray}
For the derivation of the Dirac equation we use
\begin{eqnarray}\label{X4}
&&\left[\Delta{\cal K}~,~\frac{\partial}{\partial\psi_{1,\gamma}(x)}\right]\nonumber\\
&&=-e\left(A_0(x)\frac{\partial}{\partial\psi_{2,\gamma}(x)}-A_k(x)(T_k)_{\delta\gamma}\frac{\partial}{\partial\psi_{2,\delta}(x)}\right)\nonumber\\
&&\left[ \Delta{\cal K},\frac{\partial}{\partial\psi_{2,\gamma}(x)}\right] \\
&&=e\left(A_0(x)\frac{\partial}{\partial\psi_{1,\gamma}(x)}-A_k(x)(T_k)_{\delta\gamma}\frac{\partial}{\partial\psi_{1,\delta}(x)}\right).\nonumber
\end{eqnarray}
The time evolution equation for the one particle wave function then extends eq. \eqref{H11} to
\begin{eqnarray}\label{276A}
\partial_t q_1&=&(T_k\partial_k -m\gamma^0\tilde I) q_1+e(A_0-A_k T_k)q_2,\nonumber\\
\partial_tq_2&=&(T_k\partial_k -m\gamma^0\tilde I) q_2-e(A_0-A_kT_k)q_1.
\end{eqnarray}
This is equivalent to the Dirac equation \eqref{X3} which can also be written as a matrix equation.
\begin{eqnarray}\label{276B}
&&(\gamma^\mu D_\mu-m\tilde I){q_1\choose q_2}=0,\nonumber\\
&&D_\mu= \partial_\mu+eA_\mu
{0,-1\choose 1,~~0}=\partial_\mu+eA_\mu I_2.
\end{eqnarray}
The usual complex form of the Dirac equation is recovered if we use a complex one-particle wave function
\begin{equation}\label{276C}
\varphi_D=q_1+iq_2~,~\gamma^\mu(\partial_\mu+ieA_\mu)\varphi_D=im\bar\gamma\varphi_D.
\end{equation}
This equation holds for an arbitrary representation of the Dirac matrices $\gamma^\mu$.
The derivation of the Dirac equation for a one-particle state \eqref{X2} has only used the general evolution equation for Grassmann elements $g(t)$ and the condition ${\cal K} g_0=0$. It therefore describes the dynamics for a very extended family of classical probability distributions or classical wave functions. Indeed, it is sufficient that $g_0$ is an arbitrary static state (not necessarily a priori with a fixed particle number). This reflects the physical property that isolated one-particle states can occur under a wide variety of circumstances, and that for sufficient isolation the properties of the environment do not matter for the dynamics of the isolated particle. In this context we emphasize, however, that our model only describes external electromagnetic fields while we do not account for the fields generated by the particles that may be present in the environment. In this sense our setting describes ``real physics'' only in a situation where the electromagnetic fields generated by all present particles can be approximated by a ``mean field'' that is independent of individual particle positions. This is precisely the setting of standard one-particle quantum mechanics.
The presence of electromagnetic fields influences the conditions for a static ``vacuum'' or ``environment'', ${\cal K} g_0=0$. We will concentrate here on simple vacuum states as $g_0=1$ or $g_0=|0\rangle$, where $|0\rangle$ involves a product of all Grassmann variables and obeys ${\cal N}|0\rangle=0$. These states are static for arbitrary electromagnetic fields. For simplicity we choose $g_0=|0\rangle$ such that
\begin{equation}\label{VV1}
{\cal N} g_1=g_1.
\end{equation}
(The following discussion can be easily generalized, however, to all static states $g_0$ which are eigenvalues of ${\cal N},{\cal N} g_0=m\gamma_0$, by using a modified local occupation number where a suitable constant is subtracted from ${\cal N}(x)$.)
\medskip\noindent
{\bf 2. Particle observables}
Let us consider the expectation value of the local occupation number
\begin{equation}\label{VV1A}
{\cal N}(x)=\sum_\epsilon {\cal N}_\epsilon(x).
\end{equation}
We use the identity
\begin{equation}\label{VV2}
\left[ {\cal N}(x),\int_y q_\epsilon(t,y)\frac{\partial}{\partial\psi_\epsilon(y)}\right]=
\sum_\epsilon q_\epsilon(t,x)\frac{\partial}{\partial\psi_\epsilon(x)}
\end{equation}
in order to establish the simple relation
\begin{equation}\label{VV3}
\langle {\cal N}(x)\rangle=\sum_\epsilon q^2_\epsilon(t,x).
\end{equation}
This holds for a wave function $g(t)=g_1(t)$, and we use eq. \eqref{G19} employing appropriate basis elements (a subset of $\{g_\tau\}$)
\begin{eqnarray}\label{VV4}
&&g_{1,\epsilon}(x)=\frac{\partial}{\partial\psi_\epsilon(x)}|0\rangle~,~\tilde g_{1,\eta}(y)=\psi_\eta (y),\nonumber\\
&&\int {\cal D}\psi\tilde g_{1,\eta}(y)g_{1,\epsilon}(x)=\delta_{\eta\epsilon}\delta(x-y),
\end{eqnarray}
with
\begin{equation}\label{VV5}
g_1=\int_x q_\epsilon(x,t)g_{1,\epsilon}(x)~,~\tilde g_1=\int_y q_\eta(y,t)\tilde g_{1,\eta}(y).
\end{equation}
The normalization $\int{\cal D}\psi\tilde g_1g_1=1$ implies for the one-particle wave function the normalization
\begin{equation}\label{VV6}
\int_x \sum_\epsilon\big(q_\epsilon(x)\big)^2=\int_x\varphi^\dagger_D(x)\varphi_D(x)=1,
\end{equation}
which is compatible with $\langle{\cal N}\rangle=1$, cf. eq. \eqref{VV1}. The normalization \eqref{VV6} is preserved by the unitary time evolution of the one-particle wave function. Indeed, we can write the Dirac equation \eqref{276C} as a Schr\"odinger-type equation
\begin{equation}\label{VV7}
i\hbar\partial_t\varphi_D=H\varphi_D,
\end{equation}
with hermitean Hamiltonian $(T_k=\gamma^0\gamma^k)$
\begin{equation}\label{VV8}
H=i\hbar T_k\partial_k+\hbar m\gamma^0\bar\gamma+\hbar e(A_0 -T_k A_k)=H^\dagger.
\end{equation}
For a pure one-particle state the expectation value $\langle N(x)\rangle$ can be interpreted in a natural way as the probability density to find the particle at the position $x$. This is precisely the standard interpretation of $\psi^\dagger_D(x)\psi_D(x)$ in one-particle quantum mechanics,
\begin{eqnarray}\label{VV9}
&&w(x)=\langle N(x)\rangle=\varphi^\dagger_D(x)\varphi_D(x),\nonumber\\
&&\int_xw(x)=1.
\end{eqnarray}
Since $N(x)$ is a classical observable, we can define the position of the particle as a classical observable
\begin{equation}\label{VV10}
X=\int_x xN(x).
\end{equation}
The expectation value in classical statistics coincides with the standard quantum mechanics rule
\begin{equation}\label{VV11}
\langle X\rangle=\langle \int_x xN(x)\rangle=\int_xxw(x)=\int_x\varphi^\dagger_D(x)x\varphi_D(x).
\end{equation}
This finds a natural extension to arbitrary functions of the position observable
\begin{equation}\label{VV12}
f(X)=\int_xf(x)N(x).
\end{equation}
The expectation values of this type of classical observables follow again the rule of quantum mechanics
\begin{equation}\label{VV13}
\langle f(X)\rangle=\int_x\varphi^\dagger_D(x)f(x)\varphi_D(x).
\end{equation}
In particular, one obtains the same formula for the dispersion $\langle X_kX_k\rangle-\langle X_k\rangle\langle X_k\rangle$ as in quantum mechanics. We conclude that measurements of the position of a particle, or more generally the distribution of positions in an ensemble of one-particle states, can be described equivalently in a classical statistical ensemble with classical observables, or in quantum mechanics with Hamiltonian \eqref{VV8}. The complete time evolution of the distribution of positions is identical in both descriptions. This covers, in particular, the characteristic quantum interference in a double slit experiment. The time evolution \eqref{P35}, \eqref{P44}, \eqref{231} for the classical wave function and associated classical probability distribution $\{p_\tau\}$ produces for one-particle states exactly the quantum mechanical interference pattern.
\medskip\noindent
{\bf 3. Schr\"odinger equation}
Standard quantum mechanics for an electron in a potential is recovered from the non-relativistic approximation to the Dirac equation. This is well known, and we sketch here for completeness only the case $A_k=0$. The nonrelativistic approximation becomes valid if $eA_0$ and $iT_k\partial_k$ are small compared to $m$. Since $(\gamma^0\bar \gamma)^2=1$, it is convenient to choose a basis where $\gamma^0\bar\gamma=diag (1,1,-1,-1)$. With $V(x)=\hbar eA_0(x)$ and $M=\hbar m$ the Hamiltonian \eqref{VV8} takes the form
\begin{equation}\label{VV14}
H=
\left(\begin{array}{ccc}M&,&\sigma_kp_k\\\sigma^\dagger_kp_k&,&-M\end{array}\right)
+V(x),
\end{equation}
where we use the standard quantum mechanical momentum operator
\begin{equation}\label{VV15}
p_k=-i\hbar \partial_k.
\end{equation}
The matrices
\begin{equation}\label{VV16}
\sigma_1=-1~,~\sigma_2=i\tau_2~,~\sigma_3=i\tau_3
\end{equation}
arise from
\begin{equation}\label{VV17}
T'_k=U^\dagger T_kU=-\left(\begin{array}{ccc}
0&,&\sigma_k\\\sigma^\dagger_k&,&0\end{array}\right),
\end{equation}
with
\begin{equation}\label{VV18}
U=\frac{1}{\sqrt{2}}
\left(\begin{array}{ccc}\tau_1&,&-\tau_2\\-\tau_2&,&\tau_1\end{array}\right)~,~
U^\dagger U=1~,~U^\dagger=U.
\end{equation}
(Note $T'_1=T_1,~T'_2=-T_2$.) The unitary matrix $U$ diagonalizes $\gamma^0\bar\gamma$,
\begin{equation}\label{VV19}
\gamma^0\bar\gamma=
\left(\begin{array}{ccc}0&,&-i\tau_3\\i\tau_3&,&0\end{array}\right)~,~
U^\dagger\gamma^0\bar\gamma U=
\left(\begin{array}{ccc}1&,&0\\0&,&-1\end{array}\right).
\end{equation}
We observe for the matrices $\sigma_k$ the relations
\begin{equation}\label{VV20}
\sigma^\dagger_k\sigma_l+\sigma^\dagger_l\sigma_k=\sigma_k\sigma^\dagger_l+\sigma_l\sigma^\dagger_k=2\delta_{kl},
\end{equation}
which guarantee $\{T'_k,T'_l\}=2\delta_{kl}$.
We next decompose $\varphi_D$ into two-component wave functions, $\varphi^T_D=(\chi^T,\rho^T)$, which obey
\begin{eqnarray}\label{VV21}
i\hbar\partial_t\chi&=&(M+V)\chi+\sigma_kp_k\rho,\nonumber\\
i\hbar\partial_t\rho&=&(-M+V)\rho+\sigma^\dagger_kp_k\chi.
\end{eqnarray}
For the non-relativistic electron we consider the approximate solution $\rho=A\chi$, where $A=\sigma^\dagger_kp_k/(2M)$ is determined by requiring in leading order $\partial_t\rho=A\partial_t\chi$. Insertion into eq. \eqref{E25} yields for $\psi=\exp(iMt/\hbar)\chi$ the standard Schr\"odinger equation for a particle in a potential $V$,
\begin{equation}\label{VV22}
i\hbar\partial_t\psi=\left(\frac{p_kp_k}{2M}+V\right)\psi.
\end{equation}
As usual, the Dirac equation can also describe non-relativistic positrons. For this purpose one considers a second class of solutions $\chi=B\rho$, with $B=-\sigma_kp_k/(2M)$. For $\tilde \psi=\exp(iMt/\hbar)\rho^*$ one obtains
\begin{equation}\label{VV23}
i\hbar\partial_t\tilde\psi=
\left(\frac{p_kp_k}{2M}-V\right)\tilde\psi,
\end{equation}
and we note the change of sign of the potential due to the opposite charge of the positron.
All quantum mechanical phenomena extracted from solutions of the Schr\"odinger equation are described by our time evolution equation for a classical statistical ensemble of Ising-spins on a lattice. For a potential realizing a double-slit situation the standard interference pattern is predicted for this classical statistical ensemble to appear behind the slits. This holds provided that the initial state at some time $t_0$ corresponds to a one-particle state describing a particle moving towards the slits.
\newpage\noindent
{\bf 4. Particle-wave duality}
The discreteness of measurement values in quantum mechanics can be traced back to the discrete occupation numbers of the Ising-type model. In quantum mechanics we may define an ``interval observable'' $J_{{\cal R}}$ by a function
\begin{equation}\label{VV24}
J_{\cal R}(x)=\left\{\begin{array}{ll}
1&\text{if}~x\in {\cal R}\\0&\text{otherwise}
\end{array}\right).
\end{equation}
It has the property
\begin{equation}\label{VV25}
J^2_{\cal R}=J_{\cal R},
\end{equation}
such that its spectrum consists of the discrete values $0$ and $1$. According to the rules of quantum mechanics, the possible outcomes of a measurement of $J_{\cal R}$ are $0$ or $1$. The interpretation in quantum mechanics is simple: either the particle is within the region (interval) $J_{\cal R}$, in which case the measurement value $J_{\cal R}=1$ will be found, or it is outside this region, and $J_{\cal R}=0$ will be found. Particles are discrete objects - they are either inside or outside an interval.
In our classical statistical Ising-type setting $J_{\cal R}$ is a classical observable, given by
\begin{equation}\label{VV26}
J_{\cal R}=\int_{\cal R} N(x)=\sum_{\cal R} N_L(x).
\end{equation}
The sum $\sum_{\cal R}$ extends over all lattice points within the region ${\cal R}$. The classical observable $N_L(x)$ corresponds to a normalization of occupation numbers for the discrete lattice where $\big (N_{L,\epsilon}\big)_\tau=0,1$. The sum over species $N_L(x)=\sum\limits_\epsilon N_{L,\epsilon}(x)$, cf. eq. \eqref{VV1A}, can therefore take in any classical state $\tau$ only the discrete values $\big(N_L(x)\big)_\tau=(0,1\dots, 8)$, according to the total number of eight species. In consequence, for any state $\tau$ of the classical statistical ensemble, $(J_{\cal R})_\tau$ is a positive integer or zero. According to the standard rule of classical statistics these integers describe the possible outcomes of measurements.
For a one particle state the total particle number equals one,
\begin{equation}\label{VV27}
1=\int\limits_V N(x)=\sum_VN_L(x),
\end{equation}
where $\sum\limits_V$ extends now over all lattice points in the total volume. Since ${\cal R}$ must be contained in $V$ the maximal allowed value for $J_{\cal R}$ in a one-particle state is one, such that $(J_{\cal R})_\tau=0,1$ are the only possible values of the classical observable for such a state. This is precisely the quantum mechanical rule. No new postulate is necessary for this measurement in quantum mechanics - the quantum rule is inferred from the standard rule of classical statistics. Of course, the expectation value of $J_{\cal R}$ for a one particle state is the same in the quantum mechanical and the classical statistical description
\begin{equation}\label{VV28}
\langle J_{\cal R}\rangle=\int_x\varphi^\dagger_D(x)J_{\cal R}(x)\varphi_D(x)=
\int_{\cal R}\varphi^\dagger_D(x)\varphi_D(x).
\end{equation}
While our model connects the discrete particle aspects in quantum mechanics directly to the discrete classical Ising-spins, the continuous wave aspects also arise in a natural way. The quantum wave function is continuous because probability distributions and associated classical wave functions are continuous (at least piecewise). As we have mentioned already, the characteristic interference effects for waves and the superposition arise from the linearity of the fundamental evolution equation in the classical wave function.
\section{Conclusions}
\label{Conclusions}
We have derived the Schr\"odinger equation for a quantum particle in a potential from a classical statistical ensemble for Ising-spins. The dynamics of the classical statistical system has to be specified by an appropriate evolution equation for the probability distribution. For this purpose we employ the classical wave function which is defined as the positive or negative root of the probability distribution. The proposed evolution equation is a linear differential equation for the classical wave function. The wave function at time $t$ obtains from the wave function at $t'$ by a rotation - this guarantees the preservation of the normalization of the probability distribution.
We have exploited a map between the classical wave function and a Grassmann wave function which is an element of a real Grassmann algebra. In turn, the time evolution of the Grassmann wave function can be associated to a Grassmann functional integral. This allows us to formulate the evolution equation for the classical wave function in terms of the action of a functional integral. The symmetries of the model, as Lorentz symmetry and electromagnetic gauge symmetry for our model of Dirac spinors in an external electromagnetic field, can be easily implemented in this way. Since the classical wave function is real we have to formulate the model in terms of a real Grassmann algebra.
Our model is regularized on a lattice of space points. On the one hand, this guarantees that mathematical expressions are well defined for a finite number of lattice points, with continuum limit of an infinite number of points taken at the end. On the other hand, the concept of classical Ising-spins or associated occupation numbers at every point $x$ is well defined. The discreteness of the particle aspects of quantum mechanics can be traced back to the discrete occupation numbers that can only take the values zero or one. We also have used discrete time steps, and we have formulated the lattice action such that the Grassmann wave function $g(t+2\epsilon)$ obtains from $g(t)$ by a rotation. The time evolution is unitary not only in the limit $\epsilon\to 0$, but also for finite $\epsilon$. A unitary time evolution for infinitesimal time steps is easily achieved by any antisymmetric evolution generator $K_{\tau\rho}$ for the wave function $q_\tau(t)$,
\begin{equation}\label{z1}
\partial_tq_\tau(t)=\sum_\rho K_{\tau\rho} q_\rho(t).
\end{equation}
Generating a unitary evolution also for finite smallest time steps imposes restrictions on the form of the lattice action. This, together with the requirement of a real Grassmann action, requires some care for the construction of the model and explains the specific form of the action in comparison with other possible lattice actions.
The proposed evolution equation for the classical statistical ensemble of Ising-spins does not only lead to the Schr\"odinger equation for non-relativistic one-particle states. It entails the full dynamical equations for a quantum field theory of Dirac fermions in an external electromagnetic field. The dynamics of states with an arbitrary number of fermions, including the characteristic interference patterns for indistinguishable fermions in quantum mechanics, is correctly described.
At this point it seems worthwhile to ask some questions about the origin of characteristic features of quantum mechanics in our classical statistical setting. Particle-wave duality is realized by the discreteness of Ising-spins on one side, and the continuous probability distribution or classical wave function on the other side. Interference arises from the formulation of the basic evolution law in terms of the classical wave function. While the classical wave function is real, it can nevertheless take positive and negative values which can add to zero locally. The superposition principle or linearity of the quantum evolution finds a direct origin in the formulation of a dynamical law for the classical statistical ensemble that is linear in the classical wave function. The characteristic physics of phases in quantum mechanics is connected to the presence of a complex structure within the real Grassmann algebra. Planck's constant $\hbar$ appears purely as a conversion factor of units. The uncertainty relations can be obtained directly from the possible solutions of the Schr\"odinger equation.
Finally, one of the most characteristic elements of the mathematical formulation of quantum mechanics is the presence of a non-commutative product for operators and associated observables. These structures are obviously present in our formulation and clearly very useful for a discussion of solutions of the Schr\"odinger equation or Dirac equation. We have not addressed in this paper the fundamental origin of non-commuting operators and refer in this context to related work \cite{GR}, \cite{CWP}, \cite{3A}. An essential point is the observation that a classical statistical ensemble admits many different product structures for observables, and therefore also the definition of many different correlation functions. The correct choice of a correlation function for the description of a sequence of two measurements depends on the details of the measurement process. There exist various idealizations of measurements.
The classical correlation function (``pointwise multiplication'' $A_\tau B_\tau$) is adapted to a situation of negligible mutual influence between two measurements. It is often not appropriate for the description of measurements in subsystems, that are characteristic for the quantum experiments. For an idealized separation into a subsystem and its environment many different classical observables are grouped into an equivalence class, for which the properties within the subsystem are identical, but the properties relating to the environment differ. Quantum observables are supposed to measure only properties of the subsystem. They are therefore associated to a whole equivalence class of classical observables, rather than to a single classical observable. A one-to-one correspondence between quantum and classical observables is a basic assumption of the Kochen-Specker theorem \cite{KS}, which therefore does not apply in our setting.
The classical product of observables often depends on the properties of the environment and is therefore not suitable for idealized measurements of properties of subsystems. This is the point why Bell's inequalities \cite{Bell}, \cite{BS} do not apply for idealized measurements of properties of subsystems. These inequalities implicitly assume the use of the classical correlation function. In short, Bell's inequalities are circumvented not by abandoning locality or causality, but simply by concentrating on non-classical correlation functions that are appropriate for measurements in subsystems. In other words, the experimental verification that the observed correlations violate Bell's inequalities tells us that the classical correlation function should not be used for this type of measurements. This is in accordance with our arguments that classical correlations do not correspond to idealized measurements of subsystem properties. It has been shown that in many circumstances the idealized measurements of subsystem properties correspond precisely to the quantum correlation function that is associated to the standard non-commutative operator product in quantum mechanics \cite{GR}, \cite{CWP}.
The essence of the emergence of non-commutative structures is the coarse graining of information. Again, this issue has not been addressed in the present paper. It seems reasonable to expect, however, that a system that is governed by the Dirac equation on microphysical scales - say lattice distances shorter than the Planck length - will also show similar properties at ``macroscopic scales'' associated to coarse graining. (Such macroscopic scales can still be much smaller than all characteristic scales of atom physics or elementary particle physics.) The basic reason is that the form of the Dirac equation for one-particle states is essentially fixed by the symmetries. It will not be altered if the coarse graining respects the symmetries. While it remains an interesting task to perform this coarse graining explicitly, the main message of this paper is already very clear at the present stage: Quantum field theory can be obtained from a classical statistical ensemble.
|
\section{Introduction}
Let $\mathcal{A}$ be a masa in a $II_1$ factor $\mathcal{M}$ and $E$ the normal conditional expectation from $\mathcal{M}$ to $\mathcal{A}$. Kadison, in \cite{KadPyt1} asked the following question,
\begin{question}[Kadison's carpenter problem]
Given any positive contraction $B$ in $\mathcal{A}$, does there exist a projection $P$ in $\mathcal{M}$ so that $E(P) = B$?
\end{question}
We will denote the above problem as asking if positive contractions in masas can be lifted to projections. We refer the reader to the above cited paper for the discussion leading up to this problem. The best result to date is the result of \cite{DFHS} that says the following
\begin{prop}[Dykema, Fang, Hadwin, Smith]
Any positive contraction in a generator masa in $L(F_{2})$ can be lifted to a projection. Also, for any positive contraction $B$ in a Cartan masa $\mathcal{A}$ in the hyperfinite $II_1$ factor $\mathcal{R}$, there is an automorphism $\theta$ of $\mathcal{A}$ so that $\theta(B)$ can be lifted to a projection.
\end{prop}
There are several consequences of this result that the reader can work out for herself. For general $II_1$ factors, far less is known. Indeed, everything that is known so far with the exception of the result mentioned above and some extensions proved in the same paper, is a straightforward interpretation of results for matrices. For instance, the matricial Schur-Horn theorem guarantees that $\lambda I$ can be lifted if $\lambda$ is a rational number, but it is not known if irrational multiples of the identity can be lifted to projections. In this note we show that this is indeed the case. It will follow that elements with finite spectrum can be lifted to projections.
In this note, we will work in a slightly more general context. Kadison's carpenter problem is a special case of a majorization problem for von Neumann algebras. The notion of majorization in von Neumann algebras goes back at least to Hiai's work\cite{Hiai} in the 80's.
\begin{df}[Majorization]
Given two self-adjoint operators $A, S$ in a finite factor $(\mathcal{M},\tau)$, say that $A$ is majorized by $S$, denoted by $A \prec S$ if
\[\tau(f(A)) \leq \tau(f(S))\] for every continuous convex real valued function $f$ defined on a closed interval $[c,d]$ containing the spectra of both $A$ and $S$.
\end{df}
The condition implies that $\tau(A) = \tau(S)$. Majorization can be expressed in several ways and these equivalences can be found in \cite{Hiai} and the references therein. A natural extension of Kadison's problem was formulated by Kadison and Arveson in \cite{ArvKad}.
\begin{question}[Arveson and Kadison's Schur Horn problem]
Let $\mathcal{A}$ be a positive element in $\mathcal{A}$ and $S$ a positive element in $\mathcal{M}$ such that $A \prec S$. Then, is it true that there exists an element $T$ in $\mathcal{O}(S) = \overline{\{U S U^{*},\,\, U \in \mathcal{U}(M)\}}^{||}$ such that $E(T) = A$?
\end{question}
One does need to take the norm closure; See the example following lemma(5.5) in the same paper.
This problem was solved in the affirmative for the generator and radial masas in the free group factors in \cite{DFHS}, where it was also solved modulo an automorphism of the masa for Cartain masas in the hyperfinite $II_1$. In this note, we will work with general masas inside general type $II_1$ factors. Our main result is the following theorem whose proof is an adaptation of the best known proof of the matricial Schur Horn theorem. It should come as no surprise that we do not need to take the norm closure to achieve lifting.
\begin{theorem}
Let $\mathcal{A}$ be a masa in a $II_1$ factor $\mathcal{M}$ and let $E$ be the normal conditional expectation from $\mathcal{M}$ to $\mathcal{A}$. Let $A \in \mathcal{A}$ and $S \in \mathcal{M}$ be positive operators with finite spectrum such that $A \prec S$. Then, there is a unitary $U$ in $\mathcal{M}$ so that $E(U S U^{*}) = A$.
\end{theorem}
The theorem says that the Schur-Horn problem can be solved when both elements have finite spectrum. While this result will hardly come as a surprise, it is new. Routine calculations will then allow us to adapt the above theorem to deduce an approximate Schur-Horn theorem for general operators in a $II_1$ factor.
\begin{theorem}
Let $S$ be a self-adjoint operator in $\mathcal{M}$. Then, the norm closure of $E(\mathcal{U}(S))$ equals $\{A \in \mathcal{A} \mid A \prec S\}$.
\end{theorem}
In particular, letting $\mathcal{O}(S) = \overline{\{USU^{*} \mid U \in \mathcal{U}(\mathcal{M})\}}^{||}$, we have that
\[\overline{E(\mathcal{O}(S))}^{||} = \{A \in \mathcal{A} \mid A \prec S\}\]
The conjectured Schur-Horn theorem of Arveson and Kadison says that we do not need to take the norm closure for equality, something that we are unable to prove in this note. A weaker version of our theorem, where the $\sigma-$SOT closure was taken in the place of the norm closure was proved by Argerami and Massey in \cite{MasArgIn}. Also, the above result was established for Cartan masas in the hyperfinite $II_1$ factors(and thus for general semi-regular masas, see \cite{PopKad}) in \cite{DFHS}.
The paper has four sections apart from the introduction; In section $2$, we show that scalars can be lifted to projections. In section $3$, we push this through to show that the Schur-Horn problem can be solved for operators with finite spectrum. Section $4$ contains the approximate Schur-Horn theorem. There is then a last section consisting of some remarks and observations.
Some words on notation: Given two operators $A, B$ inside a von Neumann algebra $\mathcal{M}$ such that there is a projection $P$ inside $\mathcal{M}$ such that $A = PAP$ and $B = (I-P)B(I-P)$, in order to stress the fact that $A$ and $B$ live under the auspices of orthogonal projections, we will use the expression $A\oplus B$ to denote their sum. Next, given a self-adjoint operator $A$ and a Borel measurable subset $X$ of the real line, the expression $E_{A}(X)$ will denote the spectral projection of $A$ corresponding to the subset $X$. This notation might cause confusion with the notation $E_{\mathcal{A}}(A)$ or simply $E(A)$ where $\mathcal{A}$ is a subalgebra of $\mathcal{M}$, which denotes the image under a conditional expectation $E$. We apologize for this, but retain the notations due to their provenance. Finally, lower case letters, possibly with subscripts, like $a, b$ and $s_i$ will always refer to scalars. We will always use upper case letters $S, T$ and so forth to refer to operators.
\section{Lifting Scalars}
We begin with a simple observation.
\begin{lemma}\label{simLem}
Let $P$ be a projection in a masa $\mathcal{A}$ inside a type $II_1$ factor $\mathcal{M}$ and let $\lambda, a, b$ be positive scalars such that $\tau(S) = \lambda$ where $S = a P + b (I-P)$. Then, there is a unitary $U$ in $\mathcal{M}$ and a projection $Q$ in $\mathcal{A}$ such that letting $T = U S U^{*}$, we have that
\begin{enumerate}
\item $E(Q T Q) = \lambda Q$.
\item $(I-Q) T (I - Q) = c R + d (I - Q - R)$ for some projection $R$ in $\mathcal{A}$ with $R \leq I - Q$ and positive numbers $c, d$.
\item $\tau(Q) \geq \dfrac{1}{3}$.
\end{enumerate}
\end{lemma}
\begin{proof}
The lemma is trivial if $a = b$, for then, $a = b = \lambda$ and there is nothing to prove. We assume without loss of generality that $a > b$. Since $\tau(S) = \lambda$, we must then have that $a > \lambda > b$. We may also assume that $\tau(P) \leq \dfrac{1}{2}$. For, suppose we have proved the lemma in this case, the result when $\tau(P) > \dfrac{1}{2}$ can be derived by applying the lemma to $I-S$ and $(1 -\lambda)I$. We therefore assume that $\tau(P) \leq \dfrac{1}{2}$.
Let $k$ be the largest integer such that $(k+1) \tau(P) \leq 1$. Since $\tau(P) \leq \dfrac{1}{2}$, $k$ must be at least $1$. Pick projections $Q_1, \cdots, Q_{k}$, each of trace $\tau(P)$ in $\mathcal{A}$ that are mutually orthogonal and also orthogonal to $P$. let $V_{1}, \cdots, V_{k}$ be partial isometries in $\mathcal{M}$ such that
\begin{enumerate}
\item $V_{1}^{*} V_{1} = Q_{1}$ and $V_{1} V_{1}^{*} = P$ .
\item For $2 \leq i \leq k$, $V_{i} V_{i}^{*} = Q_{i-1}$ and $V_{i}^{*} V_{i} = Q_{i}$.
\end{enumerate}
Pick $\theta_{1}$ such that $a \operatorname{cos}^{2}(\theta_{1}) + b \operatorname{sin}^{2}(\theta_1) = \lambda$ and let $U_1$ be the operator
\[U_1 = \operatorname{cos}(\theta_{1}) P + \operatorname{sin}(\theta_{1}) V_1 - \operatorname{sin}(\theta_{1}) V_1^{*} + \operatorname{cos}(\theta_{1}) Q_{1} + (I - P - Q_1)\]
We will identify the above operator with the operator matrix(using $V_1$ as the matrix unit $E_{12}$), an identification that is standard.
\[U_1 = \left( \begin{array}{ccc}
\operatorname{cos}(\theta_{1}) & \operatorname{sin}(\theta_{1}) & 0\\
-\operatorname{sin}(\theta_{1}) & \operatorname{cos}(\theta_{1}) & 0\\
0 & 0 & I\end{array} \right)\]
In this same identification, $S$ is the operator
\[S = \left( \begin{array}{ccc}
a & 0 & 0\\
0 & b & 0\\
0 & 0 & b\end{array} \right)\]
Let $S_{1} = U_1 S U_{1}^{*}$. It is easy to check that $U_1$ is a unitary and that
\[ S_1 = \left( \begin{array}{ccc}
a \operatorname{cos}^{2}(\theta_{1}) + b \operatorname{sin}^{2}(\theta_1) & \ast & 0\\
\ast & a \operatorname{sin}^{2}(\theta_{1}) + b \operatorname{cos}^{2}(\theta_{1}) & 0\\
0 & 0 & b\end{array} \right) = \left( \begin{array}{ccc}
\lambda & \ast & 0\\
\ast & a_1 & 0\\
0 & 0 & b_1\end{array} \right)\]
where $a_1 = a \operatorname{sin}^{2}(\theta_{1}) + b \operatorname{cos}^{2}(\theta_{1})$ and $b_1 = b$. By the trace condition,
\[\lambda \tau(P) + a_1 \tau(P) + b_1 (1 - 2 \tau(P)) = \lambda.\] Since $b_1 = b < \lambda$, we must have that $a_1 > \lambda$ and afortiori $a_1 > b_1$.
Now, continue as above. Pick $\theta_{2}$ such that $a_1 \operatorname{cos}^{2}(\theta_{2}) + b_1 \operatorname{sin}^{2}(\theta_2) = \lambda$ and let $U_2$ be the operator
\[U_2 = \operatorname{cos}(\theta_2) Q_1 + \operatorname{sin}(\theta_2) V_2 - \operatorname{sin}(\theta_2) V_2^{*} + \operatorname{cos}(\theta_2) Q_2 + (I - Q_1 - Q_2).\] We may write the unitary $U_2$ as
\[U_2 = \left( \begin{array}{cccc}
I & 0 & 0 & 0\\
0 &\operatorname{cos}(\theta_{2}) & \operatorname{sin}(\theta_{2}) & 0\\
0 & -\operatorname{sin}(\theta_{2}) & \operatorname{cos}(\theta_{2}) & 0\\
0 & 0 & 0 & I\end{array} \right)\]
and let $S_2 = U_{2} S_{1} U_{2}^{*}$. We have that
\[ S_2 = \left( \begin{array}{cccc}
\lambda & \ast & \ast & 0\\
\ast & \lambda & \ast & 0\\
\ast & \ast & a_2 & 0\\
0 & 0 & 0 & b_2 \end{array} \right)\]
where $a_2 = a_1 \operatorname{sin}^{2}(\theta_{2}) + b_1 \operatorname{cos}^{2}(\theta_{2})$ and $b_2 = b_1$. By the trace condition,
\[2\lambda \tau(P) + a_2 \tau(P) + b_2 (1 - 3 \tau(P)) = \lambda.\] Since $b_2 = b_1 = b < \lambda$, we must have that $a_2 > \lambda$ and afortiori $a_2 > b_2$.
Proceeding this, $k-2$ more times, we get an operator $S_k$ of the form
\[ S_k = \left( \begin{array}{cccccc}
\lambda & \ast & \hdots & \ast & \ast & 0\\
\ast & \lambda & \hdots & \ast & \ast & 0\\
\vdots & \vdots & \ddots & \vdots & \ast & 0\\
\ast & \ast & \ast & \lambda & \ast & 0\\
\ast & \ast & \ast & \ast & a_k & 0 \\
0 & 0 & 0 & 0 & 0 & b_k\end{array} \right)\]
Let $Q = P + Q_{1} + \cdots + Q_{k-1}$(if $k = 1$, let $Q = P$). We see that
\begin{enumerate}
\item $E(Q S_{k} Q) = \lambda Q$. This is because,
\begin{eqnarray*}
E(Q S_{k} Q) &=& E(P S_{k} P) + E(Q_1 S_k Q_1) + \cdots + E(Q_{k-1} S_k Q_{k-1})\\
&=& \lambda P + \lambda Q_1 + \cdots + \lambda Q_{k-1}\\
&=& \lambda Q.
\end{eqnarray*}
($S_k$ is the operator $T$ promised in the statement of the lemma).
\item $(I - Q) S_{k} (I - Q)$ has two point spectrum in $(I-Q) M (I- Q)$.
\item $\tau(Q) = k \tau(P)$. Since $(k+1) \tau(P) \leq 1 < (k+2) \tau(P)$, we see that
\[\tau(Q) = k \tau(P) = \dfrac{k}{k+2} (k+2) \tau(P) > \dfrac{k}{k+2} \geq \dfrac{1}{3}.\]
\end{enumerate}
The lemma follows.
\end{proof}
\begin{theorem}
Let $\mathcal{A}$ be a masa in a $II_1$ factor $\mathcal{M}$ and let $E$ be the normal conditional expectation from $\mathcal{M}$ to $\mathcal{A}$. Then for any $0 \leq \lambda \leq 1$, there is a projection $P$ in $\mathcal{M}$ such that $E(P) = \lambda I$.
\end{theorem}
\begin{proof}
Let $P_0$ be any projection of trace $\lambda$ in $\mathcal{A}$. Using lemma(\ref{simLem}), construct a unitary $U_1$ and a projection $Q_1$ in $\mathcal{A}$ such that, letting $P_1 = U_1 P_0 U_1^{*}$,
\begin{enumerate}
\item $\tau(Q_1) \geq \dfrac{1}{3}$.
\item $E(Q_1 P_1 Q_1) = \lambda Q_1$
\item $(I-Q_1) P_1 (I - Q_1)$ has two point spectrum in $(I-Q_1) M (I - Q_1)$.
\end{enumerate}
Let $R_{1} = Q_1$. Next, for $k = 2, 3, \cdots$, apply lemma(\ref{simLem}) to $\lambda (I-R_{k-1})$ and $(I-R_{k-1}) Q_{k-1} (I - R_{k-1})$ inside the $II_1$ factor $(I-R_{k-1})M(I-R_{k-1})$ to construct a unitary $U_{k}$ and a projection $Q_{k}$ in $(I-R_{k-1}) M (I - R_{k-1})$ and let
\[R_{k} = Q_1 \oplus Q_2 \oplus \cdots \oplus Q_{k} \quad \operatorname{and} \quad P_{k} = (R_{k-1} \oplus U_{k}) P_{k-1} (R_{k-1} \oplus U_{k})^{*}\]Here we identify $Q_{k}$ which is a projection in $(I-R_{k-1}) M (I - R_{k-1})$ with a projection in $\mathcal{M}$ dominated by $I-R_{k-1}$. Also note that $P_k$ is a projection. We have that
\begin{enumerate}
\item $E(Q_{k} P_{k} Q_{k}) = \lambda Q_{k}$ and thus,
\[E(R_{k} P_{k} R_{k}) = \sum_{m=1}^{k} E(Q_{m} P_{m} Q_{m}) = \sum_{m=1}^{k} \lambda Q_{m} = \lambda R_{k}.\]
\item $\tau(I - R_{k}) \leq \dfrac{2}{3} \tau(I - R_{k-1}) \leq (\dfrac{2}{3})^{k}$ and hence, $R_{k}$ converges to $I$ strongly.
\item $(I-R_{k}) P_{k} (I -R_{k})$ has two point spectrum in $(I-R_k) M (I - R_k)$.
\item We have that $R_{k-1} P_{k-1} R_{k-1} = R_{k-1} P_{k} R_{k-1}$ and thus,
\begin{eqnarray}\label{constancy}
R_{l} (P_{m} - P_{n}) R_{l} = 0 \quad \text{ for any } \quad n,m \geq l.
\end{eqnarray}
\end{enumerate}
We now claim that $P_{k}$ converges in the strong operator topology to a projection that we will call $P$ and also that $E(P) = \lambda I$. For the first claim, since $\tau(R_k)$ converges strongly to $I$, for any $\epsilon > 0$ there is a $N$ so that $||(I - R_N) ||_{2} < \epsilon$. For $n, m \geq N$,
\[ ||(P_{n} - P_{m})||_{2} \leq ||R_{N}(P_{n} - P_{m}) R_{N}||_{2} + 2||(I-R_{N})(P_{n} - P_{m})||_{2}\]
The first term is zero by (\ref{constancy}). For the second term,
\[||(I-R_{N})(P_{n} - P_{m})||_{2} \leq ||I - R_{N}||_{2} ||P_{n} - P_{m}|| \leq 2 \epsilon\]
Thus, $||(P_{n} - P_{m})||_{2} \leq 4 \epsilon$ and the sequence $\{P_n\}$ is strongly convergent. Let $P$ be the limit projection. Forthe second claim,
\begin{align*}
||E(P) - \lambda I||_{2} = &\operatorname{lim} ||E(P_{n}) -\lambda I||_{2}\\
= &\operatorname{lim} ||\lambda R_{n} + E((I-R_{n}) P_{n} (I-R_{n})) - \lambda I||_{2}\\
= &\operatorname{lim} ||-\lambda(I - R_{n}) + E((I-R_{n}) P_{n} (I-R_{n})||_{2}\\
\leq &\operatorname{lim}\lambda||(I-R_{n}||_{2} + ||(I-R_{n})P_n(I-R_{n})||_{2}\\
\leq &\operatorname{lim}\lambda \left(\dfrac{2}{3}\right)^{n} + ||P_n|| \left(\dfrac{2}{3}\right)^{n} \\
\leq &\operatorname{lim}(\lambda + 1)\left(\dfrac{2}{3}\right)^{n} \\
= & \,0
\end{align*}
We conclude that $E(P) = \lambda I$.
\end{proof}
We record a simple corollary
\begin{prop}\label{prop00}
Let $A$ be a positive contraction in $\mathcal{A}$ that can be written as $A = \sum_{n} \lambda_n E_{n}$, where the $E_{n}$'s are orthogonal projections summing up to $I$. Then, there is a projection $P$ in $\mathcal{M}$ such that $E(P) = A$.
\end{prop}
\begin{proof}
The element $A$ may be written as $A = \sum_{n=1}^{\infty} \lambda_n E_n$ where the $E_{n}$'s are mutually orthogonal projections in $\mathcal{A}$ summing up to $1$ and $0 \leq \lambda _n \leq 1$ for every $n$. $E_{n} M E_{n}$ is a type $II_1$ factor and we may find a projection $P_{n}$ in $E_{n} M E_{n}$ such that $E_{AE_{n}}(P_{n}) = \lambda_{n} E_{n}$ for every $n$. Let $P$ be the projection $\sum_{n=1}^{\infty} P_{n}$. Here, we are identifying $P_{n}$ which is a projection in $E_{n} M E_{n}$ with a projection in $\mathcal{M}$ that is dominated by $E_{n}$. Then,
\[E(P) = \sum_{n=1}^{\infty} E(P_{n}) = \sum_{n=1}^{\infty} E(E_{n} P_{n} E_{n}) = \sum_{n=1}^{\infty} \lambda_{n} E_{n} = A\]
\end{proof}
\section{Schur-Horn theorem for operators with finite spectrum}
We will now bootstrap the theorem in the previous section to get a Schur-Horn theorem for positive operators with finite spectrum. Recall the following reformulation of majorization in $II_1$ factors. Let $A, S$ be positive contractions in a type $II_1$ factor $\mathcal{M}$ and let $f, g : [0,1] \rightarrow [0,1]$ be the (essentially unique, right-continuous, non-increasing) spectral weight functions, which satisfy
\[\tau(A^{n}) = \int_{0}^{1} f^{n}(r) dm(r) \quad \text{ and } \quad \tau(S^{n}) = \int_{0}^{1} g^{n}(r) dm(r) \quad \text{for } n =0, 1, \cdots\]
Then $A \prec S$ if
\[\int_{0}^{t} f(r) dm(r) \leq \int_{0}^{t} g(r)dm(r), \,\, 0 \leq t \leq 1 \quad \text{and} \quad \int_{0}^{1}f(r)dm(r) = \int_{0}^{1} g(r)dm(r) \]
\begin{lemma}\label{2ptSp}
Let $A = \lambda_1 E_1 \oplus \lambda_2 E_2$ where $E_1 + E_2 = I$ and $\lambda_1 \geq \lambda_2 \geq 0$ and $S = \mu_1 F_1 \oplus \mu_2 F_2$ where $F_1 + F_2 = I$ and $\mu_1 > \mu_2 \geq 0$ be two operators in a $II_1$ factor with $\tau(A) = \tau(S)$. If $\mu_1 \geq \lambda_1$ and $\mu_2 \leq \lambda_2$, then $A \prec S$.
\end{lemma}
\begin{proof}
It is easy to see that if $B$ is a positive contraction, then $B \prec P$ for any projection $P$ with $\tau(P) = \tau(B)$. Let $c = \dfrac{1}{\mu_1 - \mu_2}$ and $d = - \dfrac{\mu_2}{\mu_1 - \mu_2}$. The operator $cS + dI$ may be checked to equal $F_1$, is hence a projection and of course, $\tau(cS+dI) = \tau(cA+dI)$.
\[cA + dI = (c \lambda_1 +d)E_{1} + (c \lambda_2 + d)E_{2} = \dfrac{\lambda_1-\mu_2}{\mu_1-\mu_2} E_{1} + \dfrac{\lambda_2-\mu_2}{\mu_1-\mu_2} E_{2}\]
Since $\lambda_2 \leq \lambda_1$, $\lambda_2 \geq \mu_2$ and $\lambda_1 \leq \mu_1$, we have that
\[0 \leq \dfrac{\lambda_2-\mu_2}{\mu_1-\mu_2} \leq \dfrac{\lambda_1-\mu_2}{\mu_1-\mu_2} \leq \dfrac{\mu_1-\mu_2}{\mu_1-\mu_2} = 1\]
And thus, $cA + dI$ is a positive contraction. By the observation in the first line of the proof, $cA + dI \prec cS + dI$ and therefore, $A \prec S$.
\end{proof}
\begin{lemma}\label{2ptMajCond}
Let $A = \lambda_1 E_1 + \lambda_2 E_2$ and $S = \mu_1 E_1 + \mu_2 E_2$ where $E_1$ and $E_2$ are orthogonal projections summing up to $I$, be positive operators in a type $II_1$ factor $\mathcal{M}$, with the same trace. If $\lambda_1 \leq \mu_1$, then $A \prec S$.
\end{lemma}
\begin{proof}
It is easy to see that we must have $\mu_2 < \lambda_2$. The lemma now follows from lemma(\ref{2ptSp}).
\end{proof}
\begin{lemma}\label{2ptMaj}
Let $A$ be a self-adjoint operator and $S$ a positive contraction in a $II_1$ factor so that $A \prec S$. Then $A$ is a positive contraction as well.
\end{lemma}
\begin{proof}
Routine verification.
\end{proof}
\begin{prop}\label{prop0}
Let $S$ be a positive operator in $\mathcal{M}$ with two point spectrum and let $A$ be a positive contraction in $\mathcal{A}$ that has finite spectrum and so that $A \prec S$. Then, there is a unitary $U$ in $\mathcal{M}$ such that $E(U S U^{*}) = A$.
\end{prop}
\begin{proof}
Write $S = \mu_1 F_1 \oplus \mu_2 F_2$ where $\mu_1 \geq \mu_2$ and $F_1 \oplus F_2 = I$. Let $c = \dfrac{1}{\mu_1 - \mu_2}$(note that $c>0$) and $d = - \dfrac{\mu_2}{\mu_1 - \mu_2}$. The operator $cS + dI$ may be checked to equal $F_1$ and is hence a projection. We also have that $cA+dI \prec cS+dI = F_1$. By lemma(\ref{2ptMaj}), $cA+dI$ must actually be a positive contraction. Also, of course, $\tau(cS+dI) = \tau(cA+dI)$. Now, by proposition(\ref{prop00}), there is a unitary $U$ so that $E(U(cS+dI)U^{*}) = cA+dI$. And hence, $E(USU^{*})=A$.
\end{proof}
When one or both operators have finite spectrum, majorization reduces to a simple condition.
\begin{lemma}\label{atoMaj}
Let $A, S$ be positive operators in a $II_1$ factor with $\tau(A) = \tau(S)$ and let $f, g$ be the spectral weight functions of $A, S$ respectively, as above. Suppose $A$ has finite spectrum, i.e, the spectral weight function $f$ has the form
\[f = \sum_{n=1}^{N} \lambda_n \chi_{[s_{n-1},s_n)}\]
for some natural number $N$ and some sequences $0 = s_0 < s_1 < \cdots < s_N = 1$ and $0 \leq \lambda_N < \lambda_{N-1} < \cdots < \lambda_1$. Then, $A \prec S$ iff for $n = 1, 2, \cdots, N$,
\[\int_{0}^{s_n} f(r) dm(r) \leq \int_{0}^{s_n} g(r)dm(r) \,\, \text{ or equivalently, } \,\, \tau(A E_{A}([0,s_n))) \leq \tau(S E_{S}([0,s_n))) \]
\end{lemma}
\begin{proof}
Routine verification.
\end{proof}
We now prove the promised special case of the Schur-Horn theorem.
\begin{theorem}[The Schur-Horn theorem for operators with finite spectrum in a $II_1$ factor]\label{SHD}
Let $A$ and $S$ be positive operators with finite spectrum in $\mathcal{A}$ and $\mathcal{M}$ respectively and so that $A \prec S$. Then, there is a unitary $U$ in $\mathcal{M}$ so that $E(U S U^{*}) = A$.
\end{theorem}
\begin{proof}
We assume that $A$ and $S$ have spectrum consisting of $N$ and $M$ points respectively. Write $A = \sum_{n=1}^{N} \lambda_n E_n$ and $S = \sum_{n=1}^{M} \mu_n F_n$ where the $\{\lambda_n\}_{1}^{N}$(respectively, the $\{\mu_n\}_{1}^{M}$) are distinct. We may assume that none of the $\lambda_i$ equal any of the $\mu_j$. For suppose $\lambda_i = \mu_j$. Assume that $\tau(E_i) \leq \tau(F_j)$, the other case is handled similarly. We may, after conjugating by a unitary, write $A = \lambda_i E_i \oplus (A-\lambda_i E_i)$ and $S = \mu_j E_i \oplus (S - \mu_j E_i) = \lambda_i E_i \oplus (S - \lambda_i E_i)$. Clearly, $A-\lambda_i E_i \prec S - \lambda_i E_i$ and it is enough to prove the theorem for $A-\lambda_i E_i$ which has at most $N-1$ point spectrum in $(I-E_i)\mathcal{A}$ and $S - \lambda_i E_i$ which has at most $M$ point spectrum inside $(I-E_i) \mathcal{M} (I-E_i)$. We therefore assume that none of the $\lambda_i$ equal any of the $\mu_j$.
Since $\mathcal{A}$ is unitarily equivalent to $L^{\infty}([0,1],dm)$, we may find a maximal nest of projections $\{P_{t} : 0 \leq t \leq 1\}$ in $\mathcal{A}$ with $P_{t} \leq P_{s}$ for $0 \leq t \leq s \leq 1$ and $\tau(P_{t}) = t$ for $0 \leq t \leq 1$. Since $A$(respectively $S$) has $N$(respectively $M$) point spectrum, we may, after conjugating $A$ and $S$ by unitaries, assume that $A$ and $S$ have the form
\[A = \sum_{n=1}^{N} \lambda_n (P_{s_{n}}-P_{s_{n-1}}) \quad \text{ and } \quad S = \sum_{n=1}^{M} \mu_n (P_{t_{n}} - P_{t_{n-1}})\]
for sequences $0 = s_{0} < s_{1} < s_{2} < \cdots < s_{N} = 1$ and $0 = t_{0} < t_{1} < t_{2} < \cdots < t_{M} = 1$ and positive scalars $\lambda_1 > \lambda_2 > \cdots > \lambda_N \geq 0$ and $\mu_1 > \mu_2 > \cdots > \mu_{M} \geq 0$.
Reindex the set $\{s_{1}, \cdots, s_{N-1}\} \cup \{t_1, \cdots, t_{M-1}\}$ by $\{r_1, \cdots, r_{L}\}$ where $r_1 < r_2 < \cdots < r_{L-1}$ and let $r_{L} = 1$. Then, we may write
\[A = \sum_{n=1}^{L} \gamma_n (P_{r_{n}} - P_{r_{n-1}}) \quad \text{ and } \quad f = \sum_{n=1}^{L} \delta_n (P_{r_{n}} - P_{r_{n-1}})\]
where $\gamma_n = \lambda_m$ for the unique value $m$ so that $[r_{n-1},r_{n}) \subset [s_{m-1},s_{m})$ and similarly for the numbers $\delta_n$.
We will prove the theorem by induction on $L$. When $L=1$, $A$ and $S$ are scalars and thus, $A = S = \tau(A)I$ and the theorem is trivial. Assume we have shown the following:
\begin{stmt}
Let $A$ and $S$ be positive operators inside a masa, which we denote by $\mathcal{A}$ inside a type $II_1$ factor, which we denote by $\mathcal{M}$, so that $A = \sum_{n=1}^{K} \gamma_{n} (P_{r_{n}} - P_{r_{n-1}})$ and $S = \sum_{n=1}^{K} \delta_{n} (P_{r_{n}} - P_{r_{n-1}})$ for some sequences $0 < r_{1} < \cdots <r_{K-1} < r_{K} = 1$, $\gamma_{1} \geq \cdots \geq \gamma_{K}$, $\delta_{1} \geq \cdots \geq \delta_{K}$, where $K$ is a natural number less than $L$. Then, there is a unitary $U$ so that $E(U S U^{*}) = A$.
\end{stmt}
We will now show that we can extend this to the case when the decompositions have length $L$ as well.
The majorization condition for the operators $A$ and $S$ that we are working with becomes the following: $\tau(A) = \tau(S)$ and for every $k = 1, \cdots, L-1$, we have that
\[\int_{0}^{r_{k}} f(r)dm(r) = \sum_{n=1}^{k} \gamma_n (r_{n}-r_{n-1}) \leq \sum_{n=1}^{k} \delta_n (r_{n}-r_{n-1}) = \int_{0}^{r_{k}} g(r)dm(r)\]
In particular, $\gamma_1 < \delta_1$. If $\gamma_n < \delta_n$ for every $n = 1, \cdots, L$, then,
\[\tau(A) = \sum_{n=1}^{L} \gamma_n (r_{n}-r_{n-1}) < \sum_{n=1}^{L} \delta_n (r_{n}-r_{n-1}) = \tau(S)\]
which contradicts the fact that $A \prec S$(which entails that $\tau(A) = \tau(S)$. Thus, there is a natural number $1 < l \leq L$ so that
\[\gamma_n < \delta_n \,\, \text{ for } \,\, n = 1, \cdots, l \quad \text{ and } \quad \gamma_{l+1} > \delta_{l+1}.\]
Suppose that $(\delta_{l} - \gamma_{l}) (r_{l}-r_{l-1}) > (\gamma_{l+1} - \delta_{l+1}) (r_{l+1}-r_{l})$(the other case is handled similarly). Pick $r$ so that $(\delta_{l} - \gamma_{l}) (r_{l}-r_{l-1}) = (\gamma_{l+1} - \delta_{l+1}) (r-r_{l})$. Let
\[A_1: = \gamma_{l} (P_{r_{l}} - P_{r_{l-1}}) + \gamma_{l+1} (P_{r}-P_{r_{l}}) \quad \text{ and } \quad S_1:= \delta_{l} (P_{r_{l}} - P_{r_{l-1}}) + \delta_{l+1} (P_{r}-P_{r_{l}})\]
Then,
\[\tau(S_1 - A_1) = (\delta_l - \gamma_l)(r_{l} - r_{l-1}) + (\delta_{l+1}-\gamma_{l+1})(r - r_{l}) = 0 \]
Combining this with the fact that $\gamma_l < \delta_l$ and using lemma(\ref{2ptMajCond}), we conclude that
\[A_{1} \prec S_{1}\] inside the $II_1$ factor $P\mathcal{M}P$ where $P$ is the projection $P = P_{r}-P_{r_{l-1}}$. Now, let
\[A_{2}:= A - A_{1} = \sum_{n\neq l,l+1} \gamma_{n} (P_{r_{n}} - P_{r_{n-1}}) + \gamma_{l+1} (P_{r_{l+1}}-P_{r})\] and similarly,
\[S_{2} := S - S_{1} = \sum_{n\neq l,l+1} \delta_{n} (P_{r_{n}} - P_{r_{n-1}}) + \delta_{l+1} (P_{r_{l+1}}-P_{r}) \]
where the operators are considered in $(I-P)\mathcal{M}(I-P)$. We have
\begin{enumerate}
\item $\sum_{n=1}^{k} \gamma_n (r_{n}-r_{n-1}) < \sum_{n=1}^{k} \delta_n (r_{n}-r_{n-1})$ for $k = 1, \cdots, l-1$.
\item And for $k \geq l+1$, (if $k = l+1$, the third term in the first expression below will not show up)
\begin{eqnarray*}
&&\sum_{n=1}^{l-1} \gamma_{n} (r_{n}-r_{n-1}) + \gamma_{l+1} (P_{r_{l+1}}-P_{r}) + \sum_{n=l+2}^{k} \gamma_{n} (r_{n}-r_{n-1})\, = \, \sum_{n=1}^{l+1} \gamma_{n} (r_{n}-r_{n-1}) \\
&&< \sum_{n=1}^{l+1} \delta_{n} (r_{n}-r_{n-1})\\
&&= \sum_{n\neq l,l+1} \gamma_{n} (P_{r_{n}} - P_{r_{n-1}}) + \gamma_{l+1} (r_{n}-r_{n-1}) + \sum_{n=l+2}^{L-1} \delta_{n} (r_{n}-r_{n-1})
\end{eqnarray*}
since $(\gamma_{l} - \delta_{l}) (r_{l}-r_{l-1}) + (\gamma_{l+1} - \delta_{l+1}) (r-r_{l}) = 0$.
\item $\tau(A_2) = \tau(A) - \tau(A_{1}) = \tau(S) - \tau(S_{1}) = \tau(S_{2})$.
\end{enumerate}
We thus conclude that we also have that
\[A_{2} \prec S_{2}\]
By proposition(\ref{prop0}), there is a unitary $U_1$ inside $P\mathcal{M}P$ so that $E(U_{1} S_{1} U_{1}^{*}) = A_{1}$. Also, the induction hypothesis holds for the operators $A_{2}$ and $M_{2}$ inside $(I-P)\mathcal{M}(I-P)$ since the partition decomposition for $A_2$ and $S_2$ has length $L-1$. We may therefore find a unitary $U_2$ inside so that $E(U_{2} S_{2} U_{2}^{*}) = A_{2}$. Thus, letting $U = U_{1} \oplus U_{2}$, we have that $E(U S U^{*}) = A$.
\end{proof}
\section{An approximate Schur-Horn theorem}
Theorem(\ref{SHD}) allows us to prove an approximate version of the Schur-Horn theorem for general operators.
\begin{theorem}
Let $S$ be a positive operator in a $II_1$ factor $\mathcal{M}$ and let $\mathcal{A}$ be a masa in $\mathcal{M}$. Then, the norm closure of $E(\mathcal{U}(S))$ equals $\{A \in \mathcal{A}^{+} \mid A \prec S\}$.
\end{theorem}
\begin{proof}
Choose $A$ in $\mathcal{A}^{+}$ so that $A \prec S$. By scaling, if needed, we assume that $A$ and $S$ are strict contractions. Fix $n > 0$ and define the mutually orthogonal projections
\[P_{k} = E_{A}([\frac{k-1}{n},\frac{k}{n})) \quad \text{ for } \quad 1 \leq k \leq n\]
Next, define $\alpha_k = \tau(AP_{k})$ for $1 \leq k \leq n$ and consider the operator $B = \sum_{k=1}^{n} \alpha_k P_{k}$. Since $\tau(C) I \prec C$ for any positive operator $C$, we have that $B \prec A$ and hence, $B \prec S$. We also have that
\[||A - B|| = ||\sum_{k=1}^{n} (A-\alpha_k) P_{k}|| \leq ||\sum_{k=1}^{n} \dfrac{1}{n} P_{k}|| = \dfrac{1}{n}\]
Choose numbers $0 = t_0 \leq t_1 \leq t_2 \cdots \leq t_n = 1$ and orthogonal projections $Q_1, \cdots, Q_n$ in $\{S^{'} \cap \mathcal{M}\}$ such that $Q_{k} \leq E_{S}([t_{k-1},t_{k}])$ and $\tau(Q_{k}) = \tau(P_{k})$ for $1 \leq k \leq n$. To see why this possible, proceed thus: Let $t_0 = 0$ and pick $t_1$ such that $\tau( E_{S}([0,t_{1})))\leq \tau(P_{1}) \leq \tau( E_{S}([0,t_{1}]))$. If $S$ has no atom at $t_1$, then let $Q_1 = E_{S}([0,t_{1}))$. If $S$ has an atom at $t_1$, pick a subprojection $R$ of $E_{S}(\{t_1\})$ such that $\tau(E_{S}([0,t_{1}))) + \tau(R) = \tau(P_1)$ and let $Q_1 = E_{S}([0,t_{1})) + R$. Continue this process for $n=2, \cdots$.
Next, pick positive operators $T_{1}, \cdots, T_{n}$ all with finite spectrum such that for $1 \leq k \leq n$,
\[ T_{k} \prec S Q_{k}, \quad \tau(T_{k}) = \tau(S Q_{k}) \quad \text{ and } ||S Q_{k} - T_{k}|| \leq \dfrac{1}{n}\]
This is done exactly in the same way as the choice of the operator $B$ given the operator $A$, in the first part of this proof. Let $T$ be the operator $T = T_{1} + \cdots + T_{n}$. Then, the above conditions imply
\[T \prec S\quad \text{ and } \quad||S - T|| \leq \dfrac{1}{n} \]
Also, for $1 \leq k \leq m$,
\[\tau(T(Q_1 + \cdots + Q_k)) = \tau(S(Q_1 + \cdots + Q_k)) \geq \tau(A (P_1 + \cdots P_k)) = \tau(B (P_1 + \cdots P_k))\]
and hence, by lemma(\ref{atoMaj}) $B \prec T$. Since $B$ and $T$ have finite spectrum, there is a unitary $U$ so that $B = E(UTU^{*})$. We calculate,
\[||A - E(USU^{*})|| \leq ||A-B|| + ||B - E(UTU^{*}|| + ||E(UTU^{*} - USU^{*})|| \leq \dfrac{1}{n} + 0 + \dfrac{1}{n}\]
and see that $A$ can be arbitrarily well approximated by elements in $E(\mathcal{U}(S))$. Since $A$ was arbitrary, we have that the norm closure of $E(\mathcal{U}(S))$ equals $\{A \in \mathcal{A}^{+} \mid A \prec S\}$.
\end{proof}
\section{Discussion}
The proofs given above can be easily adapted to masas in type $III$ factors that admit a faithful normal conditional expectation. Cartan masas, by definition satisfy this property, but not all masas do - By a result of Takesaki\cite{TakCon}, if every masa in a von Neumann algebra admits a normal conditional expectation, then it is finite. Suppose $\mathcal{A}$ is a masa in a type $III$ factor $\mathcal{M}$ admitting a normal conditional expectation $E : \mathcal{M} \rightarrow \mathcal{A}$. Let $A \in \mathcal{A}$ and $S$ be positive operators. For any self-adjoint operator $T$, let $\alpha(T) = \operatorname{min}(\{x \in \sigma(T)\}$. For any unitary $U$ in $\mathcal{M}$, we have that $||E(USU^{*})|| \leq ||S||$ and that $\alpha(E(USU^{*})) \geq \alpha(S)$. It is now easy to see that a necessary condition for the existence of an element $T \in \mathcal{O}(S)$ such that $E(T) = A$ is that $||A|| \leq ||S||$ and $\alpha(A) \geq \alpha(S)$.
The Schur-Horn problem in type $III$ factors is more tractable that in the type $II_1$ case. Standard arguments allow us to prove the following lemma
\begin{lemma}\label{typeIIILem}
Let $S = \sum_{n=1}^{N} \mu_n F_n$ be a positive contraction with finite spectrum in a type $III$ factor $\mathcal{M}$ with $||S|| = 1$ and $\alpha(S) = 0$. Then, $\mathcal{O}(S)$ contains a non-trivial projection(and thus every projection).
\end{lemma}
With this in hand, it is easy to see that if $A \in \mathcal{A}$ and $S \in \mathcal{M}$ are positive elements with finite spectrum so that $||A|| \leq ||S||$ and $\alpha(A) \geq \alpha(S)$, then we can solve the Schur-Horn problem for $A$ and $S$. There is further, a simple condition that allows us to determine when we can find a unitary so that $E(USU^{*}) = A$.
Suppose $0$ is the point spectrum of $A$, so that there is a projection $P$ in $\mathcal{A}$ so that $PAP = 0$. Suppose we write $A = E(T)$ for some positive operator $T$, then, $E(PTP) = 0$ and hence, $PTP = 0$. Thus, $0$ must be in the point spectrum of $T$. If $A = E(USU^{*})$, we get that $0$ must be in the point spectrum of $USU^{*}$ and hence in the point spectrum of $S$. Similarly, if $1$ is in the point spectrum of a positive contraction $A$ and $A = E(USU^{*})$ for some positive contraction $S$ and a unitary $U$, then, $1$ must be in the point spectrum of $S$ as well. These necessary conditions are also sufficient.
\begin{theorem}
Let $\mathcal{A}$ be a masa inside a type $III$ factor $\mathcal{M}$ admitting a faithful normal conditional expectation $E$ and let $A$ and $S$ be positive operators with finite spectrum in $\mathcal{A}$ and $\mathcal{M}$ respectively and let $E$ be the normal conditional expectation onto $\mathcal{A}$. Assume further that $\alpha(A) \geq \alpha(S)$ and $||A|| \leq ||S||$.
\begin{enumerate}
\item There is an element $T \in \mathcal{O}(S)$ such that $E(T) = A$.
\item Assume additionally that if either $0$ and $||S||$ are in the point spectrum of $A$, then they are in the point spectrum of $S$ as well. Then, there is a unitary $U$ such that $E(U^{*} S U) = A$.
\end{enumerate}
\end{theorem}
We omit the details as they are a straightforward adaptation of the proof of theorem(\ref{SHD}).
In general, we could ask,
\begin{question}
Let $A$ and $S$ be positive operators in $\mathcal{A}$ and $\mathcal{M}$ respectively, where $\mathcal{A}$ is a masa inside a type $III$ factor admitting a normal conditional expectation and so that $||A|| \leq ||S||$ and $\alpha(A) \geq \alpha(S)$. Then, is there an element $T$ in $\mathcal{O}(S)$ so that $E(T) = A$?
\end{question}
Lyapunov's theorem\cite{LyapOr}, which states that the range of any non-atomic vector valued measure taking values in $\mathbb{C}^{n}$ is compact and convex, was reformulated in operator algebraic language by Lindenstrauss\cite{LinLya} to say the following: Let $\Phi$ be a weak* continuous linear map from a non-atomic abelian von Neumann algebra into $\mathbb{C}^{n}$. Then, for any positive contraction $A$, there is a projection $P$ such that $\Phi(A) = \Phi(P)$. Anderson and Akemann, in their superb monograph\cite{LyapAA},
called any theorem concerning linear maps $\Phi: \mathcal{X} \rightarrow \mathcal{Y}$ where $\mathcal{X}$ and $\mathcal{Y}$ are subsets of linear spaces, that assures us that $\operatorname{Ran}(\Phi) = \operatorname{Ran}(\Phi\mid \partial(\mathcal{X}))$ a Lyapunov type theorem. Clearly, Kadison's carpenter problem is a Lyapunov type problem. Anderson and Akemann proved a variety of Lyapunov theorems and showed, quite surprisingly, that Lyapunov theorems are substantially more tractable when the maps considered are singular. The one of most interest to us is
\begin{theorem}[Anderson and Akemann]
Let $\mathcal{A}$ be a masa in an type $II_1$ factor $\mathcal{M}$. Let $F$ be a singular conditional expectation from $\mathcal{M}$ to $\mathcal{A}$. Then every positive contraction can be lifted to a projection $P$ under $F$.
\end{theorem}
There are plenty of singular conditional expectations onto masas in $II_1$ factors\cite{AkeShe}, though none of them are trace preserving. The corresponding Schur-Horn problem cannot be any other than
\begin{question}
Let $\mathcal{A}$ be a masa in an type $II_1$ factor $\mathcal{M}$. Let $F$ be a singular conditional expectation from $\mathcal{M}$ to $\mathcal{A}$. Suppose $A \in \mathcal{A}$ and $S \in \mathcal{M}$ positive contractions that are not multiples of the identity such that $||A|| \leq ||S||$ and $\alpha(A) \geq \alpha(S)$. Then, is there an element $T \in \mathcal{O}(S)$ such that $F(T) = A$?
\end{question}
Finally, an answer to the following related question, which we are unable to solve, should help in solving the Schur-Horn and carpenter problems in type $II_1$ factors.
\begin{question}
Let $A$ be a positive operator in a masa $\mathcal{A}$ inside a $II_1$ factor $\mathcal{M}$. Then, does the norm closure of $\mathfrak{L}(A) = \{S \in \mathcal{M} \mid \exists \, T \in \mathcal{O}(S) \text{ so that } E(T) = A\}$ equal $\{S \in \mathcal{M} \mid A \prec S\}$? Is $\mathfrak{L}(A)$ convex?
\end{question}
\subsection{Acknowledgements}
The authors would like to thank Sabanci University for a research grant that supported the visit of the first author to Sabanci University, Istanbul in September 2011, when part of this work was done. The second author would also like to thank Matt Daws for pointing out on Mathoverflow, a result of Takesaki that is mentioned in the last section.
|
\section{Introduction}
\label{sec-int}
The dominant production process of single top quarks in the Standard Model (SM)
in $ep$ collisions\footnote{Here and in the following, $e$ denotes both the
electron and the positron.} at HERA
is the charged current (CC) reaction $ep \rightarrow \nu t X$~\citesingletopCC,
which has a cross section of less than $1 \fb$~\citeStelzerMoretti.
Flavour changing neutral current (FCNC) processes could enhance single-top production, but they are
strongly suppressed in the SM by the GIM mechanism~\citeGIM. This mechanism forbids FCNCs at the
tree level, allowing only for small contributions at the one-loop level, exploiting the flavour mixing
due to the CKM matrix~\citeCKM.
Several extensions of the SM predict FCNC contributions already at the tree level~\citeFCNCtreelvl.
The search for such new interactions involving the top quark ($ut$ or $ct$ transitions mediated by neutral
vector bosons, $\gamma$ or $Z$) opens an interesting window to look for effects beyond the SM~\citesingletopBSM.
The FCNC couplings $tuV$ and $tcV$, with
$V=\gamma,Z$, have been investigated in $p\widebar{p}$ collisions at
the Tevatron, where searches for the
top-quark decays $t\rightarrow uV$ and $t\rightarrow cV$ \cite{prl:80:2525,
pl:b701:313} were carried out. The Tevatron experiments also constrained the
couplings $tug$ and $tcg$~\cite{prl:102:151801,*pl:b693:81} which induce FCNC transitions mediated by
the gluon.
The couplings $tuV$ and $tcV$ were also investigated in $e^+e^-$
interactions at LEP2 by searching for single-top production through
the reactions $e^+e^-\rightarrow t\widebar{u} \,(+{\rm c.c.})$ and
$e^+e^-\rightarrow t\widebar{c} \,(+{\rm c.c.})$
\cite{pl:b543:173,pl:b521:181,*pl:b549:290,*pl:b590:21}.
No evidence for such interactions was found
and limits were set on the branching ratios
\mbox{${\rm Br}(t\rightarrow q\gamma)$} and \mbox{${\rm Br}(t\rightarrow qZ)$},
with $q=u, c$.
The same FCNC couplings could induce single-top production in
$ep$ collisions,
$ep \rightarrow et X$~\cite{pr:d65:037501}, in which the incoming lepton
exchanges a $\gamma$ or $Z$ with an up quark in the proton, yielding
a top quark in the final state, see \fig{feynmandiagr}. Owing to the large $Z$ mass, this process
is more sensitive to a coupling of the type $tq\gamma$. Furthermore,
large values of $x$, the fraction of the proton momentum carried by the
struck quark, are needed to produce a top quark. Since the $u$-quark
parton distribution function (PDF) of the proton is dominant at large
$x$, the production of single top quark is most sensitive to the
$tu\gamma$ coupling.
In the present study, the top signal was searched for by looking for the decays
$t \rightarrow b e \nu_e$ and $t \rightarrow b \mu \nu_{\mu}$.
At HERA, such event topologies with one lepton with high transverse momentum,
$p_{T}$, and large missing transverse momentum originate predominantly from single-$W$
production, which has a cross section of about $1\pb$~\cite{pl:b672:2,*singlewH1,*singlewH1ZEUS}
and is the most important background to any top signal.
The present analysis extends the previously published ZEUS results~\citeZEUSsingletop which used data
from the HERA \Ronum{1} running period\footnote{Data collected between 1994 and 2000.}, corresponding to a total
integrated luminosity of $0.13 \fbi$. The integrated luminosity used in this analysis is about three
times larger. A combination of the results from the two running periods (total integrated luminosity $0.50 \fbi$)
has been performed.
\section{Theoretical framework}
\label{sec-theo}
The effects of the FCNC transitions induced by couplings of the type
$tuV$ are parameterised using the following
effective Lagrangian~\citeFCNClagrangian:
\begin{equation}
\Delta{\cal L}_{\rm eff} = e\ e_{t}\ \widebar{t} \frac{i \sigma_{\mu\nu} p^{\nu}}{\Lambda}\ \kappa_{\gamma}\ u\ A^{\mu} + \frac{g}{2\cos\theta_W}\ \widebar{t} \gamma_{\mu}\ v_{Z}\ u Z^{\mu}\ + {\rm h.c.}
\end{equation}
where $\kappa_{\gamma}$ and $v_{Z}$ are two FCNC couplings mediating $ut$ transitions, $e$ ($e_t$) is the electron (top quark) electric charge,
$g$ is the weak coupling constant, $\theta_W$ is the weak mixing angle,
$\sigma_{\mu\nu}=\frac{1}{2}(\gamma^{\mu}\gamma^{\nu}-\gamma^{\nu}\gamma^{\mu})$, $\Lambda$ is an effective cut-off parameter which, by convention, is set to the
mass of the $t$ quark, $M_{t}$, $p$ is the momentum of
the gauge boson and $A^{\mu}$ ($Z^{\mu}$) is the photon ($Z$)
field. In the following,
it is assumed that the magnetic coupling $\kappa_{\gamma}$ and the vector coupling
$v_{Z}$ are real and positive.
The cross section for the process $ep \rightarrow e t X$ was evaluated at the
leading order (LO) using the package CompHEP-4.5.1~\cite{nim:a534:250,*CompHEP} and was parameterised in terms
of three parameters describing the effects of the two FCNC couplings, $A_{\sigma}$ and $B_{\sigma}$,
and their interference, $C_{\sigma}$:
\begin{equation}
\sigma_{ep \rightarrow e t X}=A_{\sigma}\kappa_{\gamma}^{2} + B_{\sigma}v_{Z}^{2} + C_{\sigma}\kappa_{\gamma} v_{Z}.
\end{equation}
The decay widths of the top in the different channels were also evaluated using CompHEP-4.5.1:
\begin{equation}
\Gamma_{t\rightarrow u \gamma}=A_{\Gamma}\kappa_{\gamma}^{2},\ \ \ \
\Gamma_{t\rightarrow u Z}=B_{\Gamma}v_{Z}^{2},\ \ \ \
\Gamma_{t\rightarrow q W}=C_{\Gamma},\ \ \ \
\end{equation}
where $A_{\Gamma}$ and $B_{\Gamma}$ are the partial width of the top corresponding to $u\gamma$ and $uZ$
unitary FCNC couplings, respectively, and $C_{\Gamma}$ is the SM top width.
The above parameters, summarised in \Tab{pars}, were evaluated using the top mass
$M_{t}=172.0\pm1.6$ \gev~\cite{pdg} and the PDF set CTEQ6L1~\cite{jhep:07:012}.
The interference parameter $C_{\sigma}$ has only a small effect, producing a
cross section variation of less than $0.5\%$ in the whole range of the couplings considered in this analysis,
and was therefore neglected.
The QCD corrections to the LO cross-section were evaluated at the
approximate next-to-leading order (NLO) and next-to-next-to-leading order
(NNLO)~\cite{pr:d65:037501,jhep:12:04} for magnetic couplings both at
the $\gamma$ and $Z$ vertices. Since we considered a different
coupling (vector coupling) at the $Z$ vertex, we used such corrections
only to evaluate the limits for the $\gamma$ exchange (see~\sect{limit1d}).
Such corrections increase the LO cross-section by $15\%$ and slightly reduces
the uncertainties due to the QCD factorisation-scale (see~\sect{systematic}).
The limits involving both coupling (see~\sect{limit2d}) were evaluated using
the LO cross-section.
\section{Experimental setup}
\label{sec-exp}
The analysis is based on $ep$ collisions recorded with the ZEUS detector during the
HERA \Ronum{2} running period\footnote{Data collected between 2004 and 2007.}, using
an integrated luminosity of $0.37 \fbi$, divided into two approximately equal samples of
$e^+p$ and $e^-p$ collisions.
The lepton beams were polarised, with roughly equal luminosities
for positive and negative polarisation, such that the average
polarisation was negligible for this analysis.
A detailed description of the ZEUS detector can be found elsewhere~\cite{zeus:1993:bluebook}.
A brief outline of the components that are most relevant for this analysis is given below.
Charged particles were tracked in the central tracking detector (CTD)~\cite{nim:a279:290,*npps:b32:181,*nim:a338:254}
which operated in a magnetic field of $1.43\Tesla$ provided by a thin
superconducting solenoid. The CTD consisted of 72~cylindrical drift chamber
layers, organised in nine superlayers covering the
polar-angle\footnote{The ZEUS coordinate system is a right-handed Cartesian
system, with the $Z$ axis pointing in the proton beam direction,
referred to as the ``forward direction'', and the $X$ axis pointing towards
the centre of HERA. The coordinate origin is at
the nominal interaction point.
The pseudorapidity is defined as \mbox{$\eta=-\ln\left(\tan\frac{\theta}{2}\right)$},
where the polar angle, $\theta$, is measured with respect to the proton beam
direction.}
region \mbox{$15^\circ<\theta<164^\circ$}.
The CTD was complemented by a silicon microvertex detector
(MVD)~\cite{nim:a581:656}, consisting of three active layers in the barrel
and four disks in the forward region.
For CTD-MVD tracks that pass through all nine CTD superlayers, the momentum
resolution was $\sigma(p_{T})/p_{T} = 0.0029p_{T} \oplus 0.0081 \oplus 0.0012/p_{T}$
with $p_{T}$ in GeV.
The high-resolution uranium--scintillator calorimeter (CAL)~\cite{nim:a309:77,*nim:a309:101,*nim:a321:356,*nim:a336:23}
consisted of three parts: the forward (FCAL), the barrel (BCAL) and the rear (RCAL)
calorimeters. Each part was subdivided transversely into towers and
longitudinally into one electromagnetic section (EMC) and either one (in RCAL)
or two (in BCAL and FCAL) hadronic sections (HAC). The smallest subdivision of
the calorimeter was called a cell. The CAL energy resolutions, as measured under
test-beam conditions, were $\sigma(E)/E=0.18/\sqrt{E}$ for electrons and
$\sigma(E)/E=0.35/\sqrt{E}$ for hadrons, with $E$ in $\Gev$.
The luminosity was measured using the Bethe-Heitler reaction
$ep \rightarrow e \gamma p$ by a luminosity detector which consisted of
a lead--scintillator calorimeter~\cite{Desy-92-066,*zfp:c63:391,*acpp:b32:2025}
and an independent magnetic spectrometer~\cite{nim:a565:572}.
The fractional uncertainty on the measured luminosity was $1.9\%$.
\section{Monte Carlo simulation}
Samples of events were generated using Monte Carlo (MC)
simulations to determine the selection efficiency for
single-top events produced through FCNC processes and to estimate
background rates from SM processes. The generated events were passed
through the {\sc Geant-3.21}~\cite{tech:cern-dd-ee-84-1} ZEUS
detector- and trigger-simulation programs
\cite{zeus:1993:bluebook}. They were reconstructed and analysed by the
same program chain as the data.
Single-top samples were generated with {\sc Comphep} 4.5.1,
interfaced with {\sc Pythia} 6.14 \cite{cpc:135:238} for
parton showering, hadronisation and particle decay.
The mass of the top quark in {\sc Comphep} was set to
$M_{t} = 175 \gev$. Different sets
were produced for the two
different production processes
($\gamma$- and $Z$-mediated) and for the two decay modes
($t \rightarrow bW$ and $t \rightarrow uZ$).
Alternative sets were also generated, only for the $\gamma$-mediated process, with
the {\sc Hexf} generator~\cite{lsuhe-145-1993} assuming top-quark masses
of $170$ and $175$~GeV.
These sets were used to study the small effect of $M_{t}$ variation, in order to correct the
selection efficiency, evaluated using the {\sc Comphep} samples, for the different $M_t$ values
used in the generation and in the cross-section calculation (see~\sect{theo}).
Initial-state radiation from the lepton beam was included using the Weizs\"acker-Williams
approximation~\cite{prep:146:1}. The hadronic final state was simulated using
the matrix-element and parton-shower model of
{\sc Lepto}~\cite{cpc:101:108}
for the QCD cascade and the Lund string model~\cite{prep:97:31} as
implemented in {\sc Jetset}~\cite{cpc:39:347,*cpc:43:367} for the
hadronisation.
The results for {\sc Comphep} and the alternative samples agree within uncertainties.
Standard Model single-$W$ production is the most significant background to
top production.
Another important background in the electron-decay channel of the $W$
($t\rightarrow bW\rightarrow be\nu$) arises from neutral current (NC)
deep inelastic scattering (DIS). In addition, two-photon processes provide a
source of high-$p_T$ leptons that are a significant background in the
muon-decay channel of the $W$ ($t\rightarrow bW\rightarrow b\mu\nu$).
The CC DIS is a minor source of background for both channels.
The following MC programs were used to simulate the different background
processes.
Single-$W$ production was simulated using the event generator
{\sc Epvec}~\cite{np:b375:3} which did not include hard QCD radiation.
The $ep \rightarrow eWX$ and $ep \rightarrow \nu WX$ events from
{\sc Epvec} were scaled by a factor dependent on the transverse momentum and rapidity
of the $W$, such that the resulting cross section corresponded to a
calculation including QCD corrections at next-to-leading order
\cite{Diener:2002if,*Nason:1999xs}.
Neutral current and CC DIS events were simulated using the {\sc Lepto}~6.5
program~\cite{cpc:101:108}, interfaced to
{\sc Heracles}~4.6.1~\cite{cpc:69:155,*spi:www:heracles}
via {\sc Djangoh}~1.1~\cite{cpc:81:381,*spi:www:djangoh11}.
The {\sc Heracles} program includes photon
and $Z$ exchanges and first-order electroweak radiative
corrections. The QCD cascade was modelled with the colour-dipole
model~\cite{Azimov:1986sf,*Gustafson:1986db,*np:b306:746,*Andersson:1988gp} by
using the {\sc Ariadne}~4.08 program~\cite{cpc:71:15,*zfp:c65:285}.
Two-photon processes were simulated using the generator
{\sc Grape}~1.1~\cite{cpc:136:126}, which includes dilepton production via
$\gamma\gamma$, $Z\gamma$ and $ZZ$ processes and considers both
elastic and inelastic production at the proton vertex.
\section{Event selection}
The event selection was optimised for single-top production via photon exchange,
looking for the dominant decay $t \rightarrow bW $ and
subsequent $W$ decay to $e$ and $\mu$ and their respective neutrinos.
The selection is based on requiring an isolated high-$p_{T}$ lepton and a large missing
transverse momentum.
Cosmic background, relevant especially for the muon channel, was suppressed
using timing cuts based on calorimeter measurements and the track impact
parameter with respect to the beam spot.
Further cosmic background overlapping with $ep$ interactions was rejected
by applying a cut $E-p_{Z} < 60 \gev$, $E-p_{Z}$ being the sum
of the total and longitudinal energy deposits of the cells in the calorimeter.
For fully contained events, $E-p_{Z}$ is twice the electron-beam energy and peaks
at $55 \gev$.
Events from beam-gas interactions were rejected on the basis of the ratio of the number
of tracks pointing to the vertex to the total number of tracks in an event.
\subsection{Online selection}
\label{sub-trigcon}
A three-level trigger system was used to select events
online~\cite{uproc:chep:1992:222,*nim:a580:1257}.
At the first level, coarse calorimeter and tracking information
were available.
Events were selected using criteria based on either the transverse
energy or missing transverse momentum measured in the CAL.
Events were accepted with a low threshold on these quantities when a
coincidence with CTD tracks from the event vertex was found, while
a higher threshold was used for events with no CTD tracks.
At the second level, timing information from the CAL was used to reject
events inconsistent with an $ep$ interaction. In addition, the
topology of the CAL energy deposits was used to reject non-$ep$ background
events. In particular, a tighter cut was made on missing transverse momentum, since
the resolution in this variable was better at the second than at the first
level.
At the third level, track reconstruction and vertex finding were
performed and used to reject events with a vertex inconsistent with
$ep$ interactions.
Cuts were applied to calorimeter quantities and reconstructed tracks
to further reduce beam-gas contamination.
\subsection{Offline selection}
Jets, used in the selection to define lepton isolation, were reconstructed from CAL cells using the $k_{T}$
cluster algorithm \cite{np:b406:187} in the longitudinally invariant inclusive
mode \cite{pr:d48:3160} and were corrected
for energy loss due to the dead material in front of the CAL. The jets
were required to have a transverse energy $E_{T}^{{\rm jet}} > 4.5 \gev$ and
pseudorapidity $|\eta^{{\rm jet}}| < 2.5$.
{\bf Muon selection}
Muons were reconstructed by matching calorimeter cell-patterns compatible with a
minimum-ionising particle to CTD tracks \cite{nim:a453:336}.
Events were selected as follows:
\begin{itemize}
\item{$\mid Z_{\rm vtx} \mid < 30 \cm$, $Z_{\rm vtx}$ being the $Z$ coordinate of the interaction vertex,
to restrict to a region compatible with $ep$ interactions;}
\item{$ E-p_{Z} > 10 \gev$.
The $E-p_{Z}$ of the CAL deposit associated with
the muon was replaced by that of the muon track.
This requirement rejected photoproduction events, which populate the low $E-p_{Z}$ region;}
\item{$P_{T}^{{\rm miss}} > 10 \gev$, $P_{T}^{{\rm miss}}$ being the missing transverse momentum measured by the CAL;}
\item{at least one muon candidate with the following characteristics:}
\begin{itemize}
\item{a track from the primary vertex matched with a
CTD track with at least three hit superlayers
and a transverse momentum, $p_{T}^{\mu}$, greater than $8 \gev$;}
\item{the distance, $\Delta R$, of the muon candidate in the pseudorapidity-azimuth
($\eta$-$\phi$) plane with respect to any other track and jet in the event satisfying
$\Delta R = \sqrt{(\Delta\eta)^2+(\Delta\phi)^2} > 0.5$.}
\end{itemize}
\end{itemize}
A total of $269$ events were selected, while $260 \pm 3$ (stat.)
were expected from the SM, which is dominated by the dimuon production from
the $\gamma\gamma$ process.
The quoted uncertainty is the error on the expected SM prediction due to the MC
statistics.
\Fig{mu_prel} shows the comparison between data and
MC for the variables $p_{T}^{\mu}$, $\theta^{\mu}$,
acoplanarity ($\phi^{\rm acop}$), $P_T^{{\rm miss}}$,
transverse mass ($M_T$),
hadronic transverse momentum ($P_{T}^{{\rm had}}$). Here $P_{T}^{{\rm had}}$, $M_T$ and $\phi^{\rm acop}$ are defined
as follows:
\begin{itemize}
\item[-] $P_{T}^{{\rm had}} = \sqrt{(\sum_i{P_X^{i}})^2 + (\sum_i{P_Y^{i}})^2}$, where $P_X^{i}$ and
$P_Y^{i}$ are the $X$ and $Y$ components of the CAL
energy deposits not associated with the lepton;
\item[-] $M_T = \sqrt{2p_T^{l}p_T^{\nu}(1 - \cos{\phi^{l\nu}})}$, where $p_T^{l}$ is the
lepton transverse momentum, $p_T^{\nu}$ is the modulus of the missing $P_T$ vector
obtained from the CAL and corrected using track information to account for muons,
$\phi^{l\nu}$ is the azimuthal separation between the lepton and the
missing $P_T$ vector;
\item[-] $\phi^{\rm acop}$
is the angle
between the lepton and the vector balancing the $P_{T}^{\rm had}$ and
is defined for events with $P_{T}^{\rm had}$ greater than $1 \gev$.
\end{itemize}
Reasonable agreement is observed in all cases.
{\bf Electron selection}
Electrons were reconstructed using an algorithm that combined information from
the cluster of the energy deposits in the
calorimeter with tracks \cite{epj:c11:427}. Events were selected as follows:
\begin{itemize}
\item{$\mid Z_{\rm vtx} \mid < 30 \cm$;}
\item{$5 < E-p_{Z} < 50 \gev$, to reject NC DIS and photoproduction background;}
\item{$P_{T}^{{\rm miss}} > 12 \gev$;}
\item{at least one electron candidate with the following characteristics:}
\begin{itemize}
\item[-]{$p_{T}^{\rm el} > 10 \gev$;}
\item[-]{$ 0.3 < \theta^{\rm el} < 2 \rad$;}
\item[-]{isolated from other tracks and jets
in the event, $\Delta R >0.5$;}
\item[-]{the extrapolation of the track associated with the electron
into the CAL should have a
distance of closest approach to the CAL cluster centre
$ < 10 \cm$ and a reconstructed momentum $p > 5 \gev$;}
\end{itemize}
\item{$M_{T} > 10 \gev$, to reject events with
$P_{T}^{{\rm miss}}$ along the electron direction;}
\item{$0.1 < \phi^{\rm acop} < (\pi - 0.1) \rad$,
to reject badly reconstructed NC DIS events with $P_{T}^{{\rm miss}}$ in
the direction of the electron or of the jet.}
\end{itemize}
A total of $245$ events were selected, while $253 \pm 6$ (stat.)
were expected from the SM, which is dominated by the NC DIS process.
The quoted uncertainty is the error on the expected SM prediction due to the MC
statistics.
\Fig{el_prel} shows the comparison between data and
MC for the variables $p_{T}^{\rm el}$, $\theta^{\rm el}$,
$\phi^{\rm acop}$, $P_T^{{\rm miss}}$, $M_T$, $P_{T}^{{\rm had}}$. Reasonable agreement
is observed in all cases.
\subsection{Selection of single-top candidates}
\label{sec-singletopsel}
Since no excess of events above the SM expectation was observed,
a further selection was made to maximise the sensitivity to a possible
FCNC single top signal.
A cut on $P_{T}^{{\rm had}}$ of $40 \gev$ was applied to
both decay channels while the cuts
on $\phi^{\rm acop}$ and $P_{T}^{{\rm miss}}$ were optimised separately for
the two channels:
\begin{itemize}
\item{$P_{T}^{{\rm had}} > 40 \gev$ for both channels;}
\item[]{\textbf{muon channel:}}
\begin{itemize}
\item{$\phi^{\rm acop} > 0.05 \rad$;}
\item{events with more than one isolated muon were rejected;}
\end{itemize}
\item[]{\textbf{electron channel:}}
\begin{itemize}
\item{$\phi^{\rm acop} > 0.15 \rad$;}
\item{$P_{T}^{{\rm miss}} > 15 \gev$.}
\end{itemize}
\end{itemize}
One event survived the selection cuts in the electron channel
while three events were found in the muon channel.
\Tab{finalevts} summarises the results of the final selection.
In order to compare the MC to data,
the $P_{T}^{{\rm had}}$ cut was relaxed to $25 \gev$. Figures~\ref{fig-pthad25} (a) and (b) show
the $P_{T}^{{\rm had}}$ behaviour for data and SM expectations
for the muon and electron channels, respectively.
Good agreement between data and predictions is observed for both channels.
Also shown are the expectations for top production through FCNC, normalised to
the limit on the signal cross section obtained in~\sect{limit1d}.
The data do not support a significant contribution from this process.
\section{Systematic uncertainties}
\label{sec-systematic}
The following systematic uncertainties were taken into account:
\begin{itemize}
\item the theoretical uncertainty on the $W$ background normalisation was assumed to
be $\pm 15\%$\cite{Diener:2002if,*Nason:1999xs};
\item the statistical uncertainty on the total SM prediction after the final selection
was $\pm 13\%$ and $\pm 9\%$ for the $e$- and $\mu$-channel, respectively;
\item the uncertainty on the NC DIS background, particularly relevant for the $e$-channel,
was evaluated using a sample of events enriched in NC DIS by replacing the $E-p_{Z}$ and
acoplanarity cuts by $E-p_{Z}>40 \gev$
and $\phi^{\rm acop} < 0.3$. A systematic uncertainty of $\pm 15\%$ on this source was determined
by the level of agreement between data and MC for such a selection. The effect of this
uncertainty on the final selection SM prediction was $\pm6\%$ for the $e$-channel and
negligible for the $\mu$-channel;
\item the uncertainty on the electromagnetic and the hadronic CAL energy scale
was assumed to be $\pm 1\%$ and $\pm 2\%$, respectively.
The two scale uncertainties, summed in quadrature, produced a variation
of $\pm 6\%$ and of $\pm 5\%$ on the final SM predictions for
the $e$- and the $\mu$-channel, respectively, while the effect on the signal selection efficiencies
was below $2 \%$ and was therefore neglected;
\item the uncertainty on the top mass, $172.0\pm 1.6 \gev$ \cite{pdg}, produced a variation on the
parameters of the signal cross section and decay widths as reported in \Tab{pars} and a variation
of $\pm 2\%$ on the signal selection efficiencies;
\item the uncertainties on the signal efficiency due to the statistics of the MC samples are
reported in \Tab{eff} for the different channels and decay processes;
\item the uncertainties on the PDFs gave a variation on the parameters of the signal cross section
as reported in \Tab{pars}. Such uncertainties were evaluated as suggested by the CTEQ
group~\cite{jhep:07:012};
\item the uncertainty due to the QCD factorisation-scale affected the signal cross section
by $\pm 9\%$ for the LO calculation and by $^{+8\%}_{-7\%}$ including the approximated NLO and
NNLO QCD corrections (see~\sect{theo}).
This effect was evaluated by varying the central value, set to $M_{t}$, between
$M_{t}/2$ and $2M_{t}$;
\item the uncertainty on the luminosity determination was $\pm 1.9 \%$.
\end{itemize}
The uncertainties due to the $W$ normalisation, CAL energy scale, top mass,
PDFs and luminosity were assumed to be correlated for the different channels and datasets.
All the above uncertainties were included in the limit calculation as explained in
\sect{limit1d}.
\section{Limits on FCNC}
\label{sec-limitsonfcnc}
Since no excess over the SM prediction was observed, limits on FCNC couplings
of the type $tuV$ were evaluated using the results of \Tab{finalevts}.
As a first step, limits were evaluated
on the signal cross section and on the $\kappa_{\gamma}$ coupling assuming $v_{Z}=0$.
In a second step, the effect of a non-zero $v_{Z}$ coupling was accounted for.
Limits on the anomalous top branching ratios, {\rm Br}($t \rightarrow u \gamma$) ({\rm Br}$_{u\gamma}$) and
{\rm Br}($t \rightarrow u Z$) ({\rm Br}$_{uZ}$), were evaluated.
\subsection{Limits on the cross section and \bm{$\kappa_{\gamma}$} }
\label{sec-limit1d}
The limit on the anomalous top-production cross section was evaluated using a Bayesian approach and assuming a
constant prior in the cross section, $\sigma$:
\begin{eqnarray}
f(\sigma|{\rm data}) &=& \frac{\prod_i P(N_{i}^{\rm obs}|\sigma)f_0(\sigma)}{\int_0^{\infty} \prod_i P(N_{i}^{\rm obs}|\sigma)f_0(\sigma)d\sigma},\\
P(N_{i}^{\rm obs}|\sigma)&=&\frac{\mu_i^{N_{i}^{\rm obs}}e^{-\mu_i}}{N_{i}^{\rm obs}!},\\ \nonumber
\mu_i&=&N_{i}^{\rm sig}+N_{i}^{\rm bg},\\ \nonumber
N_{i}^{\rm sig}&=&\sigma \mathcal{L}_i \epsilon_i, \\ \nonumber
\label{eq:eq3}
\end{eqnarray}
where $f(\sigma|{\rm data})$ is the posterior probability density function (p.d.f.) of the signal cross section,
$f_0(\sigma)$ its prior, $i$ runs over the different channels and datasets, $N_{i}^{\rm obs}$ is the number
of events surviving
the event selection, $N_{i}^{\rm sig}$ and $N_{i}^{\rm bg}$ are the number of
signal events and the expected SM background, $\mathcal{L}_i$ is the
integrated luminosity and $\epsilon_i$ the signal efficiency including branching ratio for each decay channel
(see the first row in \Tab{eff}). The branching ratio of the top to $u\gamma$ was taken into account in the limits
evaluation, the selection efficiency for such channel is expected to be low and was therefore set to zero.
The systematic uncertainties were treated as nuisance parameters (NPs) and included in the limit calculation, integrating out their dependence (marginalisation) assuming Gaussian priors\footnote{In case of unphysical values, the Gaussian
priors were truncated.}.
The marginalisation over the NPs and the extraction of the posterior p.d.f. was performed using the package {\sl Bayesian Analysis Toolkit} \cite{bat}, which carries out multidimensional integration using the Markov Chain Monte Carlo technique.
The $95\%$ Credibility Level (C.L.) limit on the cross section
was evaluated by integrating the posterior p.d.f.
\begin{equation}
\centering
\int_{0}^{\sigma_{95}}f(\sigma|{\rm data})d\sigma = 0.95,
\end{equation}
and found to be
\begin{equation}
\centering
\sigma < 0.24 \pb\ (95\%~{\rm C.L.})~{\rm at}~\sqrt{s}=318\gev.
\end{equation}
The limit on the cross section was converted into a limit on the coupling
$\kappa_{\gamma}$, assuming a vanishing $v_{Z}$ coupling and using the $A_{\sigma}$
parameter described in \sect{theo} taking into account the approximated NLO and NNLO
QCD corrections (see~\sect{theo}):
\begin{equation}
\centering
\kappa_{\gamma}< 0.17\ (95\%~{\rm C.L.}).
\end{equation}
The limit is similar to that obtained by ZEUS from HERA \Ronum{1} data
\cite{singletophera1} with an integrated luminosity of $0.13 \fbi$.
In the HERA \Ronum{1} data, no events were found in either the electron
or muon channel and also the hadronic $W$-decay channel was exploited.\\
\vspace{0.1cm}
The present result was combined with the HERA \Ronum{1} limit for a total
integrated luminosity of $0.50 \fbi$, using the same Bayesian approach
as described above and assuming full correlation for the systematic uncertainties due to
the $W$ normalisation, CAL energy scale, top mass and PDFs.
The combined cross-section and $\kappa_{\gamma}$ limits are:
\begin{equation}
\centering
\sigma < 0.13 \pb\ (95\%~{\rm C.L.})~{\rm at}~\sqrt{s}=315\gev,
\end{equation}
\begin{equation}
\centering
\kappa_{\gamma} < 0.12\ (95\%~{\rm C.L.}).
\end{equation}
\vspace{0.1cm}
The combined cross-section limit corresponds to a centre-of-mass energy of $315 \gev$
since part of the HERA \Ronum{1} data was collected at $\sqrt{s}=300\gev$.
\subsection{Limits on the top anomalous branching ratios}
\label{sec-limit2d}
Following the Bayesian approach described above, a two-dimensional posterior p.d.f.,
\begin{equation}
\centering
f({\rm Br}_{u\gamma},{\rm Br}_{uZ}|{\rm data}),
\end{equation}
was evaluated combining the HERA \Ronum{1} and HERA \Ronum{2} datasets.
Such a p.d.f. was built using the parameters described in \sect{theo} (no higher-order
QCD corrections were applied in this case) to
express the FCNC cross-section in terms of the anomalous top branching ratios.
The signal efficiencies for the different production channels ($\gamma$- or
$Z$-mediated) and decay modes ($bW$ or $uZ$) were taken into account
(see \Tab{eff}). The selection efficiency of the $e$-channel is larger
for the $Z$-mediated process than the $\gamma$-mediated process, since in
this case the final-state electron is scattered at a larger angle and is
more often visible in the detector.
The decay channel $t \rightarrow u\gamma$ was not simulated since the
branching ratio is very low for the range of couplings under consideration.
In addition, the selection efficiency is expected to be low for such events
and was therefore set to zero.
The $95\%~{\rm C.L.}$ boundary in the $({\rm Br}_{u\gamma},{\rm Br}_{uZ})$ plane
was evaluated as the set of points
\begin{equation*}
\centering
f({\rm Br}_{u\gamma},{\rm Br}_{uZ}|{\rm data})=\rho_0,
\end{equation*}
where $\rho_0$ was chosen such that
\begin{equation}
\centering
{\int\int}_{f({\rm Br}_{u\gamma},{\rm Br}_{uZ}|{\rm data})>\rho_0} d{\rm Br}_{u\gamma} d{\rm Br}_{uZ} \ f({\rm Br}_{u\gamma},{\rm Br}_{uZ}|{\rm data}) = 0.95.
\end{equation}
\Fig{brlim} shows the ZEUS boundary in the (${\rm Br}_{u\gamma},{\rm Br}_{uZ}$)
plane compared to limits from H1~\cite{singletoph1} and from experiments
at other colliders:
ALEPH~\cite{pl:b543:173} at LEP (other LEP experiments~\cite{pl:b521:181}
have similar results), CDF~\cite{prl:80:2525} and D0~\cite{pl:b701:313}
at Tevatron. The $e^+e^-$ and
hadron colliders, contrary to HERA, have similar sensitivity to $u$- and
$c$-quark;
their limits are hence on both decays $t\rightarrow qV$ with $q=u,c$.
The limits set by the ZEUS experiment in the region where ${\rm Br}_{uZ}$ is less than
$4\%$ are the best to date.
\section{Conclusions}
\label{sec-conclusions}
A search for possible deviations from the Standard Model predictions due to
flavour-changing neutral current top production in events with
high-$p_T$ leptons and high missing transverse momentum was performed using an integrated
luminosity of $0.37 \fbi$, collected by the ZEUS detector in 2004--2007.
Since no significant deviation from the expectation was observed,
the results were used to put limits on the anomalous production of
single top quarks at HERA.
A $95\%$ credibility-level upper limit on the cross section of $\sigma < 0.24 \pb$ at
a centre-of-mass energy of $318 \gev$ was obtained.
The limit was combined with a previous ZEUS result, obtained using
HERA \Ronum{1} data, for a total integrated luminosity of $0.50 \fbi$,
giving a combined $95\%$ credibility-level upper limit of
$\sigma < 0.13 \pb$ at $\sqrt{s}=315 \gev$.
This limit, assuming a vanishing coupling of the top quark to the $Z$ boson, $v_{Z}$,
corresponds to a constraint on the coupling of the top to the $\gamma$, $\kappa_{\gamma}$,
of $\kappa_{\gamma} < 0.12$. Constraints on the anomalous top branching ratios
$t \rightarrow u\gamma$ and $t \rightarrow uZ$ were also evaluated assuming
a non-zero $v_{Z}$.
For low values of $v_{Z}$, resulting in branching ratios
of $t \rightarrow uZ$ of less than $4\%$, this paper provides the current best limits.
\section*{Acknowledgements}
\label{sec-ack}
\Zacknowledge
\vfill\eject
|
\section{Introduction}
Loop Quantum Gravity (LQG) is a tentative non-perturbative and background-independent
quantization of General Relativity (GR) \cite{lqg_review}. Interestingly, it has now been demonstrated that different
approaches, based on canonical quantization of GR, on covariant quantization of GR and on formal quantization
of geometry lead to the very same LQG theory. Although this is rather convincing, a direct experimental probe is
still missing. One can easily argue that cosmology is the most promising approach to search for observational features
of LQG or, more specifically, to its symmetry-reduces version, Loop Quantum Cosmology (LQC) \cite{lqc_review}.
Many efforts have been devoted to the search of possible footprints of LQC in cosmological tensor modes (see \cite{tensor}).
At the theoretical level, the situation is easier in this case as the algebra of constraints is automatically anomaly-free.
But, as far as observations are concerned, scalar modes are far more important. They have already been observed in
great details by WMAP \cite{wmap} and are currently even better observed by the Planck mission. The question of a possible
modification of the primordial scalar power spectrum (and of the corresponding TT $C_l$ spectrum) in LQC is therefore essential
in this framework.
Gravity is described by a set of constraints. However, for the (effective) theory to be consistent, it is mandatory that the
evolution generated by the constraints remains compatible with the constraints themselves. This is always true if their mutual Poisson
brackets vanish when evaluated in fields fulfilling the constraints, {\it i.e.} if they form a first class algebra. This
means that the evolution and the gauge transformations are associated with vector fields that are tangent to the manifold of
null constraints. This obviously holds at the classical level. However, when quantum modifications are added, the anomaly
freedom is not anymore automatically ensured. Possible quantum corrections must be restricted to those which close the
algebra. This means that, for consistency reasons, the Poisson brackets between any two constraints must be proportional to
one constraint of the algebra. This article is devoted to the search for such an algebra for scalar perturbations.
Our approach will follow the one developed by Bojowald {\it et al.} in \cite{Bojowald:2008gz}. There are two main quantum
corrections expected from LQC: inverse volume terms, basically arising for inverse powers of the densitized triad, which when
quantized become an operator with zero in its discrete spectrum thus lacking a direct inverse, and holonomy corrections coming
from the fact that loop quantization is based on holonomies, rather than direct connection components. In \cite{Bojowald:2008gz}
the authors focused exclusively on inverse volume corrections. Here, we extend with work to the holonomy corrections.
Scalar perturbations with holonomy corrections have been studied in \cite{Wu:2010wj}. However, the issue of anomaly freedom
was not really addressed. Recently, a new possible way of introducing holonomy corrections to the scalar perturbations was
proposed in \cite{WilsonEwing:2011es}. Although it was interestingly shown that the formulation is anomaly-free, the approach is based on
the choice of the longitudinal gauge and the extension of the method to the gauge-invariant case is not straightforward. In contrast,
the approach developed in our paper does not rely on any particular choice of gauge and the gauge-invariant
cosmological perturbations are easily constructed.
The theory of anomaly-free scalar perturbations developed in this paper is obtained on a flat FRW background, such
that the line element is given by:
\begin{equation}
ds^2 = a^2\left[ -(1+2\phi)d\eta^2+2\partial_a B d\eta dx^a+ ((1-2\psi)\delta_{ab}+2\partial_a \partial_b E)dx^adx^b \right],
\label{lineelement}
\end{equation}
where $\phi$, $\psi$, $E$ and $B$ are scalar perturbation functions. The matter content
is assumed to be a scalar field. This will allow us to investigate the generation of scalar
perturbations during the phase of cosmic inflation while taking into account the quantum gravity effects.
Our analysis of the scalar perturbations is performed in the Hamiltonian framework developed in
\cite{Bojowald:2008gz,Bojowald:2008jv}.
As it was shown there, the background variables are $(\bar{k},\bar{p},\bar{\varphi},\bar{\pi})$, while the perturbed
variables are $(\delta K^i_a,\delta E^a_i, \delta \varphi, \delta \pi)$. The Poisson bracket for the system can be decomposed
as follows:
\begin{equation}
\{ \cdot , \cdot \} = \{ \cdot, \cdot \}_{\bar{k},\bar{p}} + \{\cdot, \cdot \}_{\delta K, \delta E}
+ \{\cdot,\cdot \}_{\bar{\varphi},\bar{\pi}} + \{\cdot,\cdot \}_{\delta \varphi, \delta \pi}
\label{Poisson}
\end{equation}
where
\begin{eqnarray}
\{ \cdot, \cdot\}_{\bar{k},\bar{p}} &:=&\frac{\kappa}{3 V_0} \left[ \frac{\partial \cdot}{\partial \bar{k}} \frac{\partial \cdot}{\partial \bar{p}}
-\frac{\partial \cdot}{\partial \bar{p}} \frac{\partial \cdot}{\partial \bar{k}}\right], \\
\{\cdot, \cdot \}_{\delta K, \delta E} &:=&\kappa \int_{\Sigma} d^3x \left[ \frac{\delta \cdot}{\delta \delta K^i_a} \frac{\delta \cdot}{\delta \delta E^a_i}
-\frac{\delta \cdot}{\delta \delta E^a_i} \frac{\delta \cdot}{\delta \delta K^i_a}\right], \\
\{\cdot, \cdot \}_{\bar{\varphi},\bar{\pi}} &:=&\frac{1}{V_0} \left[ \frac{\partial \cdot}{\partial \bar{\varphi}} \frac{\partial \cdot}{\partial \bar{\pi}}
-\frac{\partial \cdot}{\partial \bar{\pi}} \frac{\partial \cdot}{\partial \bar{\varphi}}\right], \\
\{\cdot, \cdot \}_{\delta \varphi, \delta \pi} &:=&\int_{\Sigma} d^3x \left[ \frac{\delta \cdot}{\delta \delta \varphi} \frac{\delta \cdot}{\delta \delta \pi}
-\frac{\delta \cdot}{\delta \delta \pi} \frac{\delta \cdot}{\delta \delta \varphi}\right].
\end{eqnarray}
Here, $V_0$ is the volume of the fiducial cell and $\kappa=8\pi G$.
The holonomy corrections are introduced by the replacement $\bar{k} \rightarrow \mathbb{K}[n]$ in the classical Hamiltonian. We follow the notation introduced in \cite{VectorJTAJ}, where
\begin{equation}
\mathbb{K}[n] := \left\{
\begin{tabular}{ccc}
$\frac{\sin(n\bar{\mu} \gamma \bar{k})}{n\bar{\mu}\gamma}$ & for &$n \in \mathbb{Z}/\{0\}$, \\
& & \\
$\bar{k}$ & for & $n=0$,
\end{tabular}
\right.
\end{equation}
for the correction function. In cases where $\bar{k}$ appears quadratically, the integer
$n$ is fixed to two (See \cite{VectorJTAJ}). In the other cases, the integers remain to be fixed
from the requirement of anomaly freedom. The coefficient $\gamma$ is the Barbero-Immirzi parameter
and $\bar{\mu} \propto \bar{p}^{\beta}$ where $-1/2 \leq \beta \leq 0 $. In what follows, the relation
\begin{equation}
\bar{p} \frac{\partial}{\partial \bar{p}}\mathbb{K}[n] = [\bar{k} \cos(n \bar{\mu}\gamma \bar{k})-\mathbb{K}[n] ]\beta
\end{equation}
will be useful.
The organization of the paper is the following. In Sec. \ref{Sperthol}, the holonomy-corrected gravitational
Hamiltonian constraint is defined. We calculate the Poisson bracket of the Hamiltonian constraint with itself and
with the gravitational diffeomorphism constraint. In Sec. \ref{matter}, scalar matter is introduced.
The Poisson brackets between the total constraints for the system under consideration are calculated.
In Sec. \ref{anomaly}, the conditions for anomaly freedom are solved and the expressions for the
counter-terms are derived. Based on this, in Sec. \ref{equations}, equations of motion for the scalar
perturbations are derived. The system of equations is then investigated in the case of the longitudinal
gauge. Finally, gauge-invariant variables are found and the equations for the corresponding Mukhanov
variables are derived. In Sec. \ref{summary}, we summarize our results and draw out some conclusions.
\section{Scalar perturbations with holonomy corrections} \label{Sperthol}
The holonomy-modified Hamiltonian constraint can be written as:
\begin{equation}
H^Q_G[N] = \frac{1}{2 \kappa} \int_{\Sigma} d^3x \left[ \bar{N}(\mathcal{H}^{(0)}_G+\mathcal{H}^{(2)}_G)+\delta N \mathcal{H}^{(1)}_G\right],
\label{HolModHam}
\end{equation}
where
\begin{eqnarray}
\mathcal{H}^{(0)}_G = -6 \sqrt{\bar{p}} (\mathbb{K}[1])^2, \nonumber \\
\mathcal{H}^{(1)}_G = -4\sqrt{\bar{p}} \left( \mathbb{K}[s_1] +\alpha_1\right)
\delta^c_j \delta K^j_c-\frac{1}{\sqrt{\bar{p}}}
\left( \mathbb{K}[1]^2+\alpha_2 \right) \delta^j_c\delta E^c_j \nonumber \\
+\frac{2}{\sqrt{\bar{p}}} (1+\alpha_3) \partial_c \partial^j \delta E^c_j, \nonumber \\
\mathcal{H}^{(2)}_G = \sqrt{\bar{p}}(1+\alpha_4) \delta K^j_c \delta K^k_d \delta^c_k \delta^d_j
-\sqrt{\bar{p}} (1+\alpha_5)(\delta K^j_c \delta^c_j)^2 \nonumber \\
-\frac{2}{\sqrt{\bar{p}}} \left( \mathbb{K}[s_2]+\alpha_6 \right) \delta E^c_j \delta K^j_c
-\frac{1}{2\bar{p}^{3/2}} \left( \mathbb{K}[1]^2+\alpha_7\right)\delta E^c_j \delta E^d_k \delta^k_c \delta^j_d \nonumber \\
+\frac{1}{4\bar{p}^{3/2}}\left( \mathbb{K}[1]^2+\alpha_8\right) (\delta E^c_j \delta^j_c)^2
- \frac{1}{2\bar{p}^{3/2}}(1+\alpha_9) \delta^{jk} (\partial_c \delta E^c_j)(\partial_d \delta E^d_k). \nonumber
\end{eqnarray}
The standard holonomy corrections are parametrized by two integers $s_1$ and $s_2$.
The $\alpha_i$ are counter-terms, which are introduced to remove anomalies.
Those factors are defined so that they vanish in the classical limit $(\bar{\mu} \rightarrow 0)$.
The counter-terms could be, in general, functions of all the canonical variables.
We however assume here that they are functions of the gravitational background
variables only.
In our approach, the diffeomorphism constraint holds the classical form
\begin{equation}
D_G[N^a] = \frac{1}{\kappa} \int_{\Sigma} d^3x \delta N^c \left[ \bar{p} \partial_c(\delta^d_k \delta K^k_d )
-\bar{p}(\partial_k \delta K^k_c)
-\bar{k} \delta^k_c (\partial_d \delta E^d_k )\right].
\label{diffconstr}
\end{equation}
In general, the diffeomorphism constraint could also be holonomy corrected.
this possibility was studied, {\it e.g.}, in \cite{Wu:2010wj}. However, in LQG the
diffeomorphism constraint is satisfied at the classical level. Therefore, if LQC is to be considered as
a specific model of LQG, the diffeomorphism constraint should naturally hold its classical form. Because of this,
in this paper, the diffeomorphism constraint is not modified by the holonomies.
It is worth stressing, that the classicality of the diffeomorphism constraint
is also imposed by the requirement of anomaly cancelation. Namely, if one replaces $ \bar{k} \rightarrow \mathbb{K}[n]$
in (\ref{diffconstr}), the condition $n=0$ would anyway be required by the introduction of scalar matter.
In fact, the same condition was obtained for vector modes with holonomy
corrections \cite{VectorJTAJ}.
Let us now calculate the possible Poisson brackets for the constraints $H^Q_G[N]$ and $D_G[N^a]$.
\subsection{The $\left\{H^Q_G, D_G \right\}$ bracket}
Using the definition of the Poisson bracket (\ref{Poisson}), we derive:
\begin{eqnarray}
\left\{H_{G}^Q[N],D_{G}[N^a] \right\} = - H_{G}^Q\left[ \delta N^a \partial_a \delta N \right] +\mathcal{B}\ D_{G}[N^a] \nonumber \\
+ \frac{\sqrt{\bar{p}}}{\kappa}\int_{\Sigma} d^3x \delta N^a (\partial_a \delta N) \mathcal{A}_1
+ \frac{\bar{N}\sqrt{\bar{p}}\bar{k} }{\kappa}\int_{\Sigma} d^3x \delta N^a (\partial_i \delta K^i_a ) \mathcal{A}_2 \nonumber \\
+\frac{\bar{N}}{\kappa \sqrt{\bar{p}}}\int_{\Sigma} d^3x \delta N^i (\partial_a \delta E^a_i) \mathcal{A}_3
+\frac{\bar{N}}{2 \kappa \sqrt{\bar{p}}}\int_{\Sigma} d^3x (\partial_a \delta N^a) (\delta E^b_i \delta^i_b) \mathcal{A}_4,
\label{PoissHGDG}
\end{eqnarray}
where
\begin{equation}
\mathcal{B} = \frac{\bar{N}}{\sqrt{\bar{p}}} \left[ -2\mathbb{K}[2]+\bar{k}(1+\alpha_5)+\mathbb{K}[s_2]+\alpha_6 \right],
\label{Bfunction}
\end{equation}
and
\begin{eqnarray}
\mathcal{A}_1 = 2\bar{k}(\mathbb{K}[s_1]+\alpha_1)+\alpha_2-2\mathbb{K}[1]^2, \\
\mathcal{A}_2 = \alpha_5-\alpha_4, \\
\mathcal{A}_3 = -\mathbb{K}[1]^2- \bar{p} \frac{\partial}{\partial \bar{p}}\mathbb{K}[1]^2-\frac{1}{2}\alpha_7 \nonumber \\
+\bar{k}(-2\mathbb{K}[2]+\bar{k}(1+\alpha_5)+2\mathbb{K}[s_2]+2\alpha_6), \\
\mathcal{A}_4 = \alpha_8-\alpha_7.
\end{eqnarray}
The functions $\mathcal{A}_1,\dots, \mathcal{A}_4$ are the first anomalies coming from the effective
nature of the Hamiltonian constraint. Later, we will set them to zero so as to fulfill the requirement of anomaly
freedom. This will lead to constraints on the form of the counter-terms.
Beside the anomalies, the $\left\{H^Q_G, D_G \right\}$ bracket contains the $- H_{G}^Q\left[ \delta N^a \partial_a \delta N \right]$
term, which is expected classically. There is also an additional contribution from the diffeomorphism constraint $\mathcal{B}\ D_{G}[N^a]$.
This term is absent in the classical theory. This is however consistent as, for $\bar{\mu} \rightarrow 0$, the $\mathcal{B}$ function tends to zero.
\subsection{The $\left\{H^Q_G, H^Q_G \right\}$ bracket}
The next bracket is:
\begin{eqnarray}
\left\{H_{G}^Q[N_1],H_{G}^Q[N_2] \right\} =
(1+\alpha_3)(1+\alpha_5)D_G \left[ \frac{\bar{N}}{\bar{p}} \partial^a(\delta N_2 -\delta N_1) \right] \nonumber \\
+\frac{\bar{N}}{\kappa} \int_{\Sigma} d^3x \partial^a(\delta N_2 -\delta N_1)(\partial_i \delta K^i_a)(1+\alpha_3)\mathcal{A}_5 \nonumber \\
+\frac{\bar{N}}{\kappa \bar{p}} \int_{\Sigma} d^3x(\delta N_2 -\delta N_1)(\partial^i\partial_a \delta E^a_i) \mathcal{A}_6 \nonumber \\
+\frac{\bar{N}}{\kappa} \int_{\Sigma} d^3x(\delta N_2 -\delta N_1)(\delta^a_i \delta K^i_a) \mathcal{A}_7 \nonumber \\
+\frac{\bar{N}}{\kappa \bar{p}} \int_{\Sigma} d^3x (\delta N_2 -\delta N_1)(\delta^i_a \delta E^a_i) \mathcal{A}_8,
\end{eqnarray}
where
\begin{eqnarray}
\mathcal{A}_5 = \alpha_5 - \alpha_4, \\
\mathcal{A}_6 = (1+\alpha_9)(\mathbb{K}[s_1]+\alpha_1)-(1+\alpha_3)(\mathbb{K}[s_2]+\alpha_6)+\mathbb{K}[2](1+\alpha_3) \nonumber \\
-2\mathbb{K}[2]\bar{p}\frac{\partial \alpha_3}{\partial \bar{p}}
+ \frac{1}{2}\left( \mathbb{K}[1]^2+2\bar{p}\frac{\partial}{\partial \bar{p}}\mathbb{K}[1]^2 \right)\frac{\partial \alpha_3}{\partial \bar{k}}
-\bar{k}(1+\alpha_3)(1+\alpha_5), \\
\mathcal{A}_7 = 4\mathbb{K}[2] \bar{p} \frac{\partial }{\partial \bar{p}} (\mathbb{K}[s_1]+\alpha_1)
-\left( \mathbb{K}[1]^2+2\bar{p}\frac{\partial}{\partial \bar{p}}\mathbb{K}[1]^2 \right) \frac{\partial }{\partial \bar{k}} (\mathbb{K}[s_1]+\alpha_1)
\nonumber \\
+\left(1+ \frac{3}{2}\alpha_5-\frac{1}{2}\alpha_4\right)(\mathbb{K}[1]^2+\alpha_2)
-2(\mathbb{K}[s_2]+\alpha_6)(\mathbb{K}[s_1]+\alpha_1) \nonumber \\
+2\mathbb{K}[2](\mathbb{K}[s_1]+\alpha_1), \label{A7} \\
\mathcal{A}_8 = \frac{1}{2}(\mathbb{K}[s_2]+\alpha_6)(\mathbb{K}[1]^2+\alpha_2)
-(\mathbb{K}[s_1]+\alpha_1)(\mathbb{K}[1]^2+\alpha_7) \nonumber \\
+\frac{3}{2}(\mathbb{K}[s_1]+\alpha_1)(\mathbb{K}[1]^2+\alpha_8)
-\frac{1}{2} \mathbb{K}[2](\mathbb{K}[1]^2+\alpha_2) \nonumber \\
+\mathbb{K}[2]\bar{p} \frac{\partial}{\partial \bar{p}}(\mathbb{K}[1]^2+\alpha_2)
-\frac{1}{4}\left( \mathbb{K}[1]^2+2\bar{p} \frac{\partial}{\partial \bar{p}} \mathbb{K}[1]^2\right)
\frac{\partial }{\partial \bar{k}}(\mathbb{K}[1]^2+\alpha_2). \label{A8}
\end{eqnarray}
The $\mathcal{A}_5,\dots, \mathcal{A}_8$ are the next four anomalies. Moreover, the diffeomorphism
constraint is multiplied by the factor $(1+\alpha_3)(1+\alpha_5)$.
\subsection{The $\left\{D_G, D_G \right\}$ bracket}
The Poisson bracket between the diffeomorphism constraints is:
\begin{eqnarray}
\left\{D_{G}[N^a_1],D_{G}[N^a_2] \right\} = 0.
\end{eqnarray}
\section{Scalar matter} \label{matter}
In this section, we introduce scalar matter.
The scalar matter diffeomorphism constraint is
\begin{equation}
D_M[N^a] = \int_{\Sigma} \delta N^a \bar{\pi} (\partial_a \delta \varphi).
\end{equation}
The scalar matter Hamiltonian can be expressed as:
\begin{equation}
H^Q_M[N]=H_M[\bar{N}] +H_M[\delta N],\nonumber
\end{equation}
where
\begin{eqnarray}
H_M[\bar{N}] &=& \int_{\Sigma} d^3 x \bar{N} \left[ \left(\mathcal{H}^{(0)}_{\pi}+\mathcal{H}^{(0)}_{\varphi}\right)
+\left(\mathcal{H}^{(2)}_{\pi}+\mathcal{H}^{(2)}_{\nabla}+\mathcal{H}^{(2)}_{\varphi} \right) \right], \label{HMb} \\
H_M[\delta N] &=& \int_{\Sigma} d^3 \delta N \left[ \mathcal{H}^{(1)}_{\pi}+\mathcal{H}^{(1)}_{\varphi} \right]. \label{HMd} \
\end{eqnarray}
The factors in equations (\ref{HMb}) and (\ref{HMd}) are
\begin{eqnarray}
\mathcal{H}^{(0)}_{\pi} &=& \frac{\bar{\pi}^2}{2\bar{p}^{3/2}}, \nonumber \\
\mathcal{H}^{(0)}_{\varphi} &=& \bar{p}^{3/2} V(\bar{\varphi}), \nonumber \\
\mathcal{H}^{(1)}_{\pi} &=&
\frac{\bar{\pi} \delta \pi}{\bar{p}^{3/2}}-\frac{\bar{\pi}^2}{2\bar{p}^{3/2}} \frac{\delta^j_c \delta E^c_j}{2\bar{p}}, \nonumber \\
\mathcal{H}^{(1)}_{\varphi} &=&
\bar{p}^{3/2} \left[ V_{,\varphi}(\bar{\varphi}) \delta \varphi +V(\bar{\varphi}) \frac{\delta^j_c \delta E^c_j}{2\bar{p}} \right], \nonumber \\
\mathcal{H}^{(2)}_{\pi} &=& \frac{1}{2} \frac{\delta \pi^2}{\bar{p}^{3/2}}-\frac{\bar{\pi} \delta \pi}{\bar{p}^{3/2}} \frac{\delta^j_c \delta E^c_j}{2\bar{p}}
+\frac{1}{2} \frac{\bar{\pi}^2}{\bar{p}^{3/2}} \left[ \frac{(\delta^j_c \delta E^c_j )^2}{8\bar{p}^2}
+\frac{\delta^k_c \delta^j_d \delta E^c_j \delta E^d_k}{4\bar{p}^2} \right], \\
\mathcal{H}^{(2)}_{\nabla} &=& \frac{1}{2} \sqrt{\bar{p}}(1+\alpha_{10})
\delta^{ab} \partial_a \delta \varphi \partial_b \delta \varphi,\nonumber \\
\mathcal{H}^{(2)}_{\varphi} &=& \frac{1}{2} \bar{p}^{3/2} V_{,\varphi\varphi}(\bar{\varphi}) \delta \varphi^2
+\bar{p}^{3/2} V_{,\varphi}(\bar{\varphi}) \delta \varphi \frac{\delta^j_c \delta E^c_j}{2\bar{p}} \\
&+&\bar{p}^{3/2} V(\bar{\varphi}) \left[ \frac{(\delta^j_c \delta E^c_j )^2}{8\bar{p}^2}
-\frac{\delta^k_c \delta^j_d \delta E^c_j \delta E^d_k}{4\bar{p}^2} \right].
\end{eqnarray}
Here, we have introduced the counter-term $\alpha_{10}$ in the factor $\mathcal{H}^{(2)}_{\nabla}$.
Thanks to this, the Poisson bracket between two matter Hamiltonians takes the following form:
\begin{equation}
\left\{H_M^Q[N_1],H_M^Q[N_2]\right\} = (1+\alpha_{10}) D_M \left[ \frac{\bar{N}}{\bar{p}}\partial^a(\delta N_2-\delta N_1) \right].
\end{equation}
As will be explained later, the appearance of the front-factor $(1+\alpha_{10}) $ will allow us to close
the algebra of total constraints. In principle, other prefactors could have been expected, however they do not help removing anomalies.
\subsection{Total constraints}
The total Hamiltonian and diffeomorphism constraints are the following:
\begin{eqnarray}
H_{tot}[N] &=& H^Q_G[N]+H^Q_M[N], \\
D_{tot}[N^a] &=& D_G[N^a]+D_M[N^a].
\end{eqnarray}
The Poisson bracket between two total diffeomorphism constraints is vanishing:
\begin{equation}
\left\{D_{tot}[N^a_1],D_{tot}[N^a_2]\right\} = 0.
\end{equation}
The bracket between the total Hamiltonian and diffeomorphism constraints can
be decomposed as follows:
\begin{eqnarray}
\left\{ H_{tot}[N],D_{tot}[N^a] \right\} &=& \left\{ H_{M}^Q[N],D_{tot}[N^a] \right\} +\left\{ H_{G}^Q[N],D_{G}[N^a] \right\} \nonumber \\
&+&\left\{ H_{G}^Q[N],D_{M}[N^a] \right\}. \label{decompHtotDtot}
\end{eqnarray}
The first bracket in the sum (\ref{decompHtotDtot}) is given by
\begin{equation}
\left\{ H^Q_{M}[N],D_{tot}[N^a] \right\} = - H^Q_M[\delta N^a \partial_a \delta N].
\end{equation}
The second contribution to Eq. (\ref{decompHtotDtot}) is given by (\ref{PoissHGDG}),
while the last contributions is vanishing:
\begin{equation}
\left\{ H_{G}^Q[N],D_{M}[N^a] \right\} = 0.
\end{equation}
The Poisson bracket between the two total Hamiltonian constraints can be decomposed in the following way:
\begin{eqnarray}
\left\{ H_{tot}[N_1],H_{tot}[N_2] \right\} &=& \left\{ H_G^Q[N_1],H_G^Q[N_2] \right\} +\left\{ H_M[N_1],H_M[N_2] \right\} \nonumber \\
&+& \left[ \left\{H_G^Q[N_1],H_M[N_2]\right\} - (N_1 \leftrightarrow N_2) \right].
\end{eqnarray}
The contribution from the last brackets can be expressed as
\begin{eqnarray}
\left\{H_G^Q[N_1],H_M[N_2]\right\} - (N_1 \leftrightarrow N_2) = \nonumber \\
=\frac{1}{2} \int_{\Sigma}d^3x \bar{N} (\delta N_2-\delta N_1)\left( \frac{\bar{\pi}^2}{2\bar{p}^3}-V(\bar{\varphi}) \right)
(\partial_c\partial^j \delta E^c_j) \mathcal{A}_9 \nonumber \\
+ 3 \int_{\Sigma}d^3x \bar{N} (\delta N_2-\delta N_1)\left( \frac{\bar{\pi}\delta \pi}{\bar{p}^2}-\bar{p}V_{\varphi}(\bar{\varphi}) \delta \varphi \right)
\mathcal{A}_{10} \nonumber \\
+\int_{\Sigma}d^3x \bar{N} (\delta N_2-\delta N_1)(\delta^c_j \delta K^c_j) \left( \frac{\bar{\pi}^2}{2\bar{p}^3}-V(\bar{\varphi}) \right) \bar{p}
\mathcal{A}_{11} \nonumber \\
+\frac{1}{2} \int_{\Sigma}d^3x \bar{N} (\delta N_2-\delta N_1)(\delta^j_c \delta E^c_j) \left(\frac{\bar{\pi}^2}{2\bar{p}^3}\right)
\mathcal{A}_{12} \nonumber \\
+\frac{1}{2} \int_{\Sigma}d^3x \bar{N} (\delta N_2-\delta N_1)(\delta^j_c \delta E^c_j) V(\bar{\varphi})
\mathcal{A}_{13},
\end{eqnarray}
where
\begin{eqnarray}
\mathcal{A}_9 = \frac{\partial \alpha_3}{\partial \bar{k}}, \\
\mathcal{A}_{10} = \mathbb{K}[2]-\mathbb{K}[s_1]-\alpha_1, \\
\mathcal{A}_{11} = -\frac{\partial}{\partial \bar{k}}(\mathbb{K}[s_1]+\alpha_1)+\frac{3}{2}(1+\alpha_5)-\frac{1}{2}(1+\alpha_4), \\
\mathcal{A}_{12} = -\frac{1}{2}\frac{\partial}{\partial \bar{k}}(\mathbb{K}[1]^2+\alpha_2)+5(\mathbb{K}[s_1]+\alpha_1)
- 5\mathbb{K}[2]+\mathbb{K}[s_2] +\alpha_6, \label{A12} \\
\mathcal{A}_{13} = \frac{1}{2}\frac{\partial}{\partial \bar{k}}(\mathbb{K}[1]^2+\alpha_2)+\mathbb{K}[s_1]+\alpha_1
-\mathbb{K}[2]-\mathbb{K}[s_2]-\alpha_6.
\end{eqnarray}
The functions $\mathcal{A}_9,\dots, \mathcal{A}_{13}$ are the last five anomalies.
\section{Anomaly freedom} \label{anomaly}
The requirement of anomaly freedom is equivalent to the
conditions $\mathcal{A}_i=0$ for $i=1,\dots, 13$.
Let us start form the condition $\mathcal{A}_9=0$. Since $\alpha_3$ cannot be
a constant, this condition implies $\alpha_3=0$. The condition
$\mathcal{A}_{10}=0$ gives $\alpha_1 = \mathbb{K}[2]-\mathbb{K}[s_1]$.
Using this, the condition $\mathcal{A}_{1}=0$, can be written
as $\alpha_2 = 2\mathbb{K}[1]^2 -2\bar{k}\mathbb{K}[2]$. The conditions
$\mathcal{A}_{2}=0$ and $\mathcal{A}_{5}=0$ are equivalent and
lead to $\alpha_4=\alpha_5$. Based on this, the requirement $\mathcal{A}_{11}=0$,
leads to:
\begin{equation}
1+\alpha_4 = \frac{\partial \mathbb{K}[2]}{\bar{k}} = \cos(2\bar{\mu} \gamma\bar{k}) =: \Omega.
\end{equation}
For the sake of simplicity we have defined here the $\Omega$-function.
With use of this, the condition $\mathcal{A}_{6}=0$ leads to
\begin{equation}
\alpha_6 = \mathbb{K}[2](2+\alpha_9)-\mathbb{K}[s_2]-\bar{k}\Omega. \label{alpha6}
\end{equation}
So, equation (\ref{A12}) simplifies to
\begin{equation}
\mathcal{A}_{12} =\alpha_9\mathbb{K}[2].
\end{equation}
Therefore, requiring $\mathcal{A}_{12}=0$ is equivalent to the
condition $\alpha_9=0$. Furthermore, $\mathcal{A}_{4}=0$ gives $\alpha_7=\alpha_8$.
The expression for $\alpha_7$ can be derived from the condition $\mathcal{A}_{3}=0$.
Namely, using Eq. (\ref{alpha6}) one obtains:
\begin{equation}
\alpha_7 = 2(2\beta-1)\mathbb{K}[1]^2+4(1-\beta) \bar{k}\mathbb{K}[2]-2\bar{k}^2\Omega.
\end{equation}
The condition $\mathcal{A}_{13}=0$ is fulfilled by using the expressions derived
for $\alpha_1$, $\alpha_2$ and $\alpha_6$. The last two anomalies (\ref{A7}) and (\ref{A8})
can be simplified to:
\begin{eqnarray}
\mathcal{A}_7 &=& 2(1+2\beta)(\Omega\mathbb{K}[1]^2-\mathbb{K}[2]^2 ), \\
\mathcal{A}_8 &=& \bar{k} (1+2\beta)(\mathbb{K}[2]^2-\Omega\mathbb{K}[1]^2).
\end{eqnarray}
The anomaly freedom conditions for those last terms, $\mathcal{A}_{7}=0$ and $\mathcal{A}_8=0$,
are fulfilled if and only if $\beta = -1/2$.
It is also worth noticing that the function $\mathcal{B}$ given by Eq. (\ref{Bfunction})
is equal to zero when the expression obtained for $\alpha_6$ is used. There is
finally no contribution from the diffeomorphism constraint in the $\left\{H^Q_G, D_G \right\}$ bracket.
Using the anomaly freedom conditions given above, the bracket between the total Hamiltonian
constraints simplifies to
\begin{eqnarray}
\left\{ H_{tot}[N_1],H_{tot}[N_2] \right\} &=& \Omega D_{tot} \left[ \frac{\bar{N}}{\bar{p}} \partial^a(\delta N_2 - \delta N_1)\right] \nonumber \\
&+& (\alpha_{10}-\alpha_4) D_M \left[ \frac{\bar{N}}{\bar{p}}\partial^a(\delta N_2-\delta N_1) \right].
\end{eqnarray}
The closure of the algebra of total constraints implies the last condition $\alpha_{10}=\alpha_4=\Omega-1$.\\
To summarize, the counter-terms allowing the algebra to be anomaly-free are uniquely determined,
and are given by:
\begin{eqnarray}
\alpha_1 &=& \mathbb{K}[2]-\mathbb{K}[s_1], \label{alpha1} \\
\alpha_2 &=& 2\mathbb{K}[1]^2 -2\bar{k}\mathbb{K}[2], \\
\alpha_3 &=& 0, \\
\alpha_4 &=& \Omega-1, \\
\alpha_5 &=& \Omega-1, \\
\alpha_6 &=& 2\mathbb{K}[2]-\mathbb{K}[s_2]-\bar{k}\Omega, \label{alpha6} \\
\alpha_7 &=& - 4 \mathbb{K}[1]^2+6 \bar{k}\mathbb{K}[2]-2\bar{k}^2\Omega, \\
\alpha_8 &=& - 4 \mathbb{K}[1]^2+6 \bar{k}\mathbb{K}[2]-2\bar{k}^2\Omega, \\
\alpha_9 &=& 0, \\
\alpha_{10} &=& \Omega-1.
\end{eqnarray}
It is straightforward to check that the counter-terms $\alpha_1,\dots, \alpha_{10}$ are vanishing
in the classical limit ($\bar{\mu}\rightarrow 0$), as expected.
Those counter-terms are defined
up to the two integers $s_1$ and $s_2$, which appear in (\ref{alpha1}) and (\ref{alpha6}). However,
in the Hamiltonian (\ref{HolModHam}), the factor $\alpha_1$ appears with $\mathbb{K}[s_1]$
and the factor $\alpha_6$ appears with $\mathbb{K}[s_2]$. Namely, we have $\mathbb{K}[s_1]+\alpha_1=\mathbb{K}[2]$
and $\mathbb{K}[s_2]+\alpha_6= 2\mathbb{K}[2]-\bar{k}\Omega$. Therefore, the final Hamiltonian
will not depend on the parameters $s_1$ and $s_2$. No ambiguity remains to be fixed.
Moreover, the anomaly cancellation requires
\begin{equation}
\beta = -\frac{1}{2},
\end{equation}
which fixes the functional form of the $\bar{\mu}$ factor. The fact that anomaly freedom requires
$\beta=-1/2$ is a quite surprising result. The exact value of $\beta$ is highly debated
in LQC. The only {\it a priori} obvious statement is that $\beta \in [-1/2,0]$. The choice $\beta=-1/2$ is called the
$\bar{\mu}-$scheme (new quantization scheme) and is preferred by some authors for physical reasons
\cite{corisingh-beta}. Our result seems to show that the $\bar{\mu}-$scheme is embedded in the structure
of the theory and this gives a new motivation for this particular choice of quantization scheme.
The quantity $\bar{\mu}^2 \bar{p}$ can be interpreted as the physical area of an elementary
loop along which the holonomy is calculated. Because, in the $\bar{\mu}-$scheme,
$\bar{\mu}^2 \propto \bar{p}^{-1}$, the physical area of the loop remains constant.
This elementary area is usually set to be the area gap $\Delta$ derived in LQG.
Therefore, in the $\bar{\mu}-$scheme,
\begin{equation}
\bar{\mu} = \sqrt{\frac{\Delta}{\bar{p}}}.
\end{equation}
\subsection{Algebra of constraints}
Taking into account the previous conditions of anomaly-freedom, the non-vanishing Poisson brackets
for the gravity sector are:
\begin{eqnarray}
\left\{ H_G^Q[N],D_G[N^a]\right\} &=& - H_G^Q[\delta N^a \partial_a \delta N], \\
\left\{ H_G^Q[N_1],H_G^Q[N_2]\right\} &=& \Omega D_G \left[ \frac{\bar{N}}{\bar{p}} \partial^a(\delta N_2 - \delta N_1) \right].
\end{eqnarray}
This clearly shows that the \emph{gravity sector is anomaly free}. The remaining non-vanishing brackets are:
\begin{eqnarray}
\left\{ H_M[N],D_{tot}[N^a] \right\} &=& - H_M [\delta N^a \partial_a \delta N], \\
\left\{ H_M[N_1],H_M[N_2]\right\} &=& \Omega D_M\left[ \frac{\bar{N}}{\bar{p}} \partial^c(\delta N_2 - \delta N_1)\right].
\end{eqnarray}
The algebra of total constraints therefore takes the following form:
\begin{eqnarray}
\left\{ D_{tot}[N^a_1],D_{tot} [N^a_2] \right\} &=& 0, \\
\left\{ H_{tot}[N],D_{tot}[N^a] \right\} &=& - H_{tot}[\delta N^a \partial_a \delta N], \\
\left\{ H_{tot}[N_1],H_{tot}[N_2] \right\} &=& D_{tot} \left[ \Omega \frac{\bar{N}}{\bar{p}} \partial^a(\delta N_2 - \delta N_1)\right].
\label{HtotHtot}
\end{eqnarray}
Although the algebra is closed, there are however modifications with respect to the classical case,
due to presence of the factor $\Omega$ in Eq. (\ref{HtotHtot}). Therefore, not only
the dynamics, as a result of the modification of the Hamiltonian constraint, is modified but also the very structure of the space-time itself is \emph{deformed}. This is embedded
in the form of the algebra of constraints. The hypersurface deformation algebra generated
by (\ref{HtotHtot}) is pictorially represented in Fig. \ref{FIG1}.
\begin{figure}[htb]
\includegraphics[scale=0.7]{hyper.eps}
\label{FIG1}
\caption{Pictorial representation of the hypersurface deformation algebra (\ref{HtotHtot}).}
\end{figure}
As $\Omega \in [-1,1]$, the shift vector
\begin{equation}
N^a= \Omega \frac{\bar{N}}{\bar{p}} \partial^a(\delta N_2 - \delta N_1)
\label{ShiftvectOmega}
\end{equation}
appearing in (\ref{HtotHtot}) can change sign in time.
In order to see when this might
happen let us express the parameter $\Omega$ as:
\begin{equation}
\Omega = \cos(2\bar{\mu} \gamma\bar{k}) = 1 - 2\frac{\rho}{\rho_c},
\end{equation}
where $\rho$ is the energy density of the matter field and
\begin{equation}
\rho_c = \frac{3}{\kappa \gamma \bar{\mu}^2 \bar{p}} = \frac{3}{\kappa \gamma \Delta}.
\end{equation}
In the low energy limit, $\rho \rightarrow 0$, the classical case $(\Omega \rightarrow 1)$ is
correctly recovered. However, while approaching the high energy domain the situation
drastically changes. Namely, for $\rho = \rho_c/2$, the shift vector (\ref{ShiftvectOmega})
becomes null. At this point, the maximum value of the Hubble parameter is also reached.
The maximum allowed energy density is $\rho = \rho_c$ and corresponds to the
bounce. Then the shift vector (\ref{ShiftvectOmega}) fully reverses with respect to
the low energy limit. One can interpret this peculiar behavior as a geometry
change. Namely, when the universe is in its quantum stage ($\rho > \rho_c/2$), the effective
algebra of constraints shows that the space is Euclidian. At the particular value
$\rho = \frac{\rho_c}{2}$, the geometry switches to the Minkowski one \cite{private_discussion}.
This will become even clearer when analyzing the Mukhanov equation in Sec. \ref{equations}.
The consequences of this have not yet been fully understood, but it is interesting to notice that this model
naturally exhibits properties related to the Hartle-Hawking no-boundary proposal \cite{Hartle:1983ai}.
\section{Equations of motion} \label{equations}
Once the anomaly-free theory of scalar perturbations with holonomy corrections is constructed,
the equations of motion for the canonical variables can be derived. This can be achieved through
the Hamilton equation
\begin{equation}
\dot{f} = \{f ,H[N,N^a]\} \label{HamEq},
\end{equation}
where the Hamiltonian $H[N,N^a]$ is the sum of all constraints
\begin{equation}
H[N,N^a] = H^Q_{G}[N]+H_{M}[N]+D_{G} [N^a]+D_{M} [N^a].
\end{equation}
\subsection{Background equations}
Based on the Hamilton equation (\ref{HamEq}), the equations for the
canonical background variables are the following:
\begin{eqnarray}
\dot{\bar{k}} &=& -\frac{\bar{N}}{2\sqrt{\bar{p}}}\mathbb{K}[1]^2-\bar{N}\sqrt{\bar{p}} \frac{\partial}{\partial \bar{p}} \mathbb{K}[1]^2
+\frac{\kappa}{2} \sqrt{\bar{p}} \bar{N}\left[ -\frac{\bar{\pi}^2}{2\bar{p}^3}+V(\bar{\varphi}) \right], \label{dotkbar} \\
\dot{\bar{p}} &=& 2 \bar{N} \sqrt{\bar{p}}\mathbb{K}[2], \label{dotpbar} \\
\dot{\bar{\varphi}} &=& \bar{N} \frac{\bar{\pi}}{\bar{p}^{3/2}}, \label{dotbarvarphi} \\
\dot{\bar{\pi}} &=& -\bar{N} \bar{p}^{3/2} V_{, \varphi}(\bar{\varphi}). \label{dotbarpi}
\end{eqnarray}
In the following, we choose the time to be conformal by setting $\bar{N}=\sqrt{\bar{p}}$.
The $``\cdot"$ then means differentiation with respect to conformal time $\eta$.
Eqs. (\ref{dotbarvarphi}) and (\ref{dotbarpi}) can be now combined into the Klein-Gordon equation
\begin{equation}
\ddot{\bar{\varphi}}+2 \mathbb{K}[2] \dot{\bar{\varphi}}+\bar{p} V_{,\varphi} (\bar{\varphi}) = 0.
\end{equation}
Eq. (\ref{dotpbar}), together with the background part of the Hamiltonian constraint
\begin{equation}
\frac{1}{V_0} \frac{\partial H}{\partial \bar{N}} = \frac{1}{2\kappa}\left[-6 \sqrt{\bar{p}} (\mathbb{K}[1])^2\right]
+ \bar{p}^{3/2}\left[\frac{\bar{\pi}^2}{2\bar{p}^{3}}+V(\bar{\varphi})\right]= 0, \label{HamConstBack}
\end{equation}
lead to the modified Friedmann equation
\begin{equation}
\mathcal{H}^2 =\bar{p} \frac{\kappa}{3} \rho \left(1-\frac{\rho}{\rho_c}\right).
\end{equation}
Another useful expression is:
\begin{equation}
3 \mathbb{K}[1]^2 = \frac{\bar{\pi}^2}{2\bar{p}^2} + \bar{p} V(\bar{\varphi}).
\end{equation}
Here $\mathcal{H}$ stands for the conformal Hubble factor
\begin{equation}
\mathcal{H} := \frac{\dot{\bar{p}}}{2\bar{p}} = \mathbb{K}[2].
\end{equation}
The energy density and pressure of the scalar field are given by:
\begin{eqnarray}
\rho &=& \frac{\bar{\pi}^2}{2\bar{p}^{3}}+V(\varphi), \\
P &=& \frac{\bar{\pi}^2}{2\bar{p}^{3}}-V(\varphi).
\end{eqnarray}
For the purpose of further considerations, we also derive the relation
\begin{equation}
\kappa \left( \frac{\bar{\pi}^2}{2\bar{p}^2}\right) = \bar{k} \mathbb{K}[2]-\dot{\bar{k}},
\end{equation}
which comes from Eq. (\ref{dotkbar}) combined with (\ref{HamConstBack}).
\subsection{Equations for the perturbed variables}
The equations for the perturbed parts of the canonical variables are:
\begin{eqnarray}
\delta \dot{E}^a_i&=& -\bar{N}\left[
\sqrt{\bar{p}} \Omega \delta K^j_c \delta^c_i \delta^a_j- \sqrt{\bar{p}} \Omega (\delta K^j_c \delta^c_j) \delta^a_i
- \frac{1}{\sqrt{\bar{p}}} (2\mathbb{K}[2]-\bar{k}\Omega) \delta E^a_i \right] +\nonumber \\
&+&\delta N\left(2 \mathbb{K}[2] \sqrt{\bar{p}} \delta^a_i \right)
-\bar{p} (\partial_i \delta N^a -(\partial_c \delta N^c) \delta^a_i), \label{dotdE} \\
\delta \dot{K}^i_a &=& \bar{N} \left[- \frac{1}{\sqrt{\bar{p}}} (2\mathbb{K}[2]-\bar{k}\Omega) \delta K^i_a \right. \nonumber \\
&-&\frac{1}{2 \bar{p}^{3/2}}(- 3 \mathbb{K}[1]^2+6 \bar{k}\mathbb{K}[2]-2\bar{k}^2\Omega)\delta E^c_j \delta ^j_a \delta^i_c \nonumber \\
&+& \left. \frac{1}{4 \bar{p}^\frac{3}{2}}(- 3 \mathbb{K}[1]^2+6 \bar{k}\mathbb{K}[2]-2\bar{k}^2\Omega)(\delta E^c_j \delta^j_c) \delta^i_a
+\frac{\delta^{ik}}{2 \bar{p}^\frac{3}{2}} \partial_a \partial_d \delta E^d_k\right] \nonumber \\
&+& \frac{1}{2} \left[ -\frac{1}{\sqrt{\bar{p}}} (3\mathbb{K}[1]^2-2\bar{k}\mathbb{K}[2]) \delta^i_a \delta N
+\frac{2}{\sqrt{\bar{p}}} (\partial_a \partial^i \delta N) \right] \nonumber \\
&+& \delta^i_c (\partial_a \delta N^c)+ \kappa \delta N \frac{\sqrt{\bar{p}}}{2} \left[ - \frac{\bar{\pi}^2}{2\bar{p}^{3}}+V(\bar{\varphi}) \right] \delta^i_a \nonumber \\
&+& \kappa \bar{N} \left[ -\frac{\bar{\pi} \delta \pi}{2\bar{p}^{5/2}} \delta^i_a
+\frac{\sqrt{\bar{p}}}{2} \delta \varphi \frac{\partial V (\bar{\varphi}) }{\partial \bar{\varphi}} \delta^i_a
+ \left( \frac{\bar{\pi}^2}{2\bar{p}^{3/2}} + \bar{p}^{3/2} V(\bar{\varphi}) \right)\frac{\delta^j_c \delta E^c_j}{4\bar{p}^2}\delta^i_a \right. \nonumber \\
&+& \left. \left( \frac{\bar{\pi}^2}{2\bar{p}^{3/2}} - \bar{p}^{3/2} V(\bar{\varphi}) \right)\frac{\delta^i_c \delta^j_a \delta E^c_j}{2\bar{p}^2} \right],
\label{dotdK} \\
\delta \dot{\varphi} &=& \delta N \left( \frac{\bar{\pi}}{ \bar{p}^{3/2}}\right)
+\bar{N} \left(\frac{\delta \pi}{\bar{p}^{3/2}} -\frac{\bar{\pi}}{\bar{p}^{3/2}} \frac{\delta^j_c \delta E^c_j}{2\bar{p}} \right), \label{dotdeltavarphi} \\
\delta \dot{\pi} &=& -\delta N \left(\bar{p}^{3/2} V_{,\varphi}(\bar{\varphi}) \right)+\bar{\pi}(\partial_a \delta N^a) \nonumber \\
&-&\bar{N}\left[ - \sqrt{\bar{p}} \Omega \delta^{ab} \partial_a\partial_b\delta \varphi+\bar{p}^{3/2} V_{,\varphi\varphi}(\bar{\varphi}) \delta \varphi
+\bar{p}^{3/2} V_{,\varphi}(\bar{\varphi}) \frac{\delta^j_c \delta E^c_j}{2\bar{p}} \right]. \label{dotdeltapi}
\end{eqnarray}
\subsection{Longitudinal gauge}
As an example of application we will now derive the equations in the longitudinal gauge.
In this case, the $E$ and $B$ perturbations are set to zero.
The line element (\ref{lineelement}) therefore simplifies to
\begin{equation}
ds^2 = a^2\left[ -(1+2\phi)d\eta^2+(1-2\psi)\delta_{ab}dx^adx^b \right],
\end{equation}
where $\phi$ and $\psi$ are two remaining perturbation functions and $a$ is the
scale factor. From the metric above, one can
derive the laps function, the shift vector and the spatial metric:
\begin{eqnarray}
N &=& a\sqrt{1+2\phi}, \\
N^a &=& 0, \\
q_{ab} &=& a^2(1-2\psi)\delta_{ab}. \label{qab}
\end{eqnarray}
The lapse function can be expanded for the background
and perturbation part as $N=\bar{N} +\delta N$, where
\begin{eqnarray}
\bar{N} &=& \sqrt{\bar{p}} = a, \\
\delta N &=& \bar{N} \phi.
\end{eqnarray}
Using Eq. (\ref{qab}), the perturbation of the densitized triad is expressed as:
\begin{equation}
\delta E^a_i=-2\bar{p} \psi \delta^a_i. \label{dEpsi}
\end{equation}
The time derivative of this expression will also be useful and can be written as:
\begin{equation}
\delta \dot{E}^a_i = -2\bar{p} (2\mathbb{K}[2]\psi+\dot{\psi})\delta^a_i.
\end{equation}
Let us now find the expression for the perturbation of the extrinsic curvature $\delta K^i_a$
in terms of the metric perturbations $\phi$ and $\psi$. For this purpose, one can apply the
expression (\ref{dEpsi}) to the left hand side of (\ref{dotdE}). The resulting equation can be
solved for $\delta K^i_a$, leading to:
\begin{equation}
\delta K^i_a = - \delta^i_a \frac{1}{\Omega} \left( \dot{\psi}+\bar{k}\Omega \psi +\mathbb{K}[2]\phi \right).
\end{equation}
The time derivative of this variable is given by
\begin{eqnarray}
\delta \dot{K}^i_a = \delta^i_a \frac{1}{\Omega} \left[ -\ddot{\psi} -\dot{\bar{k}}\Omega \psi
+\dot{\psi}\left( \frac{\dot{\Omega}}{\Omega}-\bar{k}\Omega \right)
+\phi \mathbb{K}[2] \frac{\dot{\Omega}}{\Omega} \right. \nonumber \\
- \left. \phi \dot{\mathbb{K}}[2]-\mathbb{K}[2]\dot{\phi} \right]. \label{ddK}
\end{eqnarray}
Applying (\ref{ddK}) to the left hand side of (\ref{dotdK}), the equation containing the diagonal
part as well as the off-diagonal contribution is easily obtained. The off-diagonal part leads to
\begin{equation}
\partial_a\partial^i(\phi-\psi) = 0.
\end{equation}
This translates into $\psi=\phi$. In what follows, we will therefore consider $\phi$
only. The diagonal part of the discussed equation can be expressed as:
\begin{eqnarray}
\ddot{\phi}&+& \dot{\phi}\left[ 3\mathbb{K}[2]- \frac{\dot{\Omega}}{\Omega}\right]
+ \phi \left[ \dot{\mathbb{K}}[2] +2\mathbb{K}[2]^2 -\mathbb{K}[2] \frac{\dot{\Omega}}{\Omega} \right] \nonumber \\
&=& 4\pi G \Omega \left[ \dot{\bar{\varphi}} \delta \dot{\varphi}-\bar{p} \delta \varphi V_{,\varphi} (\bar{\varphi})\right].
\label{eqfin1}
\end{eqnarray}
One case now use the diffeomorphism constraint
\begin{equation}
\kappa \frac{\delta H[N,N^a]}{\delta (\delta N^c)} = \bar{p} \partial_c(\delta^d_k \delta K^k_d )
-\bar{p}(\partial_k \delta K^k_c)
-\bar{k}\delta^k_c (\partial_d \delta E^d_k )
+\kappa \bar{\pi} (\partial_c \delta \varphi) = 0.
\end{equation}
With the expressions for $\delta K^i_a$ and $\delta E^a_i$, it can be derived that
\begin{equation}
\partial_c \left[\dot{\phi}+\phi \mathbb{K}[2] \right] = 4\pi G \Omega \dot{\bar{\varphi}} \partial_c \delta \varphi.
\label{eqfin2}
\end{equation}
The next equation comes from the perturbed part of the Hamiltonian constraint:
\begin{eqnarray}
\frac{\delta H[N,N^a]}{\delta (\delta N)} &=&
\frac{1}{2\kappa} \left[-4\sqrt{\bar{p}} \mathbb{K}[2] \delta^c_j \delta K^j_c
-\frac{1}{\sqrt{\bar{p}}} \left( 3\mathbb{K}[1]^2-2\bar{k}\mathbb{K}[2] \right) \delta^j_c\delta E^c_j \right. \nonumber \\
&+&\left. \frac{2}{\sqrt{\bar{p}}} \partial_c \partial^j \delta E^c_j \right]
+ \frac{\bar{\pi} \delta \pi}{\bar{p}^{3/2}}-\frac{\bar{\pi}^2}{2\bar{p}^{3/2}} \frac{\delta^j_c \delta E^c_j}{2\bar{p}} \nonumber \\
&+&\bar{p}^{3/2} \left[ V_{,\varphi}(\bar{\varphi}) \delta \varphi +V(\bar{\varphi}) \frac{\delta^j_c \delta E^c_j}{2\bar{p}} \right] = 0.
\end{eqnarray}
Using the expressions for $\delta K^i_a$ and $\delta E^a_i$, this can be rewritten as:
\begin{eqnarray}
\Omega \nabla^2 \phi -3\mathbb{K}[2] \dot{\phi} - \left[ \dot{\mathbb{K}}[2]+2\mathbb{K}[2]^2 \right] \phi
= 4\pi G \Omega \left[ \dot{\bar{\varphi}} \delta \dot{\varphi}+\bar{p} \delta \varphi V_{,\varphi} (\bar{\varphi})\right].
\label{eqfin3}
\end{eqnarray}
The last equality comes from (\ref{dotdeltavarphi}) and (\ref{dotdeltapi}):
\begin{equation}
\delta \ddot{\varphi}+2\mathbb{K}[2] \delta \dot{\varphi}-\Omega \nabla^2 \delta \varphi
+\bar{p}V_{,\varphi\varphi}(\bar{\varphi} ) \delta \varphi
+2 \bar{p} V_{,\varphi}(\bar{\varphi} ) \phi -4\dot{\bar{\varphi}} \dot{\phi} = 0.
\label{eqfin4}
\end{equation}
The equations (\ref{eqfin1}), (\ref{eqfin2}) and (\ref{eqfin3}) can be now combined into:
\begin{eqnarray}
\ddot{\phi}+2\left[\mathcal{H}- \left(\frac{\ddot{\bar{\varphi}}}{\dot{\bar{\varphi}}}+\epsilon \right)\right]\dot{\phi}
+2\left[ \dot{\mathcal{H}}-\mathcal{H} \left( \frac{\ddot{\bar{\varphi}}}{\dot{\bar{\varphi}}}+\epsilon\right) \right]\phi-c_s^2\nabla^2\phi=0,
\label{finaleq}
\end{eqnarray}
with the quantum correction
\begin{equation}
\epsilon = \frac{1}{2} \frac{\dot{\Omega}}{\Omega} = 3 \mathbb{K}[2] \left(\frac{\rho+P}{\rho_c-2\rho} \right),
\end{equation}
and the squared velocity
\begin{equation}
c_s^2 = \Omega.
\end{equation}
The squared velocity of the perturbation field $\phi$ is equal to $\Omega$. Because $-1 \leq \Omega \leq 1$,
the speed of perturbations is never super-luminal. However, for $\Omega < 0$ perturbations become unstable
($c_s^2 < 0$). This corresponds to the energy density regime $\rho > \frac{\rho_{c}}{2}$, where the phase of
super-inflation is expected.
At the point $\rho=\frac{\rho_{c}}{2}$, the velocity of the perturbation field $\phi$ is vanishing. Therefore,
perturbations don't propagate anymore when approaching $\rho=\frac{\rho_{c}}{2}$, where the Hubble factor
reaches its maximal value. Moreover, at this point, the quantum correction $\epsilon \rightarrow \infty$.
Because of this, Eq. (\ref{finaleq}) diverges and cannot be used to determine the propagation
of the perturbations. However, as shown in the next section, the equation for the gauge-invariant Mukhanov variable
does not exhibit such a pathology.
It is interesting to notice that the equations of motion derived in this subsection are the same as those
found in \cite{WilsonEwing:2011es}. This is quite surprising, because they were derived
following independent paths. In our approach, we have introduced the most general form for the holonomy corrections
to the Hamiltonian. Then, by adding counter-terms, anomalies in the algebra of constraints
were removed. On the other hand, the method proposed in \cite{WilsonEwing:2011es} is based on the diagonal form of the metric in the
longitudinal gauge. This enables one to introduce holonomy corrections in almost the same way as in the case
of a homogeneous model. It was then shown that a system defined in this way stays on-shell, that is, is free
of anomalies. The non-trivial equivalence of both approaches may suggest uniqueness in defining a
theory of scalar perturbations with holonomy corrections in an anomaly-free manner.
\subsection{Gauge invariant variables and Mukhanov equation} \label{GIVaMe}
Considering the scalar perturbations, there is only one physical degree of freedom.
As it was shown in \cite{Mukhanov:1990me}, this physical variable combines
both the perturbation of the metric and the perturbation of matter. The classical expression
on this gauge-invariant quantity is:
\begin{equation}
v = a(\eta) \left( \delta \varphi^{GI} + \frac{\dot{\bar{\varphi}}}{\mathcal{H}} \Psi \right),
\label{MukhClass}
\end{equation}
and its equation of motion is given by
\begin{equation}
\ddot{v}-\nabla^2 v - \frac{\ddot{z}}{z} v = 0,
\end{equation}
where
\begin{equation}
z = a(\eta) \frac{\dot{\bar{\varphi}}}{\mathcal{H}}.
\end{equation}
In the canonical formalism with scalar perturbations, the gauge transformation of a variable
$X$ under a small coordinate transformation
\begin{equation}
x^{\mu} \rightarrow x^{\mu} + \xi^{\mu} \ \ \ \ ; \ \ \ \ \xi^{\mu} = (\xi^0, \partial^a \xi),
\end{equation}
is given by (see \cite{Bojowald:2008jv} for details):
\begin{equation}
\delta_{[\xi^0, \xi]} X \dot{=} \{X, H^{(2)}[\bar{N}\xi^0] + D^{(2)}[\partial^a \xi] \},
\end{equation}
and it is straightforward to see that, classically,
\begin{equation}
\delta_{[\xi^0, \xi]} v = 0.
\end{equation}
This means that $v$ is diffeomorphism-invariant and can be taken as an observable.
Taking into account the holonomy corrections introduced in this paper,
the $\Omega$ function will modify the gauge transformations of the time
derivative of a variable $X$, so that
\begin{equation}
\delta_{[\xi^0, \xi]} \dot{X} - (\delta_{[\xi^0, \xi]} X) \dot{} = \Omega \cdot \delta_{[0, \xi^0]} X.
\end{equation}
Using this relation and gauge transformations of the metric perturbations
\begin{eqnarray}
\delta_{[\xi^0, \xi]}\psi &=& -\mathbb{K}[2] \xi^0, \\
\delta_{[\xi^0, \xi]}\phi &=& \dot{\xi}^0+\mathbb{K}[2]\xi^0, \\
\delta_{[\xi^0, \xi]}E &=& \xi, \\
\delta_{[\xi^0, \xi]}B &=& \dot{\xi},
\end{eqnarray}
one can define the gauge-invariant variables (Bardeen potentials) as:
\begin{eqnarray}
\Phi &=& \phi +\frac{1}{\Omega} (\dot{B}-\ddot{E}) +\left( \frac{\mathbb{K}[2]}{\Omega}-\frac{\dot{\Omega}}{\Omega}\right) (B-\dot{E}), \\
\Psi &=& \psi - \frac{\mathbb{K}[2]}{\Omega} (B-\dot{E}), \\
\delta \varphi^{GI} &=& \delta \varphi + \frac{\dot{\bar{\varphi}}}{\Omega} (B-\dot{E}).
\end{eqnarray}
The normalization of these variables was set such that, in the longitudinal gauge $(B=0=E)$,
we have $\Phi = \phi $, $\Psi = \psi $ and $\delta \varphi^{GI} = \delta \varphi $.
It is possible to define the analogous of the Mukhanov variable (\ref{MukhClass}):
\begin{equation}
v := \sqrt{\bar{p}} \left( \delta \varphi^{GI} + \frac{\dot{\bar{\varphi}}}{\mathbb{K}[2]} \Psi \right).
\label{MukhQuant}
\end{equation}
Writing the equations for $\Psi$ and $\delta \varphi^{GI}$, which are
\begin{eqnarray}
\ddot{\Psi}+2\left[\mathcal{H}- \left(\frac{\ddot{\bar{\varphi}}}{\dot{\bar{\varphi}}}+\epsilon \right)\right]\dot{\Psi}
+2\left[ \dot{\mathcal{H}}-\mathcal{H} \left( \frac{\ddot{\bar{\varphi}}}{\dot{\bar{\varphi}}}+\epsilon\right) \right]\Psi-c_s^2\nabla^2\Psi=0
\end{eqnarray}
and
\begin{equation}
\delta \ddot{\varphi}^{GI}+2\mathbb{K}[2] \delta \dot{\varphi}^{GI}-\Omega \nabla^2 \delta \varphi^{GI}
+\bar{p}V_{,\varphi\varphi}(\bar{\varphi} ) \delta \varphi^{GI}
+2 \bar{p} V_{,\varphi}(\bar{\varphi} ) \Psi -4\dot{\bar{\varphi}}^{GI} \dot{\Psi} = 0,
\end{equation}
one obtains equation for the variable (\ref{MukhQuant}):
\begin{eqnarray}
&& \ddot{v}-\Omega \nabla^2 v - \frac{\ddot{z}}{z} v = 0, \label{HCMukhequation}\\
&& z = \sqrt{\bar{p}} \frac{\dot{\bar{\varphi}}}{\mathbb{K}[2]},
\end{eqnarray}
which corresponds to the Mukhanov equation for our model. As we see, the difference between the
classical and the holonomy-corrected case is the factor $\Omega$ in front of the Laplacian.
This quantum contribution leads to a variation of the propagation velocity of the perturbation $v$. This is
similar to the case of the perturbation $\phi$ considered in the previous subsection.
The main difference is that there is no divergence for $\rho = \rho_c/2$ and the evolution of perturbations
can be investigated in the regime of high energy densities. It is once again worth noticing that for $\rho > \rho_c/2$,
$\Omega$ becomes negative and Eq. (\ref{HCMukhequation}) changes from an
hyperbolic form to an elliptic one. This basically means that the time part becomes indistinguishable
from the spatial one. This can be interpreted as a transition from a Minkowskian geometry to and Euclidean
geometry, as mentioned earlier.
Finally, it is also possible to define the perturbation of curvature $\mathcal{R}$ such that
\begin{equation}
\mathcal{R} = \frac{v}{z}.
\end{equation}
Based on this, one can now calculate the power spectrum of scalar perturbations. This
opens new possible ways to study quantum gravity effects in the very early universe.
Promising applications of the derived equations
will be investigated elsewhere.
\section{Summary and conclusions} \label{summary}
In this paper, we have investigated the theory of scalar perturbations with
holonomy corrections. Such corrections are expected because of quantum
gravity effects predicted by LQG. They basically come from the
regularization of the curvature of the connection at the Planck scale. Because
of this, the holonomy corrections become dominant in the high curvature regime.
The introduction of "generic type" holonomy corrections leads to an anomalous
algebra of constraints. The conditions of anomaly freedom impose some
restrictions on the form of the holonomy corrections. However, we have shown
that the holonomy corrections, in the standard form, cannot fully satisfy the
conditions of anomaly freedom. In order to solve this difficulty, additional counter-terms
were introduced. Such counter-terms tend to zero in the classical limit,
but play the role of regularizators of anomalies in the quantum (high curvature)
regime. The method of counter-terms was earlier successfully applied to
cosmological perturbations with inverse-triad corrections \cite{Bojowald:2008gz}.
We have shown that, thanks to the counter-terms, the theory of cosmological
perturbations with holonomy corrections can be formulated in an anomaly-free
way. The anomaly freedom was shown to be fulfilled not only for the gravity
sector but also when taking into account scalar matter. The requirements of
anomaly freedom were used to determine the form of the counter-terms. Furthermore,
conditions for obtaining and anomaly-free algebra of constraints were shown to be fulfilled
only for a particular choice of the $\bar{\mu}$ function, namely for the
$\bar{\mu}-$scheme (new quantization scheme). This quantization scheme was
shown earlier to be favored because of the consistency of the
background dynamics \cite{corisingh-beta}. Our result supports these earlier claims.
In our formulation, the diffeomorphism constraint holds its classical form,
in agreement with the LQG expectations. The obtained anomaly-free gravitational
Hamiltonian contains seven holonomy modifications. It was also necessary to
introduce one counter-term into the matter Hamiltonian in order to ensure the closure
of the algebra of total constraints. There is no ambiguity in defining the holonomy
corrections after imposing the anomaly-free conditions. The only remaining free
parameter of the theory is the area gap $\Delta$ used in defining the $\bar{\mu}$
function. This quantity can however be possibly fixed with the spectrum
of the area operator in LQG. Based on the equations derived in this paper it will
also be possible to put observational constraints on the value of $\Delta$ and,
hence, on the critical energy density $\rho_c$.
Based on the studied anomaly-free formulation, equations of motion were derived.
As an example of application, we studied the equations in the longitudinal gauge.
We have also found the gauge-invariant variables, which are holonomy-corrected
versions of the Bardeen potentials. Using this, we have derived the equation for the
Mukhanov variable. This equation can be directly used to compute the power
spectrum of scalar perturbations with quantum gravitational holonomy corrections.
Similar considerations were studied in the case of inverse-triad corrections
\cite{arXiv:1011.2779}. In that case, observational consequences have been derived
and compared with CMB data \cite{Bojowald:2011hd, arXiv:1107.1540}.
\ack
The authors would like to thank M.~Bojowald, G.~Calcagni and E.~Wilson-Ewing
for the discussions. TC and JM were supported from the Astrophysics Poland-France
(Astro-PF). JM has been supported by Polish Ministry of Science and Higher Education
grant N N203 386437 and by Foundation of Polish Science award START.
\section*{References}
|
\section{Introduction}
Mappings from a sphere into a Stiefel manifold are a natural object of study in topology.
Denote by $\widetilde{V}_k(\R^n)$ the non-compact Stiefel manifold, i.e. the set of all
$k$--frames in $\R^n$, and take a polynomial mapping $\alpha:S^{n-k}\rightarrow \widetilde{V}_k(\R^n)$.
If $n-k$ is even then the homotopy group $\pi_{n-k}\widetilde{V}_k(\R^n)$ is isomorphic
to $\Z$. Let $\Lambda(\alpha)\in\Z$ be the integer associated with $\alpha$.
In Sections 2 we show that $\Lambda(\alpha)$ is equal to the topological degree
of some associated mapping
$\widetilde{\alpha}:S^{k-1}\times S^{n-k} \longrightarrow \R^n\setminus\{0\}$.
In Section 3 we prove that one may express $\Lambda(\alpha)$ in terms of
signatures of two quadratic forms (Theorem \ref{efeektywnie}), even in the case where
$S^{n-k}$ is replaced by a compact algebraic hypersurface $M\subset \R^{n-k+1}$.
These signatures may be computed using computer algebra systems.
Examples presented in this paper were calculated with the help of {\sc Singular}
\cite{GPS06}.
Assume that $m$ is even, $M\subset\R^{m+1}$ is a compact algebraic $m$--dimensional hypersurface,
and $g:M\rightarrow\R^{2m}$ is an immersion. Whitney in \cite{whitneySelfInter} introduced the intersection
number $I(g)\in\Z$. In the case where $M=S^m$, Smale in \cite{Smale} constructed a mapping
$\alpha':S^m\rightarrow \widetilde{V}_m(\R^{2m})$ such that $I(g)=\Lambda(\alpha')$.
Unfortunately, if $g$ is a polynomial immersion then $\alpha'$ is not a polynomial mapping,
so one cannot apply previous results in order to compute $I(g)$.
In Section 4 we show how to construct a polynomial mapping
$\alpha:M\rightarrow \widetilde{V}_{m+1}(\R^{2m+1})$ such that $I(g)=-\Lambda(\alpha)$ (Theorem \ref{immersje}).
Therefore $I(g)$ can be expressed and computed in terms of signatures.
Another formula expressing $I(g)$ in terms of signatures of quadratic forms,
inspired by the original definition by Whitney, was presented in \cite{KarNowSzafr}
and generalized in \cite{krzyzanowska} to the case where $M$ may have singularities.
Calculations done with the help of a computer show that the method presented in this paper
is significantly more effective.
\label{intro}
\section{Mappings into a Stiefel manifold}
\label{sec:1}
If $M,N$ are closed oriented $n$--manifolds and $f:M\longrightarrow N$ continuous,
then by $\deg (f)$ we denote the topological degree of $f$. If $(M,\partial M)$ is
a compact oriented $n$--manifold with boundary and
$f:(M,\partial M)\longrightarrow (\R^n,\R^{n}\setminus \{0\})$ is continuous,
then by $\deg (f|\partial M)$ we denote the topological degree of
$f/|f|:\partial M\longrightarrow S^{n-1}$.
Let $n-k> 0$ be an even number and $k>1$.
Denote by $\widetilde{V}_k(\R^n)$ the set of all
$k$--frames in $\R^n$, and by $V_k(\R^n)$ the Stiefel manifold,
i.e. the set of all orthonormal $k$--frames in $\R^n$.
The Stiefel manifold is a deformation retract of $\widetilde{V}_k(\R^n)$,
so that $\pi_{n-k}V_k(\R^n)=\pi_{n-k}\widetilde{V}_k(\R^n)$.
It is known (see \cite{hatcher}), that $\pi_{n-k}V_k(\R^n)\simeq\Z$.
Let $[\alpha]\in \pi_{n-k}V_k(\R^n)$ be represented by
$\alpha =(\alpha_1,\ldots ,\alpha_k):S^{n-k}\longrightarrow V_k(\R^n)$,
where $\alpha_i:S^{n-k}\longrightarrow S^{n-1}\subset\R^n$.
Since $\alpha_1(x),\ldots ,\alpha_{k}(x)$ are linearly independent,
we can define $\widetilde{\alpha}:S^{k-1}\times S^{n-k}\longrightarrow \R^n\setminus \{0\}$ by
$$\widetilde{\alpha}(\beta,x)=\beta_1\alpha_1(x)+\ldots+\beta_k\alpha_k(x),$$
where $\beta=(\beta_1\ldots ,\beta_k)\in S^{k-1}$ and $x=(x_1,\ldots ,x_{n-k+1})\in S^{n-k}$.
\begin{lemma}
The mapping
$\widetilde{\alpha}$ goes into $S^{n-1}$.
\end{lemma}
\begin{proof} We know that $|\beta|=1$, $| \alpha_i(x) |=1$
for $i=1,\ldots ,k$, and the scalar products
$\langle\alpha_i(x),\alpha_j(x)\rangle=0$ for $i\neq j$. Then
$$|\widetilde{\alpha}(\beta,x)|^2=
\sum_{i=1}^k\beta_i^2+2\sum_{i\neq j}\beta_i\beta_j\langle\alpha_i(x),\alpha_j(x)\rangle=1.$$
\end{proof}
We got $\widetilde{\alpha}:S^{k-1}\times S^{n-k}\longrightarrow S^{n-1}$,
so the topological degree $\deg(\widetilde{\alpha})$ is defined.
For $\alpha:S^{n-k}\longrightarrow \widetilde{V}_k(\R^n)$,
we also have $\widetilde{\alpha}:S^{k-1}\times S^{n-1}\longrightarrow \R^n\setminus \{0\}$.
Then $\widetilde{\alpha}/|\widetilde{\alpha}|:S^{k-1}\times S^{n-1}\longrightarrow S^{n-1}$,
and $\deg(\widetilde{\alpha}/|\widetilde{\alpha}|)$ is well defined.
Let $r:\widetilde{V}_k(\R^n)\longrightarrow V_k(\R^n)$ be the retraction
given by the Gram-Schmidt orthonormalization. Then
$\alpha^t=(1-t)\alpha+t\cdot r\circ\alpha:S^{n-k}\longrightarrow\widetilde{V}_k(\R^n)$
is a homotopy between $\alpha$ and $r\circ\alpha$. Hence
$\widetilde{\alpha}^t/|\widetilde{\alpha}^t|: S^{k-1}\times S^{n-k}\longrightarrow S^{n-1}$
is a homotopy between $\widetilde{\alpha}/|\widetilde{\alpha}|$ and $\widetilde{r\circ\alpha}$,
so that $ \deg(\widetilde{\alpha}/|\widetilde{\alpha}|)=\deg(\widetilde{r\circ \alpha})$.
\begin{lemma}\label{homot}
If $\alpha^0,\alpha^1:S^{n-k}\longrightarrow V_k(\R^n)$ represent the same element in
$\pi_{n-k}V_{k}(\R^n)$, then $\deg (\widetilde{\alpha}^0)=\deg(\widetilde{\alpha}^1)$.
\end{lemma}
\begin{proof} There is a homotopy
$\alpha^t=(\alpha_1^t,\ldots , \alpha_k^t):S^{n-k}\longrightarrow V_k(\R^n)$
between $\alpha^0$ and $\alpha^1$. Then
$$\widetilde{\alpha}^t(\beta,x)=\beta_1\alpha_1^t(x)+\ldots+\beta_k\alpha_1^t(x):
S^{k-1}\times S^{n-k}\longrightarrow S^{n-1}$$
is a homotopy between $\widetilde{\alpha}^0$ and $\widetilde{\alpha}^1$,
and so $\deg (\widetilde{\alpha}^0)=\deg(\widetilde{\alpha}^1)$.
\end{proof}
\begin{example}\label{elneutralny}
The trivial element $[e]\in \pi_{n-k}V_k(\R^n)$ is represented by a constant mapping
$e(x)=(v_1,\ldots ,v_k)\in V_k(\R^n)$. Then
$\widetilde{e}(\beta,x)=\beta_1v_1+\ldots+\beta_kv_k$ is not onto $S^{n-1}$, so $\deg(\widetilde{e})=0$.
\end{example}
\begin{proposition}\label{parzystosc}
For all $[\alpha]\in \pi_{n-k}V_k(\R^n)$, $\deg(\widetilde{\alpha})$ is an even number.
\end{proposition}
\begin{proof} Let $M_k(\R^n)$ denote the set of all $k$--tuples of vectors in $\R^n$.
According to \cite[Proposition 5.3]{golub},
$\Sigma_k(\R^n)=M_k(\R^n)\setminus \widetilde{V}_k(\R^n)$ is an algebraic subset of
$M_k(\R^n)$ with $\codim \Sigma_k(\R^n)=n-k+1$.
Take $\alpha :S^{n-k}\longrightarrow V_k(\R^n)$. There is a homotopy
$h=(h_1,\ldots ,h_k):[0,1]\times S^{n-k}\longrightarrow M_k(\R^n)$ between
$\alpha $ and $e$ given by $h(t,x)=te(x)+(1-t)\alpha(x)$.
Put $h'=(h_2,\ldots ,h_k):[0,1]\times S^{n-k}\longrightarrow M_{k-1}(\R^n)$.
According to the Elementary Transversality Theorem \cite[Corollary 4.12]{golub},
there is $g'$ arbitrarily close to $h'$ such that $g'$ intersects
$\Sigma_{k-1}(\R^n)$ transversally. As $\codim\Sigma_{k-1}(\R^n)=n-k+2>\dim([0,1]\times S^{n-k})$,
so $(g')\inv (\Sigma_{k-1}(\R^n))=\emptyset$ and then
$g':[0,1]\times S^{n-k}\longrightarrow \widetilde{V}_{k-1}(\R^n)$.
Put $g=(g_1,\ldots ,g_k)=(h_1,g')$. We can choose $g'$ such that
$g(0,\cdot):S^{n-k}\longrightarrow \widetilde{V}_{k}(\R^n)$ is homotopic to
$\alpha$ and $g(1,\cdot):S^{n-k}\longrightarrow \widetilde{V}_{k}(\R^n)$ is homotopic to $e$.
Put
$\widetilde{g}:[0,1]\times S^{k-1}\times S^{n-k}\longrightarrow \R^n$ as $\widetilde{g}(t,\beta,x)=
\beta_1g_1(t,x)+\ldots +\beta_kg_k(t,x)$.
We know that $g_2(t,x),\ldots ,g_k(t,x)$ are linearly independent,
so $\widetilde{g}(t,(0,\beta '),x)\neq 0$ for $\beta=(0,\beta ')\in \{0\}\times S^{k-2}$.
Hence
$$\widetilde{g}\inv(0)\cap ([0,1]\times (\{0\}\times S^{k-2})\times S^{n-k})=\emptyset .$$
Put $H_{\pm}=\{\beta\in S^{k-1}|\ \pm\beta_1>0\}$, and
$M_{\pm}=[0,1]\times H_{\pm}\times S^{n-k}$.
Then $\widetilde{g}\inv (0)\subset M_-\cup M_+$. Of course $i(t,\beta, x)=(t,-\beta ,x)$
is a free involution on $[0,1]\times S^{k-1}\times S^{n-k}$, such that
$$\widetilde{g}\inv(0)\cap M_-=i(\widetilde{g}\inv(0)\cap M_+).$$
We also have $\widetilde{g}\inv(0)\cap \{0,1\}\times S^{k-1}\times S^{n-k}=\emptyset$.
So there exist two $n$--dimensional compact manifolds with boundary
$N_{\pm}\subset \inter(M_{\pm})$ such that
$\widetilde{g}(0)\subset \inter(N_-)\cup \inter(N_+)$ and $N_-=i(N_+)$.
By the Excision Theorem,
$$\deg(\widetilde{g}|\partial([0,1]\times S^{k-1}\times S^{n-k}))=
\deg(\widetilde{g}|\partial N_-)+\deg(\widetilde{g}|\partial N_+).$$
It is easy to check that $i:N_+\longrightarrow N_-$ preserves the orientation when $k$ is even,
and reverses it when $k$ is odd.
We have
$$\widetilde{g}|\partial N_+=-\widetilde{g}\circ i|\partial N_+=
-id|S^{n-1}\circ \widetilde{g}|\partial N_-\circ i|\partial N_+.$$
Of course $\deg (-id|S^{n-1})=(-1)^n$. Hence
$$\deg(\widetilde{g}|\partial N_+)=(-1)^n(-1)^k\deg(\widetilde{g}|\partial N_-)=
(-1)^{n+k}\deg(\widetilde{g}|\partial N_-)=\deg(\widetilde{g}|\partial N_-).$$
On the other hand
$$\deg(\widetilde{g}|\partial([0,1]\times S^{k-1}\times S^{n-k}))$$
$$=\deg(\widetilde{g}|(\{0\}\times S^{k-1}\times S^{n-k}))+\deg(\widetilde{g}|(\{1\}\times S^{k-1}\times S^{n-k}))$$
$$=\deg (\widetilde{e})-\deg (\widetilde{\alpha})=-\deg (\widetilde{\alpha}).$$
To sum up, we get that $\deg (\widetilde{\alpha})=-2\deg(\widetilde{g}|\partial N_-)$, so
$\deg (\widetilde{\alpha})$ is even.
\end{proof}
Let us define
$\Lambda:\pi_{n-k}V_k(\R^n)\longrightarrow \Z$ by $\Lambda([\alpha])=\deg(\widetilde{\alpha})/2$.
According to Proposition \ref{parzystosc}, $\Lambda$ is well defined.
\begin{proposition}
The mapping $\Lambda$ is a group homomorphism.
\end{proposition}
\begin{proof} Take $[\alpha_0],[\alpha_1]\in\pi_{n-k}V_k(\R^n)$, represented by smooth maps such that
$\alpha_0(x_0)=\alpha_1(x_0)=y_0$. The sum $[\alpha_0]+[\alpha_1]$ in $\pi_{n-k}V_k(\R^n)$
is represented by the composition
$(\alpha_0\vee \alpha_1)\circ c:S^{n-k}\longrightarrow S^{n-k}\vee S^{n-k}\longrightarrow V_k(\R^n)$,
where $c$ collapses the equator $S^{n-k-1}$ in $S^{n-k}$ to a point,
$x_0$ lies in $S^{n-k-1}$ and $S^{n-k}\vee S^{n-k}$ is the disjoint union of $S^{n-k}$ and
$S^{n-k}$ with the identification $x_0\sim x_0$.
Then $\widetilde{\alpha_0+\alpha_1}:S^{k-1}\times S^{n-k}\longrightarrow S^{n-1}$.
It is easy to see that $\widetilde{\alpha_0+\alpha_1}(S^{k-1}\times S^{n-k-1})$
is not dense in $S^{n-1}$. So there is a regular value $y\in S^{n-1}$
such that $\widetilde{\alpha_0+\alpha_1}\inv(y)\cap (S^{k-1}\times S^{n-k-1})=\emptyset$.
Then $\widetilde{\alpha_0+\alpha_1}\inv(y)$ is the disjoint union of
$\widetilde{\alpha}_0\inv(y)$ and $\widetilde{\alpha}_1\inv(y)$, so
$\deg(\widetilde{\alpha_0+\alpha_1})=\deg(\widetilde{\alpha}_0)+\deg(\widetilde{\alpha}_1)$.
\end{proof}
\begin{proposition}
The mapping $\Lambda$ is surjective.
\end{proposition}
\begin{proof} Let us define a mapping
$\alpha=(\alpha_1,\ldots , \alpha_k) : S^{n-k}\longrightarrow V_k(\R^n)$ by
$\alpha_i(x)=(0,\ldots , 1, 0, \ldots ,0)$, with the $1$ in the $i$--th coordinate for
$i=1,\ldots ,k-1$, and $\alpha_k=(0,\ldots ,0,x_1,\ldots , x_{n-k+1})$. Then
$$\widetilde{\alpha}(\beta, x)=
(\beta_1,\ldots , \beta_{k-1},\beta_kx_1,\ldots , \beta_{k}x_{n-k},\beta_kx_{n-k+1})\in S^{n-1}.$$
It is easy to check that
$\widetilde{\alpha}\inv (0,\ldots ,0,1)=
\{(\beta,x)|\ \beta_1=\ldots=\beta_{k-1}=x_1=\ldots=x_{n-k}=0,\beta_k=x_{n-k+1}=
\pm 1\}$.
Projection onto first $(n-1)$-- coordinates in a neighbourhood of $(0,\ldots ,0,1)\in S^{n-1}$
is an orientation preserving chart if
and only if $n$ is odd. Near $(0,\ldots ,0,1,0,\ldots ,0, 1)\in S^{k-1}\times S^{n-k}$
we have an orientation preserving parametrization given by
$$(\beta',x')=(\beta_1,\ldots ,\beta_{k-1},x_1,\ldots ,x_{n-k})\longmapsto$$
$$\left((-1)^{k-1}\beta_1,\beta_2\ldots ,\beta_{k-1},\sqrt{1-|\beta'|^2},x_1,\ldots ,x_{n-k},\sqrt{1-|x'|^2}\right).$$
It is easy to check that in these coordinates the derivative matrix of
$\widetilde{\alpha}$ at $(0,\ldots ,0,1,0,\ldots ,0, 1)$ has the form
$$\left [
\begin{array}{cccc}
(-1)^{k-1}&0&\ldots &0\\
0&1&\ldots&0\\
&&\ddots&\\
0&0&\ldots &1
\end{array} \right ].$$
Near $(0,\ldots ,0,-1,0,\ldots ,0, -1)\in S^{k-1}\times S^{n-k}$ we have an orientation preserving parametrization
$$(\beta',x')=(\beta_1,\ldots ,\beta_{k-1},x_1,\ldots ,x_{n-k})\longmapsto $$
$$\left((-1)^k\beta_1,\beta_2,\ldots ,\beta_{k-1},-\sqrt{1-|\beta'|^2},-x_1,x_2,\ldots ,x_{n-k},-\sqrt{1-|x'|^2}\right).$$
It is easy to check that the derivative matrix of $\widetilde{\alpha}$ at
$(0,\ldots ,0,-1,0,\ldots ,0, -1)$ in these coordinates has the form
$$\left [
\begin{array}{ccccccc}
(-1)^k&0&\ldots &0&0&\ldots&0\\
0&1&\ldots &0&0&\ldots &0\\
&&\ddots&&&\ddots&\\
0&0&\ldots &1&0&\ldots &0\\
0&0&\ldots &0&-1&\ldots &0\\
&&\ddots&&&\ddots&\\
0&0& \ldots &0&0 &\ldots & -1\\
\end{array} \right ]$$
Then $(0,\ldots ,0,1)$ is a regular value of $\widetilde{\alpha}$, and
$\deg(\widetilde{\alpha})=(-1)^{n-1}((-1)^{k-1}+(-1)^k(-1)^{n-k-1})=2$, so $\Lambda(\alpha)=1$.
By Proposition 2, $\Lambda$ is surjective.
\end{proof}
\begin{theorem}
If $n-k$ is even then $\Lambda:\pi_{n-k}V_k(\R^n)\longrightarrow\Z$ is an isomorphism.
\end{theorem}
\begin{proof}
Since $\pi_{n-k}V_k(\R^n)\simeq\Z$, the surjective homomorphism $\Lambda$ is an isomorphism.
\end{proof}
Let $M$ be a closed oriented $(n-k)$--manifold. With any
$\alpha:M\longrightarrow \widetilde{V}_{k}(\R^n)$ we may associate the same way as above
the mapping
$\widetilde{\alpha}:S^{k-1}\times M\longrightarrow \R^n\setminus\{0\}$ given by
$$\widetilde{\alpha}(\beta,x)=\beta_1\alpha_1(x)+\ldots+\beta_k\alpha_k(x).$$
Then the topological degree of $\widetilde{\alpha}$ is well defined.
Applying the same arguments as in the proof of Proposition \ref{parzystosc},
one can prove
\begin{theorem} \label{parzystoscstopnia}
Let $\alpha:M\longrightarrow \widetilde{V}_{k}(\R^n)$ be continuous, where $n-k$ is even.
Then $\deg( \widetilde{\alpha})$ is even, and so
$\Lambda(\alpha)=\deg(\widetilde\alpha)/2$ is an integer. If
$\alpha^0,\alpha^1:M\rightarrow\widetilde{V}_k(\R^n)$ are homotopic, then
$\Lambda(\alpha^0)=\Lambda(\alpha^1)$.
\end{theorem}\hspace*{\fill}$\Box$\par\medskip
\section{Polynomial mappings into a Stiefel manifold}\label{efektywnie}
If $U$ is an open subset of an $n$-dimensional oriented manifold,
$H:U\longrightarrow \R^n$ is continuous and $p\in H\inv(0)$ is isolated in
$H\inv(0)$, then by $\deg_p H$ we denote the local topological degree of $H$ at $p$.
If $H^{-1}(0)$ is compact, then there exists a compact manifold with boundary
$N\subset U$ such that $H^{-1}(0)\subset\operatorname{int}(N)$.
If that is the case then the topological degree $\deg(H,U,0)$ is defined as the degree
of the mapping $\partial N\ni x\mapsto H(x)/|H(x)|\in S^{n-1}$.
In particular, if $H^{-1}(0)$ is finite then $\deg(H,U,0)=\sum \deg_p H$, where
$p\in H^{-1}(0)$.
Let $\alpha=(\alpha_1,\ldots ,\alpha_k):\R^{n-k+1}\longrightarrow M_k(\R^n)$
be a polynomial mapping. Denote by $\left [a_{ij}(x)\right ]$,
$1\leq i\leq n$, $1\leq j\leq k$, the matrix in which $\alpha_j(x)$
stands in the $j$--th column. Define
$$\widetilde{\alpha}(\beta, x)=
\beta_1\alpha_1(x)+\ldots +\beta_k\alpha_k(x)=\left [a_{ij}(x)\right ]\left [\begin{array}{c}
\beta_1\\
\vdots\\
\beta_k
\end{array} \right ]:\R^k\times \R^{n-k+1}\longrightarrow \R^n.$$
By $I$ we denote the ideal in $\R[x_1,\ldots ,x_{n-k+1}]$ generated by all
$k\times k$ minors of $\left [a_{ij}(x)\right ]$. Let $V(I)=\{x\in\R^{n-k+1}\ |\ h(x)=0\mbox{ for all }h\in I\}$.
\begin{lemma}\label{zaleznosc}
We have $p\in V(I)$ if and only if $\alpha_1(p), \ldots ,\alpha_k(p)$ are linearly dependent,
i.e. if $\widetilde{\alpha}(\beta, p)=0$ for some $\beta\neq 0$.
\end{lemma}\hspace*{\fill}$\Box$\par\medskip
Let $I_1$ be the ideal generated by all $(k-1)\times(k-1)$ minors of $[a_{ij}(x)]$.
Put
$$m(x)=\det\left [\begin{array}{ccc}
a_{12}(x)&\ldots &a_{1k}(x)\\ \\
a_{k-1,2}(x)&\ldots &a_{k-1,k}(x)
\end{array} \right ].$$
\begin{lemma}\label{zaleznosc1}
We have $V(I)\setminus V(I_1)=\emptyset$ if and only if
$\rank [\alpha_1(p),\ldots,\alpha_k(p)]=k-1$ at each $p\in V(I)$.
If that is the case and $V(I)$ is finite then one may choose well oriented
coordinates in $\R^{n-k+1}$ and $\R^n$
such that $m(p)\neq 0$ at each $p\in V(I)$.
\end{lemma}\hspace*{\fill}$\Box$\par\medskip
From now on we assume that $m(p)\neq 0$ at each $p\in V(I)$. Hence, if
$\alpha_1(p), \ldots ,\alpha_k(p)$ are linearly dependent
then $\alpha_2(p), \ldots ,\alpha_k(p)$ are linearly independent.
In that case there exists a uniquely determined
$\bar{\lambda}=(\bar{\lambda}_2,\ldots , \bar{\lambda}_k)\in \R^{k-1}$
such that $\alpha_1(p)+\bar{\lambda}_2\alpha_2(p)+\ldots +\bar{\lambda}_k\alpha_k(p)=0$.
Therefore we have
\begin{lemma}\label{jednoznacznosc}
Suppose that $p\in V(I)$. Then there is a uniquely determined
$\bar{\beta}\in H_+=S^{k-1}\cap \{\beta_1>0\}$ such that
$\widetilde{\alpha}(\bar{\beta},p)=\bar{\beta}_1\alpha_1(p)+\bar{\beta}_2\alpha_2(p)+\ldots +\bar{\beta}_k\alpha_k(p)=0$,
and $\bar{\lambda}_2=\bar{\beta}_2/\bar{\beta}_1,\ldots,\bar{\lambda}_k=
\bar{\beta}_k/\bar{\beta}_1$.
\end{lemma}\hspace*{\fill}$\Box$\par\medskip
Note that
$$H_+\ni (\beta_1,\beta_2,\ldots, \beta_k)\mapsto (\beta_2/\beta_1,\ldots,\beta_k/\beta_1)\in \R^{k-1}$$
is an orientation preserving diffeomorphism. For $\lambda=(\lambda_2,\ldots ,\lambda_k)\in \R^{k-1}$
and $1\leq i\leq n$
we define $F_i(\lambda, x)=a_{i1}(x)+\lambda_2a_{i2}(x)+\ldots +\lambda_ka_{ik}(x)$. Then
$$F=(F_1,\ldots , F_n)=\left [a_{ij}(x)\right ]\left [\begin{array}{c}
1\\ \lambda_2\\
\vdots\\
\lambda_k
\end{array} \right ]:\R^{k-1}\times \R^{n-k+1}\longrightarrow\R^n.$$
\begin{lemma}\label{FiLambda}
A point $(\bar{\beta}, p)\in H_+\times \R^{n-k+1}$ is an isolated zero of
$\widetilde{\alpha}|(H_+\times\R^{n-k+1})$ if and only if $(\bar{\lambda},p)\in \R^{k-1}\times \R^{n-k+1}$
is an isolated zero of $F$. If that is the case then
$$\deg_{(\bar{\beta}, p)}(\widetilde{\alpha}|(H_+\times\R^{n-k+1}))=\deg_{(\bar{\lambda},p)}(F).$$
\end{lemma} \hspace*{\fill}$\Box$\par\medskip
In such case
$$\frac{\partial (F_1,\ldots , F_{k-1})}{\partial (\lambda_2,\ldots , \lambda_k)}(\bar{\lambda},p)=m(p)\neq 0,$$
so $\partial (F_1,\ldots , F_{k-1})/\partial (\lambda_2,\ldots , \lambda_k)\neq 0$
in a neighbourhood of $(\bar{\lambda},p).$
According to the Cramer rule, there exists a uniquely determined
$\lambda(x)=(\lambda_2(x),\ldots ,\lambda_k(x))$
defined in a neighbourhood of $p$, such that
$$\left \{\begin{array}{c}
F_1(\lambda(x),x)=0\\
\vdots\\
F_{k-1}(\lambda(x),x)=0
\end{array}\right.$$
and \begin{equation}\lambda_i(x)=\frac{(-1)^{i-1}}{m(x)}\det\left[\begin{array}{ccccc}
a_{11}(x)&\ldots & \widehat{a_{1,i}(x)}&\ldots &a_{1k}(x)\\ &\ldots&&\ldots&\\
a_{k-1,1}(x)&\ldots &\widehat{a_{k-1,i}(x)}&\ldots &a_{k-1,k}(x)
\end{array} \right ]\ \label{star}\end{equation}
for $2\leq i\leq k$.
Denote $\Gamma=\{(\lambda(x),x)\}\subset \R^{k-1}\times \R^{n-k+1}$,
of course $\Gamma$ is an $(n-k+1)$-- manifold near $(\bar{\lambda},p)$.
On $\Gamma$ we take the orientation induced by equations $F_1=\ldots =F_{k-1}=0$,
i.e. vectors $v_1,\ldots, v_{n-k+1} $ tangent to
$\Gamma$ at $(\lambda(x),x)$ are well oriented if and only if
$$\nabla F_1(\lambda(x),x),\ldots ,\nabla F_{k-1}(\lambda(x),x),v_1,\ldots , v_{n-k+1}$$
are well oriented in $\R^{k-1}\times\R^{n-k+1}$. One can see that vectors
$$v_s=\frac{\partial}{\partial x_s}(\lambda(x),x)=
\Big(\frac{\partial \lambda_2}{\partial x_s},\ldots,
\frac{\partial \lambda_k}{\partial x_s},0,\ldots, 1,0,\ldots 0\Big),\ 1\leq s\leq n-k+1,$$
with the $1$ in the $(k+s-1)$--th coordinate, are tangent to $\Gamma$.
Their orientation corresponds to the standard orientation of an $\R^{n-k+1}$
\begin{lemma}\label{orientacja}
Vectors $v_1,\ldots, v_{n-k+1}$ are well oriented if and only if $m(p)>0$.
\end{lemma}
\begin{proof} We have $F_i(\lambda(x),x)\equiv 0$, for $1\leq i\leq k-1$, so
$$0\equiv\frac{\partial}{\partial x_s}F_i(\lambda(x),x)$$
$$=a_{i2}(x)\frac{\partial \lambda_2}{\partial x_s}(x)+\ldots +a_{ik}(x)\frac{\partial \lambda_k}{\partial x_s}(x)+\frac{\partial a_{i1}}{\partial x_s}(x)+\lambda_2(x)\frac{\partial a_{i2}}{\partial x_s}(x)+\ldots +\lambda_k(x)\frac{\partial a_{ik}}{\partial x_s}(x).$$
Denote by $A$ the Gramian matrix of vectors $v_1,\ldots ,v_{n-k+1}$. It is easy to see that
the determinant of the matrix having rows
$\nabla F_1(\lambda(x),x),\ldots ,\nabla F_{k-1}(\lambda(x),x),$ $v_1(x),\ldots ,v_{n-k+1}(x)$ equals
$$\det\left[\begin{array}{cccccc}a_{12}(x)&\ldots &a_{1k}(x)&0&\ldots&0\\\vdots&&\vdots&\vdots&&\vdots\\a_{k-1,2}(x)&\ldots &a_{k-1,k}(x)&0&\ldots&0\\ &*&&&A
\end{array}\right ]=m(x)\det A.$$
Vectors $v_1,\ldots ,v_{n-k+1}$ are linearly independent, so $\det A>0$.
We get that $v_1,\ldots ,v_{n-k+1}$ are well oriented in $\Gamma$ if and only if $ m(x)>0$.
\end{proof}
Denote by $\Oo_n$ the ring of germs at $(\bar{\lambda},p)$ of analytic functions
$\R^{k-1}\times\R^{n-k+1}\longrightarrow \R$, by $\Oo_{\Gamma}$ the ring of germs at
$(\bar{\lambda},p)$ of analytic functions
$\Gamma\longrightarrow \R$, and by $\Oo_{n-k+1}$ the ring of germs at $p$ of analytic functions
$\R^{n-k+1}\longrightarrow \R$. Since
$\partial (F_1,\ldots , F_{k-1})/\partial (\lambda_2,\ldots , \lambda_k)(\bar{\lambda},p)\neq 0$,
so there is a natural isomorphism
$$\Oo_{\Gamma}\simeq \Oo_n/\langle F_1,\ldots , F_{k-1}\rangle .$$
Put $\Omega_i(x)=F_i(\lambda(x),x)\in \Oo_{n-k+1}$. Because $\Gamma$ is the graph of
$x\mapsto \lambda(x)$, so $\Oo_{\Gamma}\simeq \Oo_{n-k+1}$ and
$$\Oo_n/\langle F_1,\ldots ,F_{k-1},F_k,\ldots , F_n\rangle \simeq \Oo_{\Gamma} / \langle F_k,
\ldots , F_n\rangle \simeq \Oo_{n-k+1}/\langle \Omega_1,\ldots ,\Omega_n \rangle.$$
We have $m(p)\neq 0$, so the germ of $m$ is invertible in $\Oo_n$, $\Oo_{\Gamma}$ and $\Oo_{n-k+1}$.
Denote by $J$ the ideal in $\Oo_{n-k+1}$ generated by all $k\times k$ minors of $\left [a_{ij}(x)\right ]$.
\begin{lemma}
$\Oo_n/\langle F_1,\ldots , F_n\rangle \simeq \Oo_{\Gamma} / \langle F_k,\ldots , F_n\rangle\simeq \Oo_{n-k+1}/J$.
\end{lemma}
\begin{proof}
Take $1\leq i_1<\ldots <i_k\leq n$. Of course
$\Omega_{i_1}=\ldots =\Omega_{i_k}=0$ in $\Oo_{n-k+1}/\langle \Omega_1,\ldots , \Omega_n\rangle$,
i.e.
$$\left \{\begin{array}{c}
a_{i_1,1}(x)+\lambda_2(x)a_{i_1,2}(x)+\ldots +\lambda_k(x)a_{i_1,k}(x)=0\\
\vdots\\
a_{i_k,1}(x)+\lambda_2(x)a_{i_k,2}(x)+\ldots +\lambda_k(x)a_{i_k,k}(x)=0
\end{array}\right.$$
According to the Cramer rule,
$$\det\left [\begin{array}{ccc}
a_{i_1,1}(x)&\ldots &a_{i_1,k}(x)\\ &\ldots& \\
a_{i_k,1}(x)&\ldots &a_{i_k,k}(x)
\end{array} \right ]=0\ \mbox{ in }\ \Oo_{n-k+1}/\langle \Omega_1,\ldots,\Omega_n\rangle .$$
Each generator of $J$ belongs to $\langle \Omega_1,\ldots , \Omega_n\rangle$,
so $J\subset \langle \Omega_1,\ldots , \Omega_n\rangle$.
Applying (\ref{star}) one may show
$$\Omega_i(x)=\frac{(-1)^{k-1}}{m(x)}\det\left[\begin{array}{ccc}
a_{11}(x)&\ldots &a_{1k}(x) \\
&\ldots& \\
a_{k-1,1}(x)&\ldots &a_{k-1,k}(x)\\
a_{i1}(x)&\ldots &a_{ik}(x)
\end{array} \right ]=:\frac{(-1)^{k-1}}{m(x)}\Delta_i(x).$$
In particular, germs of $\Omega_1,\ldots, \Omega_n$ in $\Oo_{n-k+1}$ belong to $J$.
Of course $\Omega_1 \equiv 0,\ldots,\Omega_{k-1}\equiv 0$.
So $\langle \Omega_1,\ldots ,\Omega_n\rangle=\langle \Omega_k,\ldots,\Omega_n \rangle=J$, and there are natural isomorphisms
$$\Oo_n/\langle F_1,\ldots , F_n\rangle \simeq \Oo_{\Gamma} / \langle F_k,\ldots ,
F_n\rangle\simeq \Oo_{n-k+1}/\langle \Omega_1,\ldots ,\Omega_n \rangle\simeq\Oo_{n-k+1}/J.$$ \end{proof}
From the previous proof we also get
\begin{lemma}
$\Oo_{n-k+1}/J\simeq \Oo_{n-k+1}/\langle \Delta_k,\ldots,\Delta_n\rangle$.
\end{lemma}\hspace*{\fill}$\Box$\par\medskip
\begin{lemma}\label{stopienF}
If $n-k$ is even and $\partial (\Delta_k,\ldots , \Delta_n)/\partial (x_1,\ldots ,x_{n-k+1})(p)\neq 0$,
then $$\deg_{(\bar{\lambda},p)}(F)=(-1)^{k-1}\sgn\frac{\partial (\Delta_k,\ldots ,
\Delta_n)}{\partial (x_1,\ldots ,x_{n-k+1})}(p).$$
\end{lemma}
\begin{proof} We have $\Omega_i(x)m(x)=(-1)^{k-1}\Delta_i(x)$ so
$$\frac{\partial\Omega_i}{\partial x_s}m+\Omega_i\frac{\partial m}{\partial x_s}=
(-1)^{k-1}\frac{\partial \Delta_i}{\partial x_s}.$$
As $\Omega_i\in J$, then
$$\frac{\partial\Omega_i}{\partial x_s}\equiv \frac{(-1)^{k-1}}{m}\frac{\partial\Delta_i}{\partial x_s}\ \mod J.$$
Because $n-k$ is even, so
$$\frac{\partial (\Omega_k,\ldots , \Omega_n)}{\partial (x_1,\ldots ,
x_{n-k+1})}\equiv \frac{(-1)^{k-1}}{m^{n-k+1}}\frac{\partial (\Delta_k,\ldots , \Delta_n)}
{\partial (x_1,\ldots ,x_{n-k+1})}\ \mod J.$$
Since $\partial (\Delta_k,\ldots , \Delta_n)/\partial (x_1,\ldots ,x_{n-k+1})(p)\neq 0$,
the mapping $(\Omega_k,\ldots ,\Omega_n)$ has an isolated regular zero at $p$, moreover
$$\deg_p(\Omega_k,\ldots,\Omega_n)=
(-1)^{k-1}\sgn (m(p))\sgn\frac{\partial (\Delta_k,\ldots , \Delta_n)}{\partial (x_1,\ldots ,x_{n-k+1})}(p).$$
So the local topological degree of
$(F_k,\ldots ,F_n):(\Gamma,(\bar{\lambda},p))\longrightarrow (\R^{n-k+1},0)$ at $(\bar{\lambda},p)$ equals
$$(-1)^{k-1}\sgn (m(p))\sgn\frac{\partial (\Delta_k,\ldots , \Delta_n)}{\partial (x_1,\ldots ,x_{n-k+1})}(p),$$
if and only if the orientation of $\Gamma$ is the same as the one induced from $\R^{n-k+1}$.
According to Lemma \ref{orientacja}, the local topological degree of
$(F_k,\ldots ,F_n):(\Gamma,(\bar{\lambda},p))\longrightarrow (\R^{n-k+1},0)$ ,
where the orientation of $\Gamma$ is the one induced by equations $F_1=\ldots=F_{k-1}=0$, equals
$$(-1)^{k-1}\sgn\frac{\partial (\Delta_k,\ldots , \Delta_n)}{\partial (x_1,\ldots ,x_{n-k+1})}(p).$$
According to \cite[Lemma 3.2]{Szafraniec}, $(\bar{\lambda}, p)$ is isolated in
$((F_k,\ldots ,F_n)|\Gamma)\inv(0)$ if and only if
$(\bar{\lambda}, p)$ is isolated in
$(F_1,\ldots ,F_n)\inv(0)$, moreover
$$\deg_{(\bar{\lambda}, p)}(F_k,\ldots ,F_n)|\Gamma=\deg_{(\bar{\lambda}, p)}(F_1,\ldots ,F_n).$$
Hence the local topological degree of
$F=(F_1,\ldots, F_{k-1},F_{k},\ldots,F_n):(\R^{k-1}\times \R^{n-k+1},(\bar{\lambda}, p))\longrightarrow
(\R^n,0)$ at $(\bar{\lambda}, p)$ equals
$$(-1)^{k-1}\sgn\frac{\partial (\Delta_k,\ldots , \Delta_n)}{\partial (x_1,\ldots ,x_{n-k+1})}(p).$$
\end{proof}
Put $\Aa=\R[x_1,\ldots,x_{n-k+1}]/I$. Let us assume that $\dim\Aa<\infty$, so that $V(I)$ is finite.
For $h\in\Aa$, we denote by $T(h)$ the trace of the linear endomorphism
$\Aa\ni a\mapsto h\cdot a\in\Aa$. Then $T:\Aa\rightarrow\R$ is a linear functional.
Let $f\in\R[x_1,\ldots,x_{n-k+1}]$ and $M=f^{-1}(0)$. Assume that $D=\{x\ |\ f(x)\geq 0\}$ is bounded and
$\nabla f(x)\neq 0$ at each $x\in M$. Then $D$ is a compact manifold with boundary
$\partial D=M$, and $\dim M=n-k$.
Put $\delta=\partial(\Delta_k,\ldots,\Delta_n)/\partial(x_1,\ldots,x_{n-k+1})$. With $f$ and $\delta$ we associate
quadratic forms $\Theta_\delta,\ \Theta_{f\cdot \delta}:\Aa\rightarrow\R$
given by $\Theta_\delta(a)=T(\delta\cdot a^2)$ and $\Theta_{f\cdot \delta}(a)=T(f\cdot\delta\cdot a^2)$.
According to \cite{becker,pedersenetal}, we have
$$\operatorname{signature}\, \Theta_\delta=\sum \operatorname{sgn}(\delta(p)),$$
$$\operatorname{signature}\, \Theta_{f\cdot \delta}=\sum \operatorname{sgn}(f(p)\delta(p)),$$
where $p\in V(I)$. Moreover, if the quadratic forms are non-degenerate then
$\delta(p)\neq 0$ and $f(p)\neq 0$ at each $p\in V(I)$, so that $V(I)\cap M=\emptyset$.
In that case vectors $\alpha_1(x),\ldots,\alpha_k(x)$ are linearly independent
at every $x\in M$, and then the restricted mapping $\alpha|M$ goes into $\widetilde{V}_k(\R^n)$.
Hence $\widetilde{\alpha}|S^{k-1}\times M$ goes into $\R^n\setminus\{0\}$.
\begin{theorem}\label{efeektywnie} If $n-k$ is even,
$\alpha=(\alpha_1,\ldots ,\alpha_k):\R^{n-k+1}\longrightarrow M_k(\R^n)$
is a polynomial mapping such that $\dim\Aa<\infty$,
$I+\langle m \rangle=\R[x_1,\ldots ,x_{n-k+1}]$ and quadratic forms
$\Theta_\delta,\, \Theta_{f\cdot\delta}:\Aa\longrightarrow\R$ are non--degenerate then
$$\Lambda(\alpha|M)=\frac{1}{2}\deg(\widetilde{\alpha}|S^{k-1}\times M)=
\frac{(-1)^{k-1}}{2}(\signature\Theta_{\delta}+\signature\Theta_{f\cdot\delta}),$$
where $\widetilde{\alpha}(\beta,x)=\beta_1\alpha_1(x)+\ldots +\beta_k\alpha_k(x)$.
\end{theorem}
\begin{proof}
The product $S^{k-1}\times D$ is a compact $n$--manifold with boundary
$\partial(S^{k-1}\times D)=S^{k-1}\times M$. The standard orientation of the boundary
coincides with the standard orientation of the product $S^{k-1}\times M$ if and only if $k$ is odd.
Then
$$\deg(\widetilde{\alpha}|S^{k-1}\times M)=(-1)^{k-1}\deg(\widetilde{\alpha},S^{k-1}\times D,0).$$
Take $(\bar{\beta},p)\in\widetilde{\alpha}^{-1}(0)\cap S^{k-1}\times D$.
Then $\alpha_1(p),\ldots,\alpha_k(p)$ are linearly dependent.
By Lemma \ref{zaleznosc}, $p$ belongs to a finite set $V(I)$. Since $I+\langle m\rangle=\R[x_1,\ldots,x_{n-k+1}]$,
we have $m(p)\neq 0$ at each $p\in V(I)$. Then,
by Lemma \ref{jednoznacznosc}, $(\bar{\beta},p)$ and $(-\bar{\beta},p)$ are isolated in
$\widetilde{\alpha}^{-1}(0)\cap S^{k-1}\times D$ and the first coordinate
$\bar{\beta}_1\neq 0$.
Because $n-k$ is even, then
$$\deg_{(\bar{\beta},p)}\widetilde{\alpha}=
\deg(-id|S^{n-1})\cdot\deg_{(-\bar{\beta},p)}\widetilde{\alpha}\cdot \deg(-id|S^{k-1})=
\deg_{(-\bar{\beta},p)}\widetilde{\alpha}.$$
Hence $\deg(\widetilde{\alpha},S^{k-1}\times D,0)=2\deg(\widetilde{\alpha},H_+\times D,0)$.
By Lemma \ref{FiLambda},
$$\deg(\widetilde{\alpha},H_+\times D,0)=\deg(F,\R^{k-1}\times D,0)=\sum\deg_{(\bar{\lambda},p)} F,$$
where $(\bar{\lambda},p)\in F^{-1}(0)\cap \R^{k-1}\times D$.
The quadratic form $\Theta_\delta$ is non-degenerate, hence
$\delta(p)\neq 0$ at each $p\in V(I)$. By Lemma \ref{stopienF},
$$\deg(\widetilde{\alpha},S^{k-1}\times D,0)=2(-1)^{k-1}\sum\sgn \delta(p),$$
where $p\in V(I)\cap D=V(I)\cap\{f>0\}$.
On the other hand,
$$\signature\Theta_\delta+\signature\Theta_{f\cdot \delta}$$
$$=\sum_{p\in V(I)}\sgn\delta(p)+\left( \sum_{p\in V(I)\cap\{f>0\}}\sgn\delta(p)- \sum_{p\in V(I)\cap\{f<0\}}\sgn\delta(p) \right)$$
$$=2\sum_{p\in V(I)\cap D}\sgn\delta(p).$$
\end{proof}
\begin{example}
Take $$A=\left[\begin{array}{cc}
2z+2&y+2\\2y+1&2y+1\\2x+1&y+2\\z+1&2y+1
\end{array}\right].$$ Let $\alpha=(\alpha_1,\alpha_2):\R^3\longrightarrow M_2(\R^4)$ be a polynomial mapping such that $\alpha_j$ is the $j$--th column of $A$.
One may check that $I$ is generated by $2y-z,$ $2x-2z-1,$ $z^2+z$ and $\Aa=\R[x,y,z]/I$ is a $2$--dimensional algebra, where $e_1=1$, $e_2=z$ is its basis.
In our case $m=y+2$, and so $I+\langle m\rangle=\R[x,y,z].$ One may check that $T(e_1)=2$, $T(e_2)=-1$, and $\delta=-24-\frac{75}{2}z$, $f\cdot \delta =-18-\frac{45}{4}z$ in $\Aa$. The matrices of $\Theta_{\delta}$ and $\Theta_{f\cdot \delta}$ are
$$\left[\begin{array}{cc}
-{21}/{2}&-{27}/{2}\\
-{27}/{2}&{27}/{2}
\end{array}\right] \ \mbox{ and }\ \left[\begin{array}{cc}
-{99}/{4}&{27}/{4}\\
{27}/{4}&-{27}/{4}
\end{array}\right].$$
Hence $\signature \Theta_{\delta}=0$, $\signature \Theta_{f\cdot \delta}=-2$. Applying Theorem \ref{efeektywnie} we get that $\alpha|S^2:S^2\longrightarrow \widetilde{V}_2(\R^4)$ and $\Lambda(\alpha|S^2)=-(0-2)/2=1.$
\end{example}
\section{Intersection number}
\label{sec:2}
By $B^n(r)$ we denote the $n$--dimensional open ball with radius $r$ centred at the origin,
by $\bar{B}^n(r)$ its closure,
and by $S^{n-1}(r)$ the $(n-1)$--dimensional sphere with radius $r$ centred at the origin.
\begin{lemma}\label{stopienNaSferze}
Let $H=(h_1,\ldots ,h_n):\R^n\longrightarrow \R^n$ be continuous. If
$Z=\{x\in \R^n|\ h_1(x)=\ldots =h_k(x)=0\}$ is compact for some $k<n$,
then the topological degree of $H/|H|:S^{n-1}(R)\longrightarrow S^{n-1}$
is equal to zero for any $R>0$ with $Z\subset B^n(R).$
\end{lemma}
\begin{proof} Suppose that $Z\subset B^n(R)$, so $Z\cap S^{n-1}(R)=\emptyset$ and
$H(x)\neq 0$ for $x\in S^{n-1}(R)$. Let us consider
$H/|H|:S^{n-1}(R)\longrightarrow S^{n-1}$. We have
$$ (H/|H|)\inv(0,\ldots ,0,1)\subset Z\cap S^{n-1}(R)=\emptyset.$$
So $(H/|H|)\inv(0,\ldots ,0,1)=\emptyset$, and the degree of $H/|H|$ equals zero.
\end{proof}
Let $H=(h_1,\ldots ,h_n):\R^n\longrightarrow \R^n$ be continuous,
let $U\subset\R^n$ be an open set such that
$H\inv (0)\cap U$ is compact, so that the topological degree
$\deg(H,U,0)$ is defined. Suppose that $h_1(x)=x_1g(x)$, where
$g(x)>0$ for $x\in U$. By $U'$ denote the set
$\{x'=(x_2,\ldots ,x_n)\in\R^{n-1}\ |\ (0,y')\in U\}$. Of course
$H\inv (0)\cap U\subset \{0\}\times U'$.
Let us define the mapping $H':\R^{n-1}\longrightarrow\R^{n-1}$ by
$H'(x')=(h_2(0,x'),\ldots ,h_n(0,x'))$.
Then $(H')^{-1} (0)\cap U'$ is compact and $\deg(H',U',0)$ is well defined.
We have
\begin{lemma}\label{stopien}
$\deg(H,U,0)=\deg(H',U',0)$.
\end{lemma} \hspace*{\fill}$\Box$\par\medskip
Let us assume that
$f :\R ^{m+1} \longrightarrow \R $
is a smooth function such that
$M=f^{-1}(0)$ is compact
and $\nabla f(p)\neq 0$ at each $p\in M$,
so that $M$ is an $m$-dimensional manifold. We shall say that vectors
$v_1, \ldots ,v_m\in T_pM$ are well oriented if vectors $\nabla
f(p),v_1,\ldots ,v_m$ are well oriented in
$\R^{m+1}$. In this way $M$ is an oriented manifold.
Let
$G=(g_1,\ldots,g_{2m}) :\R^ {m+1} \longrightarrow \R
^{2m}$
be smooth.
Put $g=G|M$ and
define
$H:\R^{m+1}\times
\R^{m+1}\longrightarrow \R^{2+2m}$ by
$$H(x,y)=(f(x),f(y),g_1(x)-g_1(y),\ldots
,g_{2m}(x)-g_{2m}(y)).$$
According to \cite[Lemma 18, Proposition 20]{KarNowSzafr} we have
\begin{proposition}\label{nasze}
The mapping $g:M\longrightarrow \R^{2m}$ is an immersion if and only if
the mapping $\R^{m+1}\ni x\mapsto (f(x),g_1(x),\ldots, g_{2m}(x))$ has rank $m+1$
at each $p\in M$.
If that is the case then there exists a compact
$2(m+1)$-dimensional manifold with boundary $N\subset\R^{m+1}\times\R^{m+1}$
such that
$$\{(x,y)\in\R^{m+1}\times\R^{m+1}\ |\ H(x,y)=0, x\neq y\}\subset N\setminus\partial N.$$
If $m$ is even, then for any such $N$ the intersection number $I(g)$ equals
$\deg(H,N,0)/2= \deg(H|\partial N)/2.$
\end{proposition}\hspace*{\fill}$\Box$\par\medskip
From now on we assume that $g=G|M$ is an immersion. Then
$$\rank\left [
\begin{array}{ccc}
\frac{\partial f}{\partial x_1}(x)&\ldots &\frac{\partial f}{\partial x_{m+1}}(x)\\
\frac{\partial g_1}{\partial x_1}(x)&\ldots &\frac{\partial g_1}{\partial x_{m+1}}(x)\\
&\ldots&\\
\frac{\partial g_{2m}}{\partial x_1}(x)&\ldots &\frac{\partial g_{2m}}{\partial x_{m+1}}(x)
\end{array} \right ]=m+1,$$
for $x\in M$. Denote by $\alpha_1(x), \ldots ,\alpha_{m+1}(x)$ the columns of the matrix above.
This way with the immersion $g$ we can associate
$\alpha=(\alpha_1,\ldots ,\alpha_{m+1}):M\longrightarrow \widetilde{V}_{m+1}(\R^{2m+1})$, and
$\widetilde{\alpha}=\beta_1\alpha_1(x)+\ldots + \beta_{m+1}\alpha_{m+1}(x):\R^{m+1}\times M\longrightarrow \R^{2m+1}$
such that $\widetilde{\alpha}|S^m\times M$ goes into $\R^{2m+1}\setminus \{0\}$.
By Theorem \ref{parzystoscstopnia}, the degree $\deg(\widetilde{\alpha}|S^m\times M)$ is well defined
and even.
Let us define $\phi:\R^{m+2}\times \R^{m+1}\longrightarrow \R^{2m+2}$ by
$$\phi(\beta,\beta_{m+2};x)=(x+\beta,x+\beta_{m+2}\nabla f(x)),$$
where $\beta=(\beta_1,\ldots ,\beta_{m+1})$, $x=(x_1,\ldots ,x_{m+1})$.
\begin{lemma}
For any $r>0$ small enough, $\phi:B^{m+2}(r)\times M\rightarrow\R^{2m+2}$ is an orientation preserving diffeomorphism
onto its image.
\end{lemma}
\begin{proof} Take a well oriented basis $v_1,\ldots , v_m$ of
$T_xM$, so that $\nabla f(x),v_1,\ldots ,v_m$ are well oriented in $\R^{m+1}$.
Let $e_1,\ldots ,e_{m+2}$ be the standard basis of $\R^{m+2}$. Take
$q=(\beta,\beta_{m+2};x)\in B^{m+2}(r)\times M$.
Then
$$(e_1,0),\ldots ,(e_{m+2},0),(0,v_1),\ldots ,(0,v_m)$$
is a well oriented basis in $T_q (B^{m+2}(r)\times M)$.
If $\beta_{m+2}=0$ then
$$\det\big(\left[ D\phi((\beta,0;x))\right]\left[(e_1,0),\ldots ,(e_{m+2},0),(0,v_1),\ldots ,(0,v_m)\right]\Big)$$
$$=\det\left[\begin{array}{ccccccc}
1&\ldots&0&0&1&\ldots &0\\
&\ddots&&0&&\ddots&\\
0&\ldots&1&0&0&\ldots&1\\
0&\ldots&0&\frac{\partial f}{\partial x_1}(x)&1&\ldots &0\\
&\ddots&&\vdots&&\ddots&\\
0&\ldots&0&\frac{\partial f}{\partial x_{m+1}}(x)&0&\ldots&1\\
\end{array}\right]\left[(e_1,0),\ldots ,(e_{m+2},0),(0,v_1),\ldots ,(0,v_m)\right]$$
$$=\det\left[\begin{array}{ccccccc}
1&\ldots&0&0&&&\\
&\ddots&&0&&*&\\
0&\ldots&1&0&&&\\
0&\ldots&0&\nabla f(x)&v_1&\ldots &v_m\\
\end{array}\right]=\det\left[\nabla f(x),v_1,\ldots ,v_m\right]>0.$$
Since $\phi(0;x)=(x,x)$ and $M$ is compact,
if $r>0$ is small enough then $\phi:B^{m+2}(r)\times M\rightarrow\R^{2m+2}$ is an orientation preserving
diffeomorphism onto its image.
\end{proof}
\begin{lemma}\label{suma}
There exist smooth functions $u_1,\ldots ,u_{m+1}:\R^{m+1}\times \R^{m+1}\longrightarrow\R$
such that $f(x+y)=f(x)+\sum_1^{m+1}y_i u_i(x,y)$, and $u_i(x,0)=\frac{\partial f}{\partial x_i}(x)$.
\end{lemma}\hspace*{\fill}$\Box$\par\medskip
\begin{theorem}\label{immersje}
If $m$ is even and $g:M\rightarrow\R^{2m}$ is an immersion then
$I(g)=-\deg (\widetilde{\alpha}|S^m\times M)/2=-\Lambda(\alpha)$.
\end{theorem}
\begin{proof}
Let $\Delta=\{(x,x)\ |\ x\in M\}$ denote the diagonal in $M\times M$.
Note that
$$H^{-1}(0)=\Delta\cup\{(x,y)\in M\times M,\ g(x)=g(y),\ x\neq y\}.$$
By Proposition \ref{nasze}, there is
$\varepsilon>0$ such that $|x-y|>\varepsilon$ for $(x,y)\in H^{-1}(0)\setminus
\Delta$.
Take $r>0$ such that $\phi:B^{m+2}(2r)\times M\rightarrow\R^{2m+2}$ is an orientation preserving diffeomorphism
onto its image.
Put $K=\phi(\bar{B}^{m+2}(r)\times M)$. Then $K$ is a closed tubular neighbourhood of $\phi(\{0\}\times M)=\Delta$ in $\R^{2m+2}$,
and so $K$ is a $(2m+2)$--dimensional compact manifold with boundary
$\partial K=\phi(S^{m+1}(r)\times M)$. Moreover we can assume that
$$|(x+\beta)-(x+\beta_{m+2}\nabla f(x))|<\varepsilon,\ \mbox{ for }(\beta,\beta_{m+2};x) \in \bar{B}^{m+2}(r)\times M.$$
In particular $\partial K\cap H\inv (0)=\emptyset$ and $K\cap H^{-1}(0)=\Delta$.
For $R>0$ big enough, $N=\bar{B}^{2m+2}(R)\setminus \phi(B^{m+2}(r)\times M)$ is a compact manifold
with boundary $S^{2m+1}(R)\cup \partial K$, where the orientation of $\partial K$ is
opposite to the one induced from $K$. We may also assume that
$N$ contains
$H\inv(0)\setminus \Delta$ in its interior. According to Proposition \ref{nasze},
$$2I(g)=\deg(H,N,0)=\deg (H|S^{2m+1}(R))-\deg (H|\partial K).$$
The hypersurface $M=f\inv(0)$ is compact, so $\{f(x)=f(y)=0\}=M\times M\subset\R^{m+1}\times\R^{m+1}$
is compact too. According to Lemma \ref{stopienNaSferze}, $\deg (H|S^{2m+1}(R))=0$. So
$$2I(g)=-\deg (H|\partial K)=-\deg (H\circ \phi|S^{m+1}(r)\times M)=-\deg(H\circ \phi, B^{m+2}(r)\times M,0).$$
Of course it holds true for any radius smaller than $r$.
According to Lemma \ref{suma}, for $(\beta, \beta_{m+2};x)\in B^{m+2}(r)\times M$
the second coordinate of $H\circ\phi$ equals
$$f(x+\beta_{m+2}\nabla f(x))=f(x)+\beta_{m+2}\sum_{i=1}^{m+1}\frac{\partial f}{\partial x_i}(x)u_i(\beta, \beta_{m+2};x)$$
$$=\beta_{m+2}\sum_{i=1}^{m+1}\frac{\partial f}{\partial x_i}(x)u_i(\beta, \beta_{m+2};x),$$
where $u_i(0;x)=\frac{\partial f}{\partial x_i}(x)$. For $r$ small enough
$\sum_{i=1}^{m+1}\frac{\partial f}{\partial x_i}(x)u_i(\beta, \beta_{m+2};x)>0$.
After permuting coordinates, by Lemma \ref{stopien}, we get
$$\deg(H\circ \phi, B^{m+2}(r)\times M,0)$$
$$=\deg\left((f(x+\beta),g_1(x+\beta)-g_1(x),\ldots,g_{2m}(x+\beta)-g_{2m}(x)),B^{m+1}(r)\times M,0\right).$$
By Lemma \ref{suma}, for $(\beta,x)\in B^{m+1}(r)\times M$ we have
$$\left(f(x+\beta),g_1(x+\beta)-g_1(x),\ldots,g_{2m}(x+\beta)-g_{2m}(x)\right)$$
$$=\left(\sum_1^{m+1}\beta_i u_i(\beta,x),\sum_1^{m+1} \beta_i w_i(\beta,x)\right)=\sum_1^{m+1}\beta_i\left(u_i(\beta,x),w_i(\beta,x)\right),$$
where $w_i(0,x)=(w_i^1(0,x),\ldots ,w_i^{2m}(0,x))=$
$(\frac{\partial g_1}{\partial x_i}(x),\ldots ,\frac{\partial g_{2m}}{\partial x_i}(x)).$
Because $g$ is an immersion, by Proposition \ref{nasze} there is small $r$ such that
$$\rank\left [
\begin{array}{ccc}
u_1(\beta,x)&\ldots &u_{m+1}(\beta,x)\\
w_1^1(\beta,x)&\ldots &w_{m+1}^1(\beta,x)\\
&\ldots & \\
w_1^{2m}(\beta,x)&\ldots &w_{m+1}^{2m}(\beta,x)\\
\end{array} \right ]=m+1,$$
for $(\beta, x)\in \bar{B}^{m+1}(r)\times M$.
Hence the columns $\alpha_1(\beta,x),\ldots,\alpha_{m+1}(\beta,x)$ of the matrix above are linearly independent.
Let
$h_t:\bar{B}^{m+1}(r)\times M\longrightarrow\R^{2m+1}$, $0\leq t\leq 1$, be a homotopy given by
$$h_t(\beta,x)=\beta_1\alpha_1(t\beta,x)+\ldots +\beta_{m+1}\alpha_{m+1}(t\beta,x),$$
so that $h_t(\beta,x)$ is a linear combination of linearly independent vectors.
Then each $h_t^{-1}(0)=\{0\}\times M$.
According to the Excision Theorem we have
$$\deg (\widetilde{\alpha}|S^m\times M)=\deg(\widetilde{\alpha},\bar{B}^{m+1}(1)\times M,0)=\deg(\widetilde{\alpha},\bar{B}^{m+1}(r)\times M,0).$$
Of course $\widetilde{\alpha}=h_0$, so $\deg(\widetilde{\alpha}|S^m\times M)=\deg(h_0,\bar{B}^{m+1}(r)\times M,0)=
\deg(h_1,\bar{B}^{m+1}(r)\times M,0)$.
By the previous arguments, $\deg (h_1,\bar{B}^{m+1}(r)\times M,0)=-2I(g)$.
To sum up we get that $2I(g)=-\deg (\widetilde{\alpha}|S^m\times M)=-2\Lambda(\alpha)$.
\end{proof}
\begin{example} Let $g=(x_3^3+x_2-x_1-3x_3,x_2^3+2x_1-x_2+x_3,x_1x_2+2x_1,x_1x_3-x_2):\R^3\longrightarrow\R^4$.
Using {\sc Singular} \cite{GPS06} and results of Theorems \ref{efeektywnie} and \ref{immersje} one may check that $I(g|S^2(10))=5$.
\end{example}
|
\section{Acknowledgement}
The authors thank Dieter Suter for helpful discussions. This work was supported by National Nature Science Foundation of China (Grants Nos. 10834005, 91021005, and 21073171), the CAS, and the National Fundamental Research Program 2007CB925200.
|
\section{introduction}
In most of the problems of quantum mechanics the Hamiltonian of the
system is time dependent and so one needs to solve the time
dependent Schr\"{o}dinger equation (TDSE). Problems with moving
boundary conditions are an interesting class of such time dependent
problem. Such a system was first considered by Fermi
\cite{Fe-PR-1949} in connection with the study of cosmic radiation.
After, several authors studied problems with moving boundaries
\cite{DoRi-AmJPhys-1969, DoKlNi-JMP-1993, SchMuRu-PRA-2009},
\cite{JaRo-PLA-2008} and references therein.
Different aspects of the problem of a particle in a one-dimensional
infinite square-well potential with one wall in uniform motion have
been discussed by earlier authors. Exact solution of the time
dependent Schr\"odinger equation for this problem at first, as far
as we know, was given by Doescher and Rice \cite{DoRi-AmJPhys-1969}.
Schlitt and Stutz \cite{SchSt-AmJPhys-1970} considered the
application of the sudden approximation to the rapid expansion of
the well. Pinder \cite{Pi-AmJPhys-1990} investigated the
applicability of both adiabatic and sudden approximation for both
expanding and contracting wells: it was shown that sudden
approximation is appropriate to the expanding well provided the wall
speed is sufficiently great, but this approximation may not be
applied for the contracting well irrespective of the rate of
contraction. Using the semiclassical approximation, Luz and Cheng
\cite{LuCh-JPA-1992} evaluated the exact propagator of the problem.
The energy gain and the transition amplitudes and probabilities
between initial and final energy eigenstates of the problem have
been calculated \cite{DoKlNi-JMP-1993}. A recent numerical study of
a particle in a box with different laws for the movement of the wall
show that physical quantities like probability density and
expectation value of position or mean value of the energy have a
smooth behavior for small speed of the moving wall. In contrast, if
this speed becomes large, many irregularities appears as sharp bumps
on the probability distribution or a chaotic shape on the averaged
values of position and energy \cite{FoGaLa-CMA-2010}.
The aim of the present paper is to probe some aspects of the time dependent boundary condition for a particle confined in an infinite square well that have remained hitherto unnoticed.
Let us focus on the effect of the time dependent boundary condition
while one of the infinite boundary walls in a box is moved where we
have a well-localized Gaussian wave packet which remains peaked at
the center of the box, $x_c$, well away from walls. Now, by
calculating the effective quantum force \cite{DoAn-PLA-2000-2002}
one can study the way this effective quantum force changes with
time. Due to such a force, the expectation value of the momentum in
the direction perpendicular to the walls gradually changes in time.
Then one can compare curves of the quantum effective force for the
static (when the wall is at rest) and the dynamic (when the wall
moves) situations, and pinpoint the instant from which the dynamic
curve deviates from the static one. Such an instant shows the time
at which the confined particle begins to feel the motion of the
wall.
Physically one expects that when the width of the initial Gaussian packet is much smaller than the initial width of the box, $\sigma_0 \ll \ell_0$, one might consider a Gaussian wavepacket to be realizable in the box trap. But, this approximation casts some doubts upon the computations (i.e. is the effect an artifact of the tails of the Gaussian at the boundaries?). So for this purpose, in addition, we consider a localized state
with a finite support embedded in the support of the expanding box trap at time $t=0$. It is most natural to consider the particle-in-a-box eigenstates of a tiny box centered at $x=x_c$, and suddenly released at $t=0$ to become the initial state without the necessary approximations to assume a Gaussian packet as an initial state. Such state is called ``tiny-box state" afterwards.
The computed Bohmian trajectories \cite{Bohmian-Mechanics, DuGoZa-JSP-1992, Holland-book-1993} for the static and the dynamic situations are also instructive in revealing the conceptual ramifications of such an example. In Bohm's model each individual particle is assumed to have a definite position, irrespective of any measurement. The pre-measured value of position is revealed when an actual measurement is done. Over an ensemble of particles having the same wave function $\psi$, these ontological positions are distributed according to the probability density $\rho=|\psi|^2$ where the wave function $\psi$ evolves with time according to the Schr\"odinger equation and the equation of motion of any individual particle is determined by the guidance equation $v = j/\rho$, where $v$ is the Bohmian velocity of the particle and $j$ is the probability current density. Solving the guidance equation one gets the trajectory of the particle.
The plan of this paper is as follows. Section \ref{Sec: BaEq} contains a very brief review of the relevant mathematical steps leading to the exact solution for the problem. In Section \ref{Sec: NuCa} numerical computations related to the effect of the time dependent boundary condition are presented. Finally, in Section \ref{Sec: SuDi} we present the concluding remarks.
\section{Basic Equations} \label{Sec: BaEq}
Consider a narrow box inside a wide box {\it with a particle inside the inner one}. Walls of the outer box are at $x_L=0$ and $x_R=\ell_0$ and walls of the inner one are at $x_1=(\ell_0-\ell_1)/2$ and $x_2=(\ell_0+\ell_1)/2$ where $\ell_1 \ll \ell_0$.
At time $t=0$ the inner box is suddenly removed and the right wall of the outer box starts to move uniformly with velocity $u$.
Infinite wall speed $u$ corresponds to a hard wall at $x=0$.
We discuss solution of TDSE for two cases. Initial wavefunction to be:\\
\begin{enumerate}[i)]
\item a Gaussian wave packet well localised in the center of the tiny box.
In fact in this case we have a truncated Gaussian packet, because of confinement of the Gaussian packet with {\it infinite} tails in a narrow region, and so the name ``truncated Gaussian packet"(TGP)). The problem concerning the tails of Gaussian packet that was mentioned in the introduction, now is translated to the truncation.\\
\item the ground state of the narrow box with kick momentum $k$ (tiny-box state, ``TBS" for abbreviation).\\
To get a picture see fig. (1).
\end{enumerate}
Using the propagator of a rigid box with the left wall at $x_L=0$ and the moving right wall in a constant velocity $u$ \cite{LuCh-JPA-1992},
\begin{eqnarray} \label{eq: propagator}
K(x, t; x^{\prime}, 0) &=& \frac{2}{\sqrt{\ell_0 \ell(t)}} e^{\frac{imu}{2\hbar}(\frac{x^2}{\ell(t)} - \frac{x^{\prime^2}}{\ell_0})}
\sum_{n=1}^{\infty} e^{\frac{in^2\pi^2 \hbar}{2mu}(\frac{1}{\ell(t)}-\frac{1}{\ell_0})} \sin{(\frac{n\pi x}{\ell(t)})}
\sin{(\frac{n\pi x^{\prime}}{\ell_0})}~,
\end{eqnarray}
and the relation,
\begin{eqnarray} \label{eq: psi_t}
\psi(x, t) &=& \int dx^{\prime} K(x, t; x^{\prime}, 0) \psi_0(x^{\prime})~,
\end{eqnarray}
one gets the wavefunction at any time having $\psi_0(x)$ in hand. $\ell(t) = \ell_0 + ut$ shows the position of the moving wall at time $t$.
At this stage, it must be mentioned that due to the Galilean invariance of the Schr\"{o}dinger equation \cite{Holland-book-1993}, the case of both moving walls is equivalent to the case of one wall in motion but with $u$ as the relative velocity of walls.
With the initial wavefunction to be a TGP well localised in the center of the box $x_c = \ell_0/2$,
\begin{eqnarray} \label{eq: ini-gauss}
\psi_0(x) &=& \frac{1}{(2\pi \sigma_0^2)^{1/4}} \exp{ \left[
ik(x-x_c)-\frac{(x-x_c)^2}{4\sigma_0^2} \right]} \Theta(x -
x_1)~\Theta(x_2 - x)~,
\end{eqnarray}
where $\Theta(x)$ is the step function; one gets,
\begin{eqnarray}
\psi(x, t) &=& \frac{2}{\sqrt{\ell_0 \ell(t)}} e^{i\frac{mux^2}{2\hbar(\ell(t))}} \sum_{n=1}^{\infty}
e^{-\frac{in^2\pi^2 \hbar t}{2m\ell_0 (\ell(t))}} \sin{\left( \frac{n\pi x}{\ell(t)} \right)} \times f_n~,
\end{eqnarray}
where
\begin{eqnarray}
f_n &=& \frac{i}{2} (\frac{\pi}{2})^{1/4} \sqrt{\frac{\sigma_0
\ell_0 \hbar}{\hbar \ell_0 + 2imu\sigma_0^2}}~ \exp{ \left[
-\frac{imu\ell_0^3 + 8n^2 \pi^2 \sigma_0^2 \hbar + 16n\pi \ell_0
\sigma_0^2 \hbar k + \ell_0^2(4in\pi\hbar + 8\sigma_0^2 k(\hbar k -
mu)) }{8\ell_0(\hbar \ell_0 + 2imu\sigma_0^2)} \right]}
\nonumber\\
&\times &
\bigg[
-e^{\frac{in\pi \ell_0 \hbar}{\hbar \ell_0 + 2imu\sigma_0^2}} \text{Erf}{\left( \frac{-2i(2n\pi \hbar - mu\ell_1) \sigma_0^2 + \ell_0[\hbar \ell_1 - 2i(2\hbar k-mu)\sigma_0^2]}{4 \sigma_0 \sqrt{\hbar \ell_0 (\hbar \ell_0 + 2imu\sigma_0^2)}}\right)}
\nonumber\\
&~~~~~~~~&~
+e^{\frac{4n\pi \hbar k \sigma_0^2}{\hbar \ell_0 + 2imu\sigma_0^2}} \text{Erf}{\left( \frac{2i(2n\pi \hbar + mu\ell_1) \sigma_0^2 + \ell_0[\hbar \ell_1 - 2i(2\hbar k-mu)\sigma_0^2]}{4\sigma_0 \sqrt{\hbar \ell_0 (\hbar \ell_0 + 2imu\sigma_0^2)}}\right)}
\nonumber\\
&~~~~~~~~&~
+e^{\frac{4n\pi \hbar k \sigma_0^2}{\hbar \ell_0 + 2imu\sigma_0^2}} \text{Erf}{\left( \frac{-2i(2n\pi \hbar - mu\ell_1) \sigma_0^2 + \ell_0[\hbar \ell_1 + 2i(2\hbar k-mu)\sigma_0^2]}{4\sigma_0 \sqrt{\hbar \ell_0 (\hbar \ell_0 + 2imu\sigma_0^2)}}\right)}
\nonumber\\
&~~~~~~~~&~
-e^{\frac{in\pi \ell_0 \hbar}{\hbar \ell_0 + 2imu\sigma_0^2}} \text{Erf}{\left( \frac{2i(2n\pi \hbar + mu\ell_1) \sigma_0^2 + \ell_0[\hbar \ell_1 + 2i(2\hbar k-mu)\sigma_0^2]}{4\sigma_0 \sqrt{\hbar \ell_0 (\hbar \ell_0 + 2imu\sigma_0^2)}}\right)}
\bigg]
\end{eqnarray}
and $\text{Erf}$ is the error function: $\text{Erf}(z) = \frac{2}{\sqrt{\pi}} \int_0^{z} e^{-t^2} dt$.
In the second case we take the initial wave function as,
\begin{eqnarray} \label{eq: ini-boxstate}
\psi_0(x) &=& \sqrt{\frac{2}{\ell_1}}
\sin[{\frac{\pi}{\ell_1}(x-x_1)}]~e^{ik(x-x_c)}~ \Theta(x -
x_1)~\Theta(x_2 - x) ~.
\end{eqnarray}
In this case the relation of $\psi(x, t)$ is cumbersome; there are eight modified error functions, $\text{Erfi}(z) = \frac{2}{\sqrt{\pi}} \int_0^{z} e^{t^2} dt$ in its summand.
In original Bohm approach to causal interpretation of quantum mechanics \cite{Bohmian-Mechanics, Holland-book-1993} to introduce the concept of particle, Schr\"{o}dinger equation is decomposed into two real equations by expressing the wavefunction in polar form $\psi = R e^{iS/\hbar}$. Then, vector filed ${\bf v} = {\bf p}/m$ is constructed from the vector filed ${\bf p} = \nabla S$ and assuming that ${\bf v}$ defines at each space-time point the tangent to a possible particle trajectory passing through that point. In this interpretation of quantum mechanics one gets,
\begin{eqnarray} \label{eq: Bmeq}
\frac{d{\bf p}}{dt} &=& -\nabla(V+Q)~,
\end{eqnarray}
where $Q = -(\hbar^2/2m) \nabla^2 R/R$ is known as quantum potential. Analogous to classical physics, in Bohm's model of quantum theory one has,
\begin{eqnarray} \label{eq: Bmeq2}
\frac{d \langle {\bf p} \rangle}{dt} &=& \langle \frac{d{\bf p}}{dt} \rangle.
\end{eqnarray}
where the mean value is defined for an ensemble of density $R^2$ and momentum ${\bf p} = \nabla S$. In the standard approach to quantum mechanics the right hand side of eq. (\ref{eq: Bmeq2}) is meaningless (ref. \cite{Holland-book-1993} pp: 111-113). Using eq. (\ref{eq: Bmeq2}) and taking the expectation value of eq. (\ref{eq: Bmeq}) one obtains,
\begin{eqnarray}
\frac{d \langle {\bf p} \rangle}{dt} &=& -\langle \nabla (V+Q) \rangle.
\end{eqnarray}
For the case of a particle within a box with one wall moving, using the integration by part one can find
\begin{eqnarray*}
-\frac{2m}{\hbar^2}\langle \nabla Q \rangle &=& \int_0^{\ell(t)} dx~R^2 \frac{\partial}{\partial x} \frac{1}{R} \frac{\partial^2 R}{\partial x^2} \\
&=&
-2 \int_0^{\ell(t)} dx~ \frac{\partial R}{\partial x} \frac{\partial^2 R}{\partial x^2}
= - 2 \int_0^{\ell(t)} dx~ \left[ \frac{\partial}{\partial x} \left( \frac{\partial R}{\partial x} \right)^2
- \frac{\partial^2 R}{\partial x^2} \frac{\partial R}{\partial x}\right] \\
&=&
+2 \int_0^{\ell(t)} dx~ \frac{\partial^2 R}{\partial x^2} \frac{\partial R}{\partial x} - 2 \left( \frac{\partial R}{\partial x} \right)^2 \bigg|_0^{\ell(t)}
= - \left( \frac{\partial R}{\partial x} \right)^2 \bigg|_0^{\ell(t)}~,
\end{eqnarray*}
where we have used the fact that wavefunction is zero on both walls.
The general case of boundary condition will be considered in the
appendix. One obtains,
\begin{eqnarray*}
\bigg| \frac{\partial \psi}{\partial x} \bigg| &=& \bigg| \frac{\partial R}{\partial x} + \frac{i}{\hbar} R \frac{\partial S}{\partial x} \bigg| = \bigg| \frac{\partial R}{\partial x} \bigg|~.
\end{eqnarray*}
where the second equality holds at the boundaries only. In
consequence we have,
\begin{eqnarray} \label{eq: qm_force}
\frac{d \langle p \rangle}{dt} &=& -\frac{\hbar^2}{2m} \left[ \bigg| \frac{\partial \psi}{\partial x}(x=\ell(t), t) \bigg| ^2 -
\bigg| \frac{\partial \psi}{\partial x}(x=0, t) \bigg| ^2 \right] \nonumber\\
&\equiv & f_{\text{qm}}(t)~,
\end{eqnarray}
where $f_{\text{qm}}(t)$ was called quantum effective force by Dodonov and Andreata \cite{DoAn-PLA-2000-2002}.
It must be mentioned that in the context of standard approach to quantum mechanics one can obtain eq. (\ref{eq: qm_force}) by simultaneous application of the Schr\"{o}dinger equation,
\begin{eqnarray} \label{eq: Sch}
-\frac{\hbar^2}{2m} \frac{\partial^2 \psi}{\partial x^2} &=& i\hbar\frac{\partial \psi}{\partial t}~,
\end{eqnarray}
and time-derivative of the expectation value of momentum operator,
\begin{eqnarray} \label{eq: t_der p}
\frac{d \langle p \rangle}{dt} &=& \frac{d}{dt} \int_0^{\ell(t)} \psi^*(x, t) \frac{\hbar}{i} \frac{\partial}{\partial x} \psi(x, t)~,
\end{eqnarray}
as it was done at first by Dodonov and Andreata \cite{DoAn-PLA-2000-2002} for the case of an impenetrable wall at $x=0$.
Now, taking the integral of both sides of eq. (\ref{eq: qm_force}) leads to
\begin{eqnarray} \label{eq: momentum-expectation}
\langle p \rangle(t) &=& \langle p \rangle(0) + \int_0^{t} f_{\text{qm}}(t) dt~.
\end{eqnarray}
where in the case of TGP $\langle p \rangle(0) = \hbar k~\text{Erf}[\frac{\ell_1}{2\sqrt{2}\sigma_0}]$ whereas for the case of TBS $\langle p \rangle(0) = \hbar k$.
\section{Numerical Calculations} \label{Sec: NuCa}
In this section we work in a unit system where $\hbar=1$ and
$m=0.5$. Other parameters are chosen as $\ell_1 = \ell_0/20$,
$\sigma_0 = \ell_1/10$ and $\ell_0 = 1$. Conservation of the
probability, given by $\int_0^{\ell(t)} \psi^* \psi~dx = 1$, can be
used as a parameter that gives us a test on the precision of the
results (It must be noted in the case of TGP because of truncation
total probability is not equal to unity, instead it is equal to
$\text{Erf}[\frac{\ell_1}{2\sqrt{2}\sigma_0}] = 0.9999994267$). We
use Simpson's rule for taking the integral of eq. (\ref{eq: psi_t}).
Using the conservation of the probability as control variable, we
got a good numerical stability by just taking the first $400$ terms
of infinite sum appearing in the relation of $\psi(x, t)$ in both
TGP and TBS cases. Using the Runge–Kutta method for solving the
guidance differential equation a selection of Bohmian trajectories
is presented.
\begin{figure}
\centering
\includegraphics[width=12cm,angle=-90]{Fig1.eps}
\caption{(Color online) Initial probability density. Vertical blue
dashed lines show the walls of the narrow box and the vertical solid
blue lines stand for the walls of the wide box.} \vspace*{0.5cm}
\label{fig: initialwave}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=12cm,angle=0]{Fig2.eps}
\caption{(Color online) A selection of Bohm trajectories, $x(t)$,
for $u = 100\pi$ and an initially motionless wavefunction: a) TBS,
static case; b) TBS, dynamic case; c) TGP, static case and d) TGP,
dynamic case. In each figure black curve starts at $x_0 = x_c -
2\sigma_0$, red one at $x_0 = x_c$ and the green one at $x_0 = x_c +
2\sigma_0$.} \vspace*{0.5cm} \label{fig: Bo-path}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=12cm,angle=-90]{Fig3.eps}
\caption{(Color online) Expectation value of position operator
versus time, $\langle x \rangle(t)$, for $u = 100\pi$ and an
initially motionless wavefunction: a) TBS and b) TGP. In each figure
the black curve shows $\langle x \rangle(t)$ for the dynamic case,
the green curve shows $\langle x \rangle(t)$ for the static case and
the red one shows the Bohm trajectory for the dynamic case which
starts at $x_0 = \langle x \rangle(0)$.} \vspace*{0.5cm} \label{fig:
x_expectation}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=12cm,angle=-90]{Fig4.eps}
\caption{(Color online) Quantum effective force versus time,
$f_{\text{qm}}(t)$, for an initially motionless wavefunction: a) TBS
and b) TGP. In each figure the black curve shows $f_{\text{qm}}(t)$
for the dynamic case and the red one shows $f_{\text{qm}}(t)$ for
the static case.} \vspace*{0.5cm} \label{fig: force-longtime}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=12cm,angle=-90]{Fig5.eps}
\caption{(Color online) Quantum effective force versus time,
$f_{\text{qm}}(t)$, for the tiny-box state and $u = 100\pi$: a) $k =
-75\pi$, b) $k = -50\pi$, c) $k = -25\pi$, d) $k = 25\pi$, e) $k =
50\pi$, f) $k = 75\pi$. In each figure the black curve shows
$f_{\text{qm}}(t)$ for the static case and the red one shows
$f_{\text{qm}}(t)$ for the dynamic case.} \vspace*{0.5cm}
\label{fig: box-force(k)}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=12cm,angle=-90]{Fig6.eps}
\caption{(Color online) Quantum effective force versus time,
$f_{\text{qm}}(t)$, for the truncated Gaussian packet and $u =
100\pi$: a) $k = -75\pi$, b) $k = -50\pi$, c) $k = -25\pi$, d) $k =
25\pi$, e) $k = 50\pi$, f) $k = 75\pi$. In each figure the dotted
black curve shows $f_{\text{qm}}(t)$ for the static case and the
red one shows $f_{\text{qm}}(t)$ for the dynamic case.}
\vspace*{0.5cm} \label{fig: gauss-force(k)}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=12cm,angle=0]{Fig7.eps}
\caption{(Color online) Quantum effective force versus time,
$f_{\text{qm}}(t)$, for a motionless a) TBS and b) TGP in dynamic
case (force is zero in static case for $k = 0$). In each figure
black curve is for $u = 20\pi$, red one is for $u = 100\pi$ and the
green one is for $u = 200\pi$.} \vspace*{0.5cm} \label{fig:
force(k=0)}
\end{figure}
Fig. \ref{fig: initialwave} show the initial wave function for both TGP and TBS and the walls of the well to give a feeling about how narrowly the initial wavefunction is.
In fig. \ref{fig: Bo-path} a selection of Bohm paths is presented for a initially motionless wavefunction, {\it i.e.,} $k = 0$ in eqs. (\ref{eq: ini-gauss}) and (\ref{eq: ini-boxstate}). From parts a) and c) one can see that Bohm particle in the static case remains at rest at the center of the box. We have checked long time behaviour of Bohm trajectories that starts at the tail of the leading half in the dynamic situation and saw that for TGP, particle eventually moves on a path approximately parallel (approximately, because it has very small oscillation around the parallel path) to the path of the moving wall, {\it i.e.,} with the velocity of wall, but for TBS it moves with a velocity less than the velocity of moving wall.
Fig. \ref{fig: x_expectation} show the expectation value of position operator for a initially motionless wavefunction.
From this figure and fig. \ref{fig: Bo-path} one finds that in the static case, {\it i.e.,} fixed wall, the Bohm path which initially placed at the centre of the motionless packet, $x_0 = \langle x \rangle(0)$ (in the case of TBS $x_0 = 0.5$ whereas $x_0 = (1/2) \text{Erf}[5/\sqrt{2}] = 0.4999997$ in the case of truncated Gaussian), moves with the centre point subsequently, $x(t) = \langle x \rangle(t)$, at least up to time $0.003$ that we have considered. But, this is not true in the dynamic case. The reason is the quantum effective force which has been shown in fig. \ref{fig: force-longtime}: In static case $f_{\text{qm}}(t)$ is zero for both tiny-box and truncated Gaussian states.
Deviation of $\langle x \rangle(t)$ from its static value $\langle x \rangle(0)$ take place sooner for the TBS. This shows that the particle begins to feel the motion of the wall sooner for the case of TBS compared to the case of TGP.
In figures \ref{fig: box-force(k)} and \ref{fig: gauss-force(k)}, we have plotted time-dependence of quantum effective force for a fixed value of the speed of the moving wall, $u = 100\pi$, but different values of kick momentum $k$. Noting these figures one finds that for a moving packet, {\it i.e.,} $k \neq 0$, quantum effective force is not zero in the static case contrary to the case of a motionless one. Comparison of these figures show that particle begins to feel the motion of the wall approximately twice sooner in the case of TBS in comparison to truncated Gaussian packet.
Fig. \ref{fig: gauss-force(k)} reveals that in the case of truncated Gaussian packet, quantum effective force is the same for both static and dynamic cases for $k < 0$ and it is zero for $k > 0$ in the dynamic case, at least for our parameters and time-domain $t \in [0, 0.0005]$. $f_{\text{qm}}(t)$
deviates from zero in positive direction for $k < 0$ but in negative direction for $k > 0$, {\it i.e.,} particle accelerates for $k<0$ but decelerates for $k>0$.
Noting fig. \ref{fig: force(k=0)} which displays $f_{\text{qm}}(t)$ for the motionless TBS and TGP but for different wall's speed, one finds:
1) in the presented region of time, quantum effective force for TGP is negligible compared to the TBS (one must note that in longer time limit opposite behaviour take places according to fig. \ref{fig: force-longtime}),
2) direction of deviation from zero changes with $u$ and
3) deviation time increases with $u$.
\section{Summary and Discussion} \label{Sec: SuDi}
In this paper we studied the solution of TDSE for a particle in a)
tiny-box state and b) truncated Gaussian packet of an infinite
square well with one wall in uniform motion. We showed that due to a
quantum effective force, which apart from a minus sign is the
expectation value of the gradient of the {\it quantum potential} in
the context of Bohmian mechanics, the expectation value of the
momentum operator changes gradually with time. We studied the
variation of this quantum effective force with time for different
values of the speed of the moving wall in the case of a motionless
packet and different values of the kick momentum but fix value of
the speed of the moving wall. Some Bohm trajectories for the
motionless packet were also plotted. We have learned from the
numerical calculations that the particle in TBS begins to feel the
motion of the wall sooner in comparison to TGP. This may be
understood by computing the speed of propagation
\cite{Holland-book-1993} for both TBS and TGP. Other ramifications
of this study, like a contracting box, dependence of quantum
effective force on related parameters like mass of the confined
particle, width of the initial packet, other initial packets like
excited particle-in-a-box eigenstates and other types of boundary
conditions, like periodic ones, call for further consideration.
\\
\\
{\bf Acknowledgments} I would like to thank Archan Majumdar for
valuable suggestions and M. R. Mozaffari for providing useful
information about numerical integration. Financial support of the
University of Qom is acknowledged.
|
\section{Fluorescence properties of the CdSe/CdS NCs deposited on a glass coverslip}
The very thick shell NCs used in the experiments were synthesized and characterized in detail by the authors. Their structural and optical properties are very well controlled. At the pump rate used in the experiments, single photon emission is always observed when NCs are deposited on a glass coverslip (see Fig. \ref{fig1_SM}).
\begin{figure}[ht]
\includegraphics[width=15cm]{fig1_SM.eps}
\caption{Coincidence counts for a typical NC deposited on a glass coverslip.} \label{fig1_SM}
\end{figure}
Their fluorescence alternate between two states. The bright state corresponds to a neutral NC (Fig. \ref{fig2_SM}). In this case, the monoexcitonic state always recombines radiatively . As for standard CdSe/ZnS NCs \cite{Brokmann04b}, the QE efficiency of the bright state of CdSe/CdS NCs has been measured very close to unity \cite{Spinicelli09}. Its radiative lifetime $\tau_{ref}$ is around 75 ns (Fig. 1c of the article).
\begin{figure}[ht]
\includegraphics[width=10cm]{fig2_SM.eps}
\caption{Schematic representation of the bright state} \label{fig2_SM}
\end{figure}
The grey state was attributed to the ionized state of the NC. In quasi-type II CdSe/CdS NCs, the hole is confined in the core. In contrast, the electron is delocalized in the whole structure due to relative positions of CdSe and CdS conduction bands. For the trion state, two recombination processes have to be considered: an electron hole-pair can recombine radiatively (rate $k_{rad}$) or through an Auger process (rate $k_{A}$), the energy being then transfered to the remaining charge (Fig. \ref{fig3_SM}). Indeed ionization of the NC does not modify its close environment and does not create new non radiative pathways except the Auger one. The total desexcitation rate $k$ is the sum $k_{rad}+k_{A}$ and corresponds to the short component of the total PL decay (12 ns).
\begin{figure}[ht]
\includegraphics[width=10cm]{fig3_SM.eps}
\caption{Schematic representation of the grey state} \label{fig3_SM}
\end{figure}
Following the method of \cite{Spinicelli09}, from the histogram of the intensity (Fig 1.b of the manuscript), the QE of the trion state $Q=k_{rad}/(k_{rad}+k_{A})$ can be deduced. It is equal to the ratio between the intensities of the grey and bright states which are given by the relative positions of bumps A (28 counts / 10 ms) and B (68 counts / 10 ms) in Fig 1.b. Since the QE of the bright state is close to 100 \%, the QE of the grey state is about 40 \%.
From $Q=k_{rad}/(k_{rad}+k_{A})=40$ \% and $k=k_{rad}+k_{A}= 1/12$ ns$^{-1}$, a straightforward calculation gives $k_{rad}=1/30$ ns$^{-1}$ and $k_A=1/20$ ns$^{-1}$.
\section{Properties of the random gold film}
Au films characteristics depend strongly on the experimental conditions and the substrate. However, the metallic structure used in the experiments is a random gold film at the percolation threshold deposited on a glass coverslip, It is realized by the authors and studied in detail by several groups including ours (\cite{Ducourtieux01,Stockman01,Krachmalnicoff10,Buil06}). It is prepared by evaporation under vacuum (10$^{-9}$ torr). The mass thickness is 61 $\AA$. A typical AFM image is shown in Fig. 2.a of the paper. Its optical properties are very well known. In particular, the disorder is at the origin of the strong enhancement of the electromagnetic field in subwavelength-sized regions. The corresponding plasmon resonances cover the visible range above 550 nm as shown by the absorption spectra of Fig. 2.b. This figure also illustrates the originality of our approach: the excitation ($\lambda$ = 485 nm) cannot be enhanced by the excitation of plasmons of the random gold film. Only the NCs emission is coupled to plasmon resonances of the gold structure. The properties of the metallic structure as well as those of the nanocrystals have been mastered in this work.
|
\section{Introduction}
What are the main physical processes that shape early-type galaxies (ETGs)?
Understanding the formation and evolution of early-type galaxies is a fundamental piece in the cosmological puzzle.
Any model that aims at providing a description of the Universe as a whole must be able to reproduce the observed characteristics of these objects.
ETGs are observed to hold tight scaling relations, such as the Fundamental Plane \citep{Dressler,DjorDavis} and the correlation between the mass of the central black hole and global galactic properties \citep{Ferrarese,Gebhardt,Marconi}.
The inner density profile of their total mass is measured to be very close to isothermal, in the so-called ``bulge-halo conspiracy'' \citep{TreuKoop,Koop2006,Koo++09,PaperX}.
Moreover, they appear to undergo a significant evolution in size, from being very compact in the early ($z\sim2$) Universe to the more diffuse objects that we observe at more recent times \citep{vanderWel2008,vanDokkum,Newman2010}.
Finally, studies of massive ETGs seem to favor a heavier stellar initial mass function \citep[IMF;][]{Grillo2009,TreuIMF,v+C10,v+C11,Auger2010L,Chiara} than for spiral galaxies \citep{B+d01,Dut++11,Suy++11}.
These characteristics should reflect a common mechanism that drives ETGs towards the tight relations observed at low redshifts ($z \lesssim 0.3$) during their formation and evolution.
Mergers with other galaxies are likely to be one of the key processes in the history of ETGs.
Mergers are believed to be at the basis of the formation \citep{Kormendy,Hernquist92,Shier}, to be involved in the black hole scaling relation \citep{Haehnelt,Peng} and to drive the size evolution \citep{Hopkins2009,Oser} of ETGs.
However, the picture is complicated by the many physical processes that are present during the evolution of galaxies, such as gas cooling, feedback from stars and AGN.
Moreover, an important fraction of the mass of ETGs is accounted for by dark matter \citep{Bertin,Franx,Gerhard,TreuKoop,Bar++11} whose nature is still unknown.
Understanding the interplay of baryonic and dark matter and how they act to produce the observed structural characteristics is essential to comprehend the evolution of ETGs, but is today a challenging task.
Theoretical models and simulations with a variety of physical ingredients have been set up to try to reproduce the observables of ETGs \citep{Gustafsson,Hopkins2010,Schaye,Duffy}.
Although simulations seem to be able to capture the general characteristics of ETGs \citep[e.g.][]{HopkinsHer,Ciotti2010}, quantitatively matching the entire set of observables proved to be difficult, often requiring an ad hoc tuning of the model parameters \citep{Nipoti,Hopkins2010}.
For example, \citet{Duffy} explored a variety of aspects of baryonic physics such as gas cooling, feedback from stars and AGN, finding that on the one hand the observed inner slopes of massive ETGs are reproduced if the feedback is weak, but on the other hand a strong feedback is needed to match the measured stellar masses.
Improving the quality of the observation of ETGs and introducing more constraints can help us to discriminate between the wealth of currently viable scenarios for their history.
Two characteristics of ETGs in particular are still not known with sufficient precision and leave room for significant improvement in their observational determination: the stellar mass and the density profile of the dark matter halo.
The stellar mass is degenerate with the stellar IMF with respect to constraints from the integrated light distribution and colors.
Breaking this degeneracy can help determining the star formation history and the content of the baryonic mass in ETGs.
The density profile of the dark matter distribution is sensitive to the physical processes that take place during the formation and evolution of ETGs.
Therefore, measuring the profile of dark matter halos in ETGs is a powerful means for testing the various theoretical models.
Part of the difficulty in comparing simulations with observations are, of course, due to the fact that dark matter, which accounts for a significant fraction of the mass of a typical galaxy, is not directly observable.
Gravitational lensing is, in this aspect, a very powerful tool, being sensitive to the gravitational pull of matter independently on its interaction with light \citep[e.g.,][and references therein]{Tre10}.
Lensing surveys indeed played a crucial role in uncovering physical
characterics of early-type galaxies, such as their average density
profile and dark matter fraction \citep{Koop2006,Koo++09,PaperX,Bar++11} or the IMF of their stellar population
\citep{Grillo2009,TreuIMF,Auger2010L}. Measurements of the inner
slope of a dark matter halo have so far been obtained for a few
cluster lenses \citep[e.g.][]{Sand08,Newman09}, for which constraints from
multiple lensed sources are available. For typical early-type galaxy
strong lenses however, there are residual degeneracies between
anistropy, stellar mass to light ratio and inner slope of the dark
matter halo and therefore the constraints are weak
\citep{K+T03,TreuKoop}. For this reason, previous studies have adopted
theoretically motivated mass density profiles for the dark matter halo
\citep{TreuIMF,Auger2010L}, rather than free power laws.
Here we present a detailed study of an early-type galaxy at redshift
$z=0.222$. The galaxy is the strong gravitational lens of the system
SDSSJ0946+1006, part of the SLACS sample \citep{Bolton2004}. This ETG
is special in that it lenses two sources at different redshifts,
creating two nearly complete Einstein rings of different radii. For
this reason, the system is also referred to as the ``Jackpot''. The
first lensed source is at redshift $z_{s1} = 0.609$, while there is no
spectroscopic measurement of the redshift of the second ring. Thanks
to the presence of the two rings, this system provides more
information than typical gravitational lenses, despite the lack of the second source redshift.
A first study of SDSSJ0946+1006 was carried out by \citet[Paper VI]{Gavazzi}.
An independent lensing analysis of this system was performed by
\citet{Vegetti}, which led to the discovery of a small satellite with no visible counterpart.
Here we include new high-quality photometry obtained
with the Hubble Space Telescope (hereafter HST) and new deep and
spatially resolved spectroscopy obtained at the Keck Telescope. The
goal of our study is to separate the contribution of dark and stellar
matter to the total mass of the lens, making as few assumptions as
possible about the density profile of the dark matter halo. This task
is achieved by combining lensing and dynamics information. Unlike
typical early-type galaxy lenses, the wealth of information provided
by this system allows us to determine both the mass of the stellar
bulge and the inner slope of the dark matter halo. Thanks to the
multi-band HST photometry we are able to obtain a photometric redshift
of the outer ring, that is necessary for improving the constraints
from the lensing data, and to infer stellar masses from stellar
population synthesis (SPS) fitting. The comparison between this
measurement of the stellar mass and the one obtained through lensing
and dynamics allows us to constrain the IMF of the stars in the lens.
This is the most robust measurement of the inner slope of the dark
matter halo and IMF of an isolated massive ETG.
The structure of this paper is the following. In Sections 2 and 3 we
describe the new photometric and spectroscopic data, respectively.
Our measurement of the photometric redshift of the outer ring is
presented in Section 4. In Section 5 we describe measurements of the
stellar mass of the lens from stellar population synthesis fitting.
Section 6 describes a lensing and dynamics model assuming a power-law
density profile for the total density profile of the lens. In Section
7 we present the bulge-halo decomposition of the lens. We discuss our
results in Section 8 and summarize in Section 9.
Throughout the paper we assume the following values for the cosmological parameters: $H_0 = 70\mbox{ km s}^{-1}\mbox{Mpc}^{-1}$, $\Omega_M = 0.3$, $\Omega_\Lambda = 0.7$. Magnitudes are expressed in the AB system, images are North-up and position angles are in degrees East of North.
In showing our results we display posterior PDFs in multiple projections wherever possible, but when
giving a point estimate of an inferred parameter we quote the position
of the peak of its one-dimensional marginalised distribution, with
uncertainties defined by the 68\% credible region.
\section{Multicolor HST photometry}
We present HST images of the lens system SDSSJ0946+1006 in five
different bands. In Paper VI we reported results based on an ACS
F814W image only. Images in WFPC2 F606W and WFC3-IR F160W (Cycle 16,
Program 11202, PI Koopmans) were available for Paper IX \citep{PaperIX}. In addition
to those data, we now have WFC3 images in F438W and F336W bands (Cycle
17, Program 11701, PI Treu). Table \ref{imgtable} summarizes the
observations. This section describes the data reduction process
(\S~\ref{Reduction}) and the photometric properties we derived for the
lens galaxy (\S~\ref{LensPhoto}). For conciseness, we sometimes refer
to the F160W, F814W, F606W, F438W, F336W bands as H, I, V, B, U
respectively. A color composite image of the lens system is shown in
Figure~\ref{fig:Jackpot}
\begin{figure}
\includegraphics[width = \columnwidth]{Jackpot_nup_compass.eps}
\caption{Gravitational lens system SDSSJ0946+1006 in a combination of F814W, F606W and F336W HST images.}
\label{fig:Jackpot}
\end{figure}
\begin{table}
\begin{center}
\caption{Summary of the HST observations.\label{imgtable}}
\begin{tabular}{ccccc}
\tableline\tableline
Instrument & Filter & Exp. time & $N_{exp}$ & Date \\
\tableline
WFC3 IR & F160W & 2397 s& 4 & 09/12/2009\\
ACS & F814W & 2096 s & 4 & 3/11/2006 \\
WFPC2 & F606W & 4400 s & 4 & 18/12/2009 \\
WFC3 UVIS & F438W & 2520 s & 4 & 20/03/2010\\
WFC3 UVIS & F336W & 5772 s & 4 & 20/03/2010\\
\tableline
\end{tabular}
\end{center}
\end{table}
\subsection{Data reduction}\label{Reduction}
The data are treated with the standard HST reduction pipeline. For each image, frames are coadded and resampled in a uniform pixel scale using the software {\sc multidrizzle} \citep{Drizzle}. Pixel sizes are $0.10''$ for the F160W image, $0.050''$ for the F814W and F606W images, and $0.0396''$ for the F438W and F336W images.
The images are then brought to the same orientation and $0.050''$ pixel scale by using the software {\sc swarp} \citep{Swarp}.
The PSF of each image is estimated from stars in the field.
\subsection{Lens galaxy properties}\label{LensPhoto}
The brightness distribution of the main lens galaxy is first obtained
by fitting S\'{e}rsic profiles to the data. This task is achieved with
the software {\sc spasmoid}, developed by M. W. Auger and described by
\citet{Bennert}. {\sc Spasmoid} fits the data in all the bands
simultaneously with a unique model, determining total magnitude and
colors of the galaxy at once.
By using a single S\'{e}rsic component we find a best-fit profile described by a S\'{e}rsic index $n=6.0$, axis ratio $q = 0.95$ and effective radius $r_{eff} = 2.93''$.
However, the residuals left by this single-component fit are rather large.
Consequently, we add a second
component, allowing for the position angle of the major axes of the
two profiles to be different but imposing a common centroid. In the fitting process, the light from
the rings is masked out manually. This procedure gives robust
estimates of the colors of the lens, rather independent from the model
adopted to describe the data. Color information will be used in
Section 5 to constrain the stellar population. In Fig. \ref{imgs1} we
show the images of the system in the five bands, before and after
subtracting the main lens.
Residuals are on the order of a few percent in the F814W band image.
Table \ref{TableFits} reports the best-fit structural parameters of the
model, while the best-fit colors are given in Table \ref{colortable}.
It is worth pointing out that the major axes of the two components are
almost perpendicular, and that the mean surface brightness within the
effective radius of component 1 is a factor $\sim30$ larger than that
of component 2.
The measured magnitude in the F814W band is consistent with the value reported by \citet{Gavazzi} for the same object.
In order to both explore model-dependent systematic errors and obtain a computationally more tractable description of the light profile for our lensing analysis, we also model the lens light with the following surface brightness distribution:
\begin{equation}\label{tNIEsb}
I(x,y) = I_cr_c\left[\frac{1}{\sqrt{r_c^2+R^2}} - \frac{1}{\sqrt{r_t^2+R^2}}\right],
\end{equation}
where $R^2 \equiv x^2/q + qy^2$.
This profile corresponds to a truncated pseudoisothermal elliptical mass distribution (tPIEMD) in 3d, with $r_c$ and $r_t$ corresponding to the core radius and truncation radius respectively.
Note that the number of parameters of the model is the same as that of a S\'{e}rsic profile.
Two components are used, as in the S\'{e}rsic case.
The best-fit parameters are reported in Table \ref{twotNIEfit}.
Both the double-S\'{e}rsic and the double-tPIEMD profiles fit well the photometry of the lens, with residuals within the outer ring on the order of a few percent in the F814W band (see Figure \ref{imgs1}).
The inferred total magnitude in the two models is different, but this is due to the different behavior at large radii, where there are no data.
In fact, the magnitude within the inner ring is the same for the two models to within 0.01 mags and the inferred colors are consistent within the errors with those reported in Table \ref{colortable}.
\begin{figure*}[!]
\begin{center}
\begin{tabular}{ccc}
\includegraphics[width=0.31\textwidth]{Jack_160_wscale.eps} &
\includegraphics[width=0.31\textwidth]{dSersic_160_res.eps} &
\includegraphics[width=0.31\textwidth]{dtPIEMD_160_res.eps} \\
\includegraphics[width=0.31\textwidth]{Jack_814_cut.eps} &
\includegraphics[width=0.31\textwidth]{dSersic_814_small.eps} &
\includegraphics[width=0.31\textwidth]{dtPIEMD_814_small.eps} \\
\includegraphics[width=0.31\textwidth]{Jack_606_cut.eps} &
\includegraphics[width=0.31\textwidth]{dSersic_606_small.eps} &
\includegraphics[width=0.31\textwidth]{dtPIEMD_606_small.eps} \\
\includegraphics[width=0.31\textwidth]{Jack_438_cut.eps} &
\includegraphics[width=0.31\textwidth]{dSersic_438_res.eps} &
\includegraphics[width=0.31\textwidth]{dtPIEMD_438_res.eps} \\
\includegraphics[width=0.31\textwidth]{Jack_336_cut.eps} &
\includegraphics[width=0.31\textwidth]{dSersic_336_res.eps} &
\includegraphics[width=0.31\textwidth]{dtPIEMD_336_res.eps} \\
\end{tabular}
\caption{From top to bottom: HST F160W, F814W, F606W, F438W and F336W images of the lens system SDSSJ0946+1006 before (left column) and after (middle and right column) light subtraction. {\em Middle column:} light distribution modeled as a double S\'{e}rsic profile, with parameters given in Table \ref{TableFits}.
{\em Right column:} light distribution modeled as a double tPIEMD profile, with parameters given in Table \ref{twotNIEfit}.}
\label{imgs1}
\end{center}
\end{figure*}
\begin{deluxetable*}{ccccccc}
\tablewidth{0pt}
\tablecaption{Lens light distribution: double S\'{e}rsic model\label{TableFits}}
\tablehead{
\colhead{Component} & \colhead{$m_{\rm{F814W}}$} & \colhead{$r_{\rm{eff}}$} & \colhead{$n$} &
\colhead{$q$} & \colhead{PA} & \colhead{$\left<\mathrm{SB}\right>_{\rm{e,F814W}}$}\\
& (mag) & (arcsec) & & & (degrees) & (mag arcsec$^{-2}$)}
\startdata
1 & $18.38\pm0.20$ & $0.50\pm0.10$ & $2.34\pm0.50$ & $0.79\pm0.10$ & $63.0\pm1.0$ & $18.87\pm0.10$\\
2 & $17.44\pm0.10$ & $4.46\pm0.50$ & $1.60\pm0.50$ & $0.64\pm0.10$ & $-23.4\pm1.0$ & $22.68 \pm 0.20$\\
\enddata
\tablenotetext{}{Best fit parameters for the double-S\'{e}rsic model surface brightness profile of the main lens: magnitude in the F814W band, effective radius, S\'{e}rsic index ($n$), axis ratio ($q$), position angle of the major axis (East of North), effective surface brightness. Each line refers to one of the S\'{e}rsic components of the model.
The errors represent the typical range of values for the parameters allowed by the model.
These errors are correlated: for example, an increase in the value of the S\'{e}rsic index $n$ results in a change of the effective radius to fit the observed slope in surface brightness.}
\end{deluxetable*}
\begin{deluxetable*}{ccccccc}
\tablewidth{0pt}
\tablecaption{Lens light distribution: double tPIEMD model\label{twotNIEfit}}
\tablehead{
\colhead{Component} & \colhead{$m_{\rm{F814W}}$} & \colhead{$r_c$} & \colhead{$r_t$} &
\colhead{$q$} & \colhead{PA} & \colhead{$\left<\mathrm{SB}\right>_{\rm{e,F814W}}$}\\
& (mag) & (arcsec) & (arcsec) & & (degrees) & (mag arcsec$^{-2}$)}
\startdata
1 & $18.75\pm0.20$ & $0.066\pm0.010$ & $0.50\pm0.05$ & $0.66\pm0.10$ & $63.0\pm1.0$ & $19.24\pm0.10$\\
2 & $17.15\pm0.10$ & $0.082\pm0.010$ & $6.05\pm0.10$ & $0.71\pm0.10$ & $-24.3\pm1.0$ & $22.48 \pm 0.20$\\
\enddata
\tablenotetext{}{Best fit parameters for the double-tPIEMD model surface brightness profile of the main lens: magnitude in the F814W band, core radius ($r_c$), truncation radius ($r_t$), axis ratio ($q$), position angle of the major axis (East of North), effective surface brightness.}
\end{deluxetable*}
\begin{table}
\begin{center}
\caption{Colors of the lens galaxy.\label{colortable}}
\begin{tabular}{cccc}
\tableline\tableline
Color & Component 1 & Component 2 & Global \\
\tableline
I - H & $1.16\pm0.05$ & $0.86\pm0.05$ & $0.96\pm0.05$\\
V - I & $0.81\pm0.05$ & $0.96\pm0.05$ & $0.91\pm0.05$\\
B - V & $2.36\pm0.20$ & $1.52\pm0.05$ & $1.73\pm0.05$\\
U - B & $2.30\pm0.30$ & $1.32\pm0.10$ & $1.44\pm0.10$\\
\end{tabular}
\end{center}
\end{table}
The infrared F160W data reveal distorsions in the shape of the light
distribution at large radii (see Figure \ref{IRtidal}), a possible
signature of tidal interactions.
As previously noted by \citet{Gavazzi}, a galaxy in the neighborhood of the lens also shows signs of a tidal interaction (see Figure \ref{IRtidal}). It is possible that the two galaxies are undergoing a merger.
This deviation from a regular light
profile is located far from the probed by our lensing and
dynamics measurements and is therefore not a concern for the accuracy
of our models.
The central part appears smooth to the few percent level and it is unlikely that the ongoing interaction would have an effect on its structure, given its deep potential well.
However, as we will discuss in \S~\ref{ssec:form} this
feature provides an interesting clue to the formation mechanism of
this galaxy.
\begin{figure}
\includegraphics[width = \columnwidth]{IRtidal_long.eps}
\caption{WFC3-IR F160W image of the lens system and its surroundings. Note the irregular shape of the faint stellar component at the outskirts of the lens galaxy (top of the image).
At the bottom, a neighbor also shows signs of tidal disruption. Both these features may be the result of a close encounter between the two objects.
}
\label{IRtidal}
\end{figure}
Another interesting feature is revealed by the image in the F336W (U) band, as there seems to be some structure in the center of the lens (see Figure \ref{imgs1}).
The fact that this feature is clearly visible only in the U band, where the lens is fainter, may suggest that it is in fact a bluer object distinct from the central galaxy, or blue emission from an active nucleus. Alternatively, the observed detail could be the result of the presence of a dust lane that separates the light of the lens into two components at shorter wavelengths. In principle it could also be an additional image of the lensed sources.
One way to discriminate between a blue object or a dust lane is to
study the position of the centroid of the lens in the different bands.
A blue object would shift the centroid towards itself at bluer
wavelengths, while a dust lane would remove blue light, causing the
apparent centroid to move away from it. When fitting for the centroid
of the lens, this latter case is observed: the centroid moves by about
one pixel towards the S-E in the F336W and F438W bands with respect to
the F814W band. This is a significant effect given the subpixel
accuracy of centroniding, and it suggests that dust is most likely the
cause of the observed feature in the F336W band.
A more detailed discussion of the dust issue is given in Appendix A.
\section{Photometric redshift of the outer ring}\label{RingsPhoto}
\subsection{Colors of the ring}
One of the main goals of this study is to constrain better the mass distribution in the lens galaxy by obtaining a photo-z of the outer ring.
This task requires a measurement of the colors of the ring.
A color map of the outer ring is obtained as follows.
For each pair of neighboring bands, $\lambda_1$, $\lambda_2$, we align the corresponding images and then convolve each image with the PSF of the neighboring band. In this way we obtain pairs of images with the same effective PSF, necessary to get an unbiased estimate of the color for each pixel.
Global colors are then measured in the following way. For a given pair of bands, we select individual pixels with flux larger than the background by more than two sigma in both of the bands considered.
We make the assumption that the source has spatially uniform colors and estimate them statistically by taking a weighted mean of the individual pixel colors.
The measured values of the colors, corrected for galactic extinction, are reported in Table \ref{TableRingColors}.
\begin{table}
\begin{center}
\caption{AB colors of the outer ring\label{TableRingColors}}
\begin{tabular}{cc}
\tableline\tableline
I - H & $0.61\pm0.10$ \\
V - I & $0.21\pm0.10$ \\
B - V & $0.15\pm0.10$ \\
U - B & $0.53\pm0.10$ \\
\end{tabular}
\end{center}
\end{table}
\subsection{Measuring the photo-z}\label{BPZ}
To estimate the photometric redshift of the outer ring we make use of
the software BPZ \citep[Bayesian Photo-z;][]{Benitez}. Photo-z
analysis consists of fitting synthetic SEDs to the observed
colors. BPZ works in a Bayesian framework that allows us to combine
the inference with that from other pieces of information: given a
prior probability distribution for the source redshift and galaxy
type, BPZ calculates the probability of the source being at redshift
$z_{s2}$ given its colors ${\bf C}$ and magnitude $m$, $P(z_{s2}|{\bf C},m)$. The stellar templates
used for the SED fitting are described by \citet{Coe}. The F814W
magnitude is taken from \citet{Gavazzi}, where the brightness
distribution of the source was reconstructed after a lens
modeling. The value adopted is therefore $m_{\rm{F814W}} =
27.01\pm0.19$
For the redshift distribution we use a prior $P(z|m_{\rm{F814W}})$ suggested by \citet{Benitez} and based on number counts from the Hubble Deep Field North (HDFN).
Figure \ref{PhotozPDF} shows the redshift posterior probability distribution function $P(z_{s2}|{\bf C},m_{\rm{F814W}})$.
The most likely redshift with 68\% confidence interval is $z_{s2} = 2.41_{-0.21}^{+0.04}$.
As will be shown later, this information is sufficient to put interesting constraints on the model of the lens system.
We also calculated the photo-z assuming a flat prior on $z_{s2}$, and found a nearly identical result.
Colors calculated with a different lens light subtraction, the double-S\'{e}rsic model, yield the same photo-z well within the quoted uncertainties.
\begin{figure}
\includegraphics[width = \columnwidth]{Probz.eps}
\caption{{\em Solid line:} posterior probability distribution function of the source redshift, as calculated with BPZ, assuming a prior on $z_{s2}$ from Hubble Deep Field North number counts.
Overplotted are the levels corresponding to 68\% and 95\% enclosed probability.
{\em Dotted line:} posterior PDF assuming a flat prior on $z_{s2}$.
}
\label{PhotozPDF}
\end{figure}
\section{Keck Spectroscopy}
The data were collected during the nights of 2006 December 23 and 24
with the LRIS instrument at the Keck Telescope I. The original goals
of the observations were to measure a velocity dispersion profile of
the foreground deflector and to measure the redshift of the outer
ring. The first goal was succesfully achieved, while we were not able
to detect any spectroscopic signature from the farthest source.
Because of the dual scope of our study, two different instrumental setups were used.
The first setup, used during the first night, was optimized for a better measurement of the velocity dispersion of the deflector.
The wavelength range in the red detector, the one used for the measurement of $\sigma$, was $\sim5700-7600\AA$, bracketing important absorption features in the rest frame of the lens at $z=0.222$.
During the second night we centered the slit on the longest arc of the outer ring, and used a setup with a broader wavelength range, up to $\sim8600\AA$.
A summary of the observations, with specifications on the setups used, is provided in Table \ref{Setups}.
\begin{deluxetable*}{ccccccccc}
\tablecaption{Spectroscopic observations: summary}
\tablehead{
\colhead{Date} & Exp. time & \colhead{Slit width} & \colhead{Dichroic} & \colhead{Blue grism} & \colhead{Red grating} & \colhead{Red $\lambda_c$} & \colhead{Weather} & \colhead{Seeing}}
\startdata
12/23/2006 & 16200 & 1.0'' & 560 & 600/4000 & 831/8200 & $6819 \AA$ & Good & $0.8''$\\
12/24/2006 & 12600 & 1.0'' & 680 & 300/5000 & 831/8200 & $7886 \AA$ & Good & $0.8''$ \\
\enddata
\label{Setups}
\end{deluxetable*}
The spectrum of the system is shown in Fig. \ref{Spectrum}.
There is no evidence for the presence of emission lines from objects other than the foreground lens and the inner ring.
Given our measurement of the photo-z of the outer ring, we would expect Ly-$\alpha$ emission to fall around $\sim4150\AA$, but it cannot be identified in our spectrum.
We can put an upper limit of $\sim5\times10^{-18}\mbox{ erg cm}^{-2}\mbox{ s}^{-1}$ to the flux in Ly-$\alpha$ from the source.
\begin{figure*}[!]
\begin{center}
\includegraphics[width=\textwidth]{plotspectrum.eps}\\
\end{center}
\caption{LRIS spectra of the Jackpot. {\em Blue:} data from the first night. {\em Red:} data from the second night. The two spectra are extracted from rectangular apertures $1''\times3.36''$.
Dotted line: noise level.}
\label{Spectrum}
\end{figure*}
\subsection{Velocity dispersion}
The velocity dispersion of the main lens is measured by fitting stellar templates convolved with a Gaussian velocity distribution to the observed spectrum.
This operation is carried out with a Monte Carlo Markov Chain approach, using a code developed by M. W. Auger, and described by \citet{Suyu2010}.
The rest frame wavelength range used for the fit is $5100-5850\AA$.
For the stellar templates we used linear combinations of nine spectra from the INDO-US library, corresponding to K,G,F and A stars.
The most prominent absorption feature in the wavelength range considered is Mgb ($5175\AA$).
However, we experienced difficulty in finding a good fit to both Mgb and the rest of the spectrum.
It is known that some galaxies have enhanced magnesium features in the spectrum that are not well reproduced in standard stellar templates \citep{Barth}.
For this reason we decided to mask the Mgb absorption line out of the fitted spectrum.
With the aim of obtaining a velocity dispersion profile, we measured
$\sigma$ in a set of apertures. The spatial position of the apertures
was determined by fitting the centroid of the trace of the lens in the
twodimensional spectra and it is accurate to $\sim0.02''$. In Table \ref{vprofile-data} and Fig. \ref{MeasProfile} we report the measured values of $\sigma$ and
of the mean velocity in each aperture, while in Fig. \ref{FitSpectrum}
we show the fit in the central $0.42''$ as an example. There is
evidence for some rotation, with $v_{\rm{rot}}^2 \ll \sigma^2$.
\begin{deluxetable}{ccc}
\tablecaption{Velocity profile measurements}
\tablehead{
\colhead{Slit offset} & \colhead{$\left<v\right>$} & \colhead{$\sigma$} \\
\colhead{(arcsec)} & \colhead{(km s$^{-1}$)} & \colhead{(km s$^{-1}$)}
}
\startdata
-1.05 & $101\pm21$ & $252\pm25$ \\
-0.84 & $85\pm16$ & $273\pm18$ \\
-0.63 & $62\pm11$ & $263\pm14$ \\
-0.42 & $30\pm10$ & $278\pm12$ \\
-0.21 & $20\pm10$ & $287\pm11$ \\
0.00 & $0\pm9$ & $287\pm11$ \\
0.21 & $-22\pm11$ & $286\pm11$ \\
0.42 & $-55\pm12$ & $299\pm13$ \\
0.63 & $-67\pm13$ & $274\pm15$ \\
0.84 & $-63\pm15$ & $272\pm19$ \\
1.05 & $-94\pm24$ & $301\pm25$
\enddata
\tablenotetext{}{Mean velocity and velocity dispersion profile. Apertures are $1.00\times0.21''$ rectangles.}
\label{vprofile-data}
\end{deluxetable}
\begin{figure}
\includegraphics[width = \columnwidth]{oneplot.eps}
\caption{Mean velocity and velocity dispersion profiles of the main lens within $1.15''$ from the centroid.}
\label{MeasProfile}
\end{figure}
\begin{figure}
\includegraphics[width = \columnwidth]{centralfit.eps}
\caption{Fit of the velocity dispersion of the lens. {\em Top:} The red curve is the best fit synthetic spectrum.
Shaded regions are masked and not used for the fit. {\em Bottom:} Residuals of the fit in fractions of the total flux.}
\label{FitSpectrum}
\end{figure}
\section{Stellar masses}\label{mstarsect}
Here we present a measurement of the stellar mass of the foreground
lens galaxy. The procedure adopted is the following: we fit stellar
population synthesis models to the observed spectral energy
distribution (SED) of the galaxy. A measurement of this kind was already
performed by \citet{Grillo2009} and \citet[Paper IX]{PaperIX} for the
same object. Their results agree within the errors. \citet{Grillo2009}
used SDSS multiband photometry ($u$, $g$, $r$, $i$, $z$ bands) as
their observed SED. In Paper IX, high resolution HST data was used,
but only in two bands (F814W and F606W). \citet{PaperIX} also
introduced a powerful statistical analysis method, based on Bayesian
statistics that allows for physically meaningful priors on the model
parameters as well as a full exploration of uncertainties and
correlation between the inferred parameters. With five band HST
photometry we can now extend the analysis of Paper IX, to obtain a
more robust estimate of the stellar mass.
The fitting method is the same as that developed by \citet{PaperIX}, and can be summarized as follows. Composite stellar population models are created from \cite{BC03} stellar templates. The star formation history is modeled with a single exponentially decaying burst.
The parameters of the model are age, metallicity, exponential burst timescale, dust reddening and stellar mass.
The parameter space is explored using a Monte Carlo Markov Chain (MCMC) routine, through which the posterior PDF is characterized.
The stellar templates used are based on either a Salpeter or a Chabrier IMF.
For the description of the photometry of the lens we use the double tPIEMD model described in Sect. \ref{LensPhoto}, that is consistent with the analyses presented in the following Sections. The stellar masses of the two components are fitted independently.
Results are listed in Table \ref{mstartable}, together with the values previously found by \citet{Grillo2009} and \citet{PaperIX}.
The analysis reveals the presence of dust for component 1, coherently with our previous findings.
Repeating the fit with the dust-corrected magnitudes yields indistinguishable stellar masses.
The logarithm of the stellar masses changes by 0.06 if we use the description of the light profile with S\'{e}rsic components instead of tPIEMDs.
This is due to the different behavior at large radii of the two profiles.
Differences in the mass within the outer Einstein radius for the two models are instead well within the measurement errors.
\begin{deluxetable}{cccc}
\tablecaption{Stellar mass of the foreground lens, from SPS models}
\tablehead{
\colhead{IMF} & \colhead{Chabrier} & \colhead{Salpeter} & \\
& $\log(M_*/M_\Sun)$ & $\log(M_*/M_\Sun)$ & }
\startdata
Comp. 1 & $10.85_{-0.06}^{+0.09}$ & $11.13_{-0.11}^{+0.05}$ & This work\\
Comp. 2 & $11.27_{-0.08}^{+0.05}$ & $11.52_{-0.08}^{+0.06}$ & This work\\
Total & $11.40 \pm 0.06$ & $11.66 \pm 0.06$ & This work\\
& $11.38_{-0.12}^{+0.04}$ & $11.61_{-0.08}^{+0.02}$ & \citet{Grillo2009} \\
& $11.34 \pm 0.12$ & $11.59 \pm 0.12$ & \citet{PaperIX} \\
\enddata
\label{mstartable}
\end{deluxetable}
\section{A single component model: measuring the average slope}\label{LensingSect}
In this Section we present a single-component lensing and dynamics study of the foreground galaxy, where the total density distribution of the lens is described with a power-law.
The goal is to obtain a measurement of the slope of the total mass profile and also to test the accuracy allowed by our data in constraining mass models.
The system, with its two Einstein rings, offers more constraints than typical single-source lenses.
However, the analysis is complicated by the presence of two different lenses along the line of sight.
Light rays from the second source are first deflected by the object corresponding to the inner ring and then by the foreground lens, with the result that, unlike the single lens case, the relation between the size of the outer Einstein ring and the enclosed projected mass of the lens is nontrivial.
Nevertheless, this can be properly accounted for as described below.
A first lens modeling of the system was carried out in Paper VI.
The procedure adopted there was a conjugate points method: multiply imaged spots in the lensed features are identified, and the lens model is determined by minimizing the distance between the corresponding points in the source plane. This is a conservative approach, since it does not make use of all of the information from the surface brightness of the rings.
The main lens was modeled as a power law ellipsoid, with dimensionless surface mass density $\kappa \equiv \Sigma/\Sigma_{cr}$ given by:
\begin{equation}
\kappa(\vec{r},z_s) = \frac{b_\infty^{\gamma'-1}}{2}(x^2 + y^2/q^2)^{(1-\gamma')/2}\frac{D_{\rm{ls}}}{D_{\rm{os}}},
\end{equation}
where $b_\infty = 4\pi(\sigma_{\rm{SIE}}/c)^2$ and $D_{\rm{ls}}$ ($D_{\rm{os}}$) is the angular diameter distance of the source relative to the lens (observer).
The second lens
(first source corresponding to the brighter arc) was modeled as a
singular isothermal sphere (SIS). The model parameter space was
explored via a MCMC. The results showed
that two types of solution are possible: a model with larger $\sigma_{\rm{SIE}}$, shallower slope $\gamma'$ and less massive second lens, or a model with a more massive second lens and steeper main lens slope
(see Figure 9 of Paper VI, or black contours of
Figure \ref{contourplots}). Part of this degeneracy was due to our
ignorance of the redshift of the outer ring.
In this Paper we use the lens model of Paper VI described above and improve it by incorporating 1) our measurement of photo-z of the outer ring and 2) a stellar dynamics analysis.
\subsection{Stellar dynamics modeling}\label{Jeanssect}
We wish to use our measurements of the velocity dispersion profile of the lens to constrain our lens models.
This is done with a procedure similar to that adopted by \citet{Suyu2010}, which can be described as follows.
For a given model provided by the lensing analysis, we compute a model velocity dispersion profile and compare it to the observed one.
The model velocity dispersion is obtained by solving the spherical Jeans equation
\begin{equation}\label{Jeans}
\frac{1}{\rho_*}\frac{d\rho_*\sigma_r^2}{dr} + 2\frac{\sigma_\theta^2}{r} = -\frac{GM(r)}{r^2},
\end{equation}
where $\rho_*(r)$ is the density distribution of the light, $\sigma_r$ and $\sigma_\theta$ are the radial and tangential components of the velocity dispersion tensor, $M(r)$ is the total mass enclosed within the spherical shell of radius $r$.
We impose spherical symmetry in the mass model by adopting a spheroidal mass distribution
\begin{equation}
\rho(r) \propto r^{-\gamma'}
\end{equation}
with normalization chosen such that the total projected mass enclosed within the Einstein radius equals that of the corresponding circularized lens model.
The light distribution is described as the sum of two tPIEMD profiles, with the same parametrization described in Section \ref{LensPhoto} (best-fit parameters are in table \ref{twotNIEfit}).
The 3d stellar distribution corresponding to the surface brightness profile (\ref{tNIEsb}) used to fit the photometry is
\begin{equation}\label{rhotNIE}
\rho(r) = \rho_cr_c^2\left[\frac{1}{r_c^2+r^2} - \frac{1}{r_t^2+r^2}\right],
\end{equation}
with $r \equiv x^2/q_* + q_*y^2 + z^2$.
Here we set the axes ratios $q_*$ to one, as we are assuming spherical symmetry.
We then assume a Osipkov-Merritt model for the velocity dispersion tensor \citep{Osipkov,Merritt}:
\begin{equation}
\frac{\sigma_{\theta}^2}{\sigma_r^2} = 1 - \frac{r^2}{r_a^2+r^2},
\end{equation}
where $r_a$ is the anisotropy radius (orbits are radially anisotropic beyond $r_a$).
Finally, we simulate the line-of-sight velocity dispersion measured in our apertures.
Rotation is neglected.
Although the lens is seen to be rotating, its mean velocity is small compared to the velocity dispersion and should not contribute much to the dynamics of the object.
The effect of this approximation will be discussed further below.
\subsection{Combining the constraints}\label{importancesect}
The models of the lens are defined by the set of parameters $\boldsymbol{\eta} \equiv\{\sigma_{\rm{SIE,lens}}, \gamma', \sigma_{\rm{SIS,s1}}, z_{s2}\}$: the strength and power-law index of the foreground lens, the strength of the background lens and the redshift of the outer ring, respectively.
Each model gives a prediction of the velocity dispersion in each aperture, $\sigma_{\rm{ap},\mathit{i}}^{\rm{(mod)}}$.
The new posterior probability distribution for the model is obtained via importance sampling: the MCMC sample corresponding to the lens modeling of Paper VI is weighted by the likelihood of the measurements ${\bf d} \equiv \{z_{s2},\sigma_{\rm{ap},\mathit{i}}^{\rm{(meas)}}\}$ given the model parameters $\boldsymbol{\eta}$.
The following likelihood function is used:
\begin{equation}
L({\bf d}|{\boldsymbol{\eta}}) = P_z(z_{s2})\prod_iG_i(\sigma_{\rm{ap},\mathit{i}}^{\rm{(meas)}}|{\boldsymbol{\eta}})
\end{equation}
where $P_z(z_{s2})$ is the PDF in Figure \ref{PhotozPDF} and
\begin{equation}
G_i(\sigma_{\rm{ap},\mathit{i}}^{\rm{(meas)}}|{\boldsymbol{\eta}}) = \frac{1}{\sqrt{2\pi\Delta_{\sigma,\mathit{i}}^2}}\exp{-\frac{(\sigma_{\rm{ap},\mathit{i}}^{\rm{(meas)}}-\sigma_{\rm{ap},\mathit{i}}^{\rm{(mod)}})}{2\Delta_{\sigma,\mathit{i}}^2}},
\end{equation}
and $\sigma_{\rm{ap},\mathit{i}}$ and $\Delta_{\sigma,\mathit{i}}$ are the zeroth and second moment of the posterior PDF of the measured velocity dispersion in aperture $i$, respectively.
In Fig. \ref{contourplots} we show the updated Posterior PDF obtained
by importance sampling with the photo-z and dynamics measurements,
both separately and jointly. It is clear that although photo-z and
stellar kinematics alone leave some degeneracies, the posterior pdfs
are almost perpendicular in this space, and therefore the combination
of the two is particularly effective.
The estimate of the slope obtained by marginalizing over the other parameters is
\begin{equation}\label{bestgamma}
\gamma' = 1.98\pm0.02.
\end{equation}
We stress that our uncertainty on this parameter is a factor of four smaller than the typical error on $\gamma'$ from studies of single-source gravitational lenses with SDSS spectroscopy \citep[see Figure \ref{comparison}]{PaperX}. Comparable precision was reached by \citet{Bar++11} for a sample of lens systems with two dimensional kinematics constraints from integral field spectroscopy.
\begin{figure}
\begin{tabular}{c}
\includegraphics[width = \columnwidth]{photozonly.eps} \\
\includegraphics[width = \columnwidth]{dynonlyOM.eps} \\
\includegraphics[width = \columnwidth]{dyn+photozOM.eps}
\end{tabular}
\caption{Posterior PDF of $\gamma'$ and $\sigma_{s1}$ of the updated (filled contours) lens model, together with the old model of Paper VI (empty contours).
The updated model includes only the photo-z measurement of the outer ring in the top panel, only the velocity dispersion profile of the lens in the middle panel, and both the photo-z and velocity dispersion profile in the bottom panel.
The levels correspond to 68\%, 95\% and 99.7\% enclosed probability.}
\label{contourplots}
\end{figure}
In order to better understand the significance of these results, we try to quantify the error introduced by our simplified model for the stellar dynamics.
Two of our assumptions are potential sources of bias: spherical symmetry and the non-rotating approximation. The uncertainty in the mass determination from kinematics data is of order $\delta\sigma^2/\sigma^2\sim10\%$.
Biases on the order of this uncertainty or smaller are unlikely to bring significant changes to the results of our analysis.
By considering only the velocity dispersion and neglecting rotation, we underestimate the mass of the galaxy by a factor $\sim (v_{\rm{rot}}/\sigma)^2$, which is within $10\%$ in all apertures but one.
To gauge the importance of this effect we perform the following test.
We fit the model velocity dispersion profiles to the following ``effective velocity dispersion'': $\sigma_{\rm{eff}} \equiv \sqrt{\sigma^2 + v_{\rm{rot}}^2})$.
We then apply the same importance sampling procedure described above to get a new constraint on the density slope $\gamma'$. The new estimate with $1\sigma$ uncertainty is:
\begin{equation}
\gamma' = 1.97\pm0.02,
\end{equation}
which is consistent with the original estimate given by (\ref{bestgamma}). On the basis of this result, we can conclude that our approximation of non-rotating halo introduces a systematic error of order 0.01 on the inferred value of the slope $\gamma'$.
Quantifying the systematics introduced by the spherical symmetry assumption is more complicated.
In a previous work, \citet{Bar++11} performed a robust dynamical modelling of 12 SLACS lenses previously analysed with a spherical Jeans equation approach by \citet{PaperX}. The slopes $\gamma'$ inferred by \citet{PaperX} are consistent with the more accurate measurements of \citet{Bar++11}, with a bias on $\gamma'$ of $0.05\pm0.04$.
However, the uncertainty on $\gamma'$ that we achieve in our work is smaller than that and an estimate of the bias requires additional work.
Two distinct effects come into play.
First, the lens has a non-circular projected shape in both its mass and light distribution.
Second, the galaxy may even have asymmetries along the line of sight.
The importance of these effects on our analysis is quantified in Appendix B.
By relaxing the assumption of spherical symmetry the additional uncertainty on the velocity dispersion is about $\delta\sigma^2/\sigma^2\sim10\%$.
It follows that none of our results change appreciably.
An independent analysis of the system was carried out by \citet{Vegetti}. The method adopted by them is more complex than the one used in Paper VI: they made use of information from all the pixels of the lensed features to reconstruct the source surface brightness as a whole.
Using data from the inner ring only, they obtained the following estimate for the density slope:
\begin{equation}\label{gammaSimona}
\gamma' = 2.20\pm0.03^{\rm{(stat)}}.
\end{equation}
This is a local estimate of the slope $\gamma'$, obtained by measuring the magnification of the arc in the radial direction.
Our measurement is instead an average slope, obtained by fitting a single power-law halo to data spanning the lens from the center (dynamics) to the outer lensed ring.
This difference may suggest that the actual mass distribution of the lens is different from a simple power-law halo.
It is also for this reason that we proceeded to model the system with a more complex model.
\section{A two-component analysis: dissecting luminous and dark matter}\label{twocompsect}
We perform a two-component lensing and dynamics study where the mass distribution is composed of a dark matter halo and a bulge of stars.
\subsection{Lensing and dynamics modeling}
We use a power-law ellipsoid for the dark matter, while the stars are described with the double tPIEMD model found from the photometry analysis.
The second lens is again modeled as a SIS.
The parameters of the stellar distribution are fixed to the best-fit values reported in Table \ref{twotNIEfit}.
The global mass-to-light ratio is left as a free parameter, but the relative contribution of the two components is fixed according to the results of the stellar population synthesis analysis presented in Section \ref{mstarsect}.
For a unit F814W-band magnitude, component 1 is measured to be a factor of 1.73 (1.77) heavier than component 2 assuming a Salpeter (Chabrier) IMF.
In our lensing model, the mass-to-light ratio of component 1 is set to be 1.75 times larger than for component 2.
We also allow for constant external shear $\gamma_{\rm{ext}}$ with position angle $\rm{PA}_{\rm{ext}}$ and constant external convergence $\kappa_{\rm{ext}}$ in the lens plane.
Issues related to the external convergence are discussed below in a dedicated subsection.
Compared to the lensing study presented in the previous section, this model has two additional free parameters: the stellar mass $M_*^{\rm{LD}}$ and the external convergence $\kappa_{\rm{ext}}$.
Given the very tight constraint on the average slope $\gamma'$ from the single component analysis, we expect to be able to determine both the slope of the dark matter halo $\gamma_{\rm{DM}}$ and the stellar mass $M_*^{\rm{LD}}$ with sufficient accuracy.
The range of values of the slope of the dark matter halo explored in this analysis is $1.0 < \gamma_{\rm{DM}} < 3.0$.
The technique adopted to fit the model to the lensing data is the same
used for Paper VI: a conjugate points method implemented with a
MCMC. The dynamics analysis is carried out with a
procedure very similar to the one described in \S~\ref{Jeanssect}: we
solve the spherical Jeans equation for our model and obtain a
synthetic velocity dispersion profile to be compared to the measured
one. The (spherically symmetric) model mass distribution is obtained
by circularizing the projected mass distribution of the lens model,
setting $q_{\rm{DM}}$ and $q_*$ to one, and by taking the
corresponding spherical deprojections. The light distribution is set
by circularizing the double tPIEMD profile specified in Table
\ref{twotNIEfit}.
We then proceed to incorporate information on stellar dynamics and on the redshift of the background source.
This is done by importance sampling, with the same method described in \S~\ref{importancesect}.
\subsection{External convergence}
Objects other than the main lens can contribute to the surface mass
density $\kappa$. This external convergence is hard to detect and is
degenerate with the total mass of the lens galaxy. Ignoring the
contribution to $\kappa$ from perturbers can lead in principle to a
bias in the measurement of the key parameters of the lens. In order
to take into account the effect of external convergence on our error
budget, we include it in our model by generating random values of
$\kappa_{\rm{ext}}$ drawn from a plausible distribution.
This procedure allows us to propagate correctly this uncertainty to the other model parameters.
Kinematics information can also help to constrain $\kappa_{\rm{ext}}$ to some extent, as it is only sensitive to the mass dynamically associated with the galaxy, in contrast to lensing that is sensitive to all mass structures along the line of sight to the source.
Insight on the actual value of $\kappa_{\rm{ext}}$ can be gained by studying the lens environment.
According to \citet{PaperVIII}, this is found to be marginally underdense with respect to average lines of sight, therefore there is no evidence for the presence of a group in the lens neighborhood.
The closest cluster known to the NASA/IPAC Extragalactic Database (NED) is MaxBCGJ146.87912+10.07800, at redshift $z=0.151$ and projected distance 8.70 arcmin from our lens \citep{PaperVIII}.
If we assume a SIS profile for the cluster with a typical value for its velocity dispersion $\sigma = 1000\mbox{ km s}^{-1}$ we obtain a contribution to the convergence $\kappa_{\rm{cl}} < 0.01$.
We also scanned the Sloan Digital Sky Survey archive looking for massive red galaxies within 5' of the lens.
Only one early-type galaxy was found, at a redshift $z=0.218$ and angular distance 2.6'.
If we assume that this object is the brightest galaxy of a group and associate it with a SIS halo of $\sigma = 500\mbox{ km s}^{-1}$ the corresponding convergence at the location of the lens is $\kappa=0.02$.
Finally, the lensing analysis of \citet{Gavazzi} quantified the external shear as $\gamma_{\rm{ext}} = 0.07$ directed $-31$ degrees East of North.
The HST images show two objects with the same alignment relative to the lens (see Fig. \ref{IRtidal}).
If we make the assumption that those objects are responsible for the shear and assume again a SIS profile we obtain $\kappa_{\rm{ext}} = |\gamma_{\rm{ext}}| = 0.07$.
\citet{Hilbert} studied the external convergence associated with strong lensing systems in cosmological simulations.
They found that for a source at redshift $z_s = 5.7$ the distribution of $\kappa_{\rm{ext}}$ is skewed with a peak at $-0.04$, has zero mean and a scatter of $0.05$.
A slightly smaller scatter and a peak at $-0.02$ is found by \citet{Suyu2010} for sources at $z_s = 1.39$.
Taking all these aspects into account, we adopt as prior for $\kappa_{\rm{ext}}$ in our analysis a Gaussian distribution peaked at $0.05$, with dispersion $\sigma_\kappa = 0.05$ and truncated to values in the interval $-0.05 < \kappa_{\rm{ext}} < 0.15$.
This range should capture the indication of a positive contribution from the object responsible for the shear and take into account the effect of random mass clumps along the line of sight.
Priors with a broader range of allowed values of $\kappa_{\rm{ext}}$ lead to larger uncertainties on the other model parameters, but none of the conclusions of our study is altered.
\subsection{Results}
Contour plots of the posterior PDF for
the model parameters are shown in Figures \ref{PDFtwocomp1} and
\ref{PDFtwocomp2}. The best-fit velocity dispersion profile is
plotted in Figure \ref{sigmaprofiles}.
The inference on the two key parameters $M_*$ and $\gamma_{DM}$ is shown in better detail in Figure \ref{Mstargamma}.
By marginalazing over the remaining parameters, our model constrains the stellar mass to
\begin{equation}
M_* = 5.5_{-1.3}^{+0.4}\times10^{11}M_\Sun.
\end{equation}
This estimate comes from lensing and dynamics data, and does not rely
on assumptions on the mass-to-light ratio of the stars. This value
will be compared with the measurement of the stellar mass obtained
independently from photometry.
\begin{figure*}
\includegraphics[width = \textwidth]{cp_masses.eps}
\caption{Posterior PDF in the multidimensional space spanned by the stellar mass $M_*^{\rm{LD}}$, slope of the dark matter halo $\gamma_{\rm{DM}}$, radial anisotropy scale radius $r_a$, strength of the second lens $\sigma_{s1}$ and redshift of the second source $z_{s2}$.
The levels correspond to 68\%, 95\%, 99.7\% enclosed probability.
{\em Solid contours:} constraints from lensing only. {\em Shaded regions:} constraints from lensing, dynamics, and photo-z.}
\label{PDFtwocomp1}
\end{figure*}
\begin{figure*}
\includegraphics[width = \textwidth]{cp_angles.eps}
\caption{Posterior PDF in the multidimensional space spanned by external convergence $\kappa_{\rm{ext}}$, strength and position angle of the external shear, $\gamma_{\rm{ext}}$, $\rm{PA}_{\rm{ext}}$, axis ratio of the dark matter halo $q_{\rm{DM}}$, position angle of the major axis of the dark matter halo, $\rm{PA}_{\rm{DM}}$.
The levels correspond to 68\%, 95\%, 99.7\% enclosed probability.
{\em Solid contours:} constraints from lensing only. {\em Shaded regions:} constraints from lensing, dynamics, and photo-z.}
\label{PDFtwocomp2}
\end{figure*}
\begin{figure}
\includegraphics[width = \columnwidth]{Mstargamma.eps}
\caption{
Posterior PDF projected in the space $M_*-\gamma_{\rm{DM}}$. The vertical shaded regions show independent measurements of the stellar mass from photometry, presented in Section 5.}
\label{Mstargamma}
\end{figure}
\begin{figure}
\includegraphics[width = \columnwidth]{bestfitsigmaboth.eps}
\caption{Best-fit velocity dispersion profile of the lens.
{\em Solid line:} two components model. {\em Dashed line:} single power-law model.}
\label{sigmaprofiles}
\end{figure}
Another important result is the constraint that we obtain on the slope of the dark matter halo:
\begin{equation}
\gamma_{\rm{DM}} = 1.7\pm0.2
\end{equation}
This result shows strong evidence for a contraction of the dark matter distribution relative to the $r^{-1}$ inner slopes typical of dark matter only simulations \citep{NFW}.
Figure \ref{profiles} shows the mean density profile of each mass component compared to the mean single power-law fit from Sect. \ref{LensingSect}.
\begin{figure}
\begin{tabular}{c}
\includegraphics[width = \columnwidth]{rho_profiles.eps} \\
\includegraphics[width = \columnwidth]{mass_profiles.eps}
\end{tabular}
\caption{Best-fit density (top) and mass (bottom) profiles. {\em Solid line:} total mass from bulge-halo decomposition. {\em Dashed line:} stellar mass. {\em Dotted line:} dark matter. {\em Dash-dotted line:} total mass from single component analysis. The shaded regions represent $1-\sigma$ uncertainties.}
\label{profiles}
\end{figure}
Our inference for the anisotropy radius constrains $r_a > 13$ kpc, meaning that radial anisotropy is ruled out in the region probed by our data.
This
is consistent with previous work \citep[e.g.][]{K+T03,TreuKoop} and
expected on theoretical grounds, because strong radial anistropy would
lead to instabilities.
In contrast to the power-law model considered in the previous section,
this mass model has a density distribution with a slope that changes
with radius. It is interesting to compare the local value of the
slope at the location of the inner ring with the measurement of
\citet{Vegetti}. \citet{Vegetti} modeled the HST F814W image using only lensing information from the inner ring. Lensing is
only sensitive to projected masses, therefore, in order for the
comparison to be meaningful, we have to consider the logarithmic slope
of the total projected mass distribution, evaluated at the inner
Einstein radius. We find
\begin{equation}
\frac{d\log{\kappa}}{d\log{r}} = -1.1\pm0.1.
\end{equation}
This value is consistent with the slope found by \citet{Vegetti}, which is given by $-(\gamma'-1) = -1.2$, where $\gamma'$ is the slope of the 3d mass distribution given in (\ref{gammaSimona}).
Finally, it is interesting to note how the inference on the stellar
mass is rather insensitive to the actual value of the redshift of the
second source, $z_{s2}$ (see Figure \ref{PDFtwocomp1}). This means
that, with the current data quality, a spectroscopic measurement of
the redshift of the outer ring would not bring significantly more
information.
\section{Discussion}\label{IMFsect}
\subsection{Luminous and dark matter in the lens}
The data in our possession allowed us to study the lens galaxy of the system SDSSJ0946+1006 under multiple aspects.
Thanks to the high resolution photometry from HST we were able to note now the light distribution is well described with two components, while single component models yield poor fits.
These two components appear to be nearly perpendicular (in projection), have significantly different effective radii and surface brightnesses.
The colors of the more compact component (component 1 from now on) are also significantly redder (see Table \ref{colortable}), indicator of an older or more metal-rich stellar population.
As we will discuss below, these characteristics suggest a particular scenario for the past evolution of this object.
In Sections \ref{mstarsect} and \ref{twocompsect} we presented two independent measurements of the stellar mass of the foreground galaxy of the system SDSSJ0946+1006 derived with a lensing+dynamics analysis and with a stellar population synthesis study.
The measured values of $M_*$, obtained by marginalizing over the other model parameters, are reported in Table \ref{mstarmeas}.
\begin{deluxetable}{ccc}
\tablecaption{Stellar mass of the foreground galaxy}
\tablehead{
\colhead{Method} & \colhead{$M_*$ $(M_\Sun)$} & \colhead{$\alpha$\tablenotemark{a}}}
\startdata
Lensing+dynamics & $5.5_{-1.3}^{+0.4}\times10^{11}$ & \\
SPS, Chabrier IMF & $(2.5\pm0.3)\times10^{11}$ & $2.0\pm0.4$ \\
SPS, Salpeter IMF & $(4.5\pm0.6)\times10^{11}$ & $1.1\pm0.2$
\enddata
\tablenotetext{a}{$\alpha$ is the IMF mismatch parameter defined as $\alpha \equiv M_*^{\rm{LD}}/M_*^{\rm{SPS}}$.}
\label{mstarmeas}
\end{deluxetable}
The stellar mass measured from gravitational lensing and dynamics, $M_*^{\rm{LD}}$, is larger than the masses obtained from the SPS study, $M_*^{\rm{SPS}}$.
This discrepancy can be quantified with the ``IMF mismatch'' parameter $\alpha \equiv M_*^{\rm{LD}}/M_*^{\rm{SPS}}$, also reported in Table \ref{mstarmeas}.
A Salpeter IMF is clearly favored,
while the probability of the IMF being heavier than Chabrier ($\alpha_{\rm{Chab}} >1$) is 95\%.
This result is in agreement with a general trend observed by
\citet{Grillo2009}, \citet{TreuIMF} and \citet{Auger2010L} for the
early-type galaxies of the SLACS sample. They find that, on average,
a Salpeter IMF better matches the measurements of stellar masses from
lensing and dynamics. A similar result is found by \citet{Chiara} for
a very massive early-type galaxy. As discussed extensively by
\citet{TreuIMF}, stellar mass and slope of the dark matter halo are
degenerate with respect to typical lensing and dynamics constraints:
given a bulge-halo decomposition, steepening the dark matter profile
and decreasing the stellar mass can result in fits to the observed
velocity dispersion and mass within the Einstein radius as good as the
original model. \citet{TreuIMF} explained how the observed trend of
increasing $\alpha$ with velocity dispersion can either be interpreted
as the effect of a correlation between IMF or dark matter inner slope
with total mass.
\citet{Auger2010L} explored this degeneracy by considering adiabatically contracted DM halos set by an imposed relation between stellar and virial mass, and found preference for a stellar mass-to-light ratio closer to a Salpeter than a Chabrier IMF.
Similarly, \citet{NRT} find that a Kroupa IMF, which has a mass-to-light ratio slightly larger than a Chabrier IMF, fits well adiabatically contracted DM halos.
In the present study we allowed the slope of the
dark matter halo of our lens galaxy to vary freely. Its measured value, $\gamma_{\rm{DM}} =
1.7_{-0.2}^{+0.2}$, is significantly steeper than the inner slope of a
NFW halo. Still, we find a stellar mass larger than what can be
accounted for with a Salpeter IMF and not compatible with a Chabrier
IMF. Our results imply that a Salpeter IMF provides a far better
description of the mass-to-light ratio of the stellar population than
a Chabrier IMF even with a steepened dark matter halo.
This result is consistent with the recent findings of \citet{Cap++12} and \citet{v+C11}.
In contrast, Salpeter-like
IMFs are typically ruled out for lower mass systems \citep{Cap++06} or
spiral galaxies \citep{B+d01,Dut++11,Suy++11,Bre++12}.
The lensing and dynamics analysis presented in Sect. \ref{twocompsect}
showed evidence for contraction of the dark matter halo with respect
to a baryonless NFW profile.
A similar result is found by \citet{Grillo2012} for an ensemble measurement of 39 massive elliptical galaxy lenses.
This result is in qualitative agreement
with many theoretical studies of the evolution of spheroidal galaxies
\citep{Blumenthal,Gnedin04,Gustafsson,Abadi,Duffy}. \citet{Duffy} in
their simulations of redshift $z=2$ galaxies find inner dark matter
slopes that span the range $1.4 < \gamma_{\rm{DM}} < 2.0$ depending on
the different prescriptions adopted to model the effect of the
baryons. Our measured value of $\gamma_{\rm{DM}}$ falls nicely in
that range, although our galaxy is at significantly lower redshift.
\citet{Gnedin04} provide a prescription to calculate the dark matter
profile of their modified adiabatic contraction (MAC) model. It is
interesting to test the MAC model on the measured slope of the dark
matter halo of our galaxy. The final dark matter density profile of
the MAC model of \citet{Gnedin04} is determined given the observed
light profile, the concentration parameter $c$ of the original (non
contracted) NFW halo and the baryon mass fraction within its virial
radius, $f_b$.
Since we do not have information on the initial properties of the dark matter
halo of our galaxy, we use a few trial values of the virial mass
$M_{\rm{vir}}$, spanning a plausible range indicated by a weak lensing
study of ellipticals \citep{Gavazzi07}, and employ a
mass-concentration relation from \citet{Maccio} based on WMAP5
cosmological parameters. We then calculate the inner slope of the
final dark matter distribution with the software Contra
\citep{Gnedin04}. The inferred inner slope for
$\log{(M_{\rm{vir}}/M_\Sun)} = 12.0, 13.0, 14.0$ is plotted in
Fig. \ref{Gnedinslope}. Despite the large range of virial mass
explored, the slopes of the contracted halos lie around $1.5 <
\gamma_{\rm{DM}} < 2.0$ over the spatial region covered by our data.
The MAC model is therefore able to reproduce our measurement of the
dark matter halo slope.
\begin{figure}
\includegraphics[width = \columnwidth]{MACprofile.eps}
\caption{{\em Solid lines:} Inner slope of the dark matter halo for modified adiabatic contraction \citep[MAC;][]{Gnedin04} models.
{\em Dashed lines:} Slope of the non-contracted (NFW) dark matter halo.
{\em Shaded region:} 68\% confidence interval of the slope measured in this Paper.}
\label{Gnedinslope}
\end{figure}
\subsection{A formation scenario}
\label{ssec:form}
As our data show, the stellar distribution in the lens galaxy consists of two components that differ in alignment, surface brightness and stellar population.
This particular structure suggests different formation histories for the two components.
The bright and compact component may have formed first, and later on accreted stellar systems in the outskirts without disrupting the structure of the original bulge.
Alternatively, component 2 might have been present originally and component 1 be formed in a star formation event following a wet merger.
We point out that in the infrared image we see evidence for tidal distortion in the outskirts of the galaxy (see Figure \ref{IRtidal}), possible indication of an ongoing merger. Part of the faint extended envelope of component 2 could be material accreted relatively recently.
The presence of the dust lane in the center of the galaxy (see \S~\ref{ssec:dust}) may also be the result of a recent merger.
We also note that \citet{Vegetti} detected a compact substructure of mass $\sim 3\times10^9M_\Sun$ located in the proximity of the inner ring image, indicating that minor mergers may still be occurring.
Let us consider our first hypothesis: the galaxy consisted initially of the compact component 1.
What are the structural parameters of component 1 and how does it relate to other elliptical galaxies?
Its effective radius is $r_{\rm{eff}} = 0.50''$ (see Tables \ref{TableFits} and \ref{twotNIEfit}), which corresponds to a physical radius of $1.79\mbox{ kpc}$.
Similar effective radii are found for high redshift ($z>1.2$) ellipticals \citep{Daddi,Trujillo,vanDokkum}.
Its stellar mass as inferred from the SPS analysis is given by $\log{(M_*^{\rm{SPS}}/M_\Sun)} = 10.85$ ($\log{(M_*^{\rm{SPS}}/M_\Sun)} = 11.13$) for a Chabrier (Salpeter) IMF.
Local ellipticals with similar values of the stellar mass have effective radii a factor of a few larger than this object \citep{Shen,Hyde}.
Analogously, the high redshift objects of \citet{Daddi}, \citet{Trujillo} and \citet{vanDokkum} are also significantly more massive than local galaxies with similar effective radii.
Finding objects in the local universe that correspond to these high redshift ``red nuggets'' is in fact a standing problem in the study of elliptical galaxies.
It is not clear how objects initially so compact evolve into the more diffuse galaxies that we observe at recent times.
Recent numerical simulations \citep{Hopkins2009,Oser} showed how minor
dry mergers can increase the size of elliptical galaxies
significantly, with the stars of the accreted objects that grow the
outskirts of the galaxy, even though the observed and predicted merger
rates are such that this mechanism might not be sufficient
\citep{New++11a}. The observational signature of this process would be
the presence of a compact core, the original red nugget, surrounded by
a more diffuse distribution of stars from the accreted systems. The
galaxy studied in this paper might be one of these objects.
\section{Summary}
We have presented a new set of photometric and spectroscopic data for the gravitational lens system SDSSJ0946+1006.
We used these data to constrain the structural properties of the foreground elliptical galaxy of the system.
On the basis of our results, the following statements can be made.
\begin{itemize}
\item The redshift of the source corresponding to the outer ring is $z_{s2} = 2.41_{-0.21}^{+0.04}$ at 68\% confidence level, as revealed by our photo-z measurement.
\item If we describe the total mass distribution with a power-law
ellipsoid $\rho \propto r^{-\gamma'}$, lensing and dynamics data give
as measured value $\gamma' = 1.98\pm 0.02\pm0.01$. This parameter
should be interpreted as an effective slope of the density profile
averaged over the region within the outer Einstein ring. The special
lensing configuration and the exquisite data quality of our data,
allowed us to measure $\gamma'$ with unprecedented precision. The
value obtained is consistent with isothermal ($\gamma' = 2$) and is in
agreement with the general trend observed for the massive early-type
galaxies of the SLACS sample, $\left<\gamma'\right> = 2.078\pm0.027$
with intrinsic scatter $\sigma_{\gamma'} = 0.16\pm0.02$
\citep{PaperX,Koo++09,Bar++11}. See Figure \ref{comparison} for a comparison of our
measurement of $\gamma'$ with measurements of the same parameter for
the SLACS sample of early-type galaxies by \citet{PaperX}.
\item We are able to decompose dark and stellar matter with lensing and dynamics data, assuming a power-law density profile for the dark matter.
The derived stellar mass is $5.5_{-1.3}^{+0.4}\times10^{11}M_\Sun$, consistent with a Salpeter IMF and inconsistent with a Chabrier IMF.
This constraint on the IMF is plotted in Figure \ref{comparison} together with similar measurements for the other SLACS lenses obtained by \citet{TreuIMF}.
Note that we achieve better precision despite using less strict assumptions on the dark matter profile.
\item The slope of the dark matter halo is found to be
$\gamma_{\rm{DM}} = 1.7\pm0.2$. This is a strong evidence for
contraction relative to the $r^{-1}$ behavior of NFW profile observed
in simulations without baryons, and is in agreement with the inner
dark matter profiles obtained by \citet{Duffy} in their simulations of
$z=2$ galaxies and with the MAC model of \citet{Gnedin04}.
Our inferred bulge-halo decomposition has a local projected slope at the inner ring
in agreement with the value measured by \cite{Vegetti} based on a completely independent technique.
\item The particular structure of the stellar distribution, with a
compact core and a misaligned faint extended envelope, might be the
result of accretion of low mass systems by a compact red nugget.
\item A spectroscopic detection of the redshift of the outer ring would still help improve the model, but would not lead to a dramatic change in the results of our analysis.
\end{itemize}
\begin{figure}
\includegraphics[width = \columnwidth]{Jack_vs_SLACS.eps}
\caption{{\em Top panel:} IMF mismatch parameter $\alpha \equiv M_*^{\rm{LD}}/M_*^{\rm{SPS}}$ relative to a Salpeter IMF vs. lens strength $\sigma_{\rm{SIE}}$ for the SLACS lenses of \citet{TreuIMF} (black crosses) and for the Jackpot (red cross). Measurements of \citet{TreuIMF} are obtained assuming a NFW dark matter halo with fixed scale radius for the lensing and dynamics analysis.
{\em Bottom panel:} average slope of the total density profile $\gamma'$ vs. lens strength $\sigma_{\rm{SIE}}$ for the SLACS lenses of \citet{PaperX} (black crosses) and for the Jackpot (red cross).}
\label{comparison}
\end{figure}
\begin{acknowledgments}
Based on observations made with the NASA/ESA Hubble Space Telescope, obtained at the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS 5-26555. These observations are associated with programs 11701, 11202, and 10886. Support for those programs 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, Inc., under NASA contract NAS 5-26555.
Some of the data presented herein were obtained at the W.M. Keck Observatory, which is operated as a scientific partnership among the California Institute of Technology, the University of California and the National Aeronautics and Space Administration. The Observatory was made possible by the generous financial support of the W.M. Keck Foundation
The authors wish to recognize and acknowledge the very significant cultural role and reverence that the summit of Mauna Kea has always had within the indigenous Hawaiian community. We are most fortunate to have the opportunity to conduct observations from this mountain.
T.~Treu acknowledges support from the Packard Foundation through a Packard Research Fellowship.
R.~Gavazzi acknowledges support from the Centre National des Etudes Spatiales. P.~J.~Marshall was given support by the Royal Society in the form of a research fellowship.
\end{acknowledgments}
|
\section{Introduction}
\label{intro}
Models of top-quark condensation~\cite{Miransky:1988xi,Miransky:1989ds,Marciano:1989mj,Marciano:1989xd,Bardeen:1989ds} are particularly appealing models of electroweak symmetry breaking. These theories are relatively compact and have the feature of automatically generating a large top Yukawa interaction with a composite Higgs field that is a bound state of a top--anti-top pair. While the simplest model is plagued by naturalness issues, subsequent embeddings of top condensation in supersymmetric~\cite{Carena:1991ky} and strongly coupled models of electroweak symmetry breaking (EWSB)~\cite{Eichten:1979ah,Hill:1991at,Hill:1994hp} can reproduce the weak scale without excessive fine tuning. However, a combination of flavor~\cite{Buras:1982ff,Buchalla:1995dp,Burdman:2000in,Simmons:2001va} and electroweak precision constraints~\cite{Peskin:1990zt,Peskin:1991sw} have consistently put tension on implementations of top condensation within strongly coupled scenarios. For a review with extensive discussion of these issues and a complete citation list, see~\cite{Hill:2002ap}.
Recent focus on extra dimensional models of EWSB, particularly those constructed on geometrically warped backgrounds~\cite{Randall:1999ee,Randall:1999vf} has shed new light on naturalness issues of the electroweak sector, and how precision tests might be addressed in a weakly coupled framework. In this paper, we explore the possibility of embedding top condensation within an extra dimensional setup. Such models with warped geometry are expected to generate natural hierarchies of scales. In this paper, we explore the 5D Nambu--Jona-Lasinio (NJL) mechanism~\cite{Nambu:1961fr, Nambu:1961tp} in a flat space toy model, with the idea that many of the results will carry over to more realistic extra dimensional scenarios utilizing a warped compactification.
In calculating the 5D effective action for fermion--antifermion bound states, we renormalize a 5D Yukawa theory compactified on an interval. The running is supplemented by an ultraviolet (UV) ``composite" boundary condition at a scale $\Lambda_0$. At the UV boundary, which we take to be at an energy greater than the compactification scale $1/L$, the theory describes 5D fermions that interact via a four-fermion interaction which arises from unspecified UV dynamics, perhaps from physics above the cutoff due to the strong coupling limit of the extra dimensional model. In deconstruction models~\cite{ArkaniHamed:2001ca,Hill:2000mu}, where the extra dimension resolves into a product gauge structure at high energies, the four-fermion operator could arise as a result of the (unspecified) dynamics which breaks the product group structure down to the Standard Model (SM) at low energies. The four-fermion operator could also arise due to intrinsically 5D dynamics such as a spontaneously broken 5D gauge theory.
Top condensation has been studied in extra dimensional contexts previously~\cite{Dobrescu:1998dg,Cheng:1999bg,Rius:2001dd,Burdman:2007sx,Bai:2008gm}, although focus has typically been on the low-energy theory below the scale of compactification. Our analysis includes the effect of 5D running up to the scale associated with the four-fermion interaction, and gives predictions for a Kaluza-Klein (KK) tower of scalar bound states corresponding to a 5D composite field. Of particular interest are the form of and role played by brane localized terms generated by fermion loops. Other top condensation models that simultaneously generate the correct top and $W$-boson masses generally supplement top condensation with a seesaw mechanism~\cite{Dobrescu:1997nm,Chivukula:1998wd}. Features of our 5D construction are similar to those found in extra dimensional top see-saw models~\cite{Cheng:2001nh,He:2001fz}, in which the lightest KK excitations of the fermions play a key role in the formation of the condensate.
We begin with a review of the 4D NJL model, which we then extend to a 5D setup compactified on an interval, or equivalently, an $S_1/\mathbb{Z}_2$ orbifold. Extension of these methods to a compactified model is relatively straightforward, although there are some complications associated with performing quantum corrections in an extra dimensional model, which we discuss. We work in the fermion bubble approximation, valid as long as the scale associated with the four-fermion operator is below the scale at which any additional 5D interactions (i.e. gauge interactions of the SM) become strongly coupled. Section~\ref{sec:5Dloops} contains a study of the relevant fermion loop graphs in 5D flat space. We then calculate the 5D quantum effective action valid at low scales. Solving the scalar equations of motion in this effective theory determines whether or not a chiral symmetry breaking condensate is formed.
We calculate the resulting light fermion and scalar spectrum, requiring a weakly gauged $SU(2)_L \times U(1)_Y$ version of the model to reproduce the observed $W$-boson mass. We find that the top quark mass and $W$ mass constraints can be simultaneously satisfied by making an appropriate choice of the fermion bulk mass parameters. The lowest lying scalar fluctuation is found to be generically heavy, due primarily to a large effective quartic coupling generated in the model. Lighter values can be generated by going to larger $N_c$ or creating a larger hierarchy between the four-fermion scale and the compactification scale $\Lambda_0 L \gg 1$. The second of these choices is made at the expense of increased fine-tuning of the interaction strength associated with the four-fermion operator and reducing the validity of the fermion bubble approximation.
\section{Extending the NJL Model to 5D}
\label{sec:5DNJL}
A toy model for spontaneous breaking of chiral symmetry in four dimensions can be constructed with a low-energy effective theory of massless fermions supplemented with a single chirally symmetric four-fermion contact operator~\cite{Nambu:1961fr, Nambu:1961tp}. The Lagrangian for this model, valid at the scale $\Lambda$ is
\begin{equation}
\label{eq:4DNJLnophi}
\mathcal{L} = \bar{\psi} i \displaystyle{\not} \partial \psi + \frac{g^2}{4\Lambda^2} \left[ (\bar{\psi} \psi)^2- (\bar{\psi} \gamma^5 \psi)^2 \right]
\end{equation}
where $\psi$ is a 4-component massless Dirac fermion. The Lagrangian is invariant under independent chiral rotations of the left- and right-handed components of $\psi$.
In two component notation, utilizing a complex auxiliary scalar field $\phi$, we can re-write this Lagrangian as
\begin{equation}
\mathcal{L} = \bar{\psi}_L i \displaystyle{\not} \partial \psi_L +\bar{\psi}_R i \displaystyle{\not} \partial \psi_R + g \phi \bar{\psi}_L \psi_R + \mathrm{h.c.} - \Lambda^2 |\phi |^2.
\label{eq:4DNJL}
\end{equation}
The field $\phi$ carries chiral charge such that this Lagrangian has the same symmetry as Eq.~(\ref{eq:4DNJLnophi}).
Running down this theory from the scale $\Lambda$ to a low scale $\mu$, taking into account only fermion loops, one finds that the scalar field $\phi$ develops dynamics and a quartic interaction. The fermion loop contribution to the scalar mass$^2$ is negative, and for sufficiently strong coupling, $g$, the quantum corrections overcome the positive $\Lambda^2 | \phi |^2$ term. In this case, the scalar field then picks a vacuum expectation value (vev), and breaks the chiral symmetry of the theory.
This mechanism was posited as a method to spontaneously break the electro-weak gauge interactions, where the fermion bound state consisted of top/anti-top pairs~\cite{Bardeen:1989ds}. A particularly appealing feature of this construction is the presence of a quasi-infrared fixed point in the top Yukawa coupling which renders the top Yukawa relatively insensitive to the compositeness scale~\cite{Pendleton:1980as,Hill:1980sq}. Above this fixed point, the top Yukawa blows up in the UV, and the coupling is in the domain of attraction for this fixed point which resides at a value of $\lambda_t \sim 1$.
We consider a 5D version of the above model, in which there is a four-fermion operator that leads to a composite five dimensional scalar field. This operator must arise from some UV dynamics, as in the case of 4D top condensation models~\cite{Hill:1991at}. In this work, we do not specify this dynamics and focus on the mechanics of the renormalization of this theory. A model with better UV behavior is currently under investigation.
The theory at a high scale $\Lambda_0$ consists of two 5D Dirac fermions, $\Psi_L$ which contains a left-handed zero mode in the spectrum, and $\Psi_R$ which contains a right handed one. Other assignments are possible, and will have different IR structure, however this theory is the one that most easily generalizes to a standard model-like low-energy spectrum. In addition, the chiral symmetries of this model are identical to those in Eq.~(\ref{eq:4DNJL}). We write the action for the theory at the scale $\Lambda > 1/L$ as defined on a circle with perimeter $2L$:
\begin{equation}
S_\text{5D NJL} = \int d^4x \int_{-L}^L dz \bar{\Psi}_L \left(i \displaystyle{\not} \partial -M_L(z) \right) \Psi_L +\bar{\Psi}_R\left( i \displaystyle{\not} \partial - M_R (z) \right) \Psi_R+ \frac{g^2}{\Lambda_0^3} \bar{\Psi}_L \Psi_R \bar{\Psi}_R \Psi_L.
\label{eq:5DNJL}
\end{equation}
where $\displaystyle{\not} \partial \equiv \gamma^\mu \partial_\mu + i \gamma^5 \partial_z$ and all fields are assigned periodic boundary conditions. The spectrum of the theory is then reduced by performing the identification $z \leftrightarrow -z$ which restricts the physical region of the space to the interval $z \in [0,L]$. The field solutions that remain can be either odd or even under this identification, although all operators in the Lagrangian must be even. The orbifold assignments that produce the spectrum described above are:
\begin{equation}
\Psi_L (z) = -\gamma^5 \Psi_L (-z) \text{, and } \Psi_R (z) = \gamma^5 \Psi_R (-z).
\end{equation}
In order for the action to be invariant, the fermion mass terms must be odd under the orbifold assignment: $M_{L,R} (z) = - M_{L,R} (-z)$.
While this procedure is equivalent to beginning with an interval and assigning boundary conditions~\cite{Csaki:2003sh,Csaki:2005vy}, we show in Appendix~\ref{app:5Dtranslation} that the orbifold language allows a simple, intuitive explanation for the presence or lack of certain brane localized terms that are induced by quantum corrections.
We assume mass profiles which are constant in the physical region, discontinuously jumping at the orbifold boundaries to satisfy the boundary condition above:
\begin{equation}
M_{L,R} (z) = \left\{ \begin{array}{cl} +m_{L,R} & z > 0 \\ - m_{L,R} & z < 0. \end{array} \right.
\label{eq:fermmasses}
\end{equation}
The zero modes are then exponentially localized, with profiles given by:
\begin{eqnarray}
\Psi_L^0 (x;z) = \sqrt{\frac{m_L}{1-e^{-2 m_L L}} } e^{- m_L |z|} \nonumber \\
\Psi_R^0 (x;z) = \sqrt{\frac{m_R}{e^{2 m_R L}-1} } e^{ m_R |z|}.
\end{eqnarray}
In the 4D low-energy effective theory and ignoring quantum effects, the zero modes couple via a four-fermion operator that has a form identical to that of Eq.~(\ref{eq:4DNJL}), with effective four-fermion coupling given by an overlap of the zero mode wave functions:
\begin{equation}
\frac{g_{4D}^2}{\Lambda_\text{eff}^2} = \frac{g^2}{\Lambda_0^3} \frac{m_L m_R}{m_L-m_R} \left( \coth m_L L - \coth m_R L \right),
\end{equation}
which is exponentially suppressed in the case that both $m_L$ and $m_R$ are the same sign, and the LH and RH zero modes are localized on opposite boundaries of the physical region. We will show that scalar bound states and chiral symmetry breaking with scales well below the scale $1/L$ can still be obtained, regardless of this suppression. In the KK mode interpretation, these scalars are presumably relativistic deeply bound states of a combination of KK modes. This strongly suggests that a full 5D calculation including all KK modes below the cutoff $\Lambda_0$ should be performed in order to properly formulate the low-energy theory.
To analyze the IR behavior of this theory, we write the 5D four-fermion interaction in terms of a complex auxiliary field $\phi$. At the scale $\Lambda_0$, the theory is then a model of Yukawa interactions in which the scalar field has no dynamics:
\begin{align}
S_\text{5D NJL} = \int d^4x \int_{-L}^L dz &\bar{\Psi}_L \left( i \displaystyle{\not} \partial -M_L(z) \right) \Psi_L +\bar{\Psi}_R \left( i \displaystyle{\not} \partial - M_R(z) \right) \Psi_R \nonumber \\
& - \Lambda_0^2 | \phi |^2 + \frac{g}{\sqrt{\Lambda_0}} \phi \bar{\Psi}_L \Psi_R + \text{h.c.} \label{eq:auxL}
\end{align}
Integrating out the field $\phi$ reduces Eq.~(\ref{eq:auxL}) to Eq.~(\ref{eq:5DNJL}). The main calculation of this paper will be on running this effective Lagrangian down to a low scale $\mu < \frac{1}{L}$, and solving the low-energy equations of motion for the scalar field. We calculate the running in the ``fermion bubble" approximation, integrating out only the fermionic contribution to the scalar effective action. This approximation is the analog of re-summing the fermion ladder diagrams in the theory written down in Eq.~\ref{eq:5DNJL}.
\section{Quantum Corrections in 5D}
\label{sec:5Dloops}
In models with compactified extra dimensions, quantum corrections are complicated by the fact that momenta along the compactified directions are discrete while the 4D momenta span a continuum. In our model, momenta along the compactified coordinate are quantized in units of $n \pi/L$, where $L$ is the size of the physical region. In this section, we compute these quantum corrections for the Yukawa theory in Eq.~(\ref{eq:auxL}).
Quantum effects in extra dimensional models have been studied in some contexts, particularly for the running of gauge couplings~\cite{Lewandowski:2002rf,Lewandowski:2004yr,Contino:2002kc}. Such calculations are often made simpler due to gauge invariance, which ensures that calculating the running of the coupling of the zero mode gauge field, which has a constant extra dimensional profile, is sufficient to describe all running effects in 5D. Our analysis of a 5D Yukawa theory must be intrinsically five dimensional, taking into account all possible external scalar states, since there is no such underlying symmetry which keeps the lowest lying mode flat.
In determining the quantum effects of the 5D theory, there is the approach of determining the KK spectrum, integrating out the extra dimension, and then truncating the effects of the tower at the desired level of accuracy. It is then a matter of computing usual 4D Feynman diagrams using these few KK modes. This approach, however, obscures 5D translation invariance, and is in fact quite complicated if more than a couple KK modes are included. This is especially the case in this construction, since there are a large number of possible scalar bound states. When the equation of motion is applied on the scalar field $\phi$ at the scale $\Lambda_0$, and the fermions are expanded in terms of their KK towers, we find:
\begin{equation}
\phi =\frac{g}{\Lambda_0^{5/2}} \bar{\Psi}_R \Psi_L = \frac{g}{\Lambda_0^{5/2}} \sum_{m,n} \bar{\psi}^m_R \psi^n_L.
\end{equation}
Quantum effects below the scale $\Lambda_0$ mix these fermion bi-linears with each other, and the effective action must then be re-diagonalized. It is much simpler and perhaps more illuminating to instead compute all quantum effects from the 5D viewpoint, and then solve the resulting 5D scalar equation of motion.
The most straightforward method is to compute all quantum corrections in momentum space, where the effects of orbifolding are taken into account in the form of the propagators. Either a hard momentum cutoff or dimensional regularization may then be used to study the divergence structure of the theory. The first of these is most suited to the 5D NJL model, since it explicitly contains information about power law divergences. Dimensional regularization, on the other hand, automatically subtracts these, leaving only poles corresponding to logarithmic divergences. We study both regulators, the former because it applies well to models with an explicit cutoff, and the latter since it is a point of interest to see how the 5D divergence structure, which contains no bulk log divergences, is obtained from the 4D KK tower which contains an infinite number of them.
It is, in principle, possible to use a mixed position-momentum space basis, where the propagators depend on the position in the extra dimensional coordinate, however in this case it is unclear how one would implement a regularization procedure which respects local 5D Lorentz invariance.
\subsection{Quantum corrections with vanishing fermion bulk masses}
In the case that the bulk fermion masses vanish, the fermion propagators are not difficult to compute. The Yukawa theory under consideration is then similar to the one examined in~\cite{Georgi:2000ks}, but with slightly different orbifold assignments and field content. In this section we utilize the notation of these authors. In particular, a derivation of the fermion propagators can be found in Section 2 of that publication.
In 5D momentum space, the fermion propagators are given by:
\begin{equation}
S_F^{(L, R)} (p ; p_5 , p'_5) = (2 L) \frac{i}{2} \left\{ \frac{ \delta_{p_5, p'_5}}{\displaystyle{\not} p + i \gamma^5 p_5} \pm \frac{ \delta_{-p_5, p'_5}}{\displaystyle{\not} p + i \gamma^5 p_5} \gamma^5 \right\}
\end{equation}
where the $+$ is for a 5D fermion in which a left-handed zero mode survives the orbifold projection, and the $-$ is for a 5D fermion which contains a right-handed zero mode in the spectrum\footnote{We have chosen a convention in which the period of the Fourier series appears in the Kronecker-$\delta$s of momentum ($2 L~\delta_{p_5,k_5}$), and in sums over unconstrained 5D momenta ($\frac{1}{2L} \sum_{k_5}$). This makes it simpler to compare with the (mostly) standard treatment in non-compact dimensions where the transformation to momentum space comes with a $\frac{1}{2 \pi}$ normalization. The dictionary between the compact and non-compact 5D theory consists of replacing sums with integrals, Kronecker-$\delta$s with $\delta$-functions, and all factors of $2L$ with $2 \pi$.}. The 5D momentum is given by $p_5 = \frac{n \pi}{L}$, where $n$ ranges over all integers. The fermion propagators conserve the magnitude of the 5D momentum, but only up to a sign. The breaking of 5D translation invariance is a manifestation of the reflection conditions at the orbifold fixed points. The remaining conservation of KK number is a tree level symmetry of the theory that is present in the limit of vanishing bulk mass.
We are interested in computing the scalar two- and four-point functions. Since interaction terms in extra dimensional theories are non-renormalizable, higher dimensional operators will be generated as well. For the purposes of illustration in this toy model, we ignore these contributions. One could, in principle, arrange for these terms to be removed via fine tuning of the coefficients of such operators against the quantum corrections to them. This tuning should then presumably be derived as a natural consequence of some UV complete model.
\begin{center}
{\bf The scalar two-point function}
\end{center}
In the massless fermion bubble approximation, the scalar two-point function at one loop consists of the diagram shown in Figure~\ref{fig:twopoint}. In the compactified 5D theory, this single diagram encapsulates the quantum corrections to the bulk kinetic and mass terms. In addition, it also contains information about brane localized terms which are quadratic in the scalar field. This diagram gives information about how to run the scalar sector of the Yukawa theory from the high scale $\Lambda_0$ down to low energies.
\begin{figure}[h]
\center{\includegraphics[width=4in]{5Dtwopoint}}
\caption{The 5D scalar two point function, where the scalar couples to two flavors of 5D Dirac fields, each of which contains either LH and RH zero mode in the KK mode spectrum.}
\label{fig:twopoint}
\end{figure}
The value for the diagram is
\begin{eqnarray}
&&- \frac{g^2}{\Lambda_0} \sum_{k_5, k'_5} \int \frac{d^d k}{(2\pi)^d}\mathrm{Tr~}\left[ \frac{ (\slashed{k} + i \gamma^5 k_5 ) ( \delta_{k_5, k'_5} - \gamma^5 \delta_{k_5, -k'_5} ) }{k^2 - k_5^2} \right. \nonumber \\
&& \left. \cdot \frac{ (\slashed{k}+\slashed{p} + i \gamma^5 [ k'_5 +p_5' ] ) ( \delta_{k_5+p_5, k'_5+p'_5} + \gamma^5 \delta_{k_5+p_5, -k'_5-p'_5} ) }{(k+p)^2 - (k'_5+p'_5)^2} \right].
\end{eqnarray}
Let us first discuss brane localized divergences of the two-point diagram. In extra dimensional theories, it is now well known that quantum effects generally violate KK-number conservation~\cite{Georgi:2000ks,Cheng:2002iz,Carena:2002me}. The presence of brane localized terms can be identified by divergences which do not conserve 5D momenta. Such divergences signal that a counterterm is necessary, and that the brane term should be included in the tree level action. Expanding the numerator of the diagram and simplifying the Kronecker-$\delta$s, there are in principle terms proportional to $\delta_{p_5, p_5'}$, $\delta_{-p_5, p_5'}$, $\delta_{2 k_5, -p_5-p'_5}$, and $\delta_{2 k_5, p'_5-p'_5}$. The first two types of terms conserve 5D momentum up to a sign and hence correspond to bulk corrections, while the second two Fourier transform into $\delta$-functions at the brane positions and so correspond to brane localized terms.
Applying the usual Dirac trace identities, the brane localized terms vanish. This is perhaps somewhat surprising at first glance. One might expect that there are brane localized quadratic divergences which renormalize the scalar mass independently on the branes versus in the bulk. One might also expect the generation of brane localized kinetic terms for the scalar field. The reason for the absence of such terms at the one-loop level is that 5D translation invariance is not broken severely enough in this process, as explained in Appendix~\ref{app:5Dtranslation}. In fact, there are a variety of scenarios in which brane localized terms are not generated at the one-loop level.
Let us now identify the bulk renormalization terms. We expect a cubically divergent mass renormalization, and a linear divergence in the 5D kinetic terms. One of the bulk renormalization terms is proportional to $\delta_{p_5,p_5'}$, the other $\delta_{p_5,-p_5'}$ (effectively reflected and transmitted waves through the orbifold fixed points).
From the trace, these have the following momentum structure:
\begin{equation}
\frac{k \cdot (k+p) - k_5 (k_5+p_5')}{\left(k^2-k_5^2 \right) \left( (k+p)^2 - (k_5+p_5')^2 \right)}
\end{equation}
The $k_5$ are quantized on $k_5 = n \pi/L$, with $n$ any integer. This means that the 5D sum cannot be shifted, while the 4D momenta can be redefined in the usual way in order to make the Wick rotated integrand spherically symmetric in Euclidean momentum.
The coefficients of the $\delta_{p_5, p'_5}$ and $\delta_{p_5, -p'_5}$ terms are identical. After combining denominators using Feynman parameters, they are given by:
\begin{equation}
- \frac{g^2}{4 \Lambda_0} \sum_{k_5} \int \frac{d^d k}{(2\pi)^d} \int_0^1 dx \frac{(l^2 - l_5^2)-x (1-x) (p^2-p_5^2) +l_5 p_5 (2x-1)}{[(l^2-l_5^2) + x (1-x) (p^2-p_5^2)]^2},
\end{equation}
where $l_5 = k_5+ x p_5$. Unfortunately, one cannot shift the 5D momentum in the sum this way since $l_5$ is not quantized on the same spectrum as $k_5$ and the above expression is only a heuristic presentation.
This lack of shift invariance highlights the fact that a naive hard cutoff for the 4D momentum integrals obscures the underlying physics. Such a procedure explicitly violates 5D Lorentz invariance, and will lead to apparent violation of the spacetime symmetries by short-distance interactions. For example, if one performs the sum over \emph{all} unconstrained five-momenta, one obtains an analytic expression as a function of the 4D loop momentum. The remaining integrand can then be performed with a hard cutoff, expanded in small external momenta, and then interpreted as a contribution to the effective action. The resulting expression contains terms proportional to $p_\mu^2$ and $p_5^2$ with coefficients which differ in general. 5D Lorentz invariance can then be restored by fine tuning separate counter terms order by order in perturbation theory, but the connection with the original 5D theory defined at the physical scale $\Lambda_0$ is then lost. To properly formulate the low energy dynamics, one must choose the regulator more carefully.
We first perform the integration utilizing dimensional regularization. Since there is no explicit cutoff scale, there are no subtleties about the regularization procedure respecting local 5D Lorentz invariance. Performing the 4D momentum integration first, we have
\begin{equation}
i \Pi (p^2, p_5) = - i \frac{g^2}{4 \Lambda_0} \sum_{k_5} \int_0^1 dx \frac{\Delta^{d/2-2}}{(4\pi)^{d/2}} \left\{ \frac{d}{2} \Delta \Gamma (1-d/2) + \left[ x (1-x) p^2 + k_5^2+p_5 k_5 \right] \Gamma (2-d/2) \right\}
\end{equation}
where $\Delta$ is given by:
\begin{equation}
\Delta = - x (1-x) (p^2 - p_5^2) + (k_5 + x p_5)^2.
\end{equation}
Using zeta-function regularization for the remaining sum over 5D internal loop momentum we have
\begin{equation}
i \Pi(p^2, p_5^2) = i\Pi (0) -\frac{i g^2}{8 \Lambda (4 \pi)^{d/2}} \left(\frac{\pi}{L} \right)^{4-d} \left(2 \zeta (4-d)+(\mu_\text{IR} L)^{d-4} \right) \Gamma (2-d/2) \left[ p^2 + p_5^2 \left( 2- d \right) \right].
\end{equation}
We have regulated the contribution of the zero mode with an IR cutoff, $\mu_\text{IR}$. The two point function for vanishing external momentum, $i\Pi(0)$, is given by:
\begin{equation}
i \Pi(0) = - i \frac{g^2}{4 \Lambda (4 \pi)^{d/2}} \left(\frac{ \pi}{L} \right)^{d-2} \zeta (2-d) \Gamma (1-d/2)
\end{equation}
Taking the limit as $d \rightarrow 4$, with $\epsilon_\text{IR} \equiv \mu_\text{IR} L$, we have the final result:
\begin{equation}
i \Pi(p^2, p_5^2) = \frac{i g^2}{4 \Lambda (4 \pi)^{2}} \left[ 2 \left(\frac{ \pi}{L} \right)^{2} \zeta'(-2) + \log (2 \pi \epsilon_\text{IR}) \left( p^2 - 2 p_5^2 \right) \right]
\end{equation}
Let us point out some aspects of these results: First, all expressions are finite as $d \rightarrow 4$. For the field strength term, the pole in the $\Gamma$ function is canceled by the sum of the zeta function and the contribution of the zero mode. That is, the UV divergences created by the zero mode are canceled by the UV divergences of the tower of KK modes. Second, note that the coefficient of the $p^2$ and $p_5^2$ terms differ in the limit $d \rightarrow 4$. These finite terms correspond to non-local contributions to violations of 5D translation invariance from the presence of the orbifold fixed points.
The finiteness of the result in this regularization scheme is expected. Since all divergences must be local, the UV structure of the bulk compactified theory should match that of the uncompactified model. All divergences in noncompact odd dimensions are power laws and are automatically subtracted when using dimensional regularization. So both the compact and uncompact models yield finite results for the two-point function in this regularization scheme.
It is possible to utilize a hard cutoff regularization scheme which respects the local spacetime symmetries. This is beneficial, since such a scheme has a better physical interpretation in terms of our physical cutoff, $\Lambda_0$. The procedure is described in detail in Appendix~\ref{app:euler-maclaurin}, but in many cases it consists simply of approximating the sum over momenta by an integral, at which point the integrand is manifestly 5D Lorentz invariant, and integration over the interior of a four-sphere in the loop momentum can be performed in the standard way. The substitution required is $\frac{1}{2L} \sum_{k_5} \rightarrow \int \frac{dk_5}{2 \pi}$.
The two-point function in this regularization scheme is then
\begin{equation}
i \Pi (p^2; p_5,p_5') = \left( \delta_{p_5,p_5'} + \delta_{p_5,-p_5'} \right) \frac{g^2L}{2 \Lambda_0} \int \frac{d^5k}{(2 \pi)^5} \int_0^1 dx \frac{(l^2 - l_5^2)-x (1-x) (p^2-p_5^2) +l_5 p_5 (2x-1)}{[(l^2-l_5^2) + x (1-x) (p^2-p_5^2)]^2},
\end{equation}
and we can now shift the full 5D loop momentum in the usual way, and use a 5D hard cutoff $\Lambda$. The result, as an expansion in $P^2 = p^2 - p_5^2$, is given by
\begin{equation}
i \Pi (p^2; p_5,p_5') = i L \left( \delta_{p_5,p_5'} + \delta_{p_5,-p_5'} \right) \left[ \frac{g^2 \Lambda^3}{18 \pi^3 \Lambda_0} + \frac{g^2 \Lambda}{10 \pi^3 \Lambda_0} P^2 \right]
\equiv L \left( \delta_{p_5,p_5'} + \delta_{p_5,-p_5'} \right) i \tilde{\Pi} (P^2).
\label{eq:2pthardcutoff}
\end{equation}
We have kept $\Lambda_0$ separate from the regulator cutoff in this expression to highlight the sensitivity to an arbitrary UV scale, although we take them to be equal in our final expression for the effective action. Implicit in Eq.~(\ref{eq:2pthardcutoff}) is an IR scale, $\mu \ll \Lambda$, which can be put into the effective action with the replacements $\Lambda^n \rightarrow \Lambda_0^n - \mu^n$.
\begin{center}
{\bf The scalar four-point function}
\end{center}
\begin{figure}[h]
\center{\includegraphics[width=4in]{4ptdiagram}}
\caption{The 5D scalar four point function.}
\label{fig:fourpoint}
\end{figure}
The quartic coupling also renormalizes, although we again find that all divergences are confined to the bulk. The relevant Feynman diagrams are shown in Figure~\ref{fig:fourpoint}, and evaluate to
\begin{align}
i V_4(0;p_5,p'_5,p''_5,p'''_5) = - \frac{g^4}{\Lambda_0^2} \sum_{ \substack{ k_5, k'_5, \\k''_5,k'''_5}} \int \frac{d^4k}{(2\pi)^4} \mathrm{Tr~} &\left[S^{R}_F(k;k_5,k_5'''+ p_5''') S^{L}_F(k;k_5''',k''_5 + p''_5) \right. \nonumber \\
& \times \left. S^{R}_F (k;k_5'',k_5' + p_5') S^L_F (k;k_5''',k_5'' + p_5'') \right].
\end{align}
Terms which contribute to bulk running of the quartic arise from an even number of insertions of the 5D momentum conserving Kronecker-$\delta$s while terms which contribute to brane running of the quartic involve an odd number of these. The potential brane terms each involve (at leading order in loop momenta) the trace of four identical Dirac matrices, $\slashed{k}$, with a $\gamma^5$, and therefore vanish.
Performing the calculation using dimensional regularization again produces a finite result, with KK modes canceling against the contribution of the zero modes. We only present the result utilizing a 5D Lorentz invariant hard cutoff.
We find
\begin{align}
i V_4(0;p_5,p'_5,p''_5,p'''_5) = \frac{- i g^4 \Lambda}{24 \pi^3 \Lambda_0^2} (2L) \sum_{\pm} \delta_{ 0, p_5 \pm p'_5 \pm p''_5 \pm p'''_5}.
\end{align}
Where the sum is over all 8 permutations of signs in the Kronecker-$\delta$.
To summarize the results of this section, we find that the bulk UV structure of the theory is as expected, where the running is purely power law. We have explicitly shown the cancellation of log divergences in the dimensional regularization scheme for the two-point function.
The one-loop brane localized divergence structure is different from naive expectations. Despite the intuition that brane localized terms should be forced by breaking translation invariance via the orbifold identification, they are not generated at one loop. As we discuss in Appendix~\ref{app:5Dtranslation}, this is due to the interplay of the left- and right-handed components of 5D fermions.
\subsection{Quantum corrections with fermion bulk masses}
The arguments that protect against brane localized terms fail when fermion mass terms are added into the theory. Under the orbifolding procedure, such masses must be odd under the projection since the fermion bilinears $\bar{\Psi} \Psi$ are odd. These masses could arise from a scalar domain wall to which the fermions are coupled via a Yukawa interaction. These domain walls are trapped at the orbifold fixed point by the orbifold quantum numbers of this scalar field and give rise to fermion localization in the extra dimension~\cite{Kaplan:1992bt,Mirabelli:1999ks,Kaplan:2001ga}. Because such fermion masses explicitly break 5D translation invariance at the orbifold fixed points, it is expected that they generate brane localized terms.
In this section, we calculate the quantum corrections in the presence of fermion bulk masses. These mass terms do not conserve even the magnitude of the 5D momenta so that the explicit form of the propagators in momentum space is rather complicated to compute. However, we can accurately capture the divergence structure of the theory by treating the 5D mass term as a perturbation to the massless scenario.
We take the fermion masses to have the profiles given in Eq.~(\ref{eq:fermmasses}). To obtain the Feynman rule in momentum space, we compute the Fourier series of the fermion mass terms in the action, and read off the interaction vertex. Since the mass term switches sign at the orbifold fixed points, its Fourier series is non-trivial. That is, the mass term acts as a source for 5D momentum which can be injected into a given diagram. The Feynman rule is:
\begin{equation}
\includegraphics[width=3in]{massinsrule},
\end{equation}
where
\begin{equation}
\delta^\text{odd}_{p_5,p'_5} \equiv \left\{ \begin{array}{cl} 1 & \text{if } p_5+p'_5 \text{ is an odd multiple of } \pi/L \\
0 & \text{if } p_5+p'_5 \text{ is an even multiple of } \pi/L. \end{array} \right.
\end{equation}
This is the familiar Fourier transform of the square wave function, with period $2L$.
The corrections to the scalar two-point function arise from two diagrams, one with a mass insertion on the fermion with a LH zero mode, the other with an insertion on the one with a RH zero mode.
\begin{equation}
\includegraphics[width=5in]{massinsdiags}.
\end{equation}
These contributions to the two-point function are linearly divergent:
\begin{equation}
i \Pi_M (0;p_5,p'_5) = i \frac{g^2 \Lambda}{3 \pi^3 \Lambda_0} \left( m_L - m_R \right) \delta^\text{odd}_{p_5,p'_5} + \text{finite terms}
\end{equation}
Adding a mass insertion diagram to the four-point function only contributes finite terms.
\section{The quantum effective action}
The two- and four-point diagrams we have calculated can now be incorporated into a quantum effective action that is valid at a low scale $\mu$. We can express this action as follows:
\begin{multline}
S_\text{effective} = \int d^4 x \int^L_{-L} dz \left[ \bar{\Psi}_L( i\displaystyle{\not} \partial-M_L(z) ) \Psi_L + \bar{\Psi}_R (i \displaystyle{\not} \partial-M_R(z) ) \Psi_R + \frac{g}{\sqrt{\Lambda_0}} H \bar{\Psi}_L \Psi_R + \text{h.c.} \right. \\
+ Z_H \partial_M H \partial^M H^\dagger - \left. \left( \Lambda_0^2 + \delta M^2 \right) |H|^2 - \frac{\lambda}{4 \Lambda_0} |H|^4 \right] \\
- \int d^4 x~\left[ m^2_{0}~|H(z=0) |^2 +m^2_{L}~|H(z=L) |^2 \right]. \\
\label{eq:effact}
\end{multline}
To map between our correlation functions and the terms in this effective action, we first note that each amplitude can be written in terms of projection operators $E_{p_5,p'_5} \equiv L \left( \delta_{p_5,p'_5} + \delta_{p_5,-p'_5} \right)$ acting on ``sub-amplitudes." The projection operators are the expression for dynamical external scalar legs when the scalar is even under the orbifolding procedure, $H(z) = H(-z)$. The sub-amplitudes represent Feynman rules arising from bulk and brane localized terms in the effective 5D action.
For the bulk contributions to the two-point function, we have
\begin{align}
i \Pi (p^2;p_5,p'_5) &= E_{p_5,p'_5} i \tilde{\Pi} (P^2) \nonumber \\
&= \frac{1}{2L} \sum_{q_5} E_{p_5,q_5} E_{q_5,p'_5} i \tilde{\Pi} (Q^2).
\end{align}
The contribution arising from the bulk mass insertion diagrams is
\begin{equation}
i \Pi_M (0;p_5,p'_5) = i \tilde{\Pi}_M \delta^{\text{odd}}_{p_5,p'_5} = i \tilde{\Pi}_M \left( \frac{1}{2L}\right)^2 \sum_{q_5,q'_5} E_{p_5,q_5} E_{p'_5,q'_5} \delta^{\text{odd}}_{q_5,q'_5}.
\end{equation}
We can identify $Z_H \equiv \tilde{\Pi}' (Q^2 = 0)$, and $\delta M^2 \equiv -\tilde{\Pi} ( Q^2 = 0)$. The mass insertion diagrams need to be Fourier transformed back into position space. We use the identities
\begin{align}
\sum_{p_5~\text{odd}} e^{i p_5 z} &= L \sum_N (-1)^N \delta( z- N L) \nonumber \\
\sum_{p_5~\text{even}} e^{i p_5 z} &= L \sum_N \delta( z- N L)
\end{align}
where the sum over $N$ spans all integers. The Fourier transform thus corresponds to opposite sign $\delta$-functions on the two branes, $\delta_{q_5,q'_5}^\text{odd} \rightarrow \frac{1}{2} \left[ \delta(z) - \delta(z-L)\right]$. The brane localized mass terms are then $m_0^2 = -m_L^2 = - \tilde{\Pi}_M/2$. Finally, the four-point function can be expressed as
\begin{align}
i V_4 (0;p_5,p'_5,p''_5,p'''_5) &= i \frac{\tilde{V}_4}{8} \sum_{\pm} \delta_{0,p_5\pm p'_5 \pm p''_5 \pm p'''_5} \nonumber \\
&= i \left( \frac{1}{2L} \right)^4 \sum_{q_5,q'_5,q''_5,q'''_5} E_{p_5,q_5} E_{p'_5,q'_5} E_{p''_5,q''_5} E_{p'''_5,q'''_5} \tilde{V}_4~\delta_{0,q_5+ q'_5 + q''_5 + q'''_5}.
\end{align}
and we make the identification $\tilde{V}_4 = \frac{\lambda}{\Lambda_0}$.
In summary, the effective action can be expressed as a function of the UV parameters as in Eq.~(\ref{eq:effact}) with coefficients given by
\begin{align}
Z_H &= \frac{N_c g^2}{10 \pi^3}\frac{\Lambda}{\Lambda_0} \nonumber \\
\delta M^2 &= - \frac{N_c g^2}{18 \pi^3} \frac{\Lambda^3}{\Lambda_0} \nonumber \\
\lambda &= \frac{N_c g^4}{3 \pi^3} \frac{\Lambda}{\Lambda_0} \nonumber \\
m^2_0 =- m^2_L & = \frac{N_c g^2}{6 \pi^3} \frac{\Lambda}{\Lambda_0} (m_R - m_L).
\end{align}
We now associate the regulator cutoff $\Lambda$ with the physical scale $\Lambda_0$. By defining the coupling constants such that they are dimensionless, with the physical scale explicitly appearing in the interaction terms, the quantum corrections (with the exception of the bulk mass term) are all seen to be independent of the scale $\Lambda_0$.
It is interesting that the scalar mass$^2$ receives brane localized contributions of opposite sign on either brane. This is a severe violation of KK parity. If this parity were preserved, the two brane localized terms are expected to be identical. However, the fermion mass terms explicitly violate KK parity. Quantum effects transmit this breaking of KK parity to the scalar sector in the form of these linear divergences.
These opposite sign, one loop, brane localized terms vanish, however, when the fermion masses are taken to be identical. In this scenario, for positive bulk masses, the LH zero mode is localized on the $z=0$ brane, whereas the RH zero mode is localized on the $z=L$ brane. If the masses are equal, then the profiles are mirror images of each other, and an ``accidental" approximate KK parity is introduced.
We now choose a convenient normalization for the 5D fields. We choose a canonical 5D scalar kinetic term, obtained by redefining $H \rightarrow H/\sqrt{Z_H}$,
\begin{align}
S = \int d^4 x &\int^L_{-L} dz \left[ \bar{\Psi}_L \left( i\displaystyle{\not} \partial -M_L (z) \right) \Psi_L + \bar{\Psi}_R \left( i \displaystyle{\not} \partial -M_R (z) \right) t_R + \frac{\tilde{g}}{\sqrt{\Lambda_0}} H \bar{\Psi}_L \Psi_R + \text{h.c.} \right. \nonumber \\
&\left. +\partial_M H \partial^M H^\dagger - \tilde{m}^2 |H|^2 - \frac{\tilde{\lambda}}{4 \Lambda_0} |H|^4 \right] - \int d^4 x \left[ \tilde{m}^2_{0} \left. |H|^2 \right|_{z=0} + \tilde{m}^2_{L} \left. |H|^2 \right|_{z=L} \right].
\end{align}
The terms in this 5D effective theory are
\begin{align}
\tilde{g}^2 &= \frac{ 10 \pi^3}{N_c} \nonumber \\
\tilde{m}^2 &= \left( \frac{10 \pi^3 }{ N_c g^2}- \frac{5}{9} \right) \Lambda_0^2 \nonumber \\
\tilde{\lambda} &= \frac{100 \pi^3}{3 N_c} \nonumber \\
\tilde{m}^2_0 =-\tilde{m}^2_L & = \frac{5}{3} (m_R - m_L).
\label{eq:finalresult}
\end{align}
Above, we have assumed $\Lambda \gg \mu$, where $\Lambda$ is the scale that our original Lagrangian with the four-fermion operator was defined, and $\mu$ is the low scale at which we evaluate our 5D effective action.
There are also finite non-local contributions that arise from quantum corrections. We have neglected these, as they are typically sub-dominant, and do not have an interpretation as terms which are local in the extra dimensional coordinate.
We note that there are no brane localized quadratic divergences at one loop. Such terms might have been expected from considerations of the field content. In the fermion bubble approximation, brane localized terms arise only from diagrams with insertions of the 5D fermion mass, whose profile explicitly violates translation invariance.
In the presence of fermion bulk masses, the conditions under which the chiral symmetry of the low-energy theory is broken are modified. In the absence of the boundary terms, the scalar bound states condense for $g^2 > 18 \pi^3/N_c$. However, the brane localized mass terms can drive condensation as well. In the next section we explore the conditions for generation of a chiral symmetry breaking condensate, and the resulting spectrum of the theory.
\section{Vacuum Solution and Mass Spectrum}
We have now shown that the low-energy effective theory is one with an additional 5D composite scalar degree of freedom. The equations of motion and the boundary conditions for this scalar field can be derived from the effective action that we have calculated. These determine the spectrum of the theory.
At the high scale, the 5D scalar Higgs field is equivalent to the fermion bilinear $H(z,x) = \bar{\psi}_L (z,x) \psi_R (z,x)$. With the fermionic orbifold assignments we have made, the orbifold parity transformation of the composite field is
\begin{equation}
H (-z) = \bar{\psi}_L (-z) \psi_R (-z) = (- \bar{\psi}_L (-z) \gamma^5)( - \gamma^5\psi_R (-z) ) = \bar{\psi}_L (z) \psi_R (z) = H(z).
\end{equation}
The scalar field is thus orbifold even, which means that when deriving the equation of motion for $H$, we cannot require that the variation itself vanish on the branes. Rather, the Higgs field is sensitive to the brane localized mass terms.
In this model, chiral symmetry breaking can occur in one of two ways. First, the coupling constant associated with the four-fermion operator may be sufficiently large that the bulk mass term is driven negative, destabilizing the origin as a vacuum solution. The bulk quartic coupling then sets the value for the scalar vacuum expectation value.
The other possibility is that the scalar bulk mass$^2$ remains positive, but a negative brane localized mass term pushes the field value away from the origin. In this case, it is still the bulk quartic coupling that stabilizes the vacuum field solution away from the origin, since we have shown that no brane localized quartic coupling is induced.
The second solution is more interesting, as it distinguishes the behavior of the compact 5D model from the non-compact one. Unlike the scalar bulk mass, the brane localized terms are sensitive to the values of the fermion bulk mass terms (and thus the relative localization of the fermion zero modes). Whether chiral symmetry breaking occurs in the extra dimensional model is thus a function of the free parameters of the model.
We now consider solutions to the composite scalar equations of motion. In the bulk, the vacuum equation for $\langle H(z,x) \rangle \equiv v(z)/(2 \sqrt{ L})$ is given by:
\begin{equation}
v''(z) = \tilde{m}^2 v(z) + \frac{\tilde{\lambda}}{8 \Lambda_0 L} v^3(z).
\end{equation}
This differential equation can be solved in terms of a Jacobi elliptic function, $\text{sc} (x | m)$. The expression for the vacuum expecation value (vev) is
\begin{equation}
v(z) = \sqrt{\frac{8 \Lambda_0 L \kappa_-}{\tilde{\lambda}}} \text{sc} \left( \left. | z - z_0 | \sqrt{\frac{\kappa_+}{2}} \right| 1- \frac{\kappa_-}{\kappa_+} \right),
\end{equation}
where we have introduced the dimensionless quantities $\kappa_\pm = \tilde{m} ^2 \pm \sqrt{ \tilde{m}^4- \frac{\tilde{\lambda} \tilde{m}^2 v_0^2}{4 \Lambda_0 L}}$. The quantities $z_0$ and $v_0$ are determined by imposing the boundary conditions. In order for the low-energy chiral symmetry to be broken, the vacuum energy for the scalar field must be minimized at a non-trivial value for $v_0$.
The only brane localized terms which survive in the large cutoff limit are scalar mass terms proportional to the difference in bulk fermion masses. These are shown in Eq.~(\ref{eq:finalresult}). These mass terms, $\tilde{m}_0^2$ and $\tilde{m}_L^2$, set the boundary conditions for the scalar vev equation:
\begin{equation}
\left. \frac{v'(z)}{v(z)} \right|_{z=0} = \frac{1}{2} \tilde{m}_0^2 ~~~~~~~~~\left. \frac{v'(z)}{v(z)} \right|_{z=L}= -\frac{1}{2} \tilde{m}_L^2.
\end{equation}
We can analytically determine the phase boundary by expanding the solution about small $v_0$. The result is $v(z) \approx v_0 \sinh( | z - z_0 | \tilde{m} )$, and the boundary conditions are then:
\begin{align}
&\left.\frac{v'(z)}{v(z)} \right|_{z=0} = \tilde{m} \coth ( |z_0| m ) = \frac{5}{6} (m_R - m_L) \nonumber \\
&\left.\frac{v'(z)}{v(z)} \right|_{z=L} = \tilde{m} \coth ( |L-z_0| m ) = \frac{5}{6} (m_R - m_L).
\end{align}
These are satisfied for $z_0 \rightarrow -\infty$, and for $\tilde{m} = \frac{5}{6} (m_R-m_L)$. We can express this phase boundary in terms of the original four-Fermi coupling $g$, which determines $\tilde{m}$ in the low-energy theory. The critical coupling is found to be:
\begin{equation}
g_\text{crit}^2 = \frac{18 \pi^3 }{N_c} \left[ 1 + \frac{5}{4} \frac{ (m_R-m_L)^2 }{\Lambda_0^2} \right]^{-1}.
\end{equation}
We now scan the parameter space of the model. For these purposes, we presume that the fermions are the 5D analogs of the LH third generation doublet and the RH top quark. In this case, the scalar field then carries the $SU(2)_L \times U(1)_Y$ quantum numbers of a SM Higgs, and when $H$ obtains a vev, the $W$ and $Z$ bosons become massive. We identify the region of parameter space in which we obtain the correct $W$-boson and top quark masses.
The $W$ mass is well approximated by assuming a flat profile for the lightest $W$-boson mode, and convoluting the flat profile with the vev$^2$:
\begin{equation}
m_W^2 = \frac{g_2^2}{4} \left[ \left(\frac{1}{2L}\right) \int_{-L}^L dz~v(z)^2 \right],
\end{equation}
where $g_2$ is the $SU(2)_L$ gauge coupling of the SM. The top quark mass is approximated from the Yukawa interaction:
\begin{equation}
m_\text{top} = \frac{\tilde{g}}{\sqrt{\Lambda_0 L}} (N_R N_L L ) \left[ \frac{1}{2L} \int_{-L}^L dz v(z) e^{(m_R - m_L) |z|}\right]
\end{equation}
where $N_{R(L)}$ are the normalization factors for the fermion zero mode profiles, $\Psi_L (z) = N_L e^{-m_L |z|}$, and $\Psi_R (z) = N_R e^{m_R |z| }$. Note that the $W$ mass depends only the difference between the fermion bulk mass terms (through the effective Higgs potential), while the top quark mass has a quite different dependence arising from the fermion normalization parameters. The $W$ and top quark masses are thus independently adjustable.
The phase boundary is shown in Figure~\ref{fig:phaseplot} along with contours of $m_W$ as a function of the original four-fermion coupling $g$ and the difference between the fermion bulk mass parameters $|m_R-m_L|$. We have set the other free parameters to $N_c=3,\text{ and }\Lambda_0 L = 10$. In Table~\ref{tab:spectrum}, values of $m_L$, $m_R$ and $L$ which give the correct top and $W$ mass are shown, along with the associated value for the Higgs mass. Additionally, we quote the value of $g^2/g_\text{crit}^2-1$, a rough measure of the fine tuning necessary in the four-fermion coupling to achieve the correct $W$-mass.
\begin{figure}[t]
\center{\includegraphics[width=4in]{phaseplot}}
\caption{The phase boundary of the model is shown, as a function of $|m_L - m_R|$ and the four-fermion coupling $g$. The size of the extra dimension is $L = 1$ TeV$^{-1}$, and $N_c = 3$. Thin solid black lines indicate contours (moving outwards) of $m_W \sim 40,~160,~320$ GeV, while the thick blue line corresponds to $m_W \sim 80$ GeV.}
\label{fig:phaseplot}
\end{figure}
We see that for the choice $\Lambda_0 L = 10$, $N_c = 3$, the Higgs is very massive. In fact, it is above the perturbative unitarity bound. This can be alleviated by increasing $\Lambda_0 L$, although the fermion ladder approximation begins to break down as $\Lambda_0$ approaches the scale at which the 5D gauge interactions become strong (about $\Lambda L \sim 30$).
\begin{table}[h]
\caption{Choices of the fermion bulk mass parameters that reproduce the SM values for $m_W$ and $m_\text{top}$, and their associated predictions for the Higgs mass. In the third column, we give a rough measure of the fine tuning necessary to achieve the weak scale from the 5D four-fermion interaction. All dimensionful parameters are given in units of TeV. We have set the other free parameters to $N_c=3,~\Lambda_0 L = 10$.}
\begin{center}
\begin{tabular}{|c|c|c|c||c|}
\hline
$m_L$ & $m_R$ & $L^{-1}$ & $g^2/g_\text{crit}^2-1$ & $m_\text{Higgs}$ \\
\hline
9.1 & 18.8 & 2 & 0.0035 & 1.4 \\
\hline
9.2 & 17.1 & 2 & 0.0031 & 1.25 \\
\hline
9.45 & 15.2 & 2 & 0.0025 & 1.1 \\
\hline
10. & 12.7 & 2 & 0.0016 & 0.85 \\
\hline
\hline
4.5 & 9.5 & 1 & 0.014 & 1.4 \\
\hline
4.6 & 8.7 & 1 & 0.012 & 1.3 \\
\hline
4.7 & 7.7 & 1 & 0.010 & 1.1 \\
\hline
5.0 & 6.5 & 1 & 0.006 & 0.85 \\
\hline
\hline
2.25 & 5.0 & 0.5 & 0.06 & 1.4 \\
\hline
2.3 & 4.5 & 0.5 & 0.053 & 1.3 \\
\hline
2.25 & 5.4 & 0.5 & 0.045 & 1.1 \\
\hline
2.4 & 3.4 & 0.5 & 0.030 & 0.9 \\
\hline
\end{tabular}
\end{center}
\label{tab:spectrum}
\end{table}
\section{Conclusions}
\label{sec:conclusions}
We have considered a compactified 5D version of a Nambu--Jona-Lasinio model. The model is studied by computing quantum corrections to a 5D Yukawa theory in which there are two species of fermions, each with a fermionic zero mode in the spectrum with opposite chiralities. The scalar field is interpreted as a bound state of the two fermion species. The classical 4D effective theory at low energies exhibits a chiral symmetry. Supplementation of the model by a 5D UV composite boundary condition renders the model equivalent at the high scale to one with a 5D bulk four-fermion operator. The quantum corrections to the low-energy Yukawa model are equivalent to a re-summation of fermion bubble diagrams in the fermion four-point function arising from the four-fermion interaction.
Both bulk and brane localized divergences are generated, although the brane localized divergences are softer than might have been expected. An accidental remnant of 5D translation invariance on the parent $S_1$ space survives, and protects against one-loop quadratically divergent contributions to the scalar mass$^2$ terms on the branes. In the presence of fermion bulk mass terms which explicitly violate translation invariance, linear divergences are generated. Under certain conditions, when the four-fermion coupling exceeds a critical value, these brane localized terms destabilize the scalar vacuum, and drive spontaneous chiral symmetry breaking.
If a portion of the chiral symmetry is weakly gauged, it is expected that this symmetry will be spontaneously broken, as in top condensation models. We numerically studied such a model, showing that it is possible to realize simultaneously the correct top quark and $W$-boson masses. This can be seen as an explicit 5D realization of top seesaw models, a deconstructed version of which was studied in~\cite{Cheng:2001nh,He:2001fz}. The Higgs mass is generically quite large in these models due to the large quartic coupling, likely in conflict with perturbative unitarity and/or electroweak precision constraints. A more realistic model implemented in warped space may alleviate both of these tensions.
\newpage
|
\section{Introduction}
Current understanding of the phase structure of strongly interacting field theories is still far from being complete, despite the fact that this problem has been at the center of active interest for many years.
The case of QCD is of particular physical relevance and many ongoing efforts are directed at mapping the geography of the phase diagram of the theory at finite temperature and density, understand the nature of the transitions between the different phases, locate the critical points, describe the morphology of the ground state.
Impediments to arrive at the desired complete characterization of the phase structure of QCD are mainly due to the notorious difficulties in performing {\it ab initio} lattice computations
under the influence of generic external conditions and, amongst the various approaches, the use of effective models is one of the most common playgrounds where the phase structure
can be discussed (see Refs.~\cite{review1,review2,review3,review4,review5,review6} for review). Within these effective models, the Nambu-Jona Lasinio (with its variants) is, possibly, the field theoretical set-up that received greatest attention.
One particular aspect that has been at the center of recent discussions is the identification of inhomogeneous phases. In fact, the appearance of modulations in strongly-coupled fermionic systems has led to several new insights concerning inhomogeneous phases in QCD. More precisely, several works indicated that, in the chiral limit, inhomogeneous phases are energetically preferred for relatively large values of the chemical potential, and the preferred inhomogeneous ground state in the vicinity of the chiral critical point assumes a crystalline structure similarly to superconductors. The problem has been analyzed in different models and seems to be quite generic (see, for example, \cite{nakano,nickel1}).
The inclusion of gauge degrees of freedom has also been discussed within models of the Nambu-Jona Lasinio class with the aid of vector interactions and Polyakov loops in Ref.~\cite{buballa2}.
The occurrence of inhomogeneous phases has also been studied in a number of different contexts including superconductivity (see for example Refs.~\cite{casalbuoni,alford,bowers,mannarelli,rajagopal,buballa,Casalbuoni:2005zp}), lower dimensional field theories and, in particular, the Gross-Neveu model with a large number of fermions (see \cite{basar1,basar2,basar3,schnetz,urlichs,boehmer,dgkl}) that allows for exact integrability. Earlier studies regarding inhomogeneous phases were also carried out in the context of the Skyrme model \cite{skyrme1,skyrme2,skyrme3,skyrme4,skyrme5} and in four-fermion models \cite{4fermi}.
Strongly coupled theories are physically relevant also in applications of astrophysical and cosmological nature (neutron stars, black holes, cosmological phase transitions).
In such cases it is natural to ask whether gravity may have some influence in the dynamics of phase transitions and, for this reason, the problem of generalizing strongly interacting fermion effective field theories to curved spacetimes triggered a large body of work (See Refs.~\cite{c1,c2,c3,c4,c5,c6,c7,c8,c9,c10,c11,c12,c13,c14,c15,c16} for a partial list of references).
In curved space, mostly for reasons of technical nature, attention was limited to consider homogeneous spacetimes and condensates, {\it i.e.} spacetime curvature and condensate were assumed not to vary in space. Even in this relatively simple situation, the interesting conclusion was that gravity may, in fact, affect the critical temperature $T_c$ of the theory. Once the theory is immersed in a spacetime of positive constant curvature, in the mean field and large-$N$ approximations, it was shown that, if temperature is fixed to some value $T_<$ smaller than the critical temperature $T_c$, chiral symmetry is broken. Keeping the temperature fixed to $T_<$ and increasing the value of the curvature, produces a phase transition into a chirally restored symmetry phase, indicating a clear analogy between thermodynamical and geometrical effects.
Although this analogy is formally interesting, in concrete physical situations (for example in the case of a neutron star) the value of the curvature is small and geometrical effects only play a marginal role. In such a case, the system can be well described as if the spacetime were effectively flat, limiting the interest of the study of effective strongly interacting theories in curved space. However, this is not a universal situation.
Clearly, geometrical effects become significant in the presence of strong gravitational sources. One natural example that comes to mind is that of black holes. This case, physically very interesting, presents several complications absent in the cases of homogeneous backgrounds. The most important one is related to the fact that approximating the condensate as a spatially constant function is simply not admissible.
Targeting a solution to this problem, we have recently initiated the analysis of inhomogeneous phases in curved spacetimes. In Ref.~\cite{flachi1}, we developed a formalism that quite naturally allows to deal with this problem in regular ultrastatic spacetimes (the assumption of ultrastaticity was simply motivated by the fact that the special geometric structure allows a direct use of the imaginary time formalism necessary to include finite temperature effects). With some additional efforts, in Ref.~\cite{flachi2}, we have analysed the case of black holes exteriors and shown that chiral phase transitions occur outside the event horizon. More precisely, if the black hole is surrounded by a gas of strongly interacting particles, an inhomogeneous condensate with a kink profile will form. Although the implementation was not straightforward, the idea is, in fact, simple. In a black hole geometry that is asymptotically flat, local quantities increase as the horizon is approached. If the asymptotic temperature, inversely proportional to the black hole mass, is lower than the critical temperature, then chiral symmetry is broken at infinity. According to the same logic, since the local temperature diverges near the horizon, we may expect that, at some finite distance from the horizon, a transition to a chirally symmetric phase may occur. This suggests that for black holes of certain mass in equilibrium with a gas of strongly interacting particles, a bubble of high temperature restored chiral symmetry phase should surround the hole. For an evaporating black hole the same holds, since asymptotically the temperature vanishes as $1/r^2$. This has been quantitatively demonstrated in \cite{flachi2} where self-consistent solutions for condensate have been constructed.
In all cases analysed so far (for both homogeneous and inhomogeneous situations), the spacetime geometry was taken to be regular (in Ref.~\cite{flachi1} we studied the case of regular ultrastatic manifolds, while in Ref.~\cite{flachi2} we focused on the region outside the event horizon that is also singularity free). The goal of this paper is to analyse more deeply the connection between the appearance of inhomogeneous phases and curvature effects, in particular when the geometry presents singularities.
Examples of this sort describe, for example, global monopoles spaces, geometries with wedges, string theory conifolds. Higher co-dimension brane models also present a similar structure \cite{kaloper}.
The first step of the present work is to compute the effective action for a strongly coupled fermion effective field theory on a conical spacetime. Since the geometry presents singularities, a suitable regularization procedure has to be used to deal with this situation. One possibility is to use the approach developed by Cheeger in Ref.~\cite{cheeger}, where the spectral geometry of differential operators on singular spaces has been discussed. Although the problem can be discussed quite generally, here, for simplicity, let us specify the set-up and consider a quantum field $\chi$ propagating on a $D$-dimensional spacetime $\mathscr{M}$ that presents a conical-type singularity. The effective action can be written in terms of the heat-kernel, $\mathscr{K}$, of some differential operator defined on the spacetime manifold in question. Schematically, we may write
\begin{eqnarray}
\mathscr{K}(t)\sim \sum_i \hat{\mathscr{C}}_{i/2} t^{i/2-D/2} + \mathscr{B} \ln t~,
\label{hka}
\end{eqnarray}\noindent
where $\hat{\mathscr{C}}_{i/2}$ are the heat-kernel coefficients and the quantity $\mathscr{B}$ is related to the value of the zeta function on the base manifold. In this case, the heat-kernel coefficients are understood as the principal part of integrals over the space manifold of geometric invariants and this is indicated by the hat. Using the above expansion, one may obtain the values $\zeta(0)$ and $\zeta'(0)$ related to the functional determinant of the operator in question. In general the zeta function may present poles and, as shown by Cheeger, by taking the principal part provides the appropriate treatment necessary in obtaining the asymptotic expansion of the heat-trace. In a sense, Cheeger's approach allows to deal directly with the singularity, however, in any practical implementation, truncating in the above expansion becomes necessary, and this affects the validity of the truncated expansion near the singularity.
Another possibility, easier to implement and that we will use here, is to resolve the singularity by excising a small region around it and substituting the excised region with a regular cap (See Fig.~\ref{figg1}) \footnote{A similar type of regularization has also been used in the context of higher co-dimension brane models \cite{kaloper,kaloper2} where the singularity was resolved by placing a brane at small distance from the singularity and by filling the interior by a regular cap.}. Adopting this way of regularization, allows to have a standard expansion for $\mathscr{K}(t)$, {\it i.e.} the coefficient $\mathscr{B}$ vanishes and the heat-kernel coefficients are the ordinary ones.
\FIGURE[h]{
\centering
\includegraphics[width=0.5\columnwidth]{figg1.eps}
\caption{Regularization by excision of the singularity. \vspace{.2cm}}
\label{figg1}
}
Adopting this regularization sche\-me, we will present the computation of the effective action for a strongly interacting fermion effective field theory on manifolds with conifold singularities in the next section.
The method is analogous to that presented in Ref.~\cite{flachi1}.
Once the explicit form of the effective action for the fermion condensate is obtained, we will proceed, in section \ref{sec3} with the numerical implementation and construct self-consistent and spatially non-trivial solutions for the condensate. In particular, we will see that even when the thermodynamical temperature is kept constant and below the critical value for the effective theory in flat space, chiral symmetry is always restored near the singularity. The condensate assumes a kink-like profile that vanishes near the singularity, thus separating the region of restored symmetry around the singularity from a region of broken symmetry.
Our conclusions will be presented in the last section.
\section{Condensate Effective Action in Conifold-type Geometries}
In this section, we will consider the simplest tractable type of geometric singularity: that of a {\it metric cone}. Given a $(d-1)$-dimensional Riemannian manifold $\mathscr{N}$, a (metric) cone, $\mathscr{C}_{\mathscr{N}}$, is defined as the space $\mathbb{R}^+ \times \mathscr{N}$. In the following we will assume that the base of the cone, $\mathscr{N}$, is compact and smooth. The formal part of the analysis will deal with this general case, while in the subsequent numerical implementation we will specify the form of the base. In local hyperspherical coordinates the line element is
\begin{eqnarray}
ds^2 = g_{ij}dx^idx^j=dr^2 +r^2 d\mathscr{N}^2~,
\label{conemetric}
\end{eqnarray}\noindent
with $r$ being a radial coordinate, $r \in \mathbb{R}^+$. This is the geometry of a infinitely long cone and it is singular at the apex, $r=0$, where a curvature singularity occurs. It is possible to cut the cone at some distance L, in which case the radial coordinate varies within the interval $I=\left[0,\mbox{L}\right]$.
Finite temperature effects may be introduced using the standard imaginary time formalism, thus considering the product of $S^1\times \mathscr{C}_{\mathscr{N}}$ where $S^1$ is a circle of radius $2\pi T$, and $T=1/\beta$ being the temperature in natural units. Anti-periodic boundary conditions will be imposed on the fermion fields on $S^1$. Generalization to scalars are almost trivial.
We will model the strongly interacting fermion effective field theory with an action of the form
\begin{eqnarray}
S= \int d^{d+1}x \sqrt{g} \left\{
\bar \psi i \gamma^\mu \nabla_\mu \psi
+
{\lambda\over 2N} \left(\bar \psi \psi\right)^2 + \cdots
\right\}~,
\label{action}
\end{eqnarray}\noindent
where the spinor $\psi$ has $(4 \times N_f\times N_c)$ components,
$N_f$ and $N_c$ are the number of flavors and colors respectively ($N\equiv N_f \times N_c$), and $\lambda$ is the coupling constant. The dots represent terms with higher mass dimension. The matrices $\gamma_\mu$ are the gamma matrices in curved space, and $g=|\mbox{Det} g_{ij}|$ with $g_{ij}$ given by (\ref{conemetric}). The theory is invariant under discrete chiral transformations, and possible phase transitions in relation to the breaking of the chiral symmetry can be discussed in terms of the appearance of a non vanishing condensate $\sigma = -{\lambda\over N}\langle \bar \psi \psi \rangle$: if the chiral symmetry is broken dynamically, $\sigma$ acquires a non-zero vacuum expectation value and a fermion mass term appears. The basics of the computation are the same as in our previous work Ref.~\cite{flachi1}, so, in recapitulating them, we will be brief.
After bosonization, the effective action at finite temperature (per fermion degree of freedom), $\Gamma$, can be expressed in the large-$N$ approximation as
\beq
\Gamma= - \int d^dx \sqrt{g} \left({\sigma^2\over 2\lambda}\right) +
\mbox{Tr} \ln \left( i \gamma^\mu \nabla_\mu - \sigma \right)~,
\label{effect}
\eeq
where the determinant acts on spinor and coordinate space. Squaring the Dirac operator gives
\begin{eqnarray}
\Gamma &=&-\int d^{d}x \sqrt{g} \left({\sigma^2\over 2\lambda}\right) + \delta \Gamma~,
\label{eff}
\end{eqnarray}\noindent
with
\begin{eqnarray}
\delta \Gamma = {1\over 2}\sum_{\epsilon =\pm}\mbox{Tr} \ln \left[
\square + {R\over 4} +\sigma^2 +\epsilon \left|\partial \sigma\right|
\right]~,
\label{diff}
\end{eqnarray}\noindent
where we have assumed the condensate to be symmetric around the axis of the cone, {\it i.e.} $\sigma=\sigma(r)$.
Zeta function regularization allows us to express $\delta\Gamma$ as
\begin{eqnarray}
\delta \Gamma ={1\over 2}\int d^dx\sqrt{g} \Big(\zeta(0) \ln \ell^2 +\zeta'(0)\Big)~,
\label{dg}
\end{eqnarray}\noindent
where $\ell$ is a renormalization scale. The quantities $\zeta(0)$ and $\zeta'(0)$ are the analytically continued values to $s=0$ of the following function
\begin{eqnarray}
\zeta(s) = {1\over \Gamma(s)} \sum_n \sum_\epsilon \int dt t^{s-1} e^{-t\omega_n^2} \mbox{Tr} e^{-t\left(-\Delta + {R\over 4} +\sigma^2 +\epsilon \left|\partial \sigma\right| \right)}~.
\end{eqnarray}\noindent
In the above expression we have used the imaginary time formalism to introduce finite temperature effects ($\omega_n= 2\pi(n+1/2)/\beta$ are the Matsubara frequencies for the fermion fields), and we expressed the zeta function in terms of the Mellin transform of the heat-trace.
In the following we will regularize the effective action by excising the singularity as this seems the most convenient way to deal with the subsequent numerical analysis. In this case the heat-kernel expansion will take the standard form and the radial coordinate is assumed to start from $r=\epsilon$, where boundary conditions will be imposed. Some algebra allows us to recast the above expression as
\begin{eqnarray}
\zeta(s)= \zeta_+(s) + \zeta_-(s)~,
\label{zpm}
\end{eqnarray}\noindent
where
\begin{eqnarray}
\zeta_\pm(s) = \int dt {t^{s-1} \over \Gamma(s)} {e^{-t \tilde\sigma^2_\pm}
\mathscr{K}_{reg}(t)\over 2\sqrt{\pi t}} g(t)~,
\nonumber
\label{zeta}
\end{eqnarray}\noindent
with
\begin{eqnarray}
g(t)&=&\left(1+2\sum_{n=1}^\infty (-1)^n e^{-{\beta^2n^2\over 4t}}\right)~,
\\
\tilde \sigma^2_\pm &=& R/12 +\sigma^2 \pm \left|\sigma'\right|~,
\end{eqnarray}\noindent
Notice that in the regular part of the heat-trace we have factorized out the exponential according to the method described in Refs.~\cite{toms,jack}.
In the above expression $s$ is regulating parameter assumed to lie in a region of the complex plane where all the above integrals are well defined. The standard expression for $\mathscr{K}_{reg}(t)$ is given by
\begin{eqnarray}
\mathscr{K}_{reg}(t) &=& {1\over (4\pi)^{d/2}}\sum_{k=0}^\infty \mathscr{C}_\pm^{(k/2)} t^{k/2-d/2}~,
\label{kreg}
\end{eqnarray}\noindent
where the coefficients ${\mathscr{C}}^{(k/2)}_\pm$ are the heat-kernel coefficients associated with the differential operator in (\ref{effect}) and are integrals of polynomials in the curvatures on the regularized spacetime manifold and therefore are all regular. Coefficient with integer index refer to volume contributions while coefficients with semi-integer indices to boundary contributions. The first coefficient is $\mathscr{C}_\pm^{(0)}=1$, while $\mathscr{C}_\pm^{(1)}=0$. Higher order coefficients will be given later as needed.
Using the above expression for $\zeta_\pm(s)$, we can express the effective action as
\begin{eqnarray}
\Gamma &=&-\int d^{3}x \sqrt{g} \left({\sigma^2\over 2\tilde g}\right) + \Gamma_+ +\Gamma_-
~,
\end{eqnarray}\noindent
where $\Gamma_\pm$ are the contributions to the effective action coming from $\zeta_\pm$.
For completeness, at the end of this section, we will illustrate how the computation can be performed also by using the full heat kernel expansion including the logarithmically singular term in (\ref{hka}).
We will start assuming $s$ to be in the region where the integrals are well defined, then analytically continue to $s=0$ and take the limit $d=3$.
The zeta function can be expressed in terms of the following integral and its derivative with respect to the regularizing parameter $s$,
\begin{eqnarray}
\mathscr{I}(a,b,c) &=& {1\over (4\pi)^{d+1\over 2}} \int {dt\over \Gamma(s)}\, t^{s-a}\,e^{{-c t }-b/t}~,\label{I}~.
\end{eqnarray}\noindent
These can be computed explicitly,
using the following expressions (see Ref.~\cite{lang}):
\begin{eqnarray}
\mathscr{I}(a,b,c) &=&
{2^{a-s}\over (4\pi)^{d+1\over 2}}
{1\over \Gamma(s)}
\left({c\over b}\right)^{(-1+a-s)/2}K_{a-s-1} \left(2\sqrt{b c}\right)~,\nonumber
\\
{d \mathscr{I}(a,b,c) \over ds} &=&
{1\over \Gamma(s)} {1\over (4\pi)^{d+1\over 2}}
\left({c\over b}\right)^{(-1+a-s)/2}
\left[
-K_{a-s-1}\left(2\sqrt{b c}\right) \left(
\ln\left({c\over b}\right)
+2 \psi^{(0)}(s) \right)\right.\nonumber\\
&&\left.-2 \left.{d\over d\nu}K_{\nu}\left(2\sqrt{b c}\right)\right|_{\nu=a-s-1}
\right]~,\nonumber
\end{eqnarray}\noindent
The above expression are valid under the following conditions:
\begin{eqnarray}
&&\Re b > 0~,\nonumber\\
&&\Re c > 0~,\nonumber\\
&&\Re \left( a -s \right) >1~.
\end{eqnarray}\noindent
The first two requirements are trivially satisfied in our case. The third is instead dealt with in the usual way of zeta function regularization, {\it i.e.} by assuming that $s$ lies in a region of the complex plane where the third inequality is satisfied and then proceed by analytical continuation. One may easily re-write the zeta function above as
\begin{eqnarray}
\zeta(s) = \sum_{\epsilon=\pm} \sum_{i=1}^2 \mathscr{Z}_\epsilon^{(i)}(s)~,
\label{zf}
\end{eqnarray}\noindent
where we have defined for notational convenience
\begin{eqnarray}
\mathscr{Z}^{(1)}_\pm(s) &=& \sum_{k=0}^\infty \mathscr{C}_\pm^{(k/2)}
\mathscr{I}({3+d-k\over 2}, 0,\tilde\sigma^2_\pm)~,\label{z1}\\
\mathscr{Z}^{(2)}_\pm(s) &=& 2 \sum_{k=0}^\infty \sum_{n=1}^\infty (-1)^n \mathscr{C}_\pm^{(k/2)}
\mathscr{I}({3+d-k\over 2},{\beta^2n^2\over 4},\tilde\sigma^2_\pm)~,\label{z2}
\end{eqnarray}\noindent
Dividing now $\Gamma_\pm$ as the sum of the contribution coming from $\zeta(0)$, $\Gamma_\pm^{(0)}$, and that coming from $\zeta'(0)$, $\Gamma_\pm^{(1)}$, allows us to write
\begin{eqnarray}
\Gamma_\pm = \Gamma_\pm^{(0)} + \Gamma_\pm^{(1)}~.
\end{eqnarray}\noindent
A direct computation shows that only $\mathscr{Z}_\pm^{(0)}$(s) gives a non vanishing contribution to $\zeta(0)$. We find
\begin{eqnarray}
\Gamma_\pm^{(0)} = {1\over 32\pi^2} \int dr\,r^2\,\sqrt{\mbox{det}_{\mathscr{N}}}\,
\left[
{1\over 2}\tilde\sigma_\pm^4 -\tilde\sigma_\pm^2\mathscr{C}^{(1)}_\pm
+\mathscr{C}^{(2)}_\pm
\right]\ln \ell^2,~
\label{g0}
\end{eqnarray}\noindent
where $\det_{\mathscr{N}}$ is the determinant of the metric on the base manifold $\mathscr{N}$. The above result is exact (no truncation in the heat-kernel expansion has been done and boundary contributions also vanish).
Also, notice that the ansatz used in the regular part of the heat-trace gives $\mathscr{C}^{(1)}_\pm =0$ (See Refs.~\cite{toms,jack,flachi1}). Therefore the second term in the above expression vanishes. The coefficient $\mathscr{C}^{(2)}_\pm$ is given by
\begin{eqnarray}
\mathscr{C}^{(2)}_{\lambda} &=& {1\over 180}R_{\mu\nu\rho\sigma} R^{\mu\nu\rho\sigma} -{1\over 180}R_{\mu\nu} R^{\mu\nu}- {1\over 120} \Delta R + {1\over 3} \left(\left(\sigma\pm {1\over r}\right)\sigma''+\sigma^{'2}+{2\over r}\sigma \sigma'\right)~,\nonumber
\end{eqnarray}\noindent
where
the quantities $R$, $R_{\mu\nu}$ and $R_{\mu\nu\lambda\rho}$ are the Ricci scalar and Ricci and Riemann tensors, respectively, for the geometry (\ref{conemetric}). All terms proportional to curvature invariants and not to the condensate will not, in fact, contribute to effective action for the condensate, in the sense that they will disappear in the equation of motion for $\sigma$. The integrand in (\ref{g0}), {\it i.e.} the contribution to the Lagrangian density for the condensate, is regular in the limit $r\rightarrow 0$.
The term that requires more work is the contribution to the effective action that comes from the derivative of the zeta function, $\Gamma_\pm^{(1)}$. Proceeding as described above, we find for the regularized effective action the following expression
\begin{eqnarray}
\Gamma^{(1)}_\pm &=&
{1\over 32\pi^2} \int dr\,r^2\,\sqrt{\mbox{det}_{\mathscr{N}}}\,
\Bigg\{
{3\over 4} \tilde\sigma^4_\pm
-\left({1\over 2}\tilde\sigma^4_\pm +\mathscr{C}^{(2)}_\pm\right)\ln \tilde \sigma^2_\pm
\nonumber\\
&&
+4\sum_{n=1}^{\infty}(-1)^{n}
\Bigg[
{4\tilde \sigma^2_\pm \over n^2 \beta^2} K_2\left(n\beta\tilde \sigma_\pm\right)
+\mathscr{C}^{(2)}_\pm
K_0\left(n\beta\tilde \sigma_\pm\right)
\Bigg] +\cdots \Bigg\}~.
\label{eapm}
\end{eqnarray}\noindent
The dots represent higher order terms in the heat kernel expansions. They can be computed explicitly, but we won't report their form here. All higher order terms give contributions in the derivatives of the condensate of order equal or greater than four, so the expression above is sufficient when keeping the analysis to second order in the derivative of the condensate. Boundary terms can also be calculated easily and the first few terms are
\begin{eqnarray}
\Gamma_{boundary}&=&{1\over 32\pi^2} \int d^3x \sqrt{g}\,\delta(r-L)\,
\Bigg\{{4\over 3}\sqrt{\pi}\tilde\sigma^3_\pm
\mathscr{C}^{(1/2)}_\pm
-2\sqrt{\pi} \tilde \sigma_\pm \mathscr{C}^{(3/2)}_\pm
\nonumber\\
&&
+4\sum_{n=1}^{\infty}(-1)^{n}
\Bigg[
+{2\sqrt{\pi}\over n^2\beta^2} \mathscr{C}^{(1/2)}_\pm
\left(\tilde \sigma_\pm+{1\over n\beta}\right)
e^{-n\beta\tilde \sigma_\pm}
+{\sqrt{\pi}\over n\beta} \mathscr{C}^{(3/2)}_\pm e^{-n\beta\tilde \sigma_\pm}
\Bigg]+\cdots \Bigg\},~
\nonumber
\end{eqnarray}\noindent
where the dots represents, as before, higher order terms in the heat-kernel expansion.
As before, one can easily verify the regularity of the expressions above.
To conclude this section, we will show how the full heat-kernel expansion (\ref{hka}), including the logarithmic terms, can be used instead of the scheme adopted above. In this case, according to (\ref{hka}) the heat-kernel will consist of two terms. The polynomial part of the heat-kernel can be treated as above with the only difference being that the heat-kernel coefficients are understood as the principal part of integrals over the space manifold. Since the manifold has singularities, the integration in the heat-kernel coefficients, naturally, leads to divergences. As shown in Ref.~\cite{cheeger}, the relevant quantity for the heat-kernel asymptotics is the principal part of these diverging integrals that can be extracted in several ways. One possibility is to cut the range of integration at finite distance from the singularity and this is equivalent to the regularization by excision used above. The additional (logarithmic) term in (\ref{hka}) can also be included in the analysis, but some attention is necessary. The quantity $\mathscr{B}$ is proportional to minus the residue of the zeta function at $s=0$, $-\mbox{Res}\,\zeta_{\mathscr{C}_{\mathscr{N}}}(0)$, that can, in turn, be related to the value of the zeta function evaluated on the base manifold analytically continued to $s=-1/2$.
The logarithmic part of the heat-kernel expansion will lead to an additional term in zeta function that can be written as
\begin{eqnarray}
\zeta(s) = \sum_{\epsilon=\pm} \sum_{i=3}^4 \mathscr{Z}_\epsilon^{(i)}(s)~,
\end{eqnarray}\noindent
with
\begin{eqnarray}
\mathscr{Z}^{(3)}_\pm(s) &=& {\mathscr{B}_\pm\over \sqrt{4\pi}}
\mathscr{H}_\pm(3/2,0,0)~,\label{z3}\\
\mathscr{Z}^{(4)}_\pm(s) &=& 2 {\mathscr{B}_\pm\over \sqrt{4\pi}}
\sum_{n=1}^\infty (-1)^n
\mathscr{H}(3/2,{\beta^2n^2\over 4},0)~,\label{z4}
\end{eqnarray}\noindent
where
\begin{eqnarray}
\mathscr{H}(a,b,c) &=& {1\over (4\pi)^{d+1\over 2}} \int {dt\over \Gamma(s)}\,t^{s-a}\,\ln t\, e^{-c t-
b/t}\label{H}~.
\end{eqnarray}\noindent
The analytical continuation of the logarithmic term can be done straightforwardly by using the following expressions
\begin{eqnarray}
\mathscr{H}(a,b) &=&
{b^{(1-a+s)/2}\over (4\pi)^{d+1\over 2}} {\Gamma(-1+a-s)\over \Gamma(s)}
\left( \psi^{(0)}\left(-1+a-s\right) -\ln b \right) ~,\nonumber
\\
{d \mathscr{H}(a,b)\over ds}
&=&
{b^{(1-a+s)/2}\over (4\pi)^{d+1\over 2}}
{\Gamma(-1+a-s)\over \Gamma(s)}
\left[
\left( \psi^{(0)}\left(-1+a-s\right) -\ln b \right)\times \right.\nonumber\\
&&
\left.\left( \psi^{(0)}\left(-1+a-s\right) +\psi^{(0)}(s) -\ln b \right)
+ \psi^{(1)}\left(-1+a-s\right) \right]
~,\nonumber
\end{eqnarray}\noindent
whose analytical continuation can be performed as before.
A straightforward computation shows that the contribution of the logarithmic term to $\zeta(0)$ vanishes, while the contribution to the effective action from $\zeta'(0)$ gives
\begin{eqnarray}
\Gamma^{log}_\pm &=& =
{\ln 2 \over (4\pi)^{d\over 2}\beta}\left(\gamma_e -\ln\left(2\beta^2\right)\right)
\int d^dx \sqrt{g}
{\mathscr{B}_\pm} ~.
\label{zsing}
\end{eqnarray}\noindent
\section{Inhomogeneous Condensates in Monopole Geometries: Numerical Construction}
\label{sec3}
This section will be devoted to implement numerically the above formalism and to construct explicit solutions for the condensate. To have a concrete working model, we need to specify the geometry of the base. In order to make contact with a case of physical interest, we will choose the base manifold to be a sphere of non-unitary radius $\rho$. The curvature is $R\propto(1-\rho^2)\rho^{-2}r^{-2}$, the space is non flat when $\rho\neq 1$. In the following we will consider $\rho<1$, therefore $R>0$. In this case, the metric (\ref{conemetric}) describes the spatial section of global monopole \cite{vilen}. For this geometry, the heat-kernel analysis of the conformal Laplacian has been done, for instance, in Ref.~\cite{bordag}.
The equation of motion for the condensate can be found by minimizing the effective action, that is by varying $\Gamma$ with respect to the condensate $\sigma$. In the previous section we have obtained the effective action to fourth order in the heat-kernel expansion ({\it i.e.} $k=4$ in (\ref{kreg})), leading to a second order non-linear equation of motion for the condensate. This can be written, with some work, in the form of non-linear Schr\"odinger-like equation, whose explicit form is very lengthly and will be omitted here.
The structure of the solution may be anticipated by looking at the form of the thermodynamic potential. At large distance from the singularity, where curvature effects are negligible, the thermodynamic potential ${U_{as}}(\sigma)$ is expected to have a behaviour similar to flat space: below (above) some critical value of the temperature, chiral symmetry is broken (restored). This can be verified by letting $\sigma'=0$, taking the limit $r\rightarrow \infty$ of the effective potential, and by computing the thermodynamic potential by numerical integration of $\partial_\sigma {U}$ with respect to $\sigma$. The profile of the asymptotic potential is shown in Fig.~\ref{fig1} (left panel).
The thermodynamic potential can also be computed locally and the result is illustrated in Fig.~\ref{fig1} (right panel). The figure shows that even if chiral symmetry is broken asymptotically (right-most (purple) curve), chiral symmetry will gradually be restored as we move towards the singularity, {\it i.e.} the minima of the potential will gradually shift towards a configuration with vanishing $\sigma$.
\FIGURE[h]{
\centering
\put(-7.,60){\rotatebox{90}{$U_{as}(\sigma)$}}
\put(100,-5.5){$\sigma$}
\put(310,-4){$\sigma$}
\put(410,65){\rotatebox{90}{$U(\sigma)$}}
\includegraphics[width=0.47\columnwidth]{asyptpot.eps}
\includegraphics[width=0.47\columnwidth]{localpot.eps}
\caption{The figure on the left illustrates how the local thermodynamic potential changes as the temperature increases. The right-most (left-most) curve shows the potential for temperature smaller (higher) than the critical one. The figure on the right illustrates how the local thermodynamic potential changes as the singularity is approached. The right-most curve shows the potential at large distance from the singularity for a set of parameters that correspond to a chirally broken symmetry phase (The minima of the potential is non vanishing). The left-most curve shows the potential close to the singularity ($r=0.1$) for which the potential has a vanishing minima at $\sigma=0$, leading to a chirally restored symmetry phase. For both figures we set $\ell=10^{6}$ and $\lambda=10^{-2}$.}
\label{fig1}
}
The solution for the condensate can be found using standard numerical methods, and here we have used a fourth order Runge-Kutta one. In solving the equation for the condensate numerically, there are several things that require some care. One has to deal with the infinite summations over Bessel functions with argument proportional to the condensate. The present situation differs from the black hole case studied earlier \cite{flachi2} as described in the following. At large distance curvature effects are negligble and the argument of the Bessel function is not small. In this case, one may conveniently truncate the summation due to the fact that the Bessel function decay exponentially for large arguments. Near the singularity the contribution from the curvature is, instead, large, and the argument of the Bessel functions is again not small. In the intermediate region for $r$ sufficiently large, we require that, when the argument of the Bessel functions becomes small, the summands are expanded for small arguments and full resummation over $n$ is performed. We then proceed by matching the solutions in the truncated and resummed domains. With the equations of motion for the condensate explicitly written, we solve them by requiring regularity for the solution. The boundary conditions on the solution are set numerically by requiring, near the singularity, that the condensate is at a minima of the potential and by fine tuning the value of the derivative to match the asymptotic value of the solution with the minima of the potential at infinity.
\FIGURE[h]{
\centering
\put(-7,60){\rotatebox{90}{$\sigma(r)$}}
\put(105,-5.5){$r$}
\includegraphics[width=0.5\columnwidth]{condsol.eps}
\caption{Condensate profile found by numerically minimizing the effective action (\ref{effect}), for four indicative values of the base manifold radius (Left to right: $\rho=0.65~\mbox{(Green)},0.57~\mbox{(Blue)},0.53~\mbox{(Black)},0.48~\mbox{(Red)}$). We set to $\ell=10^6$, $\lambda=10^{-2}$.
The superposed panel shows the solution for the condensate when higher order terms included superposed over the background solution. The continuous line refers to the background solution corresponding to $\rho=0.65\,\mbox{(Green)}$ and the dotted (red) solution corresponds to the higher order one. \vspace{.2cm}}
\label{fig2}
}
The results for the condensate profile are illustrated in Fig.~\ref{fig2} for sample values of the parameters showing that near the singularity the condensate vanishes and chiral symmetry is restored, while, as we gradually move away from the singularity, the condensate assumes a non vanishing expectation value and chiral symmetry breaks. Notice that contrary to the black hole case we have studied earlier \cite{flachi2}, in the present case, the thermodynamical temperature is constant uniformly along the spatial section of the space and the effect of breaking/restoration of the chiral symmetry is solely due to genuine curvature effects.
As in the black hole case \cite{flachi2}, this indicates that a bubble of chirally restored symetry phase surrounds the singularity.\\
The analysis can be extended to higher orders by including higher order terms in the heat-kernel expansion. The algebra becomes rather cumbersome, but it is possible to handle, with some work, all the computations automatically by using any computer algebra program. After including terms up to fourth order derivative of the condensate, we implement our analysis by perturbing the background solution, $\sigma = \sigma_{bg}+\delta \sigma$, and expanding the higher order equation for $\sigma$ keeping terms up to second derivative of the perturbation. Proceeding in this way, we are assuming that the corrections due to higher order terms only produce small changes in the background solution. This assumption will be verified {\it a posteriori} after the solution is obtained. Once higher order derivatives of the perturbation are dropped out, the equation for the $\delta \sigma$ becomes again a second order one with source term and can be solved by standard methods. This approach is similar to the case of black holes we have addressed earlier \cite{flachi2} and, as in that case, the corrrection to the solution only produces very small distortions that decrease with the thickness of the kink. A sample plot is shown in the small panel in Fig.~\ref{fig2}.
\section{Conclusions}
In this paper we have continued the analysis of the chiral symmetry breaking of strongly coupled fermion effective field theory in curved space, extending our previous work to include conifold-type geometries.
The results, in agreement with our intuition, show that, near the singularity, the condensate vanishes and chiral symmetry is restored. As we gradually move away from the singularity, the condensate assumes a non vanishing expectation value and chiral symmetry breaks. This indicates that the singularity will be surrounded by a bubble of chirally restored symmetry phase, in much the same way as black holes (See Ref.~\cite{flachi2}). We have illustrated this explicitly in the tractable example of a conifold geometry with generic base manifold (the generalized metric cone). More precisely, we have considered a strongly interacting fermion effective field theory propagating on the conifold and constructed an explicit solution for the condensate. The formal analysis has been carried out in general, and we have evaluated the effective action for the condensate introducing finite temperature effects by means of the standard Matsubara formalism. This set-up allowed us to keep the genuine thermodynamical temperature constant and to clearly single out any non-trivial effect of the curvature.
The evaluation of the effective action was carried out using a sophistication of the method we have described in Ref.~\cite{flachi1} and that makes use of zeta function regularization. Due to the presence of the singularity the regularization requires some care, however zeta function regularization proves to be very adequate in this case.
Several generalizations may be considered. First of all, one expects that the present situation is more general. That is in the vicinity of {\it any} spacetime singularity the same occurs. This can in principle be treated by studying the heat-kernel asymptotics in spacetimes with different singular structure. Although computational complications may arise, the procedure presented in this paper that regularizes the geometry by excising a small region around the singularity should work with little modifications. Generalizations to more sophisticated models may also be possible with some effort, the most interesting ones being the inclusion of gauge degrees of freedom and of a chemical potential.
It was our goal to understand how condensates and chiral symmetry may be affected by curvature singularities and for this reason we considered a purely geometrical background (\ref{conemetric}). However, the results may be easily extended to cases where a non-trivial radial dependence of the lapse function occurs, and this can be handled by using the same conformal techniques that we have adopted to study the case of black holes. In this case, if the local temperature increases as the singularity is approached we expect the same phenomena of chiral restoration described here to occur.
\acknowledgments
The financial support of the Funda\c{c}\~{a}o p\^{a}ra a Ciencia e a Tecnologia of Portugal (FCT) is gratefully acknowledged. I wish to express my gratitude to Takahiro Tanaka for his crucial help with several technical points, for carefully reading the manuscript and for many useful suggestions that helped to improve the presentation of the material of this paper. I also wish to thank Marco Ruggieri for continuous discussions on strongly interacting fermion effective field theories.
|
\section{Introduction}
Uncertainty Quantification (UQ) is an area of mathematics that
is used to quantify output distributions given parametric
uncertainty. Traditional approaches include Monte Carlo and
Quasi-Monte Carlo methods~\cite{McQMc}, response surface
methods~\cite{Response1:book,Response2:book} as well as polynomial
chaos and probabilistic collocation based
approaches~\cite{PolyReview}. The polynomial chaos approach for
uncertainty quantification was originally proposed by Norbert
Wiener~\cite{Wiener}. Assuming that one is given input
uncertainty in the form of distributions associated with various
parameters of the system, polynomial chaos/probabilistic
collocation methods provide an approach for fast uncertainty
quantification under the assumption of smooth dynamics. In
particular, polynomial chaos provides exponential convergence
for smooth systems and processes with finite
variance~\cite{PolyReview}. Polynomial chaos based methods have
been used for a multitude of
applications, see~\cite{Allen2009,Elman2011,Ghanem1998,
Najm2009,TuhinPoly,Xiu2003,Zabaras2008,Multigpc} for examples. Note that, depending on
the application, one can combine various UQ
approaches. For example, a combination of polynomial chaos and the response surface
methodology has been used to develop
probabilistic collocation methods for discrete distributions in~\cite{TuhinPoly}.
In this work, we focus on developing UQ techniques for hybrid dynamical systems.
Hybrid dynamical systems theory is used to model systems with both
discrete and continuous dynamics~\cite{SHS:book}. Examples
include the bouncing ball automaton~\cite{ZenoReg}, biological
networks~\cite{HybridCompCont:book,BioEx1}, air traffic
management systems~\cite{AirTraffic}, communication
networks~\cite{CommNet}, elevators, and robotics, to name a few.
These systems frequently display rich dynamics not seen in continuous
systems. For example, Zeno behavior in hybrid systems is characterized by an infinite number of discrete switches in
finite time~\cite{ZenoReg,Zeno}. Hybrid systems can be particularly challenging from an analysis standpoint since traditional techniques, such as polynomial chaos based methods, assume smoothness, rendering them inapplicable.
In this work, we develop polynomial chaos and transport theory based methods for
propagating uncertainty through hybrid systems. We assume that the domains associated with different modes of operation of the hybrid system do not overlap. We demonstrate that, by integrating over
appropriate time-varying regions, one can extend the polynomial
chaos framework to hybrid dynamical systems. We resolve the
issue of state resets~\cite{SHS:book} in the separable case by using boundary layer
approximations. To capture the discontinuities in the probability distributions of the output variables, we use a Haar-wavelet expansion~\cite{Haar1910}. This expansion has previously been used in the polynomial chaos setting to propagate uncertainty through dynamical systems close to bifurcation points~\cite{Najm2009,LeMaitre2004,LeMaitre2004a}. Here we develop a methodology to propagate uncertainty through systems with discontinuities in dynamics and output along with state resets. We also develop a transport theory based approach that allows one to propagate the uncertainties through the various modes of the hybrid dynamical system.
Our paper is organized as follows: in section~\ref{Sec:problem} we define hybrid dynamical systems and the problem of uncertainty quantification. In section~\ref{Sec:hyb_poly} we first construct the framework for polynomial chaos in the hybrid dynamical system setting~(\ref{Sec:hpc}). We then demonstrate the Haar wavelet expansion for hybrid polynomial chaos in~\ref{Sec:wavelets}. The handling of state resets is considered in section~\ref{Sec:resets}. Finally, the results on hybrid polynomial chaos are presented in~\ref{Sec:Results}. The transport operator theory based method for propagating uncertainty through hybrid dynamical systems is developed in section~\ref{sec:transporttheory} and conclusions are drawn in section~\ref{Sec:conclusions}.
\section{Problem definition}\label{Sec:problem}
Let $S=(q, X, f, x(0), D, E, G, R)$ denote a hybrid system $S$, where \\
\centerline{
\begin{tabular}{l l}
$q$ & Set of discrete variables \\
$X$ & Set of continuous variables\\
$f:X\times q \rightarrow TX$& Vector field\\
$x(0)$ & Set of initial conditions\\
$D:q\rightarrow P(X)$ & Domain\\
$E$ & Set of discrete transitions\\
$G: E\rightarrow P(X)$ & Guard conditions\\
$R: E\times X \rightarrow P(X)$ & Reset map.\\
\end{tabular}}
In the above table, $TX$ is the tangent bundle of $X$ and $P(X)$
is the power set of $X$. For more details on the definition of
hybrid systems, see~\cite{SHS:book}.
We can use the following representation for hybrid systems,
\begin{equation}
\dot x = f(x,\lambda,q), \label{eq:fullsys}
\end{equation}
where $x \in X$ is a vector of state variables and the form of
$f(x,\lambda,q)$ is dictated by~$q$, which represents the mode
of operation of the hybrid dynamical system. The discrete state~$q$ is
determined by the guard conditions that dictate transitions
between modes (see Fig.~\ref{Fig:HySchematic}). The reset
functions $h_{i}(x)$ are a part of the reset map~$R$. In
Eqn.~\ref{eq:fullsys} let~$\lambda$ denote the vector of system
parameters and $x(0)$ the initial condition for the
system.
\begin{figure}
\centering
\includegraphics[width=0.7\hsize]{Hybridtest}\\
\caption{Schematic for hybrid (switching) systems.
\label{Fig:HySchematic} }
\end{figure}
If the system parameters~$\lambda$ in Eqn.~\ref{eq:fullsys} are
uncertain (i.e., each $\lambda_{i}$ has an associated
distribution) then one typically desires to quantify the
time-varying moments (such as mean and variance) of~$x(t)$ (note that $x(0)$ may also be uncertain). As
mentioned earlier, although one can use Monte Carlo based
sampling methods~\cite{McQMc}, they are plagued by slow
convergence. In particular, the mean is expected to converge as
$1/\sqrt{N}$, where $N$ is the number of samples. Quasi-Monte
Carlo based sampling methods are expected to give a convergence
rate of $\log^d (N)/N$, where $d$ is the dimensionality of the
random space~\cite{QMC}, making these methods attractive for
problems in low dimensions. Polynomial chaos based methods
provide an alternative framework for uncertainty quantification
with exponential convergence for processes with finite variance,
but they too suffer from the curse of
dimensionality~\cite{PolyReview}. In the next section we extend
the polynomial chaos framework to hybrid systems.
\section{Polynomial chaos for hybrid dynamical systems}\label{Sec:hyb_poly}
Starting with a complete probability space $\Gamma$ given by
$(\Omega, \mathcal{F},\mathbb{P})$, where $\Omega$ is the sample
space, $\mathcal{F}$ is the $\sigma$-algebra on $\Omega$ and
$\mathbb{P}$ is a probability measure, let $L_{2}(\Gamma,X)$
denote the Hilbert space of square-integrable,
$\Gamma$-measurable, $X$-valued random elements. Then one can,
in general, define a polynomial chaos basis
$\{H_{\alpha}(\lambda(\omega))\}$, where $\lambda(\omega)$ is a
random vector and $\alpha = (\alpha_{1},\alpha_{2},\dots)$ is a
vector of non-negative indices. We denote the probability
density function of the random vector $\lambda$ by
$\rho(\lambda)$.
Generalized polynomial chaos (gPC)~\cite{BeyondWienerAskey}, provides a framework for
representing second-order stochastic processes $r\in
L_{2}(\Gamma,X)$ for arbitrary distributions of $\lambda$ by the following expansion:
\begin{equation}
r(\lambda) =
\displaystyle\sum_{|\alpha|=0}^{\infty}a_{\alpha}H_{\alpha}(\lambda),
\label{eq:expan1}
\end{equation}
where $|\alpha| = \sum_{i} \alpha_{i}$ is the sum of the indices of $\alpha$ and
$H_{\alpha}(\lambda)$ are orthogonal polynomials on $\Gamma$
with respect to $\rho(\lambda)$, i.e.
\begin{equation}
\displaystyle\int_{\Gamma}
\rho(\lambda)H_{\alpha}(\lambda)H_{\beta}(\lambda)d\lambda =
\delta_{\alpha\beta}, \label{eq:ortho}
\end{equation}
where $\delta_{\alpha\beta}$ is the Kronecker delta product.
Depending on $\rho(\lambda)$ one can generate an appropriate
orthogonal basis for representing~$r(\lambda)$. For example, if $\rho$ is
Gaussian, then the appropriate polynomial chaos basis is the set
of Hermite polynomials; if $\rho$ is the uniform distribution,
then the basis is the set of Legendre polynomials. For details
on the correspondence between distributions and polynomials
see~\cite{PolyReview,Ogura}. A framework to generate polynomials
for arbitrary distributions has been developed
in~\cite{BeyondWienerAskey}.
In practice, the expansion in Eqn.~\ref{eq:expan1} is truncated
at a particular order, say, $p$. One can then use Galerkin projections to
obtain a set of differential equations for the coefficients
$a_{\alpha}$ in Eqn.~\ref{eq:expan1}~\cite{PolyReview}.
We now extend the standard polynomial chaos framework to hybrid
dynamical systems despite the presence of switching and state resets. To the best of our knowledge, it is the first attempt
to develop tools for fast uncertainty quantification for this class of systems.
\subsection{Hybrid Polynomial Chaos}
\label{Sec:hpc}
Without loss of generality, consider the following two-mode
hybrid dynamical system as representative of systems in which the different operating modes are associated with non-overlapping regions:
\begin{equation}
\dot{x} = \begin{cases} f(x,\lambda) & \text{if $x\geq 0$} \\
g(x,\lambda) & \text{otherwise}.
\end{cases}
\label{eq:wlogsys}
\end{equation}
Here one desires to quantify $x(t;\lambda)$, i.e., determine
$x$ as a function of time $t$ and parameters $\lambda$. The system above has two
modes of operation determined by its state. One
can parameterize these modes in the following way:
\begin{eqnarray}
\1_{R_{1}}(x) &=& \begin{cases}
1 & \text{if $x\geq 0$} \\
0 & \text{otherwise}
\end{cases}
\label{eq:indict_r1} \\
\1_{R_{2}}(x) &=& 1 - \1_{R_{1}}(x). \label{eq:indict_r2}
\end{eqnarray}
When $\1_{R_{1}}(x) = 1$ (corresponding to $x\geq 0$) the governing
differential equations are $f(x,\lambda)$, and when $\1_{R_{2}}(x)=1$
($x < 0$) the governing differential equations are
$g(x,\lambda)$. Thus, one can rewrite Eqn.~\ref{eq:wlogsys} as,
\begin{equation}
\dot x = \1_{R_{1}}(x) f(x,\lambda) + \1_{R_{2}}(x) g(x,\lambda).
\label{eq:homotop}
\end{equation}
This equation extends easily to $k$ modes of operation by constructing indicator functions for each mode of operation $\1_{R_{1}},\1_{R_{2}},\hdots,\1_{R_{k}}$ of the hybrid system. We now expand $x(t;\lambda)$ in the appropriate orthogonal
polynomial chaos basis,
\begin{equation}
x(t;\lambda) =
\displaystyle\sum_{|\alpha|=0}^{p}a_{\alpha}(t)H_{\alpha}(\lambda).
\label{eq:expx}
\end{equation}
Dropping the arguments of $a_{\alpha}(t)$ and
$H_{\alpha}(\lambda)$ for simplicity and using the above
relation with Eqn.~\ref{eq:homotop}, one gets
\begin{eqnarray}
\sum_{|\alpha|=0}^{p}\dot a_{\alpha} H_{\alpha} &=&
\1_{R_{1}}(\sum_{|\alpha|=0}^{p}a_{\alpha}H_{\alpha})f(\sum_{|\alpha|=0}^{p}a_{\alpha}H_{\alpha},\lambda)
\nonumber\\
&+& \1_{R_{2}}(\sum_{|\alpha|=0}^{p}a_{\alpha}H_{\alpha})
g(\sum_{|\alpha|=0}^{p}a_{\alpha}H_{\alpha},\lambda). \nonumber
\label{eq:exphomotop}
\end{eqnarray}
By multiplying the above relation by
$\rho(\lambda)H_{k}(\lambda)$, integrating over $\Gamma$, and
using orthogonality conditions, we get
\begin{eqnarray}
\dot a_{k}(t) &=&
\int_{R_{1}(t)}f(\sum_{|\alpha|=0}^{p}a_{\alpha}H_{\alpha},\lambda)\rho(\lambda)H_{k}(\lambda)d\lambda
\nonumber\\ &+&
\int_{R_{2}(t)}g(\sum_{|\alpha|=0}^{p}a_{\alpha}H_{\alpha},\lambda)\rho(\lambda)H_{k}(\lambda)d\lambda.
\label{eq:intregionpc}
\end{eqnarray}
Note that $R_{1}(t) = \{\lambda:
\sum_{|\alpha|=0}^{p}a_{\alpha}H_{\alpha} \geq 0\}$ and
$R_{2}(t) = \Gamma - R_{1}(t)$. Thus by evaluating the two
integrals one can evolve $a_{k}(t)$ for any index vector~$k$.
Note, however, that the regions of integration $R_{1}$
and~$R_{2}$ are time-dependent quantities and must be evaluated
at every instant in time.
For a pictorial depiction of $R_{1}$ and $R_{2}$, see Fig.~\ref{Fig:Region}.
\begin{figure}
\centering
\psfrag{a}[][]{$\sum_{|\alpha|=0}^{p}a_{\alpha}(t)H_{\alpha}(\lambda)$}
\psfrag{b}[][]{$R_{1}(t)$}
\psfrag{c}[][]{$R_{2}(t)$}
\includegraphics[width=0.7\hsize]{R1R2}\\
\caption{Pictorial representation of the regions of integration for the two integrals in
Eqn.~\ref{eq:intregionpc}. In this example, the regions are not simply connected.
\label{Fig:Region}}
\end{figure}
\subsection{Hybrid Polynomial Chaos and wavelet expansions}
\label{Sec:wavelets}
Hybrid systems can display discontinuous behavior as a function of the uncertain parameters. In view of this, a smooth polynomial chaos expansion is expected to degrade as the discontinuities become more severe. In Ref.~\cite{LeMaitre2004} the authors develop a wavelet-based Wiener-Haar expansion to treat bifurcating (but smooth) dynamical systems with uncertain initial conditions that result in discontinuous behavior. In this section we adapt the Wiener-Haar expansion to hybrid dynamical systems.
In~\cite{LeMaitre2004, LeMaitre2004a}, output variables are expanded in terms of Wiener-Haar wavelets expressed as functions of the Cumulative Distribution Function (CDF) of the uncertain parameters. For simplicity, consider the univariate case. Here we denote the CDF of the uncertain parameter~$\lambda$ as $u(\lambda)$ and expand the state vector~$x$ as\footnote{For the multivariate case, see Ref.~\cite{LeMaitre2004}.}
\begin{equation}
x(t; \lambda) = x_0(t) + \sum_{j=0}^P \sum_{k=0}^{2^j-1} x_{jk}(t) \psi_{jk} (u(\lambda)),
\label{eq:x_waveletexp}
\end{equation}
where,
\[
\psi_{jk}(u) = 2^{j/2} \psi(2^j u - k), \quad \text{(with $j=0,1,\ldots$ and $k=0,\ldots,2^j-1$)}
\]
is a family of Haar wavelets~\cite{Haar1910}, defined in terms of the \emph{mother wavelet}:
\[
\psi(u) = \begin{cases}
1 & 0 \leq u < 1/2 \\
-1 & 1/2 \leq u < 1 \\
0 & \text{otherwise.}
\end{cases}
\]
The index~$j$ determines the scale of the wavelet and $k$ its displacement. Note that $\{\psi_{jk}\}$ is a family of orthonormal functions on the interval $[0,1]$ with respect to the uniform density. This makes the family $\{\psi_{jk} \circ u\}$ automatically orthonormal with respect to the probability density of~$\lambda$:
\[
\delta_{jl} \delta_{km} = \int_0^1 \psi_{jk}(u) \psi_{lm}(u) du
= \int (\psi_{jk} \circ u)(\lambda) \, (\psi_{lm} \circ u) (\lambda) \rho(\lambda) d\lambda.
\]
Additionally, all $\psi_{jk}$'s are orthogonal to the constant function on $[0,1]$, which implies that the mean of $x$ is given by the first term in the expansion
\[
x_0(t) = \int x(t;\lambda) \rho(\lambda) d\lambda
\]
and that the variance is
\[
\sigma^2(t) = \sum_{j=0}^P \sum_{k=0}^{2^j-1} x_{jk}^2(t).
\]
We now use this expansion on a switching oscillator example:
\begin{align}
\ddot x + c\dot x + x + \lambda &= 0\,\,\,\text{if $x \geq 0$} \nonumber \\
\ddot x + c\dot x + x - \lambda &= 0\,\,\,\text{if $x < 0$},
\label{eq:SHMswitch}
\end{align}
which can be rewritten as,
\begin{align*}
\dot{x} &= y \\
\dot{y} &= - cy -x - \lambda \mathbf{1}_{R_1}(x) + \lambda \mathbf{1}_{R_2}(x).
\end{align*}
The expansion in this case is
\begin{align*}
x(t;\lambda) &= x_0(t) + \sum_{j=0}^P \sum_{k=0}^{2^j-1} x_{jk}(t) \psi_{jk} (u(\lambda)) \\
y(t;\lambda) &= y_0(t) + \sum_{j=0}^P \sum_{k=0}^{2^j-1} y_{jk}(t) \psi_{jk} (u(\lambda)).
\end{align*}
Projecting these equations onto the basis functions yields,
\begin{align*}
\dot{x}_0 &= y_0 \\
\dot{y}_0 &= -c y_0 - x_0 - \int_0^1 \lambda(u) \mathbf{1}_{R_1}(x(u)) du + \int_0^1 \lambda(u) \mathbf{1}_{R_2}(x(u)) du \\
\dot{x}_{jk} &= y_{jk} \\
\dot{y}_{jk} &= -c y_{jk} - x_{jk} - \int_0^1 \lambda(u) \mathbf{1}_{R_1}(x(u)) \psi_{jk}(u) du
+ \int_0^1 \lambda(u) \mathbf{1}_{R_2}(x(u)) \psi_{jk}(u) du
\end{align*}
Note that to compute $\lambda(u)$ we invert the CDF~$u(\lambda)$.
To compute the integrals needed to evolve these equations numerically, we take advantage of the fact that Haar wavelets are piecewise constant. Namely, for a given truncation order~$P$, if we divide $[0,1]$ into $2^{P+1}$ equal subintervals, both $\psi_{jk}$ (with $j \leq P$) and the truncated expansion for $x$ (and therefore the indicator functions) are constant in each subinterval. This implies that in each subinterval we only need to calculate the integral of $\lambda(u)$, which is known a priori. For the case of a Gaussian $\lambda \sim N(\mu,\sigma^2)$, we have
\[
\lambda(u) = \mu + \sigma \sqrt{2} \erf^{-1} (2u-1),
\]
which has a primitive,
\[
\int \lambda(u) du = \mu u - \sigma
\frac{1}{\sqrt{2\pi}} \exp\left\{ -[\erf^{-1}(2u-1)]^2 \right\}.
\]
Therefore the contribution to $\dot{y}_{jk}$ of the integrals in each subinterval $l=0,\ldots,2^{P+1}-1$ is either zero or $\pm 2^{j/2}$ times the precomputed value
\[
\int_{l/2^{P+1}}^{(l+1)/2^{P+1}} \!\!\!\!\!\! \lambda(u) du.
\]
Section~\ref{Sec:Results} presents the results obtained with the Wiener-Haar wavelet expansion in Eq.~\ref{eq:x_waveletexp} for hybrid dynamical systems.
\subsection{Modeling state resets}
\label{Sec:resets}
A significant challenge that hybrid dynamical systems present is
the possibility of state resets~\cite{SHS:book}. When a hybrid
system switches from one mode to another, the state of the
system can, in general, be reset discontinuously. For example,
in the case of the bouncing ball~\cite{SHS:book}, the velocity
of the ball changes discontinuously after every impact. When the
hybrid system transitions from mode $q$ to~$q'$ the state resets are typically represented as,
\begin{equation}
x^{+} = h(x^{-}),
\label{eq:reset}
\end{equation}
where $x^{-}$ and $x^{+}$ are the states of the system before and after the reset. Such discontinuities cannot be
easily accommodated within the hybrid polynomial chaos framework as described in the previous sections.
To circumvent this problem, one can construct a boundary layer in the vicinity of the guard condition. We also introduce a dummy vector ($z$) that tracks the state $x$ outside the boundary layer and is set to $x^{-}$ within the boundary layer. Note that we assume \emph{separability} of the states, i.e. the guard conditions (which determine the switching between modes of operation) can be written independently of the state reset conditions in Eqn.~\ref{eq:reset}. In other words, the states that determine the guard conditions do not participate in the state reset. Let the reset condition be in terms of vector $x$ in Eqn.~\ref{eq:reset}, and the guard conditions be in terms of vector $y$ (given by $\left\{y: g(y) = 0\right\}$). Note that, $\left[x \quad y\right]^{T}$ represents the entire state vector with the following governing equation,
\begin{eqnarray}
\dot x &=& f_{1}(x,y) \nonumber \\
\dot y &=& f_{2}(x,y).
\label{eq:sep_sys}
\end{eqnarray}
We now construct a boundary layer around the guard condition for vector $y$ as follows:
\begin{equation}
\begin{pmatrix}
\dot x\\
\dot y\\
\dot z
\end{pmatrix} = \begin{cases}
\begin{pmatrix}
f_{1}(x,y)\\
f_{2}(x,y)\\
(x - z)/\epsilon
\end{pmatrix} & \mbox{if}: |g(y)|\geq\epsilon \\
\\
\begin{pmatrix}
[h(z) - x]/\epsilon\\
\epsilon f_{2}(x,y)\\
0
\end{pmatrix} & \mbox{otherwise.}
\end{cases}
\label{eq:blayer}
\end{equation}
The above dynamical system is constructed such that $x^{-}$
evolves to $h(x^{-})$ in $\Delta t\approx\epsilon$, where $\epsilon$ is a small parameter.
By replacing each reset condition with an equation of the form
given by Eqn.~\ref{eq:blayer}, one obtains a new dynamical system without resets that approximates the original. On this new dynamical system one can use the expansion from Sec.~\ref{Sec:hpc} and evolve it using Eqn.~\ref{eq:intregionpc}. In other words, the framework
generalizing polynomial chaos to hybrid systems can be augmented using
Eqn.~\ref{eq:blayer} to include state resets.
To illustrate the procedure presented above we turn to the classic bouncing ball example~\cite{ZenoReg}: we consider the dynamics of a ball bouncing on a floor with coefficient of restitution $\gamma < 1$ under the action of gravity of uncertain magnitude ($\mu(g) = 9.8\,m/s^{2}$ and $\sigma(g) = 0.2\,m/s^2$). Thus, every time
the ball makes contact with the floor the velocity $v^{-}$ is reset to a new value given by $v^{+}=-\gamma v^{-}$. The guard condition for resetting the velocity is given by $y(t)=0$ (where $y(t)$ is the height of the ball above the floor at time $t$). The equations for the bouncing ball are given by,
\begin{eqnarray}
\dot y &=& v, \nonumber\\
\dot v &=& -g,
\label{eq:balleqs}
\end{eqnarray}
with the reset condition at $y=0$: $v^{+}=-\gamma v^{-}$. Thus, if one uses the boundary layer approximation in Eqn.~\ref{eq:blayer} we get,
\begin{equation}
\begin{pmatrix}
\dot y\\
\dot v\\
\dot z
\end{pmatrix} = \begin{cases}
\begin{pmatrix}
v\\
-g\\
(v - z)/\epsilon
\end{pmatrix} & \mbox{if}: |y|\geq\epsilon \\
\\
\begin{pmatrix}
\epsilon v\\
(\gamma z - v)/\epsilon \\
0
\end{pmatrix} & \mbox{otherwise.}
\end{cases}
\label{eq:blayer_ball}
\end{equation}
We now use the hybrid polynomial chaos expansion in Eqn.~\ref{eq:intregionpc} along with the Wiener-Haar wavelet basis functions. The Monte Carlo simulations on the bouncing ball are shown in Fig.~\ref{Fig:MonteBall}. The average or nominal trajectory is also shown. In Fig.~\ref{Fig:ByLayerBall} we compare the nominal trajectory (mean trajectory from Monte Carlo) with the mean predicted using the boundary layer expansion with $\epsilon=0.01$. As shown, the boundary layer accurately approximates the mean over multiple state resets events (in this case, each impact with the floor).
\begin{figure}
\centering
\subfigure[]{\includegraphics[scale=0.15]{BallMonteCarlo2_CDC.eps}\label{Fig:MonteBall}}
\subfigure[]{\includegraphics[scale=0.15]{ByLayerComp2_CDC.eps}\label{Fig:ByLayerBall}}
\caption{a) Monte Carlo simulation of the bouncing ball system with uncertain gravitational acceleration. b) Nominal bouncing ball trajectory compared with the mean trajectory obtained through the wavelet-based hybrid UQ approach with the boundary layer approximation ($\epsilon = 0.01$). \label{Fig:DistBall}}
\end{figure}
\subsection{Results}\label{Sec:Results}
To demonstrate the hybrid polynomial chaos approach on hybrid dynamical systems we
consider the simple yet challenging example of a switching
oscillator given by Eqn.~\ref{eq:SHMswitch}.
\begin{align}
\ddot x + c\dot x + x + \lambda &= 0\,\,\,\text{if $x \geq 0$} \nonumber \\
\ddot x + c\dot x + x - \lambda &= 0\,\,\,\text{if $x < 0$}. \nonumber
\end{align}
The value of $c$ is deterministic and equal to~0.5. Here we
consider three cases with $\lambda$ normally distributed with:
$\mu(\lambda)=-10$ and $\sigma(\lambda)=2$ (case~1), $\mu(\lambda)=10$ and
$\sigma(\lambda)=2$ (case~2), and $\mu(\lambda)=0$ and $\sigma(\lambda)=1$ (case~3). In all cases we assume that the initial
conditions are deterministic and given by $\left[x(0),\dot
x(0)\right] = \left[10^{-2},1.0\right]$.
\subsubsection{Case 1: $\mu(\lambda)=-10$, $\sigma(\lambda)=2$}
Let us start with the case when $\mu(\lambda)=-10$ and
$\sigma(\lambda)=2$ in Eqn.~\ref{eq:SHMswitch}. A representative
trajectory for the dynamics of the system is shown in
Fig.~\ref{Fig:Trajcase1}. The corresponding histogram for
$x(20.0;\lambda)$ is shown in Fig.~\ref{Fig:Histcase1}. Most
importantly, one desires to compute the mean and variance of
$x(t;\lambda)$ as a function of time. In the system given by
Eqn.~\ref{eq:SHMswitch}, we expand $x(t;\lambda)$ using
Eqn.~\ref{eq:expx} and perform a Galerkin projection as shown in
Eqn.~\ref{eq:intregionpc}. One then gets a system of equations
for the coefficients of expansion in
Eqn.~\ref{eq:expx}. These coefficients, once computed, can be
used to calculate the moments of the distribution of
$x(t;\lambda)$.
We compare the results obtained from hybrid polynomial chaos with those
obtained using Monte Carlo and Quasi-Monte Carlo based methods.
In particular, we use a Weyl sequence~\cite{Niederreiter1992}
along with inverse transform sampling~\cite{InvTran:book} to
generate the Quasi-Monte Carlo samples. We find that the results
(in the first two moments) from $5000$ samples of Monte Carlo,
$3000$ samples of Quasi-Monte Carlo and the Wiener-Haar hybrid PC
expansion with $P=3$ are visually indistinguishable (see
Figs.~\ref{Fig:Meanxcase1} and~\ref{Fig:Varxcase1}). Treating
$5000$ Monte Carlo samples as baseline, we find that the hybrid PC expansion has a maximum error of $5\times
10^{-2}$ in the prediction of $\mu(x(t;\lambda))$.
\begin{figure}
\centering
\subfigure[]{\includegraphics[scale=0.4]{SingleTrajcase1.eps}\label{Fig:Trajcase1}}
\subfigure[]{\includegraphics[scale=0.4]{HistCase2MC.eps}\label{Fig:Histcase1}}
\caption{a) A representative trajectory for case $1$. b) Histogram of $x(20;\lambda)$ for case $1$.}
\end{figure}
\begin{figure}
\centering
\subfigure[]{\includegraphics[scale=0.3]{Case2MeanP3.eps}\label{Fig:Meanxcase1}}
\subfigure[]{\includegraphics[scale=0.3]{Case2VarP3.eps}\label{Fig:Varxcase1}}
\caption{ Comparison of a) predicted mean and b) predicted variance of $x(t;\lambda)$ for various UQ methods for case $1$.}
\end{figure}
\subsubsection{Case 2: $\mu(\lambda)=10$, $\sigma(\lambda) = 2$}
The case of $\mu(\lambda)=10$ and $\sigma(\lambda)=2$ is
significantly more challenging. A representative trajectory of
the system (for $\lambda = 10$) is shown in
Fig.~\ref{Fig:Trajcase2}. When $\lambda > 0 $ the system
switches back and forth between modes. The reason for this is that when the system is in the right half-plane, the equilibrium
of the system is in the left half-plane and vice versa. A
histogram for $x(3.0;\lambda)$ is depicted in
Fig.~\ref{Fig:Histcase2}.
We again compare hybrid Wiener-Haar polynomial chaos to Monte Carlo sampling in
Fig.~\ref{Fig:case2}. We find that hybrid Wiener-Haar polynomial chaos
($p=5$) accurately computes the mean and the variance of the distribution of $x$. The maximum absolute error of hybrid polynomial chaos in mean is $\mu(x(t;\lambda) = 1.8\times10^{-3}$ and variance is $7\times10^{-5}$. Note that expansions
in terms of standard basis functions such as Hermite and Legendre polynomials
are unable to compute the moments of $x(t;\lambda)$ beyond a
threshold time that depends weakly on the order of
expansion~$p$ (see Fig.~\ref{Fig:failcase2}). The solution in this case is particularly challenging because it becomes more oscillatory in
terms of $\lambda$ at $t$ increases (Fig.~\ref{Fig:Oscfail}). The Wiener-Haar basis functions are naturally oscillatory and hence more accurate than Hermite polynomials in capturing the solution $x(t;\lambda).$ Note that, for large time simulations the Wiener-Haar
expansions will also fail since the solution will eventually become too oscillatory for the order of expansion. This problem is well known in the polynomial chaos literature~\cite{Longterm}.
\begin{figure}
\centering
\subfigure[]{\includegraphics[scale=0.4]{DynamicsCase1.eps}\label{Fig:Trajcase2}}
\subfigure[]{\includegraphics[scale=0.4]{HistCase1MC.eps}\label{Fig:Histcase2}}
\caption{a) A representative trajectory for case $2$. b) Histogram of $x(3.0;\lambda)$ for case~2.\label{Fig:Dycase2}}
\end{figure}
\begin{figure}
\subfigure[]{\includegraphics[scale=0.4]{Case1MeanP5.eps}}
\subfigure[]{\includegraphics[scale=0.4]{Case1VarP5.eps}\label{Fig:case2var}}
\caption{Comparison of a) predicted mean and b) predicted variance of $x(t;\lambda)$ by various UQ methods for case $2$. \label{Fig:case2}}
\end{figure}
\begin{figure}
\centering
\subfigure[]{\includegraphics[scale=0.3]{MeanxCase1CompLong.eps}\label{Fig:failcase2mean}}
\subfigure[]{\includegraphics[scale=0.3]{VarxCase1CompLong.eps}\label{Fig:failcase2var}}
\caption{a) Mean and b) Variance predicted by standard (Hermite) polynomial chaos basis functions for case $2$.\label{Fig:failcase2}}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.5\hsize]{increased_lambda_oscillations.eps}
\caption{Case~2: $x(t;\lambda)$ becomes more oscillatory in $\lambda$ as $t$ increases.
The bottom plot shows the distribution for~$\lambda$. \label{Fig:Oscfail}}
\end{figure}
\subsubsection{Case 3: $\mu(\lambda)=0$, $\sigma(\lambda) = 1.0$}
We also consider the case of $\mu(\lambda)=0$ with $\sigma(\lambda)=1.0$. This case is particularly challenging because there is a concentration of probability of $x(t;\lambda)$ as shown in Fig.~\ref{Fig:Histcase3}. The reason for this is as follows: when $\mu(\lambda)=0$, the nominal trajectory converges to $0$ and so do all trajectories with $\lambda>0$. Indeed, for trajectories with $\lambda>0$ the equilibrium lies in the opposite half-plane with respect to the current state. This gives rise to decaying switching trajectories, as case 2 in Fig.~\ref{Fig:Trajcase2}. Note that for $\lambda<0$, the trajectories are similar to the ones in case 1 (Fig.~\ref{Fig:Trajcase1}).
The Wiener-Haar basis functions along with the hybrid polynomial chaos approach accurately capture the moments of the distribution for $x(t;\lambda)$ (see Fig.~\ref{Fig:case3}). In fact, an expansion to just $P=3$ captures the first two moments. The step-function nature of the Wiener-Haar basis allows it to perform well in this scenario. Standard basis functions like Hermite polynomials are completely incapable of accurately capturing the moments of the distribution for $x(t;\lambda)$ shown in Fig.~\ref{Fig:Histcase2}.
\begin{figure}
\centering
\subfigure[]{\includegraphics[scale=0.4]{TrajMu0.eps}\label{Fig:Trajcase3}}
\subfigure[]{\includegraphics[scale=0.4]{HistCase3MC.eps}\label{Fig:Histcase3}}
\caption{a) Nominal trajectory for case~3. b) Histogram of $x(20;\lambda)$ for case~3.}
\end{figure}
\begin{figure}
\subfigure[]{\includegraphics[scale=0.4]{Case3MeanP3.eps}}
\subfigure[]{\includegraphics[scale=0.4]{Case3VarP3.eps}}
\caption{Comparison of a) predicted mean and b) predicted variance of $x(t;\lambda)$ for various UQ methods for case $3$.\label{Fig:case3}}
\end{figure}
\section{Transport theory approach for uncertainty quantification in hybrid systems}
\label{sec:transporttheory}
In this section we present a qualitatively different approach to UQ in hybrid systems based on transport equations. We write an advection equation for the probability density of the state and expand this equation in an appropriate basis, as is done in polynomial chaos. The resulting equation is equivalent to the Fokker-Planck equation~\cite{Cit:FPE:book} in the absence of a diffusion term. Though significant effort has been put into computing solutions for the Fokker-Planck equation in various applications~\cite{Cit:FPE:book,Cit:FPE:book2}, our setting is particularly challenging due to the switching dynamics of hybrid systems. We note that advection equations for probability distribution functions have been used to propagate uncertainty through heterogeneous porous media with uncertain properties~\cite{Tartakovsky2011} and for hyperbolic conservation laws with noise~\cite{Luo2006}. Recently, similar methods have been extended to cumulative distribution functions in hyperbolic conservation laws~\cite{Wang2012}.
The polynomial chaos expansion in this setting yields a system of hyperbolic partial differential equations for the coefficients of the expansion, which are then solved by integrating along characteristics. The hybrid nature of the original system is reflected in that the characteristics exhibit switching. Even though we only consider systems without resets, we can use the results of Sec.~\ref{Sec:resets} to treat systems with resets.
As in Sec.~\ref{Sec:hyb_poly}, let us consider a hybrid system without resets and uncertain parameters~$\lambda$ with guard conditions independent
of~$\lambda$:\footnote{Note that this embodies the constraint of having no overlap in the domains for different modes.}
\[
\dot{x} = f_i(x,\lambda) \quad \text{when $G_i(x)$ is true.}
\]
The system has uncertain initial conditions described by the
probability density~$\rho_{x0} (x)$ and the uncertain parameters
$\lambda$ follow~$\rho_\lambda (\lambda)$.
We describe the system by the time evolution of the distribution
function $\rho(x,\lambda;t)$, which has initial condition
\[
\rho(x,\lambda;0) = \rho_{x0}(x) \rho_\lambda(\lambda) \quad
\text{(initial uncertainties are independent)}
\]
and normalization
\[
\int \rho(x,\lambda;t) dx d\lambda = 1.
\]
Note that, for all time, we have
\begin{equation}
\rho_\lambda(\lambda) = \int \rho(x,\lambda;t) dx.
\label{rho_lambda_is_constant}
\end{equation}
Our goal is to compute the evolution of the density in~$x$ (the
marginal distribution):
\[
\rho_x(x,t) = \int \rho(x,\lambda;t) d\lambda.
\]
However, without introducing assumptions on $\rho_\lambda$, the equation for $\rho_x$ is not closed. We therefore focus on computing the evolution of $\rho$ directly through an expansion. From this evolution, $\rho_x$ can then be calculated at every instant.
\subsection{Equation for $\rho$}
Let us define the sets $S_i = \left\{ x \; : \; G_i(x) \;
\text{is true} \right\}$ and the indicator functions
\[
\mathbf{1}_i(x) = \begin{cases} 1 & \text{if $x \in S_i$} \\ 0 &
\text{if not}. \end{cases}
\]
With this notation, and because $\lambda$ is constant along a
trajectory, we have
\begin{equation}
\frac{\partial \rho}{\partial t} + \nabla \cdot (\rho f) = 0
\qquad \forall \lambda \label{liouville}
\end{equation}
where $f(x,\lambda) = \sum_i \mathbf{1}_i(x) f_i(x,\lambda)$ and the gradient operator acts only on~$x$ and not on~$\lambda$.
\subsection{Boundary conditions at the interfaces}
The discontinuity in the equation implies that mass may
accumulate at the boundaries between zones where different guard
conditions are valid. Integrating on a cylinder that
crosses one such boundary we obtain the matching condition
\begin{equation}
\frac{\partial \sigma}{\partial t} + \nabla_s \cdot (\sigma f) =
\rho_i f_i \cdot \hat{n}_{ik} - \rho_k f_k \cdot \hat{n}_{ik},
\label{boundary_condition}
\end{equation}
where $\sigma$ is a surface probability density between regions
$S_i$ and~$S_k$, $\hat{n}_{ik}$ is the surface normal from $S_i$
to~$S_k$, $\nabla_s$ is the divergence in the space tangent to
the surface, and $f$ is the flow \emph{at the
surface}.\footnote{Whether $f$ is $f_i$ or $f_k$ on the surface
will depend on how the guard conditions are expressed.} This may
lead to a cascade, with probability condensing into
progressively lower dimensional structures: where two
hypersurfaces meet (the boundary between three guard conditions)
the same scenario repeats, until we have mass accumulating at
points. To solve Eqn.~\ref{liouville} the initial condition must
include initial values for~$\sigma$ and the
probability density on any lower dimensional structure where
mass may accumulate.
The discontinuity of $f$ does not \emph{necessarily} imply
accumulation. In fact, at an interface we have several options:
\begin{enumerate}
\item Both $f_i \cdot \hat{n}_{ik}$ and $f_k \cdot \hat{n}_{ik}$ are nonzero and have the same sign. In this case, there is no accumulation. If we assume that $\rho$ has a singularity at the interface, the flow will move the singularity away from it.
\item $f_i \cdot \hat{n}_{ik} > 0$ and $f_k \cdot \hat{n}_{ik} \leq 0$: accumulation occurs.
\item $f_i \cdot \hat{n}_{ik} \geq 0$ and $f_k \cdot \hat{n}_{ik} < 0$: accumulation occurs.
\item $f_i \cdot \hat{n}_{ik} = f_k \cdot \hat{n}_{ik} = 0$: no accumulation.
\item $f_i \cdot \hat{n}_{ik} \leq 0$ and $f_k \cdot \hat{n}_{ik} \geq 0$: no accumulation.
\end{enumerate}
If we are in the case without accumulation and without initial concentration of density in lower dimensional structures, then $\sigma=0$ and
Eqn.~\ref{boundary_condition} becomes
\begin{equation}
\rho_i f_i \cdot \hat{n}_{ik} = \rho_k f_k \cdot \hat{n}_{ik}.
\label{flow_matching}
\end{equation}
\begin{thm}
\label{thm:noaccumulation}
Any second order ODE of the form
\[
\ddot{X} = F_i(X,\dot{X},\lambda) \quad \text{when $G_i(X)$ is
true}
\]
satisfies the conditions for no accumulation at the interface
where the ODE is discontinuous.
\end{thm}
\begin{proof}
Without loss of generality we can focus on just two regions
$S_i$ and~$S_k$ and rewrite the problem as the first-order ODE
\[
\dot{x} \equiv \left(\begin{matrix} \dot{X} \\ \dot{Y}
\end{matrix} \right) = f(x) = \left(\begin{matrix} Y \\
F_i(X,Y,\lambda) \end{matrix} \right) \quad \text{when $G_i(X)$
is true.}
\]
To find the normal $\hat{n}_{ik}$ we consider a ${\cal C}^1$
function $b(x) = b(X,Y) = B(X)$ that is positive in $S_k$ and
negative $S_i$ so that the interface is given by the locus of
$b(x)=0$. The gradient of this function is proportional to the
normal:
\[
\hat{n}_{ik} \propto \nabla b = \left(\begin{matrix} \nabla_X B
\\ 0 \end{matrix} \right)
\]
and therefore
\[
\hat{n}_{ik} \cdot f = \frac{Y \cdot \nabla_X B}{||\nabla_X
B||},
\]
which is continuous at the interface.
\end{proof}
\subsection{Expansion of the equation for $\rho$}
At every point $x$ we expand the distribution in $\lambda$:
\begin{equation}
\rho(x,\lambda;t) = \sum_k a_k(x,t) w(\lambda) \psi_k(\lambda),
\label{expansion}
\end{equation}
where $\{\psi_k\}$ forms an orthogonal basis with respect to
$w$:
\[
\int \psi_i(\lambda) \psi_k(\lambda) w(\lambda) d\lambda = w_k
\delta_{ik}.
\]
We keep $w_k$ to allow for a non-normalized weight function
$w(\lambda)$. We replace the expansion in Eqn.~\ref{expansion}
into Eqn.~\ref{liouville} and project onto $\psi_i$ to obtain a set of partial differential equations for the coefficients $a_i(x,t)$:
\begin{equation}
\frac{\partial a_i(x,t)}{\partial t} + \frac{1}{w_i}\nabla \cdot
\sum_k a_k(x,t) \int w(\lambda) \psi_i(\lambda) \psi_k(\lambda)
f(x,\lambda) d\lambda = 0. \label{liouville_expanded}
\end{equation}
Since this equation is local in~$x$ there is no question as to
which $f(x,\lambda)$ must be used at any given point.
\subsection{Example: switching oscillator}
Here we revisit the switching oscillator system
\[
\ddot{x} = \begin{cases} -x - \gamma \dot{x} - \lambda & \text{if $x \geq 0$} \\
-x - \gamma \dot{x} + \lambda & \text{otherwise,} \end{cases}
\]
which can be expressed as the 2--D system
\[
\left(\begin{matrix} \dot{x} \\ \dot{y} \end{matrix}\right) =
\left( \begin{matrix} f_x \\ f_y \end{matrix} \right)
\]
where $f_x = y$ and
\[
f_y =
\begin{cases} -x - \gamma y - \lambda & \text{if $x \geq 0$} \\
-x - \gamma y + \lambda & \text{otherwise}.
\end{cases}
\]
The transition points are located at $x=0$ and therefore
\[
f \cdot \hat{n} = f \cdot \left(\begin{matrix}1 \\ 0
\end{matrix} \right) = f_x = y
\]
which is continuous and therefore has the same sign on both
sides. Therefore there is no mass accumulation at the interface
for this system. This is a special case of theorem~\ref{thm:noaccumulation}.
\subsubsection{Case~3 revisited}
To connect with the example presented in case~3 ($\mu=0$, $\sigma=1$) we choose $w(\lambda)= e^{-\lambda^2/2}$ and $\psi_k$'s as the probabilist's Hermite polynomials $H_k$, with the following properties
\begin{align*}
\int H_i(\lambda) H_k (\lambda) w(\lambda) d\lambda &= k! \sqrt{2\pi} \delta_{ik} \\
H_{k+1}(\lambda) &= \lambda H_k(\lambda) - H_k'(\lambda) \\
H_k'(\lambda) &= k H_{k-1}(\lambda).
\end{align*}
Calculating the terms in Eqn.~\ref{liouville_expanded}:
\begin{align*}
\int w H_i H_k f_x d\lambda &= y \, i! \sqrt{2\pi} \delta_{ik} \\
\int w H_i H_k f_y d\lambda &= - (x+\gamma y) \, i! \sqrt{2\pi}
\delta_{ik} \mp i! \sqrt{2\pi} (\delta_{i,k+1} + k
\delta_{i,k-1}),
\end{align*}
where the upper sign ($-$) is for $x\ge 0$ and the lower sign
($+$) is for~$x < 0$. Substituting into Eqn.~\ref{liouville_expanded} we
obtain the equations
\begin{equation}
\begin{split}
0 &= \partial_t a_0 + \partial_x (y a_0) + \partial_y \left[-(x + \gamma y)a_0 \mp a_1 \right] \\
0 &= \partial_t a_i + \partial_x (y a_i) + \partial_y \left[-(x
+ \gamma y)a_i \mp (i+1) a_{i+1} \mp a_{i-1} \right] \quad (i\ge
1).
\end{split}
\label{oscillator_hyperbolic_system}
\end{equation}
Note that, instead of using the Hermite polynomials, one can use the Haar wavelet expansion~\cite{LeMaitre2004} to represent the solution as was done in previous sections. We plan to present this calculation in future work.
\begin{thm}
\label{thm:oscillatorishyperbolic}
The system of PDEs for the switching oscillator is hyperbolic.
\end{thm}
\begin{proof}
Consider a system of PDEs of the form
\[
\partial_t u + \sum_\nu A_\nu \partial_\nu u = B,
\]
where $u(x_1, \ldots, x_n, t) \in \mathbb{R}^m$ and the $A_\nu$
are $m\times m$ matrices. The system is hyperbolic if for any
$\alpha_\nu \in \mathbb{R}$ the linear combination $A = \sum_\nu
\alpha_\nu A_\nu$ has real eigenvalues.
For the switching oscillator the system of PDEs can be written
as
\begin{multline*}
\partial_t \begin{pmatrix} a_0 \\ a_1 \\ \vdots \end{pmatrix}
+ \begin{pmatrix} y & 0 & \hdotsfor{2} \\ 0 & y & & \\ \vdots
& & \ddots & \end{pmatrix}
\partial_x \begin{pmatrix} a_0 \\ a_1 \\ \vdots \end{pmatrix} \\
+ \begin{pmatrix}
\beta & \mp 1 & 0 & \hdotsfor{2} \\
\mp 1 & \beta & \mp 2 \\
0 & \mp 1 & \beta & \mp 3 \\
\vdots & & \ddots & \ddots & \ddots
\end{pmatrix}
\partial_y \begin{pmatrix} a_0 \\ a_1 \\ \vdots \end{pmatrix}
= \gamma \begin{pmatrix} a_0 \\ a_1 \\ \vdots \end{pmatrix},
\end{multline*}
where $\beta = - (x+ \gamma y)$. Thus, any combination of the
matrices $A_\nu$ is going to be of the tridiagonal form
\[
A = \begin{pmatrix}
a & b & 0 & \ldots & \text{\huge{0}} \\
b & a & 2b & 0 & \ldots\\
0 & b & a & 3b \\
\vdots & 0 & b & a \\
\text{\huge{0}} & \ldots & & & \ddots
\end{pmatrix}.
\]
This tridiagonal non-symmetric matrix is similar to a
tridiagonal symmetric matrix with a diagonal similarity matrix:
$S = D A D^{-1}$, where
\[
D = \begin{pmatrix} \sqrt{0!} & \\ & \sqrt{1!} & & & \text{\huge{0}} \\ & &
\sqrt{2!} \\ \text{\huge{0}}& & & \sqrt{3!} \\ & & & & \ddots
\end{pmatrix}.
\]
$A$ is similar to $S$, a symmetric and real matrix, and
therefore $A$~has real eigenvalues.
\end{proof}
The issue of hyperbolicity is discussed in more depth in~\cite{Tryoen2010a}. To solve the hyperbolic system from Eqn.~\ref{oscillator_hyperbolic_system} we write it in the form
\[
\partial_t a + y \partial_x a - (x + \gamma y) \partial_y a \mp A \partial_y a = \gamma a,
\]
where $A$ is the tridiagonal matrix
\begin{equation}
A =
\begin{pmatrix}
0 & 1 \\
1 & 0 & 2 & & \text{\huge{0}}\\
& 1 & 0 & 3 \\
\text{\huge{0}} & & 1 & 0 \\
& & & & \ddots
\end{pmatrix}.
\label{hermite_matrix}
\end{equation}
We diagonalize $A = P \Lambda P^{-1}$ and define $b = P^{-1} a$
to obtain the set of uncoupled hyperbolic PDEs
\begin{equation}
\partial_t b_i + y \partial_x b_i + \left[-(x + \gamma y) \mp \lambda_i\right] \partial_y b_i = \gamma b_i,
\label{diagonalized_system}
\end{equation}
where $\lambda_i$ is the $i$-th eigenvalue. We will now prove that when the expansion
is truncated up to $a_{n-1}$, i.e., $A$ is truncated to a $n\times n$
matrix, the eigenvalues $A$ are the zeros of~$H_n$.
\begin{thm}
\label{thm:hermite_zeros}
The eigenvalues of $A_n$, the $n\times n$ truncated version of
the matrix in Eqn.~\ref{hermite_matrix}, are the zeros of the $n$-th
order probabilist's Hermite polynomial~$H_n$.
\end{thm}
\begin{proof}
We proceed by induction to prove that $\det (A_n - \lambda I) =
H_n(-\lambda)$, which will then, by the symmetry of $H_n$, prove
our result.
Let $B_n = A_n - \lambda I$. Indeed, $\det B_1 = -\lambda$ and
$\det B_2 = \lambda^2 - 1$. For the general case,
\[
B_{n+1} =
\begin{pmatrix}
& & & & 0 \\
& B_n & & & \vdots \\
& & & & 0 \\
& & & & n \\
0 & \cdots & 0 & 1 & -\lambda
\end{pmatrix}
\]
Therefore,
\[
\det B_{n+1} = -\lambda \det B_n + (-1)^n n (-1)^{n-1} \det
B_{n-1} = - \lambda \det B_n - n \det B_{n-1},
\]
which is the recurrence relation satisfied by $H_n(-\lambda)$.
\end{proof}
The characteristic curves of Eqn.~\ref{diagonalized_system} are
given by
\begin{align*}
\dot{x} &= y \\
\dot{y} &= - x - \gamma y \mp \lambda_i \\
\dot{b}_i &= \gamma b_i.
\end{align*}
In other words, the characteristics are damped oscillators where
the equilibrium position is given by the eigenvalues of~$\mp A$.
The exponential growth of $b_i$ along a trajectory is due to the
contraction in phase space produced by the dissipation~$\gamma$.
\subsubsection{Results}
We now show results obtained by using transport theory approach on case~3 ($\mu=0$, $\sigma=1$) for the switching oscillator (Eqn.~\ref{eq:SHMswitch}).
In Fig.~\ref{Fig:MC_vs_transport_colormaps}, we show a series of probability distribution snapshots for Monte Carlo ($5000$ samples) and the transport operator method (for case~3), gridded in the $(x,y)$ plane. Note that we set $y=\dot x$, as defined in the first part of the paper. The Monte Carlo color map snapshots show that the distribution lies on a one-dimensional manifold (in two dimensional space). The one-dimensional nature of the distribution arises because we chose deterministic initial conditions. As discussed previously, all trajectories in case~3 with $\lambda \geq 0$ converge to the origin, resulting in a jump in the cumulative distribution function (CDF).
\begin{figure
\centerline
\includegraphics[width=\columnwidth]{colormaps_MC_5ksamples_vs_PC_order70.eps}
\caption{Snapshots at $t=2,4,6,8,10$ of the gridding from 5000 samples of Monte Carlo and an order 70 expansion using the transport theory based method. The right column compares the one-dimensional cumulative distribution function for the corresponding times.
\label{Fig:MC_vs_transport_colormaps
\end{figure}
\begin{figure
\centerline
\includegraphics[width=0.7\columnwidth]{CDF_MC_5ksamples_vs_PC_order70_t_18.eps}
\caption{Comparison at $t=18$ of the cumulative distribution function from 5000 samples of Monte Carlo (solid) and an order~70 expansion using the transport theory based method (dashed).
\label{Fig:MC_vs_transport_CDF
\end{figure}
For $\lambda < 0$, however, the trajectories converge asymptotically to either $+\lambda$ or $-\lambda$. This convergence to $+\lambda$ or $-\lambda$ is highly dependent on the individual trajectory and results in a fragmentation of the output distribution. Close to convergence ($t=18$), very few trajectories converge to a point in the range $0.2 < x < 0.4$, as seen in the flat region of the CDF in Fig.~\ref{Fig:MC_vs_transport_CDF}. The transport-based method captures the singularity at the origin accurately, but is unable to accurately capture the fragmentation. This is because the method samples the distribution sparsely (determined by the order of expansion), resulting in UQ acceleration. However, this sparsity makes the method miss such fine details. On using a high order of expansion ($n=70$), some samples partially capture the structure around $x=0$. Note that a much lower order expansion accurately captures the jump at the origin and the asymptotic ($x > 1$) shape of the CDF.
As in Fig.~\ref{Fig:case3}, we compare Monte Carlo (5000 samples) with the transport theory approach (orders~$15$ and~$70$ expansion) in Fig.~\ref{Fig:pde_approach_mean_and_var_case3}. The $\mbox{L}_{\infty}$ (maximum) errors for $n=15$ are $9.56\times 10^{-2}$ (mean) and $7.95\times 10^{-2}$ (variance). For $n=70$ we get $\mbox{L}_{\infty}$ errors of $5.28\times 10^{-2}$ and $4.92\times 10^{-2}$ for mean and variance respectively. As shown in Fig.~\ref{Fig:case3}, the method performs reasonably well, however, the results are not nearly as good as those obtained using hybrid polynomial chaos with the Wiener-Haar wavelet expansion in section~\ref{Sec:hpc}. However, with a better choice of basis functions, one does expect better results. The transport operator theory is attractive as it appears to be more versatile. In general, the transport operator approach is applicable to hybrid dynamical systems with overlapping modes of operation (by constructing multiple PDEs for the overlapping mode). In contrast, the hybrid polynomial chaos method suffers from the disadvantage of being inapplicable to such systems.
\begin{figure}
\subfigure[]{\includegraphics[scale=0.4]{mean_MC_5ksamples_vs_PC.eps}}
\subfigure[]{\includegraphics[scale=0.4]{variance_MC_5ksamples_vs_PC.eps}}
\caption{Comparison of a) predicted mean and b) predicted variance of $x(t;\lambda)$ for case~3. The curves show a 5000-sample Monte Carlo run, and 15 \& 70 term expansions using the transport theory approach. \label{Fig:pde_approach_mean_and_var_case3}}
\end{figure}
\section{Conclusions}\label{Sec:conclusions}
As the modeling of hybrid dynamical systems becomes increasingly
important for modern day engineering applications such as
electrical and biological networks, air traffic systems,
communication networks, etc., quantifying uncertainty in these
systems is going to become a central concern. Since uncertainty
quantification allows one to compute moments of output
distributions in the presence of parametric
uncertainty, these techniques will be used to aid decisions related to robust system design and performance.
In this work, we have made the first attempts to develop fast
uncertainty quantification methods targeted for hybrid dynamical
systems. In particular, we extended polynomial chaos methods, a
popular technique for propagating uncertainty through smooth systems, to hybrid dynamical systems. We also
developed methods to handle state resets within the polynomial
chaos framework by using boundary layer approximations. We then
applied this new approach to perform uncertainty quantification
on switching harmonic oscillators and the bouncing ball examples. We also
demonstrated the efficacy of using Wiener-Haar expansions~\cite{Najm2009, LeMaitre2004,LeMaitre2004a} with our
hybrid polynomial chaos approach for quantifying uncertainty in hybrid systems that
give rise to multi-modal distributions or become increasingly
oscillatory in time. Finally, we showed how a transport theory based approach can capture naturally-emerging discontinuities in the distribution. Future efforts involve providing rigorous error bounds for Wiener-Haar expansions in the hybrid polynomial chaos setting with boundary layer expansions. We are also extending our hybrid polynomial chaos approach to networks of hybrid dynamical systems using our recent work on propagating uncertainty through complex networks~\cite{Tuhin_iter}. We also intend to extend the transport operator based UQ method to hybrid systems with overlapping modes of operation.
\section{Acknowledgements}
The authors thank Habib Najm for pointing us to his work on Wiener-Haar based polynomial chaos expansion and his insightful input. We also thank Alessandro Pinto and George Mathew for valuable discussions related to hybrid dynamical systems.
\bibliographystyle{unsrt}
|
\section{\@startsection {section}{1}{\z@}{-3.5ex plus-1ex minus
-.2ex}{2.3ex plus.2ex}{\reset@font\Large\bf}}
\makeatother
\renewenvironment{equation}{\refstepcounter{subsection}\refstepcounter{prop}$$}{\leqno{\bf (\theprop)}$$}
\def\thesection.\arabic{prop}{\thesection.\arabic{prop}}
\def\mar[#1]{\ar@{-}[#1]|-{\object@{<}}}
\def\marb[#1]{\ar@{-}[#1]|{\object+{ }}}
\newcommand{\fleche}[2]{\xymatrix{*!U(-0.25){#1}&*!U(-0.25){#2}\mar[l]}}
\newcommand{\perm}[1]{\underline{\mathsf{Perm}}_k(#1)}
\newcommand{\operm}[1]{\mathsf{Perm}_k(#1)}
\newcommand{\permalg}[1]{\underline{\mathsf{PermAlg}}_k(#1)}
\newcommand{\opermalg}[1]{\mathsf{PermAlg}_k(#1)}
\usepackage[all]{xy}
\def\rm Inn{\rm Inn}
\def\longrightarrow{\longrightarrow}
\def\stackrel{{_{\cong}}}{\to}{\stackrel{{_{\cong}}}{\to}}
\renewcommand{\oplusb}[2]{\mathop{\oplus}_{{\scriptstyle #1}\atop{\scriptstyle #2}}}
\def\overrightarrow{\Delta}{\overrightarrow{\Delta}}
\def\overleftarrow{\Delta}{\overleftarrow{\Delta}}
\def\mathop{{\hbox{\raisebox{-1.7ex}{\LARGE\rm *}}}}{\mathop{{\hbox{\raisebox{-1.7ex}{\LARGE\rm *}}}}}
\newcommand{\monast}[2]{{{\raisebox{-.5ex}{$\scriptstyle #2$}}\atop{\rule{0ex}{4ex}\hbox{\Huge\rm *}\atop{\smash{\raisebox{1.5ex}{$\scriptstyle#1$}}}}}\!}
\def\medskip\par{\medskip\par}
\begin{document}
\centerline{\Large\bf The Roquette category of finite $p$-groups}\vspace{.5cm}\par
\centerline{\bf Serge Bouc
}\vspace{1cm}\par
{\footnotesize {\bf Abstract :} Let $p$ be a prime number. This paper introduces the {\em Roquette category} $\mathcal{R}_p$ of finite $p$-groups, which is an additive tensor category containing all finite $p$-groups among its objects. In $\mathcal{R}_p$, every finite $p$-group $P$ admits a canonical direct summand $\partial P$, called {\em the edge} of $P$. Moreover $P$ splits uniquely as a direct sum of edges of {\em Roquette $p$-groups}, and the tensor structure of $\mathcal{R}_p$ can be described in terms of such edges.}\par
{\footnotesize The main motivation for considering this category is that the additive functors from~$\mathcal{R}_p$ to abelian groups are exactly the {\em rational $p$-biset functors}. This yields in particular very efficient ways of computing such functors on arbitrary $p$-groups~: this applies to the representation functors $R_K$, where $K$ is any field of characteristic 0, but also to the functor of units of Burnside rings, or to the torsion part of the Dade group.
}\vspace{2ex}\par
{\footnotesize {\bf AMS Subject classification :} 18B99, 19A22, 20C99, 20J15.\vspace{2ex}}\par
{\footnotesize {\bf Keywords :} $p$-group, Roquette, rational, biset, genetic.}
\section{Introduction}
Let $p$ be a prime number. This article introduces {\em the Roquette category} $\mathcal{R}_p$ of finite $p$-groups, which is an additive tensor category with the following properties~:
\begin{itemize}
\item Every finite $p$-group can be viewed as an object of $\mathcal{R}_p$. The tensor product of two finite $p$-groups $P$ and $Q$ in $\mathcal{R}_p$ is the direct product $P\times Q$.
\item In $\mathcal{R}_p$, any finite $p$-group has a direct summand $\partial P$, called {\em the edge} of~$P$, such that
$$P\cong\dirsum{N\mathop{\trianglelefteq} P}\partial (P/N)\;\;.$$
Moreover, if the center of $P$ is not cyclic, then $\partial P=0$.
\item In $\mathcal{R}_p$, every finite $p$-group $P$ decomposes as a direct sum
$$P\cong\dirsum{R\in \mathcal{S}}\partial R\;\;,$$
where $\mathcal{S}$ is a finite sequence of {\em Roquette groups}, i.e. of $p$-groups of normal $p$-rank 1, and such a decomposition is essentially unique. Given the group~$P$, such a decomposition can be obtained explicitly from the knowledge of a {\em genetic basis} of $P$ (Theorem~\ref{Roquette category} and Proposition~\ref{isomorphism}).
\item The tensor product $\partial P\times\partial Q$ of the edges of two Roquette $p$-groups $P$ and $Q$ is isomorphic to a direct sum of a certain number $\nu_{P,Q}$ of copies of the edge $\partial (P\diamond Q)$ of another Roquette group (where both $\nu_{P,Q}$ and $P\diamond Q$ are known explicitly - see Theorem~\ref{structure diamond} and Corollary~\ref{edge product}).
\item The additive functors from $\mathcal{R}_p$ to the category of abelian groups are exactly the {\em rational $p$-biset functors} introduced in~\cite{bisetsections}.
\end{itemize}
The latter is the main motivation for considering this category~: any structural result on $\mathcal{R}_p$ will provide for free some information on such rational functors for $p$-groups, e.g. the representation functors $R_K$, where $K$ is a field of characteristic 0 (see \cite{doublact}, \cite{fonctrq}, and L. Barker's article~\cite{rhetoric}), the functor of units of Burnside rings~(\cite{burnsideunits}), or the torsion part of the Dade group~(\cite{dadegroup}). \par
In particular, the above results on $\mathcal{R}_p$ yield isomorphisms describing the structure of some $p$-groups as objects of this category, and this is enough to compute the evaluations of rational $p$-biset functors. For example
$$(D_8)^n\cong {\bf 1}\oplus (5^n-1)\cdot\partial C_2$$
in $\mathcal{R}_2$ (Equation~\ref{D2n}). More generally, Proposition~\ref{D2nk} gives a formula for $(D_{2^m})^n$. A straightforward consequence, applying the functor $R_\mathbb{Q}$, is the following
\begin{mth*}{{\bf Example}} For any $n\in N$, the group $(D_8)^n$ has $5^n$ conjugacy classes of cyclic subgroups.
\end{mth*}
Another important by-product of the above result giving the tensor structure of $\mathcal{R}_p$ is the explicit description of a genetic basis of a direct product $P\times Q$, in terms of a genetic basis of $P$ and a genetic basis of $Q$ (Theorem~\ref{genetic basis product}). This allows in particular for a quick computation of the torsion part of the Dade group of some $p$-groups, e.g. (Theorem~\ref{product of dihedral groups}, Assertion~1 and Assertion~3)~:
\begin{mth*}{{\bf Theorem}} \begin{itemize}
\item Let $P$ be an arbitrary finite direct product of groups of order 2 and dihedral 2-groups. Then the Dade group of any factor group of $P$ is torsion free.
\item Let $n$ be a positive integer. For any integer $m\geq 4$, let $P=SD_{2^m}$ be a semidihedral group of order $2^m$, and let $P^{*n}$ denote the central product of $n$-copies of $P$. Then the torsion part of the Dade group of $P^{*n}$ is isomorphic to $(\mathbb{Z}/2\mathbb{Z})^{2^{(n-1)(m-3)}}$.
\end{itemize}
\end{mth*}
This also yields similar results on groups of units of Burnside rings of these groups (Remark~\ref{units burnside}), or on representations of central products of $p$-groups, as in Examples~\ref{edges} and~\ref{extraspecial}~:
\begin{mth*}{\bf Example} Let $p$ be a prime, let $X$ be an extraspecial $p$-group, and let~$Q$ be a non-trivial $p$-group. Let $K$ be a field of characteristic 0. Then $Q$ has the same number (possibly 0) of isomorphism classes of faithful irreducible representations over $K$ as any central product $X*Q$.
\end{mth*}
Another possibly interesting phenomenon is that some non-isomorphic $p$-groups may become isomorphic in the category $\mathcal{R}_p$. This means that some non-isomorphic $p$-groups cannot be distinguished using only rational $p$-biset functors. When $p=2$, there are even examples where this occurs for groups of different orders (Example~\ref{different orders}). When $p>2$, saying that the $p$-groups $P$ and $Q$ are isomorphic in $\mathcal{R}_p$ is equivalent to saying that the group algebras $\mathbb{Q} P$ and $\mathbb{Q} Q$ have isomorphic centers (Proposition~\ref{isomorphism Roquette}).\par
The category $\mathcal{R}_p$ is built as follows~: consider first the category $\mathcal{R}_p^\sharp$, which is the quotient category of the biset category of finite $p$-groups (in which objects are finite $p$-groups and morphisms are virtual bisets) obtained by killing a specific element $\delta$ in the Burnside group of the Sylow $p$-subgroup of $PGL(3,\mathbb{F}_p)$. Then take idempotent completion, and additive completion of the resulting category.\par
In particular, this construction relies on bisets, and related functors. Consequently, the paper is organized as follows~: Section~\ref{rational p-biset functors} is a (not so) quick summary of the background on biset functors, Roquette groups, genetic bases of $p$-groups, and rational $p$-biset functors. The category $\mathcal{R}_p$ is introduced in Section~\ref{section Roquette category}, and in Section~\ref{the tensor structure}, its tensor structure is described. Finally Section~\ref{examples and applications} gives some examples and applications.
\bigskip\par\noindent
{\bf Acknowledgments :}
Even though the idea of considering the category $\mathcal{R}_p$ was implicit in~\cite{rationnel}, the present paper wouldn't exist without the illuminating conversations I had with Paul Balmer in november 2010, during which I gradually understood I was a kind of {\em Monsieur Jourdain} of tensor categories\ldots ~I~wish to thank him for this revelation.
\section{Rational $p$-biset functors}\label{rational p-biset functors}
\masubsect{Biset functors} The {\em biset category} $\mathcal{C}$ of finite groups is defined as follows~:
\begin{itemize}
\item The objects of $\mathcal{C}$ are the finite groups.
\item Let $G$ and $H$ be finite groups. Then
$${\rm Hom}_{\mathcal{C}}(G,H)=B(H,G)\;\;,$$
where $B(H,G)$ denotes the Grothendieck group of the category of finite $(H,G)$-bisets, i.e. the Burnside group of the group $H\times G^{op}$.
\item Let $G$, $H$, and $K$ be finite groups. The composition of morphisms
$$B(K,H)\times B(H,G)\to B(K,G)$$
in the category $\mathcal{C}$ is the linear extension of the product induced by the product of bisets $(V,U)\mapsto V\times_HU$, where $V$ is a $(K,H)$-biset, and $U$ is an $(H,G)$-biset.
\item The identity morphism of the finite group $G$ is the image in $B(G,G)$ of the set $G$, endowed with its $(G,G)$-biset structure given by left and right multiplication.
\end{itemize}
\begin{mth}{Definition} A {\em biset functor} is an additive functor from $\mathcal{C}$ to the category of abelian groups. A {\em morphism} of biset functors is a natural transformation of functors.
\end{mth}
Morphisms of biset functors can be composed, and the resulting category of biset functors is denoted by $\mathcal{F}$. It is an abelian category.
\begin{rem}{Examples} \label{basic bisets}\begin{enumerate}
\item The correspondence $B$ sending a finite group $G$ to its Burnside group $B(G)$ is a biset functor, called the {\em Burnside functor}~: indeed $B(G)={\rm Hom}_{\mathcal{C}}({\bf 1},G)$, so $B$ is in fact the Yoneda functor ${\rm Hom}_{\mathcal{C}}({\bf 1},{-})$.
\item The formalism of bisets gives a single framework for the usual operations of induction, restriction, inflation, deflation, and transport by isomorphism via the following correspondences~:
\begin{itemize}
\item If $H$ is a subgroup of $G$, then let ${\rm Ind}_H^G\in B(G,H)$ denote the set~$G$, with left action of $G$ and right action of $H$ by multiplication.
\item If $H$ is a subgroup of $G$, then let ${\rm Res}_H^G\in B(H,G)$ denote the set~$G$, with left action of $H$ and right action of $G$ by multiplication.
\item If $N\mathop{\trianglelefteq} G$, and $H=G/N$, then let ${\rm Inf}_H^G\in B(G,H)$ denote the set $H$, with left action of~$G$ by projection and multiplication, and right action of $H$ by multiplication.
\item If $N\mathop{\trianglelefteq} G$, and $H=G/N$, then let ${\rm Def}_H^G\in B(H,G)$ denote the
set~$H$, with left action of $H$ by multiplication, and right action of~$G$ by projection and multiplication.
\item If $\varphi: G\to H$ is a group isomorphism, then let ${\rm Iso}_G^H={\rm Iso}_G^H(\varphi)\in B(H,G)$ denote the set $H$, with left action of $H$ by multiplication, and right action of $G$ by taking image by $\varphi$, and then multiplying in $H$.
\item When $H$ is a subgroup of $G$, let ${\rm Defres}_{N_G(H)/H}^G\in B\big(N_G(H)/H,G\big)$ denote the set $H\backslash G$, viewed as a $\big(N_G(H)/H,G\big)$-biset. It is equal to the composition ${\rm Def}_{N_G(H)/H}^G\circ {\rm Res}_{N_G(H)}^G$.
\item When $H$ is a subgroup of $G$, let ${\rm Indinf}_{N_G(H)/H}^G\in B\big(G,N_G(H)/H\big)$ denote the set $G/H$, viewed as a $\big(G,N_G(H)/H\big)$-biset. It is equal to the composition ${\rm Ind}_{N_G(H)}^G\circ {\rm Inf}_{N_G(H)/H}^{N_G(H)}$.
\end{itemize}
\end{enumerate}
\end{rem}
\vspace{2ex}
\masubsect{$p$-biset functors} From now on, the symbol $p$ will denote a prime number.
\begin{mth}{Definition and Notation}
\vspace{2ex}
\begin{itemize}
\item The {\em biset category} $\mathcal{C}_p$ of finite $p$-groups is the full subcategory of $\mathcal{C}$ consisting of finite $p$-groups.
\item A $p$-biset functor is an additive functor from $\mathcal{C}_p$ to the category of abelian groups. A morphism of $p$-biset functors is a natural transformation of functors.
\item $p$-biset functors form an abelian category $\mathcal{F}_p$.
\end{itemize}
\end{mth}
\vspace{2ex}
\pagebreak[3]
\masubsect{Roquette $p$-groups}
\begin{mth}{Definition} A finite group $G$ is call a {\em Roquette group} if it has normal rank 1, i.e. if all the normal abelian subgroups of $G$ are cyclic.
\end{mth}
The Roquette $p$-groups have been first classified by\ldots Roquette (\cite{roquette}, see also~\cite{gorenstein})~: these are the cyclic groups, if $p>2$. The Roquette 2-groups are the cyclic groups, the generalized quaternion groups, the dihedral and semidihedral groups of order at least 16. \par
More generally, the $p$-hyperelementary Roquette groups have been classified by Hambleton, Taylor and Williams (Theorem~3.A.6 of \cite{htw}). \par
The following schematic diagram represents the lattice of subgroups of the dihedral group $D_{16}$, the quaternion group $Q_{16}$, and the semi-dihedral group $SD_{16}$ (an horizontal dotted link between two vertices means that the corresponding subgroups are conjugate)~:
\def\xar[#1]{\ar@{-}[#1]}
\newcommand\boitext[1]{\makebox[0pt][l]{$\scriptstyle #1$}}
$$\xymatrix@C=6pt@R=12pt{
&&&&\bullet\boitext{D_{16}}&&&&\\
&&\bullet\xar[urr]\boitext{D_8}&&\bullet\xar[u]\boitext{C_8}&&\bullet\xar[ull]\boitext{D_8}&\\
&\bullet\xar[ur]\ar@{::}[r]&\bullet\xar[u] &&\bullet\xar[ull]\xar[u]\xar[urr]&&\bullet\xar[u]&\bullet\xar[ul]\ar@{::}[l]\\
\bullet\xar[ur]\ar@{::}[r]&\bullet\xar[u]\ar@{::}[r]&\bullet\xar[u]\ar@{::}[r]&\bullet\xar[ul]&\bullet\xar[ull]\xar[u]\xar[urr]\xar[ulll]\xar[urrr]&\bullet\xar[ur]&\bullet\xar[u]\ar@{::}[l]&\bullet\xar[u]\ar@{::}[l]&\bullet\xar[ul]\ar@{::}[l]\\
&&&&\bullet\xar[ullll]\xar[ulll]\xar[ull]\xar[ul]\xar[u]\xar[ur]\xar[urr]\xar[urrr]\xar[urrrr]\\
&&&&*+[F]{D_{16}}\\
}
\xymatrix@C=6pt@R=12pt{
&&&&\bullet\boitext{Q_{16}}&&&&\\
&&\bullet\xar[urr]\boitext{Q_8}&&\bullet\xar[u]\boitext{C_8}&&\bullet\xar[ull]\boitext{Q_8}&\\
&\bullet\xar[ur]\ar@{::}[r]&\bullet\xar[u] &&\bullet\xar[ull]\xar[u]\xar[urr]&&\bullet\xar[u]&\bullet\xar[ul]\ar@{::}[l]\\
&&&&\bullet\xar[ull]\xar[u]\xar[urr]\xar[ulll]\xar[urrr]&&&&\\
&&&&\bullet\xar[u]\\
&&&&*+[F]{Q_{16}}\\
}$$
$$
\xymatrix@C=6pt@R=12pt{
&&&&\bullet\boitext{SD_{16}}&&&&\\
&&\bullet\xar[urr]\boitext{D_8}&&\bullet\xar[u]\boitext{C_8}&\ \ &\bullet\xar[ull]\boitext{Q_8}&\\
&\bullet\xar[ur]\ar@{::}[r]&\bullet\xar[u] &&\bullet\xar[ull]\xar[u]\xar[urr]&&\bullet\xar[u]&\bullet\xar[ul]\ar@{::}[l]\\
\bullet\xar[ur]\ar@{::}[r]&\bullet\xar[u]\ar@{::}[r]&\bullet\xar[u]\ar@{::}[r]&\bullet\xar[ul]&\bullet\xar[ull]\xar[u]\xar[urr]\xar[ulll]\xar[urrr]&&&&\\
&&&&\bullet\xar[ullll]\xar[ulll]\xar[ull]\xar[ul]\xar[u]\\
&&&&*+[F]{SD_{16}}\\
}$$
These diagrams give a good idea of the general case.
\begin{mth}{Definition} \label{axial}Let $G$ be a finite group, of exponent $e$. An {\em axis} of $G$ is a cyclic subgroup of order $e$ in $G$. An {\em axial} subgroup of $G$ is a subgroup of an axis of $G$.
\end{mth}
With these definitions, let us recall without proof the following properties of Roquette $p$-groups~:
\begin{mth}{Lemma}\label{en vrac} Let $P$ be a non-trivial Roquette $p$-group, of exponent $e_P$.
\begin{enumerate}
\item The center of $P$ is cyclic, hence $P$ admits a unique central subgroup $Z_P$ of order $p$.
\item There exists a non-trivial subgroup $Q$ of $P$ such that $Q\cap Z(P)={\bf 1}$ if and only if $p=2$, and $P$ is dihedral or semidihedral. In this case moreover $|Q|=2$, and $N_P(Q)=QZ_P$.
\item If $P$ is not cyclic, then $p=2$ and $e_P=|P|/2$.
\item There is a unique axis in $P$, except in the case $P\cong Q_8$, where there are three of them. Any axis of $P$ is normal in $P$.
\item If $R$ is a non-trivial axial subgroup of $P$, then $R\geq Z_P$, and $R\mathop{\trianglelefteq} P$. If moreover $|R|\geq p^2$, then $C_P(R)$ is the only axis of $P$ containing $R$.
\end{enumerate}
\end{mth}
Let us also recall the following~:
\begin{mth}{Lemma} Let $P$ be a finite Roquette $p$-group. Then there is a unique simple faithful $\mathbb{Q} P$-module $\Phi_P$, up to isomorphism.
\end{mth}
\noindent{\bf Proof~:}\ See \cite{bisetfunctors} Proposition 9.3.5.~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\begin{rem}{Example} \label{endo Phi}Let $P$ be a cyclic group of order $p^m$, and suppose first that $m>1$. The algebra $\mathbb{Q} P$ is isomorphic to the algebra $A=\mathbb{Q}[X]/(X^{p^m}-1)$. As $$X^{p^m}-1=(X^{p^{m-1}}-1)\Psi_{p^m}(X)\;\;,$$
where $\Psi_{p^m}$ denotes the $p^m$-th cyclotomic polynomial, it follows that there is a split exact sequence of $A$-modules
$$0\mapsto \mathbb{Q}[X]/(\Psi_{p^m})\to A\to \mathbb{Q}[X]/(X^{p^{m-1}}-1)\to 0\;\;,$$
which can be viewed as the following sequence of $\mathbb{Q} P$-modules
$$0\to \Phi_P\to \mathbb{Q} P\to \mathbb{Q} (P/Z)\to 0\;\;,$$
where $Z$ is the unique subgroup of order $p$ of $P$. It follows that there is an isomorphism of $\mathbb{Q}$-algebras
$${\rm End}_{\mathbb{Q} P}\Phi_P\cong \mathbb{Q}[X]/(\Psi_{p^m})\cong\mathbb{Q}(\zeta_{p^m})\;\;,$$
where $\zeta_{p^m}$ is a primitive $p^m$-th root of unity in $\mathbb{C}$. \par
Now if $m=0$, then $P={\bf 1}$, $\Phi_P=\mathbb{Q}$, and ${\rm End}_{\mathbb{Q} P}\Phi_P\cong \mathbb{Q}$, also.
\end{rem}
\masubsect{Expansive and genetic subgroups}
\begin{mth}{Definition} A subgroup $H$ of a group $G$ is called {\em expansive} if for any $g\in G$ such that $H^g\neq H$, the group $\big(H^g\cap N_G(H)\big)H/H$ contains a non trivial normal subgroup of $N_G(H)/H$, i.e. if
$$g\in G-N_G(H)\Longrightarrow \bigcap_{n\in N_G(H)}\big(H^{gn}\cap N_G(H)\big)H>H\;\;.$
\end{mth}
\begin{rem}{Example} \label{normalisateur normal}If $H\mathop{\trianglelefteq} G$, then $H$ is expansive in $G$. More generally, if $N_G(H)\mathop{\trianglelefteq} G$, then $H$ is expansive in $G$~: indeed $N_G(H^g)=N_G(H)$, for any $g\in G$. Hence for $g\in G-N_G(H)$
$$\bigcap_{n\in N_G(H)}\big(H^{gn}\cap N_G(H)\big)H=\big(H^g\cap N_G(H)\big)H=H^g\cdot H>H\;\;.$$
\end{rem}
\begin{mth}{Notation} When $H$ is a subgroup of the group $G$, denote by $Z_G(H)$ the subgroup of $N_G(H)$, containing $H$, defined by
$$Z_G(H)/H=Z\big(N_G(H)/H\big)\;\;.$$
\end{mth}
The following is an easy consequence of well known properties of $p$-groups~:
\begin{mth}{Lemma} {\rm [\cite{bisetfunctors} Lemma 9.5.2]} Let $Q$ be a subgroup of a finite $p$-group~$P$. Then $Q$ is expansive in $P$ if and only if
$$\forall g\in P,\;\;Q^g\cap Z_P(Q)\leq Q\Longrightarrow Q^g=Q\;\;.$$
\end{mth}
\begin{mth}{Example} \label{diagonal}Let $P$ be a $p$-group, and let
$$\Delta(P)=\big\{(x,x)\in (P\times P)\mid x\in P\big\}$$
denote the diagonal subgroup of $(P\times P)$. Then $N_{P\times P}\big(\Delta(P)\big)/\Delta(P)\cong Z(P)$, and $\Delta(P)$ is expansive in $(P\times P)$ if and only if
$$\forall x\in P,\;\;[P,x]\cap Z(P)=\{1\}\implies x\in Z(P)\;\;,$$
where $[P,x]$ is the {\em set} of commutators $[y,x]=y^{-1}x^{-1}yx=y^{-1}y^x$, for $y\in P$.\par
In particular, if $[P,P]\leq Z(P)$, then $\Delta(P)$ is expansive in $(P\times P)$.
\end{mth}
\noindent{\bf Proof~:}\ Indeed $N_{P\times P}\big(\Delta(P)\big)$ consists of pairs $(a,b)\in (P\times P)$ such that $ab^{-1}\in Z(P)$. This shows that $N_{P\times P}\big(\Delta(P)\big)/\Delta(P)\cong Z(P)$, and that
$$Z_{P\times P}\big(\Delta(P)\big)=N_{P\times P}\big(\Delta(P)\big)=\big({\bf 1}\times Z(P)\big)\Delta(P)\;\;.$$
Now $\Delta(P)^{(u,v)}=\Delta(P)^{(1,x)}$, for any $(u,v)\in (P\times P)$, where $x=u^{-1}v$, and
$$\Delta(P)^{(1,x)}\cap Z_{P\times P}\big(\Delta(P)\big)=\big\{(t,t^x)\mid t\in P,\;t^{-1}t^x\in Z(P)\big\}\;\;.$$
Hence $\Delta(P)^{(1,x)}\cap Z_{P\times P}\big(\Delta(P)\big)\leq \Delta(P)$ if and only if for any $t\in P$, the assumption $t^{-1}t^x\in Z(P)$ implies $t=t^x$, i.e. $[t,x]=1$, in other words if $[P,x]\cap Z(P)=\{1\}$. Hence $\Delta(P)$ is expansive in $(P\times P)$ if and only if for any $x\in P$, the assumption $[P,x]\cap Z(P)=\{1\}$ implies $(1,x)\in N_{P\times P}\big(\Delta(P)\big)$, i.e. $x\in Z(P)$, as claimed. The last assertion follows trivially.~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\begin{mth}{Definition} Let $Q$ be a subgroup of the finite $p$-group $P$. Then $Q$ is called a {\em genetic} subgroup of $P$ if $Q$ is expansive in $P$ and if the group $N_P(Q)/Q$ is a Roquette group.
\end{mth}
\begin{mth}{Definition} Define a relation $\bizlie{P}$ on the set of subgroups of the finite $p$-group $P$ by
$$Q\bizlie{P}R\;\;\Leftrightarrow \;\;\exists g\in P,\;\;Q^g\cap Z_P(R)\leq R\;\;\hbox{and}\;\;{^gR}\cap Z_P(Q)\leq Q\;\;.$$\vspace{-4ex}
\end{mth}
\begin{mth}{Lemma} \label{normalisateur normal bizlie}Let $P$ be a finite $p$-group. If $Q$ and $R$ are subgroups of $P$ such that $N_P(Q)=N_P(R)\mathop{\trianglelefteq} P$, then
$$Q\bizlie{P}R\;\;\Leftrightarrow\;\; Q=_PR\;\;.$$
\end{mth}
\noindent{\bf Proof~:}\ Indeed, since $P$ is a $p$-group, saying that $Q^g\cap Z_P(R)\leq R$ is equivalent to saying that the subgroup $\big(Q^g\cap N_P(R)\big)R/R$ of $N_P(R)/R$ contains no non-trivial normal subgroup of $N_P(R)/R$, i.e. that
$$\bigcap_{n\in N_P(R)}\big(Q^{gn}\cap N_P(R)\big)R=R\;\;.$$
But if $N_P(Q)\mathop{\trianglelefteq} P$, then $N_P(Q)=N_P(Q^g)=N_P(R)$, for any $g\in G$. Hence
$$\bigcap_{n\in N_G(R)}\big(Q^{gn}\cap N_P(R)\big)R=Q^g\cdot R\;\;.$$
This is equal to $R$ if and only if $Q^g\leq R$. Similarly ${^gR}\cap Z_P(Q)\leq Q$ if and only if $^gR\leq Q$. Hence $Q^g=R$.~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\begin{mth}{Definition} Let $G$ be a group.
\begin{enumerate}
\item A {\em section} of $G$ is a pair $(T,S)$ of subgroups of $G$ such that $S\mathop{\trianglelefteq} T$. The quotient $T/S$ is called the corresponding {\em subquotient} of $G$.
\item Two sections $(T,S)$ and $(Y,X)$ of $G$ are said to be {\em linked} (notation $(T,S)\raisebox{.5ex}{\;\rule{2.5ex}{.2ex}\;} (Y,X)$) if
$$S(T\cap Y)=T,\;\;X(T\cap Y)=Y,\;\;T\cap X=S\cap Y\;\;.$$
They are said to be {\em linked modulo $G$} (notation $(T,S)\estliemod{G}(Y,X)$) if there exists $g\in G$ such that $(T,S)\raisebox{.5ex}{\;\rule{2.5ex}{.2ex}\;} ({^gY},{^gX})$.
\end{enumerate}
\end{mth}
Observe in particular that if $(T,S)\estliemod{G}(Y,X)$, then the corresponding subquotients $T/S$ and $Y/X$ are isomorphic.
\begin{mth}{Theorem} \label{genetic recall}Let $P$ be a finite $p$-group.
\begin{enumerate}
\item If $S$ is a genetic subgroup of $P$, then the module
$$V(S)={\rm Ind}_{N_P(S)}^P{\rm Inf}_{N_P(S)/S}^{N_P(S)}\Phi_{N_P(S)/S}$$
is a simple $\mathbb{Q} P$-module. Moreover, the functor ${\rm Ind}_{N_P(S)}^P{\rm Inf}_{N_P(S)/S}^{N_P(S)}$ induces an isomorphism of $\mathbb{Q}$-algebras
$${\rm End}_{Q P}V(S)\cong{\rm End}_{\mathbb{Q} N_P(S)/S}\Phi_{N_P(S)/S}\;\;.$$
\item If $V$ is a simple $\mathbb{Q} P$-module, then there exists a genetic subgroup $S$ of $P$ such that $V\cong V(S)$.
\item If $S$ and $T$ are genetic subgroups of $P$, then
$$V(S)\cong V(T) \;\;\Leftrightarrow \;\;S\bizlie{P}T\;\;\Leftrightarrow\;\; \big(N_P(S),S\big)\estliemod{P}\big(N_P(T),T\big)\;\;.$$
In particular, if $S\bizlie{P}T$, then $N_P(S)/S\cong N_P(T)/T$. Moreover, the relation $\bizlie{P}$ is an equivalence relation on the set of genetic subgroups of $P$, and the corresponding set of equivalence classes is in one to one correspondence with the set of isomorphism classes of simple $\mathbb{Q} P$-modules.
\end{enumerate}
\end{mth}
\noindent{\bf Proof~:}\ See \cite{bisetfunctors}, Theorem 9.6.1.~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\begin{mth}{Definition} Let $P$ be a finite $p$-group. A {\em genetic basis} of $P$ is a set of representatives of equivalence classes of genetic subgroups of $P$ for the relation $\bizlie{P}$.
\end{mth}
\masubsect{Faithful elements} (\cite{bisetfunctors} Sections 6.2 and 6.3) Let $G$ be a finite group. If $N$ is a normal subgroup of $G$, recall from Examples~\ref{basic bisets} that ${\rm Inf}_{G/N}^G$ denotes the set~$G/N$, viewed as a $(G,G/N)$-biset for the actions given by (projection to the factor group and) multiplication in $G/N$. Similarly ${\rm Def}_{G/N}^G$ denotes the same set~$G/N$, considered as a $(G/N,G)$-biset.\par
There is an isomorphism of $(G/N,G/N)$-bisets
\begin{equation}\label{definf}
{\rm Id}_{G/N}\cong {\rm Def}_{G/N}^G\circ {\rm Inf}_{G/N}^G\;\;.
\end{equation}
More generally, if $M$ and $N$ are normal subgroups of $G$, there is a isomorphism of $(G/M,G/N)$-bisets
\begin{equation}\label{infdef}
{\rm Def}_{G/M}^G\circ{\rm Inf}_{G/N}^G={\rm Inf}_{G/MN}^{G/M}\circ{\rm Def}_{G/MN}^{G/N}\;\;.
\end{equation}
It follows that if $j_N^G$ is defined by
$$j_N^G={\rm Inf}_{G/N}^G\circ{\rm Def}_{G/N}^G\;\;,$$
then $j_M^G\circ j_N^G=j_{MN}^G$. In particular $j_N^G$ is an idempotent of $B(G,G)$. Moreover, by a standard orthogonalization procedure, the elements $f_N^G$ defined for~$N\mathop{\trianglelefteq} G$ by
$$f_N^G=\sum_{N\leq M\mathop{\trianglelefteq} G}\mu_{\mathop{\trianglelefteq} G}(N,M)j_M^G\;\;,$$
where $\mu_{\mathop{\trianglelefteq} G}(N,M)$ is the M\"obius function of the poset of normal subgroups of~$G$, are orthogonal idempotents of $B(G,G)$, and their sum is equal to ${\rm Id}_{G/N}$. The idempotent $f_{\bf 1}^G$ is of special importance~:
\begin{mth}{Lemma} \label{funG}Let $G$ be a finite group, and $N$ be a normal subgroup of~$G$.
\begin{enumerate}
\item $f_N^G={\rm Inf}_{G/N}^G\circ f_{\bf 1}^{G/N}\circ {\rm Def}_{G/N}^G$ in $B(G,G)$.
\item If $N\neq 1$, then ${\rm Def}_{G/N}^G\circ f_{\bf 1}^G=0$ in $B(G/N,G)$, and $f_{\bf 1}^G\circ{\rm Inf}_{G/N}^G=0$ in $B(G,G/N)$.
\end{enumerate}
\end{mth}
\noindent{\bf Proof~:}\ Assertion~1 is Remark~6.2.9 of~\cite{bisetfunctors}, and Assertion~2 is a special case of Proposition~6.2.6.~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
If $F$ is a biset functor, the set $\partial F(G)$ of {\em faithful elements} of $F(G)$ is defined by
$$\partial F(G)=F(f_{\bf 1}^G)F(G)\;\;.$$
It can be shown (\cite{bisetfunctors} Lemma 6.3.2) that
$$\partial F(G)=\mathop{\bigcap}_{{\bf 1}<N\mathop{\trianglelefteq} G}\limits{\rm Ker}\; F({\rm Def}_{G/N}^G)\;\;.$$
\begin{rem}{Example} \label{representation edge}Let $F=R_K$ be the representation functor over a field $K$ of characteristic 0. Then for a finite group $G$, the group $\partial R_K(G)$ is the direct summand of $R_K(G)$ with basis the set of (isomorphism classes of) faithful irreducible $KG$-modules.
\end{rem}
\begin{mth}{Lemma}\label{inflation} Let $G$ be a group, and $S$ be a subgroup of $G$, such that $S\cap Z(G)\neq {\bf 1}$. Then ${\rm Defres}_{N_G(S)/S}^Gf_{\bf 1}^G=0$ in $B\big(N_G(S)/S,G\big)$.
\end{mth}
\noindent{\bf Proof~:}\ Set $N=S\cap Z(G)$, $\sur{G}=G/N$, and $\sur{S}=S/N$. Then
$${\rm Defres}_{N_G(S)/S}^G={\rm Defres}_{N_{\sur{G}}(\sur{S})/\sur{S}}^{\sur{G}}{\rm Def}_{G/N}^G\;\;.$$
Now ${\rm Def}_{G/N}^Gf_{\bf 1}^G=0$ if $N\neq {\bf 1}$, by Lemma~\ref{funG}. ~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\begin{mth}{Theorem} {\rm [\cite{bisetfunctors} Theorem 10.1.1]} \label{roquette plus}Let $P$ be a finite $p$-group, and $\mathcal{B}$ be a genetic basis of $P$. Then, for any $p$-biset functor $F$, the map
$$\mathcal{I}_\mathcal{B}=\dirsum{S\in\mathcal{B}}{\rm Indinf}_{N_P(S)/S}^P:\dirsum{S\in\mathcal{B}}\partial F\big(N_P(S)/S\big)\to F(P)$$
is split injective. A left inverse is the map
$$\mathcal{D}_\mathcal{B}=\dirsum{S\in\mathcal{B}}f_{\bf 1}^{N_P(S)/S}\circ {\rm Defres}_{N_P(S)/S}^P:F(P)\to \dirsum{S\in\mathcal{B}}\partial F\big(N_P(S)/S\big)\;\;.$$
\end{mth}
One can show (\cite{bisetfunctors} Lemma 10.1.2) that if $\mathcal{B}$ and $\mathcal{B}'$ are genetic bases of $P$, the map $\mathcal{I}_\mathcal{B}$ is an isomorphism if and only if the map $\mathcal{I}_{\mathcal{B}'}$ is an isomorphism. This motivates the following definition~:
\begin{mth}{Definition} \label{rational functors} A $p$-biset functor $F$ is called {\em rational} if for any finite $p$-group $P$, there exists a genetic basis $\mathcal{B}$ of $P$ such that the map $\mathcal{I}_\mathcal{B}$ is an isomorphism.
\end{mth}
So $F$ is rational if and only if for any finite $p$-group $P$ and {\em any} genetic basis $\mathcal{B}$ of $P$, the map $\mathcal{I}_\mathcal{B}$ is an isomorphism.
\begin{rem}{Examples} \medskip\par\noindent
$\bullet$ The functor $R_\mathbb{Q}$ of rational representations, which sends the finite $p$-group $P$ to the group $R_\mathbb{Q}(P)$, is a rational $p$-biset functor. This example is of course the reason for calling {\em rational} the $p$-biset functors of Definition~\ref{rational functors}. This choice has proved rather unfortunate, since the $p$-biset functor $R_\mathbb{C}$ of {\em complex} representations is also a rational functor\ldots More generally, if $K$ is a field of characteristic 0, then the functor $R_K$ is a rational $p$-biset functor.\medskip\par\noindent
$\bullet$ The functor of units of the Burnside ring, sending a $p$-group $P$ to the group of units $B^\times(P)$ of its Burnside ring, is a rational $p$-biset functor (see~\cite{burnsideunits}).\medskip\par\noindent
$\bullet$ Let $k$ be a field of characteristic $p$. The correspondence sending a finite $p$-group $P$ to the torsion part $D_k^t(P)$ of the Dade group of $P$ over $k$ is not a biset functor in general, because of phenomenons of {\em Galois twists}, but still the maps $\mathcal{I}_{\mathcal{B}}$ and $\mathcal{D}_{\mathcal{B}}$ can be defined for $D^t_k$, and Theorem~\ref{roquette plus} holds (see~\cite{dadegroup}).
\end{rem}
\section{The Roquette category} \label{section Roquette category}
\begin{mth}{Notation} \label{def X}Let $\pi$ be a projective plane over $\mathbb{F}_p$, and let $X$ denote a Sylow $p$-subgroup of ${\rm Aut}(\pi)\cong{\rm PGL}(3,\mathbb{F}_p)$. Let $\mathbb{L}$ be the set of lines of $\pi$, and let $\mathbb{P}$ be the set of points of $\pi$, both viewed as elements of $B(X)$. Let $\delta=\mathbb{L}-\mathbb{P}\in B(X)$. Equivalently
$$\delta=(X/I-X/IZ)-(X/J-X/JZ)\;\;,$$
where $I$ and $J$ are non-conjugate non-central subgroups of order $p$ of $X$, and $Z$ is the center of $X$.\par
Let $B_\delta$ denote the $p$-biset subfunctor of $B$ generated by $\delta$.
\end{mth}
\begin{rem}{Remark} When $p=2$, the group $X$ is dihedral of order 8, and $\delta$ is well-defined up to a sign. When $p>2$, the group $X$ is an extraspecial $p$-group of order $p^3$ and exponent $p$, and there are several possible choices for the element $\delta$. However, in any case, the functor $B_\delta$ does not depend on the choice of $\delta$.
\end{rem}
\begin{mth}{Definition} The {\em Roquette category} $\mathcal{R}_p$ of finite $p$-groups is defined as the {\em idempotent additive completion} of the category $\mathcal{R}_p^\sharp$, quotient of the biset category $\mathcal{C}_p$, defined as follows~:
\begin{enumerate}
\item the objects of $\mathcal{R}_p^\sharp$ are the finite $p$-groups.
\item if $P$ and $Q$ are finite $p$-groups, then
$${\rm Hom}_{\mathcal{R}_p^\sharp}(P,Q)=(B/B_\delta)(Q,P)$$
is the quotient of $B(Q\times P^{op})$ by $B_\delta(Q\times P^{op})$.
\item the composition in $\mathcal{R}_p^\sharp$ is induced by the composition of bisets.
\item the identity morphism of the finite $p$-group $P$ in $\mathcal{R}_p^\sharp$ is the image of ${\rm Id}_P$ in $(B/B_\delta)(P,P)$.
\end{enumerate}
\end{mth}
\begin{rem}{Remark} It was shown in~\cite{rationnel} that $\mathcal{R}_p^\sharp$ is indeed a category. It was also shown there that if $p>2$, the functor $B_\delta$ is equal to the kernel $K$ of the linearization morphism $B\to R_\mathbb{Q}$. It follows that in this case, for any two finite $p$-groups $P$ and $Q$
$${\rm Hom}_{\mathcal{R}_p}(P,Q)\cong R_\mathbb{Q}(Q\times P^{op})$$
is isomorphic to the Grothendieck group of $(\mathbb{Q} Q,\mathbb{Q} P)$-bimodules, or, equivalently by the Ritter-Segal theorem, to the Grothendieck group of the subcategory of $(\mathbb{Q} Q,\mathbb{Q} P)$-permutation bimodules. In other words in this case, the category $\mathcal{R}_p$ is the full subcategory category of the category considered by Barker in~\cite{rhetoric}, consisting of finite $p$-groups. The construction of the category $\mathcal{R}_p$ is also very similar to the construction of the category $\mathbb{Q} G$-Morita by Hambleton, Taylor, and Williams in~\cite{htw} (Definition 1.A.4).\par
In the case $p=2$, the situation is more complicated~: the functor $B_\delta$ is a proper subfunctor of the kernel $K$, and there is a short exact sequence
$$0\to K/B_\delta \to B/B_\delta \to R_\mathbb{Q}\to 0$$
of $p$-biset functors. Moreover, for each $p$-group $P$, the group $(K/B_\delta)(P)$ is a finite elementary abelian 2-group of rank equal to the number of groups $S$ in a genetic basis of $P$ for which $N_P(S)/S$ is dihedral.
\end{rem}
\begin{mth}{Lemma} The direct product $(P,Q)\mapsto P\times Q$ of $p$-groups induces a well defined symmetric monoidal structure on $\mathcal{R}_p^\sharp$.
\end{mth}
\noindent{\bf Proof~:}\ Let $P$, $P'$, $Q$ and $Q'$ be finite $p$-groups. If $U$ is a finite $(P',P)$-biset and $V$ is a finite $(Q',Q)$-biset, then $U\times V$ is a $(P'\times Q',P\times Q)$-biset. This induces a bilinear map
$$\pi : B(P',P)\times B(Q',Q)\to B(P'\times Q',P\times Q)\;\;,$$
and this clearly induces a symmetric monoidal structure on the biset category~$\mathcal{C}_p$. This induces a monoidal structure on the quotient category if
$$\pi\big(B_\delta(P',P),B(Q',Q)\big)\subseteq B_\delta(P'\times Q',P\times Q)\;\;.$$
But this is a consequence of the following~: let $X$ be as defined in Notation~\ref{def X}, let $U$ be a finite $(P',P\times X)$-set, let $D$ be an $X$-set, and $V$ be a finite $(Q',Q)$-biset. Clearly, there is an isomorphism of $(P'\times Q',P\times Q)$-sets
$$(U\times_XD)\times V\cong (U\times V)\times_XD\;\;,$$
where the right action of $X$ on $U\times V$ is defined in the obvious way
$$\forall (u,v)\in U\times V,\;\forall x\in X,\;\;(u,v)x=(ux,v)\;\;.$$
The lemma follows.~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\smallskip\par\noindent\pagebreak[2]\refstepcounter{subsection}\refstepcounter{prop}{\bf \thesection.\arabic{prop}.\ \ } Recall that the objects of the idempotent additive completion $\mathcal{R}_p$ are by definition formal finite sums $\mathop{\oplus}_{(P,e)\in\mathcal{P}}\limits(P,e)$ of pairs of the form $(P,e)$, where $P$ is a finite $p$-group, and $e$ is an idempotent in ${\rm Hom}_{\mathcal{R}_p^\sharp}(P,P)=(B/B_\delta)(P,P)$. A morphism $\varphi:\mathop{\oplus}_{(P,e)\in\mathcal{P}}\limits(P,e)\to \mathop{\oplus}_{(Q,f)\in\mathcal{Q}}\limits(Q,f)$ in $\mathcal{R}_p$ is a matrix indexed by $\mathcal{P}\times\mathcal{Q}$, where the coefficient $\varphi_{(P,e),(Q,f)}$ belongs to $f{\rm Hom}_{\mathcal{R}_p^\sharp}(P,Q)e$. The composition of morphisms is given by matrix multiplication. In particular~:
\pagebreak[3]
\begin{mth}{Definition and Notation} Let $P$ be a finite $p$-group.
\begin{enumerate}
\item The object $(P,{\rm Id}_P)$ of $\mathcal{R}_p$ is denoted by $P$. Similarly, when $Q$ is a finite $p$-group, and $f\in B(Q,P)$, the corresponding morphism from $(P,{\rm Id}_P)$ to $(Q,{\rm Id}_P)$ in the category $\mathcal{R}_p$ is simply denoted by $f$.
\item The {\em edge} $\partial P$ of $P$ is the pair $(P,f_{\bf 1}^P)$ of $\mathcal{R}_p$.
\end{enumerate}
\end{mth}
The category of additive functors from $\mathcal{R}_p$ to abelian groups is equivalent to the category of additive functors from $\mathcal{R}_p^\sharp$ to abelian groups. It was shown in~\cite{rationnel} that the latter is exactly the category of rational $p$-biset functors. If $F^\sharp$ is such a functor, then $F^\sharp$ extends to a functor $F$ on $\mathcal{R}_p$ defined as follows~:
$$F\big(\mathop{\oplus}_{(P,e)\in\mathcal{P}}\limits(P,e)\big)=\mathop{\oplus}_{(P,e)\in\mathcal{P}}\limits F^\sharp(e)\big(F^\sharp(P)\big)\;\;,$$
with the obvious definition of $F(\varphi)$ for a morphism $\varphi$ in the category $\mathcal{R}_p$. In particular, with the above notation
$$F(\partial P)=\partial F^\sharp(P)\;\;.$$
In the sequel, we will use the same symbol for $F$ and $F^\sharp$, writing in particular $F(\partial P)=\partial F(P)$.
\pagebreak[3]
\begin{mth}{Proposition} \label{sum of edges of quotients}Let $P$ be a finite $p$-group. Then
$$P\cong\dirsum{N\mathop{\trianglelefteq} P}\partial (P/N)$$
in the category $\mathcal{R}_p$.
\end{mth}
\noindent{\bf Proof~:}\ Let
$$a:P\to\dirsum{N\mathop{\trianglelefteq} P}\partial (P/N)$$
be the direct sum of the morphisms induced by the elements $f_{\bf 1}^{P/N}{\rm Def}_{P/N}^P$ of $B(P/N,P)$, and let
$$b: \dirsum{N\mathop{\trianglelefteq} P}\partial (P/N)\to P$$
be defined similarly from the elements ${\rm Inf}_{P/N}^Pf_{\bf 1}^{P/N}$ of $B(P,P/N)$.\par
By Lemma~\ref{funG}
$$\sum_{N\mathop{\trianglelefteq} P}{\rm Inf}_{P/N}^Pf_{\bf 1}^{P/N}{\rm Def}_{P/N}^P=\sum_{N\mathop{\trianglelefteq} P}f_N^P={\rm Id}_P$$
in $B(P,P)$, thus $a\circ b$ is equal to the identity morphism of $P$ in $\mathcal{R}_P$. Conversely, for normal subgroups $N$ and $M$ of $P$
$$f_{\bf 1}^{P/N}{\rm Def}_{P/N}^P{\rm Inf}_{P/M}^Pf_{\bf 1}^{P/M}=f_{\bf 1}^{P/N}{\rm Inf}_{P/NM}^{P/N}{\rm Def}_{P/NM}^{P/M}f_{\bf 1}^{P/M}\;\;,$$
by Equation~\ref{definf}. This is equal to 0 if $N\neq M$, by Lemma~\ref{funG}. And if $N=M$, this is equal to $f_{\bf 1}^{P/N}$. It follows that $b\circ a$ is equal to the identity morphism of $\dirsum{N\mathop{\trianglelefteq} P}\partial (P/N)$, and this completes the proof.~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\begin{mth}{Corollary} \label{edge cyclic center}If $P$ is non-trivial, with cyclic center, then
$$P\cong \partial P\oplus (P/Z)$$
in $\mathcal{R}_p$, where $Z$ is the unique central subgroup of order $p$ in $P$.
\end{mth}
\noindent{\bf Proof~:}\ Indeed, if $N$ is a non-trivial normal subgroup of $P$, then $N\geq Z$. Thus
$$P\cong\partial P\oplus\dirsum{N\geq Z}\partial(P/N)\cong \partial P\oplus (P/Z)$$
in the category $\mathcal{R}_p$.~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\begin{rem}{Remark} \label{factor group}More generally, let $P$ be a finite $p$-group, and let $N$ be a normal subgroup of~$P$. Since in $\mathcal{R}_p$
$$P/N\cong \dirsum{N\leq M\mathop{\trianglelefteq} P}\partial \big((P/N)\big/(M/N)\big)\cong\dirsum{N\leq M\mathop{\trianglelefteq} P}\partial (P/M)\;\;, $$
it follows that $P/N$ is isomorphic to a direct summand of $P$ in the category~$\mathcal{R}_p$.
\end{rem}
\begin{mth}{Theorem} \label{Roquette category}\begin{enumerate}
\item The Roquette category $\mathcal{R}_p$ is an additive tensor category.
\item Let $P$ be a finite $p$-group, and $\mathcal{B}$ be a genetic basis of $P$. Then, in the category $\mathcal{R}_p$,
$$P\cong \mathop{\oplus}_{S\in\mathcal{B}}\partial \sur{N}_P(S)\;\;,$$
where $\sur{N}_P(S)=N_P(S)/S$.
\item Let $P$ be a finite $p$-group, and $\mathcal{B}$ be a genetic basis of $P$. Then, in the category $\mathcal{R}_p$,
$$\partial P\cong \oplusb{S\in\mathcal{B}}{S\cap Z(P)={\bf 1}}\partial \sur{N}_P(S)\;\;.$$
\end{enumerate}
\end{mth}
\noindent{\bf Proof~:}\ Assertion 1 results from standard results~: in particular, the tensor product of $\dirsum{(P,e)\in\mathcal{P}}(P,e)$ and $\dirsum{(Q,f)\in\mathcal{Q}}(Q,f)$ is defined by
$$\big(\dirsum{(P,e)\in\mathcal{P}}(P,e)\big)\times\big(\dirsum{(Q,f)\in\mathcal{Q}}(Q,f)\big)=\oplusb{(P,e)\in\mathcal{P}}{(Q,f)\in\mathcal{Q}}(P\times Q,e\times f)\;\;.$$
For Assertion 2, by Proposition 10.7.2 of \cite{bisetfunctors}, if $F$ is a rational $p$-biset functor, the functor $F_P$ obtained from $F$ by the Yoneda-Dress construction at $P$ is also a rational $p$-biset functor. This applies in particular to the functor $Y=B/B_\delta$, so the functor $Y_P$ is rational. Hence, if $Q$ is any finite $p$-group and $\mathcal{B}_Q$ is a genetic basis of $Q$, there are mutual inverse isomorphisms
$$\xymatrix{
Y_P(Q)\ar[r]<.5ex>^-{\mathcal{D}_Q}&*!U(0.5){\mathop{\oplus}_{S\in\mathcal{B}_Q}\limits\partial Y_P\big(\sur{N}_Q(S)\big)}\ar[l]<.5ex>^-{\mathcal{I}_Q}
}
$$
where $\mathcal{I}_Q=\mathop{\oplus}_{S\in\mathcal{B}_Q}\limits{\rm Indinf}_{\sur{N}_P(S)}^P$ and $\mathcal{D}_Q=\mathop{\oplus}_{S\in\mathcal{B}_Q}\limits f_{\bf 1}^{\sur{N}_P(S)}\circ{\rm Defres}_{\sur{N}_P(S)}^P$. Thus for any $f\in Y_P(Q)$
$$f=\big(\sum_{S\in\mathcal{B}_Q}{\rm Indinf}_{\sur{N}_P(S)}^Pf_{\bf 1}^{\sur{N}_P(S)}{\rm Defres}_{\sur{N}_P(S)}^P\big)\circ f\;\;.$$
Applying this to $Q=P$, $\mathcal{B}_Q=\mathcal{B}$, and $f={\rm Id}_P$ gives that $\mathcal{I}_P\circ\mathcal{D}_P={\rm Id}_P$. On the other hand, by Proposition~6.4.4 and Theorem~9.6.1 of~\cite{bisetfunctors}, for $S,T\in\mathcal{B}$, the composition
$$f_{\bf 1}^{\sur{N}_P(S)}{\rm Defres}_{\sur{N}_P(S)}^P\circ {\rm Indinf}_{\sur{N}_P(S)}^Pf_{\bf 1}^{\sur{N}_P(S)}$$
is equal to $f_{\bf 1}^{\sur{N}_P(S)}$ in $B\big(\sur{N}_P(S),\sur{N}_P(S)\big)$ if $T=S$, and to 0 if $T\neq S$. It follows that $\mathcal{D}_P\circ\mathcal{I}_P$ is also equal to the identity map of the direct sum $\mathop{\oplus}_{S\in\mathcal{B}}\limits\partial Y_P\big(\sur{N}_P(S)\big)$ in the category $\mathcal{R}_p$.\par
For Assertion~3, observe that by Lemma~\ref{inflation}
$$f_{\bf 1}^{\sur{N}_P(S)}{\rm Defres}_{\sur{N}_P(S)}^Pf_1^P=0$$
if $S\cap Z(P)\neq {\bf 1}$. Taking opposite bisets, this gives also
$$f_1^P{\rm Indinf}_{\sur{N}_P(S)}^Pf_{\bf 1}^{\sur{N}_P(S)}=0\;\;,$$
so the isomorphism of Assertion~2 restricts to an isomorphism
$$\partial P\cong \oplusb{S\in\mathcal{B}}{S\cap Z(P)={\bf 1}}\partial \sur{N}_P(S)\;\;,$$
as the diagram
$$\xymatrix{
P\ar[r]^-{\cong}&\mathop{\oplus}_{S\in\mathcal{B}}\partial \sur{N}_P(S)\\
\partial P\ar[u]^{f_{\bf 1}^P}\ar[r]&*!U(0.5){\oplusb{S\in\mathcal{B}}{S\cap Z(P)={\bf 1}}\limits\partial \sur{N}_P(S)}\ar@{^{(}->}[u]
}
$$
is commutative. ~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\begin{mth}{Corollary} \label{zero edge} Let $P$ be a finite $p$-group. If $Z(P)$ is non-cyclic, then $\partial P=0$ in $\mathcal{R}_p$.
\end{mth}
\noindent{\bf Proof~:}\ This follows from Assertion~3~: suppose indeed that there exists a genetic subgroup $S$ of $P$ such that $S\cap Z(P)\neq {\bf 1}$. Then the group $Z(P)$ maps injectively in the center of the Roquette group $\sur{N}_P(S)$, which is cyclic. Hence $Z(P)$ is cyclic.~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\pagebreak[3]
\begin{rem}{Examples} \label{D8Q8}
\begin{enumerate}
\item Let $P=D_8$ be a dihedral group of order 8. Let $A$, $B$, and $C$ be the subgroups of index $2$ in $P$, and let $I$ be a non-central subgroup of order~2 in $P$. Then the set $\{P,A,B,C,I\}$ is a genetic basis of $P$, and there is an isomorphism
$$P\cong {\bf 1}\oplus 4\cdot \partial C_2$$
in the category $\mathcal{R}_2$, where $4\cdot \partial C_2$ denotes the direct sum of 4 copies of $\partial C_2$~: indeed, for $S\in\{A,B,C,I\}$, the group $\sur{N}_P(S)$ is isomorphic to~$C_2$.
\item Let $P=Q_8$ be a quaternion group of order 8. Let $A$, $B$, and $C$ be the subgroups of index 2 in $P$. Then the set $\{P,A,B,C,{\bf 1}\}$ is a genetic basis of $P$ (such a basis is unique in this case), and there is an isomorphism
$$P\cong {\bf 1}\oplus 3\cdot \partial C_2\oplus\partial Q_8$$
in the category $\mathcal{R}_2$.
\item Let $P=(C_p)^n$ be an elementary abelian $p$-group of rank $n$. Then $P$ has a unique genetic basis, consisting of $P$, and all its subgroups of index $p$. Hence
$$P\cong {\bf 1}\oplus\frac{p^n-1}{p-1}\cdot\partial C_p$$
in the category $\mathcal{R}_p$.
\end{enumerate}
\end{rem}
\section{The tensor structure} \label{the tensor structure}
\begin{mth}{Notation} Let $G$ and $H$ be groups. When $L$ is a subgroup of $G\times H$, set
\begin{eqnarray*}
p_1(L)&=&\{g\in G\mid \exists h\in H,\;(g,h)\in L\}\\
p_2(L)&=&\{h\in H\mid \exists g\in G,\;(g,h)\in L\}\\
k_1(L)&=&\{g\in G\mid (g,1)\in L\}\\
k_2(L)&=&\{h\in H\mid (1,h)\in L\}\\
\end{eqnarray*}
Recall (\cite{bisetfunctors} 2.3.18 and 2.3.21) that $k_i(L)\mathop{\trianglelefteq} p_i(L)$, for $i\in\{1,2\}$, and that $\big(k_1(L)\times k_2(L)\big)\mathop{\trianglelefteq} L$. Set $q(L)=L/\big(k_1(L)\times k_2(L)\big)$, and recall that there are canonical group isomorphisms
$$q(L)\cong p_1(L)/k_1(L)\cong p_2(L)/k_2(L)\;\;.$$
\end{mth}
\pagebreak[3]
\begin{mth}{Definition} Let $G$ and $H$ be groups. A subgroup $L$ of $(G\times H)$ will be called {\em diagonal} if
$$L\cap(G\times{\bf 1})=L\cap ({\bf 1}\times H)={\bf 1}\;\;,$$
i.e. equivalently, if $k_1(L)={\bf 1}$ and $k_2(L)={\bf 1}$.\par
The subgroup $L$ will be called {\em centrally diagonal} if
$$L\cap\big(Z(G)\times{\bf 1}\big)=L\cap\big({\bf 1}\times Z(H)\big)={\bf 1}\;\;,$$
i.e. equivalently, if $k_1(L)\cap Z(G)={\bf 1} \;\;\hbox{and}\;\;k_2(L)\cap Z(H)={\bf 1}$.
\end{mth}
\begin{mth}{Notation} Let $G$ and $H$ be groups. When $K$ is a subgroup of $G$, and $\varphi:K\to H$ is a group homomorphism, set
$$\overrightarrow{\Delta}_\varphi(K)=\{\big(x,\varphi(x)\big)\mid x\in K\}\leq G\times H\;\;,$$
and
$$\overleftarrow{\Delta}_\varphi(K)=\{\big(\varphi(x),x\big)\mid x\in K\}\leq H\times G\;\;.$$
\end{mth}
\begin{rem}{Remark} The subgroup $L$ of $(G\times H)$ is diagonal if and only if there exists a subgroup $K\leq G$ and an {\em injective} group homomorphism $\varphi:K\hookrightarrow H$ such that $L=\overrightarrow{\Delta}_\varphi(K)$.
\end{rem}
\begin{mth}{Lemma} Let $P$ and $Q$ be $p$-groups, and let $L$ and $L'$ be genetic subgroups of $(P\times Q)$ such that $L\bizlie{P\times Q}L'$. Then $L$ is centrally diagonal in $(P\times Q)$ if and only if $L'$ is centrally diagonal in $(P\times Q)$.
\end{mth}
\noindent{\bf Proof~:}\ Let $(x,y)\in (P\times Q)$ such that
$$L'^{(x,y)}\cap Z_{P\times Q}(L)\leq L\;\;.$$
Since $Z(P)\times Z(Q)\leq Z_{P\times Q}(L)$, it follows that
\begin{eqnarray*}
L'\cap \big(Z(P)\times 1\big)&=&\Big(L'\cap \big(Z(P)\times 1\big)\Big)^{(x,y)}\\
&=&L'^{(x,y)}\cap \big(Z(P)\times 1\big)\\
&=&L'^{(x,y)}\cap Z_{P\times Q}(L)\cap\big(Z(P)\times 1\big)\\
&\leq& L\cap \big(Z(P)\times 1\big)={\bf 1}\;\;.
\end{eqnarray*}
A similar argument shows that $L'\cap\big({\bf 1}\times Z(Q)\big)={\bf 1}$.~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
Recall from Definition~\ref{axial} that an {\em axial} subgroup of a finite group $G$ is a subgroup of a cyclic subgroup of maximal order of $G$~:
\begin{mth}{Theorem} \label{genetic product Roquette}Let $P$ and $Q$ be non-trivial Roquette $p$-groups, let $e_P$ (resp. $e_Q$) denote the exponent of $P$ (resp. of $Q$), and let $Z_P$ (resp. $Z_Q$) denote the central subgroup of order $p$ in $P$ (resp. in $Q$). \par
\begin{enumerate}
\item Let $L$ be a centrally diagonal genetic subgroup of $(P\times Q)$.
Then $L=\overrightarrow{\Delta}_\varphi(H)$, where $H\leq P$ and $\varphi:H\hookrightarrow Q$ is an injective group homomorphism. Moreover, either $P\cong Q\cong Q_8$, and $H=P$, or $H$ is an axial subgroup of $P$ of order $\min(e_P,e_Q)$, such that $\varphi(H)$ is an axial subgroup of $Q$.
\item Conversely~:\begin{enumerate}
\item if $P\cong Q\cong Q_8$, let $L=L_\varphi=\overrightarrow{\Delta}_\varphi(P)$, where $\varphi:P\to Q$ is a group isomorphism. Then $L$ is a centrally diagonal genetic subgroup of $(P\times Q)$, and $N_{P\times Q}(L)/L\cong C_2$.
\item in all other cases, let $L=L_\varphi=\overrightarrow{\Delta}_\varphi(H)$, where $H$ is an axial subgroup of $P$ of order $\min(e_P,e_Q)$, and $\varphi:H\hookrightarrow Q$ is an injective group homomorphism such that $\varphi(H)$ is an axial subgroup of $Q$. Then $L$ is a centrally diagonal genetic subgroup of $(P\times Q)$. Moreover, the isomorphism class of the group $N_{P\times Q}(L)/L$ depends only on $P$ and $Q$.
\end{enumerate}
\end{enumerate}
\end{mth}
\noindent{\bf Proof~:}\ $\bullet$ Observe first that the group $(Z_P\times Z_Q)L/L$ is a central subgroup of the Roquette group $N_{P\times Q}(L)/L$, hence it is cyclic. Hence $(Z_P\times Z_Q)\cap L\neq 1$. Since both $(Z_P\times {\bf 1})\cap L$ and $({\bf 1}\times Z_Q)\cap L$ are trivial, is follows that $(Z_P\times Z_Q)\cap L$ is equal to
\begin{equation}\label{delta}
\overrightarrow{\Delta}_\psi(Z_P)=\big\{\big(z,\psi(z)\big)\mid z\in Z_P\big\}\;\;,
\end{equation}
where $\psi:Z_P\stackrel{{_{\cong}}}{\to} Z_Q$ is some group isomorphism. In particular $p_1(L)$ contains $Z_P$, and $p_2(L)$ contains $Z_Q$.\bigskip\par\noindent
$\bullet$ Let us prove now that $L$ is diagonal, i.e. that there exists a subgroup $H$ of~$P$ and an injective group homomorphism $\varphi:H\hookrightarrow Q$ such that
$$L=\overrightarrow{\Delta}_\varphi(H)=\big\{\big(h,\varphi(h)\big)\mid h\in H\big\}\;\;.$$
Otherwise, at least one of the groups $k_1(L)$ or $k_2(L)$ is non trivial. But the assumption $L\cap \big(Z(P)\times{\bf 1}\big)={\bf 1}$ is equivalent to $L\cap(Z_P\times{\bf 1})={\bf 1}$, i.e. $k_1(L)\cap Z_P={\bf 1}$, and similarly, the assumption $L\cap ({\bf 1}\times Z(Q))={\bf 1}$ is equivalent to $k_2(L)\cap Z_Q={\bf 1}$. But if there exists a non trivial subgroup $X$ of $P$ such that $X\cap Z_P={\bf 1}$, then $p=2$, $X$ has order 2, and $P$ is dihedral or semidihedral (Lemma~\ref{en vrac}). So if $L$ is not diagonal, then $p=2$, and at least one of $P$ or $Q$ is dihedral or semidihedral.\par
Up to exchanging $P$ and $Q$, one can assume that the group $C=k_1(L)$ is non-trivial, hence non-central of order 2 in $P$. Set $A=p_1(L)$. Since $A\leq N_P(C)=CZ_P$ (Lemma~\ref{en vrac}), it follows that $q(L)=A/C$ has order 1 or~2. \par
If $q(L)={\bf 1}$, then $A=C$, and $L=C\times D$, where $D=k_2(L)=p_2(L)$. In this case
$$N_{P\times Q}(L)/L=\big(N_P(C)/C\big)\times \big(N_Q(D)/D\big)\cong C_2\times \big(N_Q(D)/D\big) $$
cannot be a Roquette group, since $N_Q(D)/D$ is non-trivial (as $D\cap Z_Q={\bf 1}$).\par
And if $|q(L)|=2$, then $A=CZ_P$. If $(a,b)\in N_{P\times Q}(L)$, then in particular $a\in N_P(A,C)=A$. Thus $N_{P\times Q}(L)\leq A\times Q$.
Now the group $N_P(A)$ is a proper subgroup of $P$, since $A$ is elementary abelian of rank 2, and $P$ is a Roquette group. Choose $x\in P-N_P(A)$, whence $A^x\cap A=Z_P$. If $(a,b)\in L^{(x,1)}\cap (A\times Q)$, then $a\in A^x\cap A=Z_P$, hence $(a,b)={^{(x,1)}(a,b)}\in L$. Thus $L^{(x,1)}\cap(A\times Q)\leq L$, and
\begin{eqnarray*}L^{(x,1)}\cap Z_{P\times Q}(L)&\leq& L^{(x,1)}\cap N_{P\times Q}(L)\\
&\leq& L^{(x,1)}\cap(A\times Q)\\
&\leq& L
\end{eqnarray*}
but $L^{(x,1)}\neq L$, since $A^x\neq A$. It follows that $L$ is not expansive in $(P\times Q)$, hence $L$ is not a genetic subgroup of $(P\times Q)$.\bigskip\par\noindent
$\bullet$ Hence $L$ is diagonal in $(P\times Q)$, i.e.
$$L=\overrightarrow{\Delta}_\varphi(H)=\big\{\big(h,\varphi(h)\big)\mid h\in H\big\}\;\;,$$
for some subgroup $H\geq Z_P$ of $P$ and some $\varphi:H\hookrightarrow Q$, such that $\varphi(H)\geq Z_Q$. Then
$$N_{P\times Q}(L)=\Big\{(x,y)\in N_P(H)\times N_Q\big(\varphi(H)\big)\mid \forall h\in H,\;\;\varphi({^xh})={^y\varphi(h)}\Big\}\;\;.$$
The unique central subgroup of order $p$ of the Roquette group $N_{P\times Q}(L)/L$ is equal to $Z/L$, where
\begin{equation}\label{omega1Z}
Z=(Z_P\times{\bf 1})L=({\bf 1}\times Z_Q)L\;\;.
\end{equation}
For any $(x,y)\in (P\times Q)$, saying that $L^{(x,y)}\cap Z_{P\times Q}(L)$ is contained in $L$ is equivalent to saying that the group $I=L^{(x,y)}\cap Z$ is contained in $L$. In particular, for $y=1$
\begin{eqnarray*}
I&=&L^{(x,1)}\cap (Z_P\times{\bf 1})L\\
&=&\Big\{\big(h^x,\varphi(h)\big)\mid h\in H,\;\exists z\in Z_P,\;\exists h'\in H,\; \big(h^x,\varphi(h)\big)=\big(zh',\varphi(h')\big)\Big\}\\
&=&\Big\{\big(h^x,\varphi(h)\big)\mid h\in H,\;h^{-1}h^x\in Z_P\Big\}\;\;,
\end{eqnarray*}
since $\varphi(h)=\varphi(h')$ implies $h=h'$, and since $Z_P$ is central in $P$. Denoting by $[h,x]=h^{-1}h^x$ the commutator of $h$ and $x$, it follows that $I\leq L$ if and only if
$$\forall h\in H,\;\;[h,x]\in Z_P\Longrightarrow h^x\in H,\;\big(h^x,\varphi(h)\big)=\big(h^x,\varphi(h^x)\big)\;\;.$$
In other words $[h,x]\in Z_P$ implies $h^x=h$. Thus $I\leq L$ if and only if
$$\forall h\in H,\;\;[h,x]\in Z_P\Longrightarrow [h,x]=1\;\;.$$
Equivalently $[H,x]\cap Z_P=\{1\}$, where $[H,x]$ denotes {\em the set} of commutators $[h,x]$, for $h\in H$.\par
Since $L=\overrightarrow{\Delta}_\varphi(H)$ is expansive in $(P\times Q)$, it follows that
$$[H,x]\cap Z_P=\{1\}\Longrightarrow L^{(x,1)}=L\;\;.$$
Now $(x,1)$ normalizes $L$ if and only if $x\in C_P(H)$, i.e. if $[H,x]=\{1\}$. Hence
\begin{equation}\label{expansive diagonal}
[H,x]\cap Z_P=\{1\}\Longrightarrow [H,x]=\{1\}\;\;.
\end{equation}
$\bullet$ Let us show now that unless $H=P\cong Q\cong Q_8$, the group $H$ is an axial subgroup of $P$, and the subgroup $\varphi(H)$ is an axial subgroup of $Q$.\par
Let $X$ be a cyclic subgroup of $P$ of order $e_P$, let $x$ be a generator of $X$, and suppose that $H\nleq X$. Then in particular $P$ is not cyclic, so $p=2$, the group $X$ is a (normal) subgroup of index 2 of $P$ (Lemma~\ref{en vrac}), and $X$ is equal to its centralizer in~$P$. Moreover $|H:H\cap X|=2$ since $H\cdot X=P$. The set $[H,x]$ is equal to $\{1,x^2\}$ if $P$ is cyclic or generalized quaternion, or to $\{1,x^{2+2^{n-2}}\}$ if $P$ is semidihedral~: indeed, the image of $H$ in the group of automorphisms of $X$ has order 2, as $H\cap X$ centralizes $X$, and $H$ does not. Since $Z_P$ is generated by $x^{2^{n-2}}$, it follows that $[H,x]\cap Z_P=\{1\}$ if $n\geq 4$, i.e. if $|P|\geq 16$. But $[H,x]\neq\{1\}$, hence $L$ is not expansive in $(P\times Q)$, if $P\geq 16$.\par
So if $H\nleq X$, then $p=2$, and $P$ is non-cyclic, of order at most $8$. Hence $P\cong Q_8$. If $H\neq P$, then $H$ is cyclic, and $H\nleq X$. Thus $|H|=4=e_P$, and in particular $H$ is an axis of $P$. Since $H$ embeds into~$Q$, it follows that $|H|=\min(e_P,e_Q)$. The same argument applied to $\varphi(H)$ shows that $\varphi(H)$ is an axial subgroup of $Q$, as claimed.\par
In this case moreover, the group $Q$ cannot be isomorphic to $Q_8$~: indeed otherwise, one can assume that $P=Q$ and $L=\Delta(H)$ is the diagonal embedding. Then
$$N_{P\times P}(L)=\{(a,b)\mid a^{-1}b\in H\}\;\;.$$
The group $N_{P\times P}(L)/L$ has order 8, generated by the cyclic subgroup
$$C=\{(a,1)L\mid a\in H\}$$
of index 2, and the involution $(b,b)L$, where $b\in P-H$. Hence $N_{P\times P}(L)/L\cong D_8$ is not a Roquette group, and $L$ is not a genetic subgroup of $(P\times P)$.\par
If $H$ is non-cyclic, then $H=P$, and the same argument applied to $\varphi(H)$ shows that $Q\cong Q_8$. And indeed $\overrightarrow{\Delta}_\varphi(P)$ is a genetic subgroup of $P\times Q$~: this follows from Example~\ref{diagonal}, since the map $(x,y)\mapsto \big(x,\varphi^{-1}(y)\big)$ is a group isomorphism from $(P\times Q)$ to $(P\times P)$, sending $\overrightarrow{\Delta}_\varphi(P)$ to $\Delta(P)$. Moreover $[P,P]\leq Z(P)$, and $Z(P)$ has order~2. In particular
\begin{equation}\label{diamond 1}
N_{P\times Q}(L)/L\cong C_2
\end{equation}
does not depend on $\varphi$, up to isomorphism. This proves part a) of Assertion~2.\bigskip\par\noindent
$\bullet$ In the remaining cases $L=\overrightarrow{\Delta}_\varphi(H)$, where $H$ is a non-trivial axial subgroup of $P$, and $\varphi:H\hookrightarrow Q$ is such that $\varphi(H)$ is an axial subgroup of~$Q$. In particular $H$ is cyclic, and non-trivial. As $H\cong\varphi(H)\leq Q$, it follows that $|H|\leq\min(e_P,e_Q)$. \par
Let $C_{e_P}$ be an axis of $P$ containing $H$, and $C_{e_Q}$ be an axis of $Q$ containing $\varphi(Q)$. Then $(C_{e_P}\times C_{e_Q})/\Delta_\varphi(H)$ is an abelian normal subgroup of the Roquette group $N_{P\times Q}(L)/L$, hence it is cyclic. Thus $L=\Delta_\varphi(H)$ is not contained in the Frattini subgroup $(C_{e_P/p}\times C_{e_Q/p})$ of $(C_{e_P}\times C_{e_Q})$. In particular $p_1(L)=C_{e_Q}$, or $p_2(L)=C_{e_Q}$. In other words $H=C_{e_P}$, or $\varphi(H)=C_{e_Q}$, hence $|H|=\min(e_P,e_Q)$. This completes the proof of Assertion~1.\bigskip\par\noindent
$\bullet$ Now assume that at least one of the groups $P$ or $Q$ is not isomorphic to~$Q_8$. Assume also that $e_P\leq e_Q$, and let $H$ be an axis of~$P$~: then $H$ is unique if $P\not\cong Q_8$, and there are three possibilities for $H$ if $P\cong Q_8$ (Lemma~\ref{en vrac}). In any case $H\mathop{\trianglelefteq} P$. Let $K$ denote an axial subgroup of $Q$ of order~$e_P$. Such a group is unique, except if $p=2$, $e_P=4$, and $Q\cong Q_8$ (thus $P\cong C_4$ as $P$ has exponent 4, and is not isomorphic to~$Q_8$). In any case $K\mathop{\trianglelefteq} Q$.\par
Let $\varphi: H\stackrel{{_{\cong}}}{\to} K$ be any group isomorphism, and set $L_\varphi=\overrightarrow{\Delta}_\varphi(H)\leq (P\times Q)$. Then $L_\varphi$ is obviously centrally diagonal, and
$$N_{P\times Q}(L_\varphi)=\big\{(a,b)\in (P\times Q)\mid \forall h\in H,\;\;\varphi({^ah})={^b\varphi(h)}\big\}\;\;.$$
Since $H$ is cyclic of order $e_P$, the map
$$\pi_H:r\in(\mathbb{Z}/e_P\mathbb{Z})^\times\mapsto (x\mapsto x^r)\in{\rm Aut}(H)$$
is a canonical group isomorphism. Similarly, the map
$$\pi_K:r\in(\mathbb{Z}/e_P\mathbb{Z})^\times\mapsto (x\mapsto x^r)\in{\rm Aut}(K)$$
is a canonical group isomorphism.\par
Let $\alpha:P\to {\rm Aut}(H)\stackrel{\pi_H^{-1}}{\longrightarrow}(\mathbb{Z}/e_P\mathbb{Z})^\times$ denote the group homomorphism obtained from the action of $P$ on its normal subgroup~$H$ by conjugation, and $\beta:Q\to {\rm Aut}(K)\stackrel{\pi_K^{-1}}{\longrightarrow}(\mathbb{Z}/e_P\mathbb{Z})^\times$ denote the group homomorphism obtained from the action of $Q$ on its normal subgroup~$K$. Then
$$N_{P\times Q}(L_\varphi)=\big\{(a,b)\in (P\times Q)\mid \alpha(a)=\beta(b)\big\}\;\;.$$
Now the group $(\mathbb{Z}/e_P\mathbb{Z})^\times$ is abelian. The map
$$\Theta: (a,b)\in (P\times Q)\mapsto \beta(b)^{-1}\cdot\alpha(a)\in(\mathbb{Z}/e_P\mathbb{Z})^\times$$
is a group homomorphism, and $N_{P\times Q}(L_\varphi)={\rm Ker}\;\Theta$. In particular, it is a normal subgroup of $(P\times Q)$, which does not depend on $\varphi$, once $H$ and $K=\varphi(H)$ are fixed. In particular $L_\varphi$ is an expansive subgroup of $(P\times Q)$, by Example~\ref{normalisateur normal}.
Moreover setting $I_{P,Q}={\rm Im}(\alpha)\cap{\rm Im}(\beta)$, there is an exact sequence
\begin{equation}\label{simple ses}
{\bf 1}\to C_P(H)\times C_Q(K)\to N_{P\times Q}(L_\varphi)\stackrel{\Psi}{\longrightarrow} I_{P,Q}\to{\bf 1}\;\;,
\end{equation}
where $\Psi(a,b)=\alpha(a)=\beta(b)$, for $(a,b)\in N_{P\times Q}(L_\varphi)$.\medskip\par
\S.~Suppose first that $e_P=p$, i.e. that $P\cong C_p$. In this case $H=Z_P=P$, and $K=Z_Q$, so $L_\varphi$ is central in $(P\times Q)$. Moreover
\begin{equation}\label{diamond 2}
N_{P\times Q}(L_\varphi)/L_\varphi=(P\times Q)/L_\varphi=(P\times Q)/\overrightarrow{\Delta}_\varphi(P)\cong Q
\end{equation}
is a Roquette group, independent of $\varphi$, up to isomorphism. In particular $L_\varphi$ is a genetic subgroup of $(P\times Q)$.\medskip\par
\S.~Assume from now on that $e_P\geq p^2$. Then $C_P(H)\cong C_{e_P}$, and $C_Q(K)\cong C_{e_Q}$, by Lemma~\ref{en vrac}.\medskip\par
\S\S.~If $p>2$, then $P$ and $Q$ are cyclic, hence $H=P$, and
\begin{equation}\label{diamond 3}N_{P\times Q}(L_\varphi)/L_\varphi\cong (P\times Q)/\overrightarrow{\Delta}_\varphi(P)\cong Q
\end{equation}
as above. It is a Roquette group, independent of $\varphi$, up to isomorphism. In particular $L_\varphi$ is genetic in $(P\times Q)$.
\medskip\par
\S\S.~Assume now that $p=2$. The image of $\alpha$ has order $|P:C_P(H)|$, which is equal to 1 if $P$ is cyclic, and to 2 otherwise. Similarly, the image of $\beta$ has order $|Q:C_Q(K)|$, which is equal to 1 if $K$ is central in $Q$, i.e. if $Q$ is cyclic (since $|K|=e_P\geq 4$ by assumption), and to 2 otherwise. Set $I_{P,Q}={\rm Im}(\alpha)\cap{\rm Im}(\beta)$. Then $I_{P,Q}$ has order 1 or 2, and there is an exact sequence
\begin{equation}\label{type}
1\to C_{e_P}\times C_{e_Q}\to N_{P\times Q}(L_\varphi)\to I_{P,Q}\to 1\;\;.
\end{equation}
Note that $I_{P,Q}$ does not depend on $\varphi:H\stackrel{{_{\cong}}}{\to} K$~: more precisely
$${\rm Im}(\alpha)=\left\{\begin{array}{cl}\{1\}&\hbox{if $P$ is cyclic,}\\
\{1,-1\}&\hbox{if $P$ is dihedral or generalized quaternion,}\\
\{1,e_P/2-1\}&\hbox{if $P$ is semidihedral.}
\end{array}\right.$$
So ${\rm Im}(\alpha)$ only depends on the type of $P$.\par
Similarly
$${\rm Im}(\beta)=\left\{\begin{array}{cl}\{1\}&\hbox{if $Q$ is cyclic,}\\
\{1,-1\}&\hbox{if $Q$ is dihedral or generalized quaternion,}\\
\{1,e_Q/2-1\}&\hbox{if $Q$ is semidihedral.}
\end{array}\right.$$
Moreover if $Q$ is semidihedral and if $e_Q>e_P$, then $e_Q/2-1\equiv -1 ({\rm mod.} e_P)$, hence ${\rm Im}(\beta)=\{1,-1\}$. In other words, the group $I_{P,Q}$ is trivial in one of the following cases~:
$$\!\!\left\Arrowvert\begin{array}{l}
\bullet\;\hbox{ $P$ or $Q$ is cyclic,}\\
\bullet\;\hbox{ $P$ is dihedral or generalized quaternion, $Q$ is semidihedral, and $e_P=e_Q$,} \\
\phantom{\bullet}\;\hbox{(i.e. equivalently $|P|=|Q|$),}\\
\bullet \;\hbox{ $P$ is semidihedral, and $Q$ is dihedral or generalized quaternion,}\\
\bullet\;\hbox{ $P$ and $Q$ are semidihedral, and $e_P<e_Q$ (i.e. equivalently $|P|<|Q|$),}
\end{array}\right.$$
and the group $I_{P,Q}$ has order 2 in all other cases, i.e. in one of the following cases~:
$$\!\!\left\Arrowvert\begin{array}{l}
\bullet\;\hbox{ $P$ and $Q$ are dihedral or generalized quaternion,}\\
\bullet\;\hbox{ $P$ is dihedral or generalized quaternion, $Q$ is semidihedral, and $|Q|{>}|P|$,}\\
\bullet\;\hbox{ $P$ and $Q$ are semidihedral, and $P\cong Q$.}
\end{array}\right.$$
As $L_\varphi\leq C_{e_P}\times C_{e_Q}$, the exact sequence~\ref{type} yields the exact sequence
\begin{equation}\label{exact diamond}
1\to (C_{e_P}\times C_{e_Q})/L_\varphi\to N_{P\times Q}(L_\varphi)/L_\varphi\to I_{P,Q}\to 1\;\;.
\end{equation}
\sou{Case 1~:} If $I_{P,Q}$ is trivial, then $N_{P\times Q}(L_\varphi)\cong C_{e_P}\times C_{e_Q}$, and
\begin{equation}\label{diamond 4}
N_{P\times Q}(L_\varphi)/L_\varphi\cong C_{e_Q}\;\;,
\end{equation}
which is a Roquette group, independent of $\varphi$, up to isomorphism. In particular $L_\varphi$ is a genetic subgroup of $(P\times Q)$. \bigskip\par\noindent
\sou{Case 2~:} Suppose now that $I_{P,Q}$ has order 2, i.e. that ${\rm Im}(\alpha)={\rm Im}(\beta)=\{1,\epsilon\}$, where $\epsilon$ is either $-1$ or $e_Q/2-1$ (in the case where $P$ and $Q$ are semidihedral and isomorphic). One can choose an element $u\in P$, of order 2 if $P$ is dihedral or semidihedral, and of order 4 if $P$ is generalized quaternion, such that $\alpha(u)=\epsilon$. Similarly, one can choose an element $v\in Q$, of order 2 if $Q$ is dihedral or semidihedral, and of order 4 if $Q$ is generalized quaternion, such that $\beta(v)=\epsilon$. These choices imply that $(u,v)\in N_{P\times Q}(L_\varphi)$.\par
In the exact sequence~\ref{exact diamond}
$$
1\to (C_{e_P}\times C_{e_Q})/L_\varphi\to N_{P\times Q}(L_\varphi)/L_\varphi\to \{1,\epsilon\}\to 1\;\;,
$$
the group $C=(C_{e_P}\times C_{e_Q})/L_\varphi$ is cyclic, isomorphic to $C_{e_Q}$. The element $\pi=(u,v)L_\varphi$ of $N_{P\times Q}(L_\varphi)/L_\varphi$ acts on $C$ in the same way that $v$ acts on the subgroup $C_{e_Q}$ of $Q$, namely by inversion if $Q$ is dihedral or generalized quaternion, and by raising elements to the power $e_Q/2-1$ if $Q$ is semidihedral. Finally $\pi^2=(u^2,v^2)L_\varphi=L_\varphi$ if none of $P$ and $Q$ are generalized quaternion. If $P$ is generalized quaternion and $Q$ is not, then $\pi^2=(z_P,1)L_\varphi\in C-\{1\}$, where $z_P$ is a generator of $Z_P$. Similarly, if $Q$ is generalized quaternion and $P$ is not, then $\pi^2=(1,z_Q)L_\varphi\in C-\{1\}$, where $z_Q$ is a generator of $Z_Q$. In these two cases $\pi^4=(1,1)L_\varphi=L_\varphi$, so $\pi$ has order 4 in $N_{P\times Q}(L_\varphi)/L_\varphi$. And finally, if both $P$ and $Q$ are generalized quaternion, then $\pi^2=(z_P,z_Q)L_\varphi=L_\varphi$, since $\varphi(Z_P)=Z_Q$. \par
\begin{equation}\label{diamond 5}
\hbox{It follows that the group $N_{P\times Q}(L_\varphi)/L_\varphi$ has order $2e_Q=|Q|$, and that it is~:}
\end{equation}
$$
\!\!\left\Arrowvert\begin{array}{l}
\bullet\;\hbox{ dihedral if $P$ and $Q$ are both dihedral, or both generalized quaternion,}\\
\bullet\;\hbox{ generalized quaternion if one of $P$, $Q$ is generalized quaternion, and the}\\
\phantom{\bullet}\;\hbox{ other is dihedral,}\\
\bullet\;\hbox{ semidihedral if $Q$ is semidihedral.}
\end{array}\right.
$$
So $N_{P\times Q}(L_\varphi)/L_\varphi$ is a Roquette group, independent of $\varphi$, up to isomorphism. In particular, it follows that $L_\varphi$ is a genetic subgroup of $(P\times Q)$. This completes the proof of Theorem~\ref{genetic product Roquette}.~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\begin{mth}{Notation}\label{diamond}\begin{itemize}
\item Let $P$ and $Q$ be Roquette $p$-groups. If $P$ and $Q$ are non-trivial, set $P\diamond Q=N_{P\times Q}(L)/L$, where $L$ is a centrally diagonal genetic subgroup of $(P\times Q)$. Set moreover ${\bf 1}\diamond P=P\diamond{\bf 1}=P$.
\item Let $P$ and $Q$ be Roquette $p$-groups. If $P$ and $Q$ are non-trivial, let $\nu_{P,Q}$ denote the number of equivalence classes of centrally diagonal genetic subgroups of $(P\times Q)$ for the relation $\bizlie{P\times Q}$. Set moreover $\nu_{{\bf 1},P}=\nu_{P,{\bf 1}}=1$.
\end{itemize}
\end{mth}
\vspace{1ex}
\begin{rem}{Remark} if $P={\bf 1}$, and $Q$ is a Roquette $p$-group, then $P\times Q\cong Q$, and the centrally diagonal genetic subgroups of $P\times Q$ are the subgroups ${\bf 1}\times R$, where $R$ is a genetic subgroup of $Q$ such that $R\cap Z(Q)={\bf 1}$. The only such subgroup is $R={\bf 1}$, so $N_Q(R)/R\cong Q\cong N_{P\times Q}({\bf 1}\times R)/({\bf 1}\times R)$. Hence the above definition of $P\diamond Q$ and $\nu_{P,Q}$ is consistent, in the case $P={\bf 1}$.
\end{rem}
\pagebreak[3]
\begin{mth}{Theorem}\label{structure diamond} Let $P$ and $Q$ be Roquette $p$-groups, of exponents $e_P$ and $e_Q$, respectively. Suppose $e_P\leq e_Q$, and set $q=|Q|$. Then
$$P\diamond Q\cong\left\{\begin{array}{cl}
Q&\hbox{if $P={\bf 1}$ or $P\cong C_p$,}\\
C_2&\hbox{if $P\cong Q\cong Q_8$,}\\
D_q&\hbox{if $q\geq 16$ and $P$ and $Q$ are both dihedral,}\\
&\hbox{\ \ \ or both generalized quaternion,}\\
Q_q&\hbox{if one of $P, Q$ is dihedral,}\\
&\hbox{\ \ \ and the other one is generalized quaternion,}\\
SD_q&\hbox{if $Q$ is semidihedral, and}\\
&\hbox{- either $P$ is dihedral or generalized quaternion,}\\
&\hspace{\parindent}\hbox{and $|P|<|Q|$,}\\
&\hbox{- or $P\cong Q$,}\\
C_{e_Q}&\hbox{otherwise.}
\end{array}\right.$$
\end{mth}
\noindent{\bf Proof~:}\ The case $P={\bf 1}$ is trivial, the case $P\cong C_p$ follows from~\ref{diamond 2}, the case $P\cong Q\cong Q_8$ follows from~\ref{diamond 1}, the three next cases in the list follow from~\ref{diamond 5}, and the last case follows from~\ref{diamond 3} and~\ref{diamond 4}.~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\begin{mth}{Theorem} \label{tensor}Let $P$ and $Q$ be Roquette $p$-groups, of exponent $e_P$ and~$e_Q$, respectively, and let $m=\min(e_P,e_Q)$. \begin{enumerate}
\item If $p=2$ and and one of the groups $P$ or $Q$ is isomorphic to $Q_8$, then then $L\bizlie{P\times Q}L'$ for any centrally diagonal genetic subgroups $L$ and $L'$ of $P\times Q$. In other words $\nu_{P,Q}=1$.
\item In all other cases, if $L$ and $L'$ are centrally diagonal genetic subgroups of $P\times Q$, then $L\bizlie{P\times Q}L'$ if and only if $L$ and $L'$ are conjugate in $P\times Q$. In particular $\nu_{P,Q}=\phi(m)m\displaystyle\frac{|P\diamond Q|}{|P||Q|}$, where $\phi$ is the Euler function.
\end{enumerate}
\end{mth}
\noindent{\bf Proof~:}\ Assume $e_P\leq e_Q$, without loss of generality.\medskip\par\noindent
1) If $P={\bf 1}$, then $\nu_{P,Q}=1$, and $P\diamond Q=Q$ by definition, and $e_P=1$, so there is nothing to prove.\medskip\par\noindent
2) If $p=2$ and $P\cong Q\cong Q_8$, then by Theorem~\ref{genetic product Roquette}, a genetic centrally diagonal subgroup~$L$ of $P\times Q$ is of the form $L_\varphi=\overrightarrow{\Delta}_\varphi(P)$, where $\varphi:P\to Q$ is some group isomorphism. Moreover $N_{P\times Q}(L)/L\cong C_2$, so $P\diamond Q\cong C_2$.\par
Now let $\varphi,\psi:P\to Q$ be two group isomorphisms. Then $L_\varphi\bizlie{P\times Q}L_\psi$ if and only if there exists $(x,y)\in (P\times Q)$ such that
\begin{eqnarray*}
L_\varphi^{(x,y)}\cap Z_{P\times Q}(L_\psi)&\leq& L_\psi\\
^{(x,y)}L_\psi\cap Z_{P\times Q}(L_\varphi)&\leq& L_\varphi\;\;.
\end{eqnarray*}
These conditions depend only on the double coset $N_{P\times Q}(L_\varphi)(x,y)N_{P\times Q}(L_\psi)$, which admits a representative of the form $(u,1)$. \par
Now the condition $L_\varphi^{(u,1)}\cap Z_{P\times Q}(L_\psi)\leq L_\psi$ is equivalent to
$$\forall h\in P,\;\;\varphi(h)^{-1}\psi(h^u)\in Z_Q \Longrightarrow \varphi(h)^{-1}\psi(h^u)=1\;\;,$$
and similarly, the condition $^{(u,1)}L_\psi\cap Z_{P\times Q}(L_\varphi)\leq L_\varphi$ is equivalent to
$$\forall h\in P,\;\;\psi(h)^{-1}\varphi({^uh})\in Z_Q \Longrightarrow \psi(h)^{-1}\varphi({^uh})=1\;\;.$$
Applying $\psi^{-1}$ to the first condition and $\varphi^{-1}$ to the second one, and setting $\theta=\psi^{-1}\varphi$, these two conditions become
\begin{eqnarray*}
\forall h\in P,\;\;\theta(h)^{-1}h^u\in Z_P&\Longrightarrow& \theta(h)=h^u\\
\forall h\in P,\;\;\theta^{-1}(h)^{-1}{^uh}\in Z_P&\Longrightarrow&\theta^{-1}(h)={^uh}\;\;.
\end{eqnarray*}
Since $h^uh^{-1}\in[P,P]=Z_P$, there are equivalences
\begin{eqnarray*}
\theta(h)^{-1}h^u\in Z_P\;\; \Longleftrightarrow&\theta(h)^{-1}h\in Z_P&\Longleftrightarrow\;\; h^{-1}\theta(h)\in Z_P\;\;,\\
\theta^{-1}(h)^{-1}{^uh}\in Z_P\;\;\Longleftrightarrow&\theta^{-1}(h)^{-1}{h}\in Z_P&
\Longleftrightarrow\;\; h^{-1}\theta(h)\in Z_P\;\;.
\end{eqnarray*}
Hence, in order to prove that $L_\varphi\bizlie{P\times Q}L_\psi$ for any $\varphi, \psi:P\stackrel{{_{\cong}}}{\to} Q$, it is enough to prove that
\begin{equation}\label{condautQ8}\forall \theta\in {\rm Aut}(P),\;\exists u\in P,\;\forall h\in P,\;\;\theta(h)h^{-1}\in Z_P\implies \left\{\begin{array}{l}\theta(h)=h^u\\\theta^{-1}(h)={^uh}\end{array}\right.\;\;.
\end{equation}
If $h$ has order 1 or 2, then $\theta(h)=h=h^u$ for any $u\in P$, hence the conditions $\theta(h)=h^u$ and $\theta^{-1}(h)={^uh}$ only have to be checked for $|h|=4$. Now the group ${\rm Aut}(P)$ permutes the three cyclic subgroups of order 4 of $P$, and this gives an exact sequence
$$1\to\rm Inn(P)\to {\rm Aut}(P)\to S_3\to 1\;\;,$$
where $S_3$ is the symmetric group on three symbols. Saying that $\theta(h)h^{-1}\in Z_P$ is equivalent to saying that $\theta({<}h{>})={<}h{>}$. Hence, either $\theta$ stabilizes the three subgroups of order 4 of $P$, and in this case $\theta$ is inner, hence there exists $u\in P$ such that $\theta(h)=h^u$, for any $h\in P$, hence $\theta^{-1}(h)={^uh}$, for any $h\in P$. \par
Or there exists a unique subgroup $C$ of order $4$ of $P$ such that $\theta(C)=C$. Then either $\theta(h)=h$ for any $h\in C$, or $\theta(h)=h^{-1}$, for any $h\in C$. In the first case, take $u=1$, and in the second case take $u\in P-{<}h{>}$, and then $\theta(h)=h^u$ for any $h\in C$, hence $h=\theta^{-1}(h)^u$ for $h\in C$, since $C=\theta(C)$. Hence~\ref{condautQ8} holds. This completes the proof in this case.\medskip\par\noindent
3) If $P\cong Q_8$ and $Q\ncong Q_8$, then a centrally diagonal genetic subgroup of $(P\times Q)$ is of the form $L=\overrightarrow{\Delta}_\varphi(H)$, where $H$ is one of the 3 subgroups of order 4 of $P$, and $\varphi$ is some isomorphism from $H$ to the unique axial subgroup $K$ of order 4 of $Q$. Moreover $Z_{P\times Q}(L)=L({\bf 1}\times Z_Q)$.\par
Let $H$ and $H'$ be subgroups of order 4 of $P$. Let $\varphi: H\to K$ and $\varphi':H'\to\nolinebreak K$ be group isomorphisms, and set $L=\overrightarrow{\Delta}(\varphi)$ and $L'=\overrightarrow{\Delta}_{\varphi'}(H')$. Suppose first that $H\neq H'$, and let $(a,b)\in L'\cap Z_{P\times Q}(L)=L({\bf 1}\times Z_Q)$. It means that $a\in H'\cap H=Z_P$, and that there exists $z\in Z_Q$ such that $\varphi'(a)=\varphi(a)z$. But the restrictions of $\varphi$ and $\varphi'$ to $Z_P$ are equal, so $\varphi'(a)=\varphi(a)$, hence $z=1$. It follows that $L'\cap Z_{P\times Q}(L)\leq L$, hence $L\cap Z_{P\times Q}(L')\leq L'$ by symmetry, so $L\bizlie{P\times Q} L'$ in this case.\par
Now if $H=H'$, choose a subgroup $H''$ of order 4 in $P$, different from~$H$, and a group isomorphism $\varphi'':H''\to K$. Set $L''=\overrightarrow{\Delta}_{\varphi''}(H'')$. Then $L\bizlie{P\times Q}L''\bizlie{P\times Q}L'$, by the previous argument, thus $L\bizlie{P\times Q}L'$. \par
Hence $\nu_{P,Q}=1$ in this case, as was to be shown.\medskip\par\noindent
4) In the case there are several choices for $K$, i.e. if $Q\cong Q_8$ and $K$ has order $4=\min(e_P,e_Q)$, it follows that $P\cong C_4$, since $P\ncong Q_8$. In this case, we can exchange $P$ and $Q$, and use the previous argument. Hence $\nu_{P,Q}=1$ in this case as well.\medskip\par\noindent
5) In all other cases, by Theorem~\ref{genetic product Roquette}, a centrally diagonal genetic subgroup $L$ of $(P\times Q)$ is of the form $\overrightarrow{\Delta}_\varphi(H)$, where $H\leq P$ is the unique axis of $P$, and $\varphi:H\hookrightarrow Q$ is a group isomorphism to the unique axial subgroup $K$ of order $e_P$ of $Q$. The normalizer of $L$ in $P\times Q$ does not depend on $\varphi$, by~\ref{simple ses}, so it does not depend on $L$, since $H$ and $K$ are also unique.\par
Let $L$ and $L'$ be two such centrally diagonal genetic subgroups of $P\times Q$. Then $N_{P\times Q}(L)=N_{P\times Q}(L')$, thus $L\bizlie{P\times Q}L'$ if and only if $L$ and $L'$ are conjugate in $(P\times Q)$, by Lemma~\ref{normalisateur normal bizlie}. Moreover, it follows from the definition of $P\diamond Q$ that
$$|N_{P\times Q}(L)=|L||P\diamond Q|=e_P|P\diamond Q|\;\;,$$
so the conjucacy class of $L$ in $(P\times Q)$ has cardinality $\displaystyle\frac{|P||Q|}{e_P|P\diamond Q|}$. Since there are $\phi(e_P)$ possible choices for the isomorphism $\varphi:H\to K$, i.e. $\phi(e_P)$ centrally diagonal subgroups of $(P\times Q)$, it follows that
$$\nu_{P,Q}=\phi(e_P)e_P\frac{|P\diamond Q|}{|P||Q|}\;\;,$$
as was to be shown.~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\begin{rem}{Remark} Suppose that $P\cong Q_8$ and $Q\ncong Q_8$. Then $|P\diamond Q|=|Q|$, by Theorem~\ref{structure diamond}. Hence
$$\phi(e_P)e_P\frac{|P\diamond Q|}{|P||Q|}=2\times 4\times\frac{|Q|}{8|Q|}=1=\nu_{P,Q}\;\;,$$
so the formula for $\nu_{P,Q}$ holds in this case. The only case where $\nu_{P,Q}$ is not equal to $\phi(m)m\displaystyle\frac{|P\diamond Q|}{|P||Q|}$ (where $m=\min(e_P,e_Q)$) is the case $P\cong Q\cong Q_8$~: in this case $\nu_{P,Q}=1$, but
$$\phi(m)m\frac{|P\diamond Q|}{|P||Q|}=2\times 4\times\frac{2}{8\times 8}=\frac{1}{4}\;\;.$$
\end{rem}
\pagebreak[3]
\begin{mth}{Corollary} \label{edge product}Let $P$ and $Q$ be Roquette $p$-groups. Then, in the category $\mathcal{R}_p$
$$\partial P\times\partial Q\cong \nu_{P,Q}\cdot\partial(P\diamond Q)\;\;.$$
In other words, if $P$ has exponent $e_P$, if $Q$ has order $q$ and exponent $e_Q$, and if $e_P\leq e_Q$~:
$$\partial P\times\partial Q=\left\{\begin{array}{ll}\partial Q&\hbox{if $P={\bf 1}$ or $P\cong C_2$,}\\
\rule{0ex}{4ex}\partial C_2&\hbox{if $P\cong Q\cong Q_8$,}\\
\rule{0ex}{4ex}\displaystyle\frac{\phi(e_P)}{2}\cdot\partial D_q&\hbox{if $q\geq 16$ and $P$ and $Q$ are both dihedral,}\\
&\hbox{\ \ \ or both generalized quaternion,}\\
\displaystyle\frac{\phi(e_P)}{2}\cdot\partial Q_q&\hbox{if one of $P, Q$ is generalized quaternion,}\\
&\hbox{\ \ \ and the other one is dihedral,}\\
\displaystyle\frac{\phi(e_P)}{2}\cdot\partial SD_q&\hbox{if $Q$ is semidihedral, and}\\
&\hbox{- either $P$ is dihedral or generalized}\\
&\hspace{\parindent}\hbox{quaternion, and $|P|<|Q|$,}\\
&\hbox{- or $P\cong Q$,}\\
\displaystyle\rule{0ex}{4ex}\frac{\phi(e_P)e_Pe_Q}{|P||Q|}\cdot\partial C_{e_Q}&\hbox{otherwise.}
\end{array}\right.$$
\end{mth}
\noindent{\bf Proof~:}\ Let $\mathcal{B}$ be a genetic basis of the group $R=P\times Q$. In the category $\mathcal{R}_p$, the product $\partial P\times\partial Q$ is equal to $(P\times Q,f_{\bf 1}^P\times f_{\bf 1}^Q)$, and it is a summand of $R$. By Theorem~\ref{Roquette category}, there are mutual inverse isomorphisms
$$\xymatrix{
R\ar[r]<.5ex>^-{\mathcal{D}}&*!U(0.4){\mathop{\oplus}_{S\in \mathcal{B}}\limits\partial\sur{N}_{R}(S)}\ar[l]<.5ex>^-{\mathcal{I}}\;\;,
}
$$
where $\mathcal{I}$ is the direct sum of the maps ${\rm Indinf}_{\sur{N}_{R}(S)}^{R}$, and $\mathcal{D}$ is the direct sum of the maps $f_{\bf 1}^{\sur{N}_{R}(S)}{\rm Defres}_{\sur{N}_{R}(S)}^{R}$. Corollary~\ref{edge product} follows from the fact that
\begin{equation}\label{restricted iso}f_{\bf 1}^{\sur{N}_R(S)}{\rm Defres}_{\sur{N}_{R}(S)}^{R}(f_{\bf 1}^P\times f_{\bf 1}^Q)=0
\end{equation}
unless $S$ is centrally diagonal in $R=P\times Q$~: indeed ${\rm Defres}_{\sur{N}_{R}(S)}^{R}$ is given by the $\big(\sur{N}_{R}(S),R\big)$-biset $S\backslash R$. On the other hand
\begin{eqnarray*}
f_{\bf 1}^P\times f_{\bf 1}^Q&=&(P/{\bf 1}-P/Z_P)\times (Q/{\bf 1}-Q/Z_Q)\\
&=&R/({\bf 1}\times {\bf 1})-R/({\bf 1}\times Z_Q)-R/(Z_P\times {\bf 1})+R/(Z_P\times Z_Q)\;\;,
\end{eqnarray*}
hence ${\rm Defres}_{\sur{N}_{R}(S)}^{R}(f_{\bf 1}^P\times f_{\bf 1}^Q)$ is equal to
$$S\backslash R/({\bf 1}\times {\bf 1})-S\backslash R/({\bf 1}\times Z_Q)-S\backslash R/(Z_P\times {\bf 1})+S\backslash R/(Z_P\times Z_Q)$$
which is
\begin{equation}\label{zero}
S\backslash R-S({\bf 1}\times Z_Q)\backslash R-S(Z_P\times {\bf 1})\backslash R+S(Z_P\times Z_Q)\backslash R\;\;.
\end{equation}
If $S$ is not centrally diagonal in $R=P\times Q$, then either $S=S({\bf 1}\times Z_Q)$ or $S=S(Z_P\times{\bf 1})$. In each case the sum~\ref{zero} vanishes.\par
And if $S$ is centrally diagonal in $R$, then
$$S({\bf 1}\times Z_Q)=S(Z_P\times {\bf 1})=S(Z_P\times Z_Q)\;\;,$$
since the image of these groups in the Roquette group $\sur{N}_R(S)$ is equal to its unique central subgroup $\widehat{S}/S$ of order $p$. In this case
$${\rm Defres}_{\sur{N}_{R}(S)}^{R}(f_{\bf 1}^P\times f_{\bf 1}^Q)=S\backslash R-\widehat{S}\backslash R\;\;.$$
Since $f_{\bf 1}^{\sur{N}_R(S)}=N_R(S)/S-N_R(S)/\widehat{S}$, it follows that ${\rm Defres}_{\sur{N}_{R}(S)}^{R}(f_{\bf 1}^P\times f_{\bf 1}^Q)$ is invariant by composition with $f_{\bf 1}^{\sur{N}_R(S)}$. \par
Conversely, if $S$ is not centrally diagonal in $(P\times Q)$, then
$$(f_{\bf 1}^P\times f_{\bf 1}^Q){\rm Indinf}_{\sur{N}_R(S)}^Rf_{\bf 1}^{\sur{N}_R(S)}=0\;\;,$$
as can be seen by taking opposite bisets in equation~\ref{restricted iso}. And if $S$ is centrally diagonal, then
$$(f_{\bf 1}^P\times f_{\bf 1}^Q){\rm Indinf}_{\sur{N}_R(S)}^Rf_{\bf 1}^{\sur{N}_R(S)}=R/S-R/\widehat{S}={\rm Indinf}_{\sur{N}_R(S)}^Rf_{\bf 1}^{\sur{N}_R(S)}\;\;.$$
Hence the isomorphisms $\mathcal{D}$ and $\mathcal{I}$ restrict to mutual inverse isomorphisms between $\partial P\times \partial Q$ and the direct sum of the edges $\partial \sur{N}_R(S)$, where $S$ is a centrally diagonal genetic subgroup of $R$. But for all such subgroups $S$, the group $\sur{N}_R(S)$ is isomorphic to $P\diamond Q$, and there are $\nu_{P,Q}$ centrally diagonal subgroups in a genetic basis of $R=P\times Q$. This completes the proof.~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\vspace{-2ex}
\section{Examples and applications} \label{examples and applications}
\vspace{-1ex}
\smallskip\par\noindent\pagebreak[2]\refstepcounter{subsection}\refstepcounter{prop}{\bf \thesection.\arabic{prop}.\ \ } Suppose first that $p$ is odd. Then the Roquette $p$-groups are just the cyclic groups $C_{p^n}$, for $n\geq 0$. The ``multiplication rule" of the edges $\partial C_{p^n}$ is the following
\begin{equation}\label{product cyclic edges}
\forall m,\forall n\in \mathbb{N},\;\partial C_{p^m}\times \partial C_{p^n}=\phi(p^{\min(m,n)})\partial C_{p^{\max(m,n)}}\;\;,
\end{equation}
where $\phi$ is the Euler function (thus $\phi(p^k)=p^{k-1}(p-1)$ if $k>0$, and $\phi(1)=1$).
\pagebreak[3]
\smallskip\par\noindent\pagebreak[2]\refstepcounter{subsection}\refstepcounter{prop}{\bf \thesection.\arabic{prop}.\ \ } Some surprising phenomenons occur when $p=2$~:
\begin{mth}{Proposition}\label{edge C2}
In $\mathcal{R}_2$, the edge $\partial C_2$ is isomorphic to the trivial group~${\bf 1}$ (or its edge $\partial {\bf 1}$).
\end{mth}
\noindent{\bf Proof~:}\ Indeed Corollary~\ref{zero edge} implies that if $E\cong (C_2)^2$, then $\partial E=0$ in $\mathcal{R}_2$. Let $X$, $Y$ and $Z$ denote the subgroups of order 2 of $E$. The element
$$u={\rm Res}_X^E\times_Ef_{\bf 1}^E\times_E{\rm Ind}_Y^E$$
of $B(X,Y)$ can be viewed as a morphism from $Y$ to $X$ in the category $\mathcal{R}_2$, which factors through $\partial E$. So this morphism is equal to 0. Since
$$f_{\bf 1}^E=E/{\bf 1}-E/X-E/Y-E/Z+2E/E\;\;,$$
it follows that
$$u={\rm Ind}_{\bf 1}^X{\rm Res}_{\bf 1}^Y-{\rm Inf}_{\bf 1}^X{\rm Res}_{\bf 1}^Y-{\rm Ind}_{\bf 1}^X{\rm Def}_{\bf 1}^Y-{\rm Iso}(\varphi)+2{\rm Inf}_{\bf 1}^X{\rm Def}_{\bf 1}^Y\;\;,$$
where $\varphi$ is the unique group isomorphism from $Y$ to $X$.
Thus
$$0=u\;{\rm Iso}(\varphi^{-1})={\rm Ind}_{\bf 1}^X{\rm Res}_{\bf 1}^X-{\rm Inf}_{\bf 1}^X{\rm Res}_{\bf 1}^X-{\rm Ind}_{\bf 1}^X{\rm Def}_{\bf 1}^X-{\rm Id}_X+2{\rm Inf}_{\bf 1}^X{\rm Def}_{\bf 1}^X\;\;.$$
Hence in the category $\mathcal{R}_2$
$${\rm Id}_X=({\rm Ind}_{\bf 1}^X-{\rm Inf}_{\bf 1}^X)({\rm Res}_{\bf 1}^X-{\rm Def}_{\bf 1}^X)+{\rm Inf}_{\bf 1}^X{\rm Def}_{\bf 1}^X\;\;.$$
It follows that
\begin{equation}\label{ab}f_{\bf 1}^X=f_{\bf 1}^X({\rm Ind}_{\bf 1}^X-{\rm Inf}_{\bf 1}^X)({\rm Res}_{\bf 1}^X-{\rm Def}_{\bf 1}^X)f_{\bf 1}^X\;\;.
\end{equation}
But on the other hand $f_{\bf 1}^X={\rm Id}_X-{\rm Inf}_{\bf 1}^X{\rm Def}_{\bf 1}^X$, so
\begin{eqnarray*}
f_{\bf 1}^X({\rm Ind}_{\bf 1}^X-{\rm Inf}_{\bf 1}^X)&=&f_{\bf 1}^X{\rm Ind}_{\bf 1}^X\\
&=&{\rm Ind}_{\bf 1}^X-{\rm Inf}_{\bf 1}^X{\rm Def}_{\bf 1}^X{\rm Ind}_{\bf 1}^X\\
&=&{\rm Ind}_{\bf 1}^X-{\rm Inf}_{\bf 1}^X\;\;.
\end{eqnarray*}
It follows that
\begin{eqnarray*}
({\rm Res}_{\bf 1}^X-{\rm Def}_{\bf 1}^X)f_{\bf 1}^X({\rm Ind}_{\bf 1}^X-{\rm Inf}_{\bf 1}^X)&=&({\rm Res}_{\bf 1}^X-{\rm Def}_{\bf 1}^X)({\rm Ind}_{\bf 1}^X-{\rm Inf}_{\bf 1}^X)\\
&=&2{\rm Id}_{\bf 1}-{\rm Id}_{\bf 1}-{\rm Id}_{\bf 1}+{\rm Id}_{\bf 1}\\
&=&{\rm Id}_{\bf 1}\;\;.
\end{eqnarray*}
Thus, setting
\begin{eqnarray*}
a&=&f_{\bf 1}^X({\rm Ind}_{\bf 1}^X-{\rm Inf}_{\bf 1}^X)\in{\rm Hom}_{\mathcal{R}_2}({\bf 1},\partial X)\\
b&=&({\rm Res}_{\bf 1}^X-{\rm Def}_{\bf 1}^X)f_{\bf 1}^X\in{\rm Hom}_{\mathcal{R}_2}(\partial X,{\bf 1})\;\;,
\end{eqnarray*}
the composition $b\circ a$ is equal to ${\rm Id}_{\bf 1}$, and Equation~\ref{ab} shows that the composition $a\circ b$ is equal to the identity of~$\partial X$. So $a$ and $b$ are mutual inverse isomorphisms between ${\bf 1}$ and $\partial X$.~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\begin{mth}{Corollary} Let $F$ be a rational $2$-biset functor. Then for any finite 2-group $P$
$$F(C_2\times P)\cong F(P)\oplus F(P)\;\;.$$
$$F(D_8\times P)\cong F(P)^{\oplus 5}\;\;.$$
\end{mth}
\noindent{\bf Proof~:}\ Indeed rational $p$-biset functors are exactly those $p$-biset functors which factor through the category $\mathcal{R}_p$. And in the category $\mathcal{R}_2$, by Theorem~\ref{Roquette category}, there is an isomorphism
$$C_2\cong {\bf 1}\oplus \partial C_2\cong {\bf 1}\oplus{\bf 1}\;\;.$$
Thus $C_2\times P\cong P\oplus P$, and the first assertion follows. The second one follows from Example~\ref{D8Q8}, which shows that in $\mathcal{R}_2$
$$D_8\cong {\bf 1}\oplus4\cdot\partial C_2\cong 5\cdot {\bf 1}\;\;.$$
Hence $D_8\times P\cong 5\cdot P$, thus $F(D_8\times P)\cong F(P)^{\oplus 5}$.
~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\begin{mth}{Proposition} \label{Q8 involution}The edge $\partial Q_8$ is an involution~: more precisely
$$\partial Q_8\times\partial Q_8=\partial C_2\cong{\bf 1}\;\;.$$
\end{mth}
\noindent{\bf Proof~:}\ Indeed $Q_8\diamond Q_8=C_2$, and $\nu_{Q_8,Q_8}=1$, by Theorem~\ref{tensor}.~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\begin{rem}{Remark} The ``action" of this involution on the edges of the other Roquette 2-groups (that is, different from ${\bf 1}$, $C_2$, and $Q_8$) is as follows~: it stabilizes cyclic and semidihedral groups, and exchanges dihedral and generalized quaternion groups. More precisely, it follows from Corollary~\ref{edge product} that
\begin{eqnarray*}
\forall n\geq 2,\;\partial Q_8\times \partial C_{2^n}&=&\partial C_{2^n}\\
\forall n\geq 4,\;\partial Q_8\times \partial D_{2^n}&=&\partial Q_{2^n}\\
\forall n\geq 4,\;\partial Q_8\times \partial Q_{2^n}&=&\partial D_{2^n}\\
\forall n\geq 4,\;\partial Q_8\times \partial SD_{2^n}&=&\partial SD_{2^n}\\
\end{eqnarray*}\vspace{-5ex}
\end{rem}
\smallskip\par\noindent\pagebreak[2]\refstepcounter{subsection}\refstepcounter{prop}{\bf \thesection.\arabic{prop}.\ \ } By Theorem~\ref{Roquette category}, any finite $p$-group is isomorphic to a direct sum of edges of Roquette $p$-groups in the category $\mathcal{R}_p$. The following result shows that the summands of such an arbitrary direct sum are unique, up to group isomorphism, with the possible exception of the isomorphism ${\bf 1}=\partial {\bf 1}\cong \partial C_2$ of Proposition~\ref{edge C2}~:
\begin{mth}{Proposition} \label{unique}Let $\mathcal{S}$ and $\mathcal{T}$ be finite sequences of Roquette $p$-groups, such that there exists an isomorphism
\begin{equation}\label{iso}
\dirsum{S\in\mathcal{S}}\partial S\cong \dirsum{T\in\mathcal{T}}\partial T
\end{equation}
in the category $\mathcal{R}_p$. If $p=2$, replace any occurrence of $C_2$ in $\mathcal{S}$ and $\mathcal{T}$ by the trivial group, which does not change the existence of the isomorphism~\ref{iso}, by Proposition~\ref{edge C2}.\par
Then there exists a bijection $\varphi:\mathcal{S}\to\mathcal{T}$ such that the groups $S$ and $\varphi(S)$ are isomorphic, for any $S\in\mathcal{S}$.
\end{mth}
\noindent{\bf Proof~:}\ By~\cite{fonctrq}, the simple biset functors $S_{R,\mathbb{F}_p}$, where $R$ is a Roquette $p$-group different from $C_p$, are rational biset functors. Moreover, if $|R|\geq p^2$, then for any finite $p$-group $P$, the dimension of $S_{R,\mathbb{F}_p}(P)$ is equal to the number of groups $S$ in a genetic basis of $P$ such that $\sur{N}_P(S)\cong R$. On the other hand, the $\mathbb{F}_p$-dimension of $S_{{\bf 1},\mathbb{F}_p}(P)$ is equal to the number of groups $S$ in a genetic basis of $P$ such that $|\sur{N}_P(S)|\leq p$.\par
The functor $S_{R,\mathbb{F}_p}$ extends to an additive functor from $\mathcal{R}_p$ to the category of $\mathbb{F}_p$-vector spaces, and the value of this functor at the edge $\partial P$ is by definition equal to $\partial S_{R,\mathbb{F}_p}(P)$. If $|R|\geq p^2$, then the $\mathbb{F}_p$-dimension of $\partial S_{R,\mathbb{F}_p}(P)$ is equal to the number of groups $S$ in a genetic basis of $P$ such that $\sur{N}_P(S)\cong R$ and $S\cap Z(P)={\bf 1}$. In particular, if $P$ itself is a Roquette group, then
$$\dim_{\mathbb{F}_p}S_{R,\mathbb{F}_p}(\partial P)=\dim_{\mathbb{F}_p}\partial S_{R,\mathbb{F}_p}(P)=\left\{\begin{array}{cl}1&\hbox{if }\;P\cong R\\0&\hbox{otherwise}\end{array}\right.$$
Applying the functor $S_{R,\mathbb{F}_p}$ to the isomorphism~\ref{iso}, this implies that the number of terms in the sequence $\mathcal{S}$ which are isomorphic to $R$ is equal to the corresponding number in the sequence $\mathcal{T}$.\par
Similarly, for any finite $p$-group $P$, the $\mathbb{F}_p$-dimension of $\partial S_{{\bf 1},\mathbb{F}_p}(P)$ is equal to the number of groups $S$ in a genetic basis of $P$ such that $\sur{N}_P(S)\leq p$ and $S\cap Z(P)={\bf 1}$. If $P$ itself is a Roquette group, this gives
$$\dim_{\mathbb{F}_p}S_{{\bf 1},\mathbb{F}_p}(\partial P)=\dim_{\mathbb{F}_p}\partial S_{{\bf 1},\mathbb{F}_p}(P)=\left\{\begin{array}{cl}1&\hbox{if}\;P\cong C_p\\1&\hbox{if }\;P\cong {\bf 1}\\0&\hbox{otherwise}\end{array}\right.$$
Hence the number of terms in the sequence $\mathcal{S}$ which are isomorphic to ${\bf 1}$ or $C_p$ is equal to the corresponding number in the sequence $\mathcal{T}$. If $p=2$, there are no $S$ in $\mathcal{S}\cup\mathcal{T}$ such that $S\cong C_2$, by assumption. It follows that for any Roquette $p$-group $R$, the number of terms in the sequence $\mathcal{S}$ which are isomorphic to $R$ is equal to the corresponding number in the sequence $\mathcal{T}$. The proposition follows in this case.\par
If $p>2$, the above argument shows that
$$\oplusb{S\in\mathcal{S}}{|S|\geq p^2}\partial S\cong\oplusb{T\in\mathcal{T}}{|T|\geq p^2}\partial T\;\;.$$
Let $M$ denote this direct sum. The isomorphism~\ref{iso} can be rewritten as
\begin{equation}\label{iso2}
m_{\bf 1}\un\oplus m_{C_p}\partial C_p\oplus M\cong
n_{\bf 1}\un\oplus n_{C_p}\partial C_p\oplus M\;\;,
\end{equation}
for some integers $m_{\bf 1}, m_{C_p}, n_{\bf 1}, n_{\mathbb{C}_p}$ such that $m_{\bf 1}+m_{C_p}=n_{\bf 1}+n_{C_p}$.\par
Now, let $\zeta$ be a primitive (i.e. non-trivial, since $p$ is prime) character $(\mathbb{Z}/p\mathbb{Z})^\times\to \mathbb{C}$. Such a character exists since $p>2$. The functor $S_{C_p,\zeta}$ is a rational $p$-biset functor, as it is a summand of $\mathbb{C} R_\mathbb{C}$ (\cite{bisetfunctors} Corollary~7.3.5). Applying this functor to the isomorphism~\ref{iso2} and taking dimensions gives
$$m_{C_p}+\dim_{\mathbb{C}}S_{C_p,\zeta}(M)=n_{C_p}+\dim_{\mathbb{C}}S_{C_p,\zeta}(M)\;\;,$$
since $S_{C_p,\zeta}({\bf 1})=0$ and $S_{C_p,\zeta}(C_p)=\partial S_{C_p,\zeta}(C_p)\cong \mathbb{C}$. \par
It follows that $m_{C_p}=n_{C_p}$, hence $m_{\bf 1}=n_{\bf 1}$, which completes the proof.~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\vspace{2ex}
\pagebreak[3]
\begin{mth}{Corollary} \label{simplifiable}Let $X$, $Y$, and $Z$ be objects of $\mathcal{R}_p$, isomorphic to direct sums of edges of Roquette $p$-groups.
\begin{enumerate}
\item If $X\oplus Z\cong Y\oplus Z$ in $\mathcal{R}_p$, then $X\cong Y$.
\item If $n$ is a positive integer, and if $n\cdot X\cong n\cdot Y$ in $\mathcal{R}_p$, then $X\cong Y$.
\end{enumerate}
\end{mth}
\noindent{\bf Proof~:}\ Decompose $X$ as $X\cong\dirsum{R}n_R(X)\cdot\partial R$, where $R$ runs through the set of isomorphism classes of Roquette $p$-groups, and the function $R\mapsto n_R(X)\in\nolinebreak\mathbb{N}$ has finite support. Choose similar decompositions $Y\cong\dirsum{R}n_R(Y)\cdot\partial R$ and $Z\cong\dirsum{R}n_R(Z)\cdot\partial R$.\par
\pagebreak[3]
For Assertion~1, if $p>2$, it follows from Proposition~\ref{unique} that
$$n_R(X)+n_R(Z)=n_R(Y)+n_R(Z)\;\;,$$
for each $R$. Thus $n_R(X)=n_R(Y)$ for each $R$, hence $X\cong Y$ in $\mathcal{R}_p$.\par
If $p=2$, and if $R$ is a Roquette $p$-group different from ${\bf 1}$ and $C_2$, Proposition~\ref{unique} shows that $n_R(X)+n_R(Z)=n_R(Y)+n_R(Z)$, hence $n_R(X)=n_R(Y)$. Proposition~\ref{unique} also implies that
$$n_{\bf 1}(X)+n_{C_2}(X)+n_{\bf 1}(Z)+n_{C_2}(Z)=n_{\bf 1}(Y)+n_{C_2}(Y)+n_{\bf 1}(Z)+n_{C_2}(Z)\;\;,$$
whence $n_{\bf 1}(X)+n_{C_2}(X)=n_{\bf 1}(Y)+n_{C_2}(Y)$, and $X\cong Y$ in $\mathcal{R}_2$ again, since ${\bf 1}\cong C_2$.\par
The proof of Assertion~2 is similar~: if $p>2$, Proposition~\ref{unique} shows that $n\, n_R(X)=n\, n_R(Y)$, for any $R$, thus $n_R(X)=n_R(Y)$, and $X\cong Y$. And if $p=2$, the conclusion $n_R(X)=n_R(Y)$ is valid for $R$ different from ${\bf 1}$ and $C_2$. Moreover $n \big(n_{\bf 1}(X)+n_{C_2}(X)\big)= n \big(n_{\bf 1}(Y)+n_{C_2}(Y)\big)$, hence $n_{\bf 1}(X)+n_{C_2}(X)= n_{\bf 1}(Y)+n_{C_2}(Y)$, and $X\cong Y$, since ${\bf 1}\cong C_2$. ~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
In the case of the decomposition of a $p$-group as a direct sum of edges of Roquette groups, the above isomorphism $\partial C_2\cong \partial{\bf 1}$ doesn't matter, and the decomposition is unique~:
\begin{mth}{Proposition} \label{isomorphism}Let $P$ and $Q$ be finite $p$-groups. The following assertions are equivalent~:
\begin{enumerate}
\item The groups $P$ and $Q$ are isomorphic in the category $\mathcal{R}_p$.
\item There exist genetic bases $\mathcal{B}_P$ and $\mathcal{B}_Q$ of $P$ and $Q$, respectively, and a bijection $\sigma:\mathcal{B}_P\stackrel{\cong}{\longrightarrow}\mathcal{B}_Q$ such that
\begin{equation}\label{isobases}
\forall S\in\mathcal{B}_P,\;\;N_{Q}\big(\sigma(S)\big)/\sigma(S)\cong N_P(S)/S\;\;.
\end{equation}
\item For any genetic bases $\mathcal{B}_P$ and $\mathcal{B}_Q$ of $P$ and $Q$, respectively, there exists a bijection $\sigma:\mathcal{B}_P\stackrel{\cong}{\longrightarrow}\mathcal{B}_Q$ such that~\ref{isobases} holds.
\end{enumerate}
\end{mth}
\noindent{\bf Proof~:}\
Assertion~2 implies Assertion~1 by Theorem~\ref{Roquette category}. Now suppose that Assertion~1 holds. Then in particular $F(P)\cong F(Q)$, for any rational $p$-biset functor $F$. Let $\mathcal{B}_P$ and $\mathcal{B}_Q$ be genetic bases of $P$ and $Q$, respectively. If $R$ is a Roquette $p$-group, set
$$m_P(R)=\big|\{S\in\mathcal{B}\mid N_P(S)/S\cong R\}\big|\;\;,$$
and define similarly $m_{Q}(R)$ for the group $Q$. The integers $m_P(R)$ and $m_Q(R)$ do not depend on the choices of the genetic bases $\mathcal{B}_P$ and $\mathcal{B}_Q$.\par
If $R$ is not isomorphic to $C_p$, then the simple functor $S_{R,\mathbb{F}_p}$ is rational. Moreover, the $\mathbb{F}_p$-dimension of $S_{R,\mathbb{F}_p}(P)$ is equal to $m_P(R)$ if $|R|>p$, and to $1+m_{P}(C_p)$, if $R={\bf 1}$. Since $P$ is the only element $S$ of $\mathcal{B}$ such that $N_P(S)/S={\bf 1}$, it follows $m_P({\bf 1})=1$, and then $m_P(R)=m_{Q}(R)$ for any Roquette $p$-group $R$. Assertion~2 follows. The equivalence of Assertions 2 and 3 follows from Theorem~\ref{genetic recall}.~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\begin{rem}{Examples} \label{extra}\medskip\par\noindent
$\bullet$ Let $p>2$, and let $X^+$ (resp. $X^-$) denote the extraspecial $p$-group of order $p^3$ and exponent $p$ (resp. $p^2$). Then $X^+\cong X^-$ in~$\mathcal{R}_p$, for if~$P$ is one of these groups, each genetic basis of $P$ consists of $S=P$, for which $N_P(S)/S={\bf 1}$, of the $p+1$ subgroups $S$ of index $p$ in $P$, for which $N_P(S)=P/S\cong C_p$, and an additional non-normal genetic subgroup~$S$ such that $N_P(S)/S\cong C_p$. In other words
$$X^+\cong X^-\cong{\bf 1}\oplus(p+2)\cdot\partial C_p$$
in the category $\mathcal{R}_p$. \medskip\par\noindent
$\bullet$ Similar examples exist for $p=2$~: if $P$ is one of the groups labelled 6 or~7 in the GAP list of groups of order 32 (see \cite{GAP4}), with respective structure $((C_4\times C_2)\rtimes C_2)\rtimes C_2$ and $(C_8\rtimes C_2)\rtimes C_2$, then in any genetic basis of $P$, there is a unique group $S(=P)$ such that $N_P(S)/S={\bf 1}$, there are 6 groups $S$ such that $N_P(S)/S\cong C_2$, and 2 groups $S$ such that $N_P(S)/S\cong C_4$.\medskip\par\noindent
$\bullet$ \label{different orders}Some 2-groups {\em with different orders} may become isomorphic in the category~$\mathcal{R}_2$~: using GAP, one can show that the elementary abelian group of order~16 is isomorphic to each of the groups labelled 134, 138, and 177 in GAP's list of groups of order 64. These groups have respective structure
$$\big((C_4\times C_4)\rtimes C_2)\rtimes C_2,\;\;\Big(\big((C_4\times C_2)\rtimes C_2\big)\rtimes C_2\Big)\rtimes C_2,\;\;\hbox{and}\;\;(C_2\times D_{16})\rtimes C_2\;\;.$$
\vspace{-3ex}
\end{rem}
I couldn't find any similar example for $p>2$. In this case however, the following result characterizes those $p$-groups which become isomorphic in the category $\mathcal{R}_p$~:
\begin{mth}{Proposition} \label{isomorphism Roquette} Let $p$ be a prime number, and let $P$ and $Q$ be finite $p$-groups. \begin{enumerate}
\item If $P\cong Q$ in the category $\mathcal{R}_p$, then the $\mathbb{Q}$-algebras $Z\mathbb{Q} P$ and $Z\mathbb{Q} Q$ are isomorphic.
\item If $p>2$, and if $Z\mathbb{Q} P$ and $Z\mathbb{Q} Q$ are isomorphic $\mathbb{Q}$-algebras, then $P\cong Q$ in the category $\mathcal{R}_p$.
\end{enumerate}
\end{mth}
\noindent{\bf Proof~:}\ Let $\mathcal{G}$ be a genetic basis of $P$. For $S\in\mathcal{G}$, let $V(S)$ denote the corresponding simple $\mathbb{Q} P$-module, defined by
$$V(S)={\rm Indinf}_{\sur{N}_P(S)}^P\Phi_{\sur{N}_P(S)}\;\;.$$
The multiplicity $v_S$ of $V(S)$ in the $\mathbb{Q} P$-module $\mathbb{Q} P$ is equal to
$$v_S=\frac{\dim_\mathbb{Q} V(S)}{\dim_\mathbb{Q}{\rm End}_{\mathbb{Q} P}\big(V(S)\big)}\;\;.$$
As $S$ is a genetic subgroup of $P$, there is an isomorphism of (skew-)fields
$${\rm End}_{\mathbb{Q} P}\big(V(S)\big)\cong {\rm End}_{\mathbb{Q} \sur{N}_P(S)}\big(\Phi_{\sur{N}_P(S)}\big)\;\;.$$
It follows that there is an isomorphism of $\mathbb{Q}$-algebras
$$\mathbb{Q} P\cong\prod_{S\in\mathcal{G}}M_{v_S}\Big({\rm End}_{\mathbb{Q} \sur{N}_P(S)}\big(\Phi_{\sur{N}_P(S)}\big)\Big)\;\;.$$
Hence
$$Z\mathbb{Q} P\cong\prod_{S\in\mathcal{G}}Z\Big({\rm End}_{\mathbb{Q} \sur{N}_P(S)}\big(\Phi_{\sur{N}_P(S)}\big)\Big)\;\;.$$
It shows that the isomorphism type of the $\mathbb{Q}$-algebra $Z\mathbb{Q} P$ depends only on the genetic basis $\mathcal{G}$~: more precisely, it is determined by the isomorphism type of $P$ in $\mathcal{R}_p$. This proves Assertion~1.\par
Now if $p>2$, the group $\sur{N}_P(S)$ is cyclic, of order $p^{m_S}$, say. By Example~\ref{endo Phi}, there is an isomorphism of (skew-)fields
$${\rm End}_{\mathbb{Q} \sur{N}_P(S)}\big(\Phi_{\sur{N}_P(S)}\big)\cong \mathbb{Q} (\zeta_{p^{m_S}})\;\;,$$
where $\zeta_{p^{m_S}}$ is a primitive root of unity of order $p^{m_S}$. Hence
$$Z\mathbb{Q} P\cong\prod_{S\in\mathcal{G}}\mathbb{Q}(\zeta_{p^{m_S}})\;\;.$$
Similarly, if $\mathcal{H}$ is a genetic basis of $Q$
$$Z\mathbb{Q} Q\cong\prod_{T\in\mathcal{H}}\mathbb{Q}(\zeta_{p^{n_T}})\;\;,$$
where $p^{n_T}=|\sur{N}_Q(T)|$. \par
Let $l$ be an integer bigger than all the $m_S$'s, for $S\in\mathcal{G}$, and all the $n_T$'s, for $T\in\mathcal{H}$. Set $K=\mathbb{Q}(\zeta_{p^l})$, and let $G$ be the Galois group of $K$ over $\mathbb{Q}$. By Galois theory (\cite{szamuely} Theorem 1.5.4 and Remark 1.5.5), the $\mathbb{Q}$-algebras $Z\mathbb{Q} P$ and $Z\mathbb{Q} Q$ are isomorphic if and only if there is an isomorphism of $G$-sets
$${\rm Hom}_{\rm alg}(Z\mathbb{Q} P,K)\cong {\rm Hom}_{\rm alg}(Z\mathbb{Q} Q,K)\;\;.$$
When $r\leq l$ is an integer, let $G_r$ denote the Galois group of $K$ over $\mathbb{Q}(\zeta_{p^{r}})$. Then the $G$-set ${\rm Hom}_{\rm alg}(Z\mathbb{Q} P,K)$ is isomorphic to
$$\mathop{\bigsqcup}_{S\in\mathcal{G}}\limits G/G_{n_S}\;\;.$$
The isomorphism $Z\mathbb{Q} P\cong Z\mathbb{Q} Q$ implies that for any $r\leq l$, the number of $S\in\mathcal{G}$ such that $\sur{N}_P(S)$ has order~$p^r$ is equal to the number of $T\in \mathcal{H}$ such that $\sur{N}_Q(T)$ has order $p^r$. Now Assertion~2 of the proposition follows from Proposition~\ref{isomorphism}.~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\begin{rem}{Remark} \label{not true for p=2}Assertion~2 of Proposition~\ref{isomorphism Roquette} is not true for $p=2$~: let $P=D_8$ and $Q=Q_8$ denote a dihedral group of order 8 and a quaternion group of order 8, respectively. By Example~\ref{D8Q8}, in a genetic basis of $P$, there is one group $S$ such that $N_P(S)/S={\bf 1}$ (namely $S=P$), and 4 subgroups $S$ such that $N_P(S)/S\cong C_2$ (the 3 subgroups of index 2 in $P$, and a non-central subgroup of order 2 of $P$). It follows easily that
$$\mathbb{Q} D_8\cong \mathbb{Q}\oplus\mathbb{Q}\oplus\mathbb{Q}\oplus\mathbb{Q}\oplus M_2(\mathbb{Q})\;\;.$$
On the other hand, a genetic basis of $Q$ contains one subgroup $S$ such that $N_Q(S)/S={\bf 1}$ (namely $S=Q$), 3 subgroups $S$ such that $N_Q(S)/S\cong C_2$ (the 3 subgroups of index 2 in $Q$), and one subgroup $S$ such that $N_Q(S)/S\cong Q_8$ (the trivial subgroup of $Q$). Hence
$$\mathbb{Q} Q_8\cong \mathbb{Q}\oplus\mathbb{Q}\oplus\mathbb{Q}\oplus\mathbb{Q}\oplus \mathbb{H}_\mathbb{Q}\;\;,$$
where $\mathbb{H}_\mathbb{Q}$ is the field of quaternions over $\mathbb{Q}$. Then
$$Z \mathbb{Q} D_8\cong \mathbb{Q}^5\cong Z\mathbb{Q} Q_8\;\;.$$
But $D_8$ and $Q_8$ are not isomorphic in $\mathcal{R}_2$, by Proposition~\ref{isomorphism}.
\end{rem}
\masubsect{Genetic bases of direct products} Theorem~\ref{tensor} yields a way to compute a genetic basis of a direct products of $p$-groups. More precisely:
\begin{mth}{Theorem} \label{genetic basis product}Let $P$ and $Q$ be finite $p$-groups, let $\mathcal{B}_P$ be a genetic basis of $P$, and let $\mathcal{B}_Q$ be a genetic basis of $Q$.
\begin{enumerate}
\item For each pair $(S,T)\in\mathcal{B}_P\times\mathcal{B}_Q$, let $\sur{R}$ be a centrally diagonal genetic subgroup of $\sur{N}_P(S)\times \sur{N}_Q(T)$, and let
$$R=\big\{(x,y)\in N_P(S)\times N_Q(T)\mid (xS,yT)\in\sur{R}\big\}\;\;.$$
\nopagebreak
Then $R$ is a genetic subgroup of $P\times Q$, such that
$$\sur{N}_{P\times Q}(R)\cong \sur{N}_P(S)\diamond \sur{N}_Q(T)\;\;.$$
\pagebreak[3]
\item For $(S,T)\in\mathcal{B}_P\times\mathcal{B}_Q$, let $\mathcal{E}_{S,T}$ denote the set of subgroups $R$ obtained in Assertion~1, when $\sur{R}$ runs through a set of representatives of centrally diagonal genetic subgroups of $\sur{N}_P(S)\times \sur{N}_Q(T)$, for the relation $\bizlie{\sur{N}_P(S)\times \sur{N}_Q(T)}$, as described in Theorem~\ref{tensor}. \par
Then the sets $\mathcal{E}_{S,T}$ consist of mutually inequivalent genetic subgroups of $P\times Q$, for the relation $\bizlie{P\times Q}$, and the (disjoint) union
$$\mathcal{B}_{P\times Q}=\mathop{\bigsqcup}_{(S,T)\in\mathcal{B}_P\times\mathcal{B}_Q}\mathcal{E}_{S,T}$$
is a genetic basis of $(P\times Q)$.
\end{enumerate}
\end{mth}
\noindent{\bf Proof~:}\ Assertion~1 is straightforward if the group $\sur{N}_P(S)$ is trivial, i.e. if $S=P$, or if the group $\sur{N}_Q(T)$ is trivial, i.e. if $T=Q$. So we can assume that $S<P$ and $T<Q$. \par
By Theorem~\ref{genetic product Roquette}, the group $\sur{R}$ is diagonal in $\sur{N}_P(S)\times\sur{N}_Q(T)$. It follows that $k_1(R)=S$, $k_2(R)=T$, and $\sur{R}=R/(S\times T)$. This implies in particular that $N_{P\times Q}(R)\leq N_P(S)\times N_Q(T)$. More precisely
$$N_{P\times Q}(R)=\big\{(a,b)\in N_P(S)\times N_Q(T)\mid (aS,bT)\in N_{\sur{N}_P(S)\times\sur{N}_Q(T)}(\sur{R})\big\}\;\;,$$
and the map $(a,b)\mapsto (aS,bT)$ induces a group isomorphism
$$N_{P\times Q}(R)/R\cong N_{\sur{N}_P(S)\times\sur{N}_Q(T)}(\sur{R})/\sur{R}\;\;.$$
It follows that $N_{P\times Q}(R)/R$ is a Roquette group, and by Theorem~\ref{genetic product Roquette} again, and Notation~\ref{diamond}
$$N_{P\times Q}(R)/R\cong\sur{N}_P(S)\diamond \sur{N}_Q(T)\;\;.$$
Let $\widehat{S}\geq S$ denote the subgroup of $N_P(S)$ such that $\widehat{S}/S$ is the unique central subgroup of order $p$ of the Roquette group $\sur{N}_P(S)$. Define $\widehat{T}\geq T$ similarly, and let $\widehat{R}/R$ be the unique central subgroup of order $p$ of $N_{P\times Q}(R)/R$. Then
$$\widehat{R}=(\widehat{S}\times{\bf 1})R=({\bf 1}\times\widehat{T})R\;\;.$$
Let $(x,y)\in P\times Q$ such that $R^{(x,y)}\cap \widehat{R}\leq R$. Intersecting this inclusion with $P\times {\bf 1}$ gives
$$(S^x\cap\widehat{S})\times {\bf 1}\leq S\times {\bf 1}\;\;,$$
thus $S^x\cap \widehat{S}\leq S$, and it follows that $x\in N_P(S)$, since $S$ is an expansive subgroup of $P$. Similarly, intersecting the inclusion $R^{(x,y)}\cap \widehat{R}\leq R$ with ${\bf 1}\times Q$ gives $T^y\cap\widehat{T}\leq T$, hence $y\in N_Q(T)$.\par
Now $S\times T\leq R^{(x,y)}\cap \widehat{R}\leq R$, and taking the quotient by $S\times T$ gives
$$\sur{R}^{(xS,yT)}\cap \big(\widehat{R}/(S\times T)\big)\leq \sur{R}\;\;.$$
As $\sur{R}$ is a genetic subgroup of $\sur{N}_P(S)\times\sur{N}_Q(T)$, it follows that $\sur{R}^{(xS,yT)}$ is equal to $\sur{R}$, hence $R^{(x,y)}=R$. Thus $R$ is an expansive subgroup of $P\times Q$. Since $N_{P\times Q}(R)/R$ is a Roquette group, the group $R$ is a genetic subgroup of $P\times Q$, and this completes the proof of Assertion~1.\par
For Assertion~2, let $(S,T)$ and $(S',T')$ in $\mathcal{B}_P\times\mathcal{B}_Q$, and let $R\in\mathcal{E}_{S,T}$ and $R'\in\mathcal{E}_{S',T'}$, such that $R\bizlie{P\times Q}R'$. It means that there exists $(x,y)\in P\times Q$ such that
\begin{equation}\label{intersect}
R^{(x,y)}\cap \widehat{R}'\leq R'\;\;,\;\;^{(x,y)}R'\cap\widehat{R}\leq R\;\;.
\end{equation}
Intersecting these two inclusions with $P\times {\bf 1}$ gives
$$S^x\cap \widehat{S}'\leq S'\;\;,\;\;^xS'\cap \widehat{S}\leq S\;\;.$$
Hence $S'\bizlie{P}S$, thus $S'=S$, since $S$ and $S'$ are in the same genetic basis of $P$. Moreover $x\in N_P(S)$. Similarly, intersecting~\ref{intersect} with ${\bf 1}\times Q$ implies $T=T'$, and $y\in N_Q(T)$. \par
Quotienting the inclusions~\ref{intersect} by $(S\times T)$ gives that $\sur{R}'\bizlie{\sur{N}_P(S)\times\sur{N}_Q(T)} \sur{R}$. Hence $R'=R$, as was to be shown.\par
Now setting
$$\mathcal{B}_{P\times Q}=\mathop{\bigsqcup}_{(S,T)\in\mathcal{B}_P\times\mathcal{B}_Q}\mathcal{E}_{S,T}$$
yields a set of genetic subgroups of $P\times Q$, which are inequivalent to one other for the relation $\bizlie{P\times Q}$. But
$$|\mathcal{E}_{S,T}|=\nu_{\sur{N}_P(S),\sur{N}_Q(T)}\;\;,$$
and $N_{P\times Q}(R)/R\cong \sur{N}_P(S)\diamond\sur{N}_Q(T)$, for any $R\in\mathcal{E}_{S,T}$
it follows that
\begin{eqnarray*}
\dirsum{R\in\mathcal{B}_{P\times Q}}\partial N_{P\times Q}(R)/R&\cong&\oplusb{S\in\mathcal{B}_P}{T\in\mathcal{B}_Q}\nu_{\sur{N}_P(S),\sur{N}_Q(T)}\partial\big(\sur{N}_P(S)\diamond\sur{N}_Q(T)\big)\\
&\cong&\big(\dirsum{S\in\mathcal{B}_P}\partial \sur{N}_P(S)\big)\times\big(\dirsum{T\in\mathcal{B}_Q}\partial \sur{N}_Q(T)\big)\\
&\cong& P\times Q\;\;.
\end{eqnarray*}
In particular, the rank $l_\mathbb{Q}(P\times Q)$ of the group $R_\mathbb{Q}(P\times Q)$ is equal to $|\mathcal{B}_{P\times Q}|$. Since $\mathcal{B}_{P\times Q}$ is contained in a genetic basis of $P\times Q$, which has cardinality $l_\mathbb{Q}(P\times Q)$, it follows that $\mathcal{B}_{P\times Q}$ is a genetic basis of $P\times Q$.~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\begin{rem}{Remark} Theorem~\ref{genetic basis product} does not mean that any genetic subgroup of $P\times Q$ can be obtained by the construction of Assertion~1. For example, if $[P,P]\leq Z(P)$ and $Z(P)$ is cyclic, then the diagonal $R=\Delta(P)$ is a genetic subgroup of $P\times P$, by Example~\ref{diagonal}. But $k_1(R)={\bf 1}$ is not a genetic subgroup of $P$, if $P$ is not a Roquette group.
\end{rem}
\masubsect{Example of application} As explained in Example~\ref{D8Q8}, the dihedral group $D_8$ splits as
\begin{equation}\label{iso D_8}D_8\cong {\bf 1}\oplus 4\partial C_2
\end{equation}
in the category $\mathcal{R}_2$. By Proposition~\ref{edge C2}, it follows that $D_8\cong 5\cdot{\bf 1}$ in $\mathcal{R}_2$. Hence $(D_8)^n\cong 5^n\cdot{\bf 1}$, for any $n\in\mathbb{N}$. In particular, if $F$ is a rational $2$-biset functor such that $F({\bf 1})=\{0\}$, then $F\big((D_8)^n\big)=\{0\}$. Hence $F(P)=\{0\}$ for any quotient of a direct product of copies of $D_8$, by Remark~\ref{factor group}.\par
Actually, one can be more precise~: since $\partial C_2\times \partial C_2\cong\partial C_2$ by Corollary~\ref{edge product}, it follows that for any $n\in\mathbb{N}$
\begin{equation}\label{D2n}(D_8)^n\cong\dirsum{i=0}^{n}\binom{n}{i}4^{i}\cdot(\partial C_2)^i={\bf 1}\oplus \dirsum{i=1}^{n}\binom{n}{i}4^{i}\cdot\partial C_2\cong {\bf 1}\oplus (5^n-1)\cdot\partial C_2\;\;.
\end{equation}
It means that a genetic basis of the group $P=(D_8)^n$ is made of the group $S=P$, for which $N_P(S)/S={\bf 1}$, and of $(5^n-1)$ subgroups $S$ for which $N_P(S)/S\cong C_2$. \par
In particular, by Theorem~9.5 of~\cite{dadegroup} (or Corollary~12.10.3 of~\cite{bisetfunctors}), the Dade group of $P$ is torsion free, and so is the Dade group of any factor group of~$P$, by Remark~\ref{factor group} again. It shows that the Dade group of a central product of any number of copies of~$D_8$ is torsion free (see Theorem~\ref{product of dihedral groups} for a generalization of this result)~: this was proved by Nadia Mazza and me (Theorem~9.2 of~\cite{boma}). However, the above argument cannot be considered as a new proof of this result, since Theorem~9.5 of~\cite{dadegroup} relies on Theorem~9.2 of~\cite{boma}.
\masubsect{Edges of central products}
Let $P$ and $Q$ be non-trivial finite $p$-groups. Recall that {\em a central product} $P*_\varphi Q$ of $P$ and $Q$ is by definition a group of the form $(P\times Q)/\overrightarrow{\Delta}_\varphi(Z_P)$, where $Z_P$ is a central subgroup of order~$p$ of $P$, and $\varphi : Z_P\hookrightarrow Z(Q)$ is some isomorphism from $Z_P$ to some central subgroup $Z_Q$ of $Q$. \par
In the case where $p=2$ and the groups $P$ and $Q$ both have cyclic center, the group $Z_P$ is unique, as well as the morphism $\varphi$, so the central product is simply denoted by $P*Q$ in this case.
\begin{mth}{Proposition} \label{edge of central products}Let $p$ be a prime number, and let $P$ and $Q$ be non-trivial finite $p$-groups. Let $Z_P$ (resp. $Z_Q$) denote a central subgroup of order~$p$ of $P$ (resp. $Q$).
\begin{enumerate}
\item If one of the groups $Z(P)$ or $Z(Q)$ is non-cyclic, or if $|Z(P)|>p$ and $|Z(Q)|>p$, then $\partial (P*_\varphi Q)=0$ in $\mathcal{R}_p$, for any group isomorphism $\varphi:Z_P\to Z_Q$.
\item If $Z(P)$ and $Z(Q)$ are cyclic, and if moreover $Z(P)$ or $Z(Q)$ has order~$p$, then, in $\mathcal{R}_p$,
$$\dirsum{\varphi:Z_P\stackrel{\cong}{\to}Z_Q}\partial(P*_\varphi Q)\cong \partial P\times\partial Q\;\;.\vspace{-4ex}$$
\end{enumerate}
\end{mth}
\noindent{\bf Proof~:}\ The center of the group $P*_\varphi Q$ is equal to $Z(P)*_\varphi Z(Q)$. It is cyclic if and only if both $Z(P)$ and $Z(Q)$ are cyclic, and if one of them has order~$p$. This proves Assertion~1.\par
For Assertion~2, suppose that $Z(P)$ and $Z(Q)$ are cyclic, and that one of them has order $p$. Then the subgroups $Z_P$ and $Z_Q$ are uniquely determined, and there are $p-1$ group isomorphisms $\varphi:Z_P\to Z_Q$. For each of them, the only central subgroup $Z_\varphi$ of order $p$ of $P*_\varphi Q$ is equal to $(Z_P\times Z_Q)/\overrightarrow{\Delta}_\varphi(Z_P)$, and
$$(P*_\varphi Q)/Z_\varphi\cong (P\times Q)/(Z_P\times Z_Q)\cong\sur{P}\times\sur{Q}\;\;,$$
where $\sur{P}=P/Z_P$ and $\sur{Q}=Q/Z_Q$.\par
By Proposition~\ref{sum of edges of quotients}
\newlength{\longl}
\settowidth{\longl}{$\dirsum{(Z_P\times Z_Q)\leq N\mathop{\trianglelefteq} (P\times Q)}$}
\begin{eqnarray*}
P\times Q&\cong&\makebox[\longl]{$\dirsum{{\bf 1}\leq N\mathop{\trianglelefteq} (P\times Q)}$}\partial\big((P\times Q)/N\big)\\
P*_\varphi Q&\cong&\makebox[\longl]{$\dirsum{\Delta_\varphi(Z_P)\leq N\mathop{\trianglelefteq} (P\times Q)}$}\partial\big((P\times Q)/N\big)\\
\sur{P}\times Q&\cong&\makebox[\longl]{$\dirsum{(Z_P\times{\bf 1})\leq N\mathop{\trianglelefteq} (P\times Q)}$}\partial\big((P\times Q)/N\big)\\
P\times \sur{Q}&\cong&\makebox[\longl]{$\dirsum{({\bf 1}\times Z_Q)\leq N\mathop{\trianglelefteq} (P\times Q)}$}\partial\big((P\times Q)/N\big)\\
\sur{P}\times \sur{Q}&\cong&\makebox[\longl]{$\dirsum{(Z_P\times Z_Q)\leq N\mathop{\trianglelefteq} (P\times Q)}$}\partial\big((P\times Q)/N\big)\;\;.\\
\vspace{-2ex}\end{eqnarray*}
Set \vspace{-2ex}
\begin{equation}\label{S egale}S=\partial(P\times Q)\oplus\left(\dirsum{\varphi:Z_P\stackrel{\cong}{\to}Z_Q}(P*_\varphi Q)\right)\oplus(\sur{P}\times Q)\oplus(P\times \sur{Q})\;\;.
\end{equation}
\pagebreak[3]
Then $S$ is equal to the direct sum of the edges $\partial\big((P\times Q)/N\big)$, for $N\mathop{\trianglelefteq} (P\times Q)$, with multiplicity $(p+1)$ if $N\geq (Z_P\times Z_Q)$, and multiplicity 1 otherwise. Hence
\begin{equation}\label{S egale2}S\cong (P\times Q)+p\cdot (\sur{P}\times \sur{Q}\big)\;\;.
\end{equation}
Now $\partial (P\times Q)=0$ by Corollary~\ref{zero edge}. Moreover, by Corollary~\ref{edge cyclic center}, for each $\varphi:Z_P\stackrel{\cong}{\to}Z_Q$,
$$P*_\varphi Q=\partial(P*_\varphi Q)\oplus (\sur{P}\times\sur{Q})\;\;.$$
But $P\cong\partial P\oplus \sur{P}$ and $Q\cong\partial Q\oplus \sur{Q}$, by Corollary~\ref{edge cyclic center} again. Replacing $P$, $Q$, and $P*_\varphi Q$ by these values in Equation~\ref{S egale} gives
\begin{equation}\label{S1}S\cong\left(\dirsum{\varphi:Z_P\stackrel{\cong}{\to}Z_Q}\partial(P*_\varphi Q)\right)\oplus (p+1)\cdot(\sur{P}\times\sur{Q})\oplus \big(\sur{P}\times(\partial Q)\big)\oplus \big((\partial P)\times \sur{Q}\big)\;\;.
\end{equation}
But replacing $P$ by $\partial P\oplus \sur{P}$ and $Q$ by $\partial Q\oplus \sur{Q}$ in Equation~\ref{S egale2} gives
\begin{equation}\label{S2}S\cong(\partial P\times\partial Q)\oplus (p+1)\cdot(\sur{P}\times\sur{Q})\oplus \big(\sur{P}\times(\partial Q)\big)\oplus \big((\partial P)\times \sur{Q}\big)\;\;.
\end{equation}
Comparing Equations~\ref{S1} and~\ref{S2} gives
$$\dirsum{\varphi:Z_P\stackrel{\cong}{\to}Z_Q}\partial(P*_\varphi Q)\cong \partial P\times\partial Q\;\;,$$
by Corollary~\ref{simplifiable}. This completes the proof.~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\begin{mth}{Corollary} \label{edge of 2-groups}\begin{enumerate}
\item Let $P$ and $Q$ be non-trivial finite 2-groups with cyclic center, and assume that $Z(P)$ or $Z(Q)$ has order 2. Then
$$\partial(P*Q)\cong \partial P\times \partial Q$$
in the category $\mathcal{R}_2$.
\item For each $i\in\{1,\ldots,n\}$, let $P_i$ be a finite 2-group with center $Z_i$ of order 2. Let $\rule{0ex}{4ex}\monast{i=1}{n}P_i=P_1*P_2*\cdots*P_n$ denote the central product of the groups $P_i$. Then
$$\monast{i=1}{n}P_i\cong \prod_{i=1}^n(\partial P_i)\oplus \prod_{i=1}^n(P_i/Z_i)$$
in the category $\mathcal{R}_2$.
\item In particular, for any positive integer $n$, and any integer $m\geq 4$, there are isomorphisms
\begin{eqnarray*}
(D_{2^m})^{*n}&\cong&2^{(n-1)(m-3)}\cdot\partial D_{2^m}\oplus (D_{2^{m-1}})^n\\
(SD_{2^m})^{*n}&\cong&2^{(n-1)(m-3)}\cdot\partial SD_{2^m}\oplus (D_{2^{m-1}})^n\\
(Q_{2^m})^{*n}&\cong&\left\{\begin{array}{l}2^{(n-1)(m-3)}\cdot\partial D_{2^m}\oplus (D_{2^{m-1}})^n\;\hbox{if $n$ is even}\\
2^{(n-1)(m-3)}\cdot\partial Q_{2^m}\oplus (D_{2^{m-1}})^n\;\hbox{if $n$ is odd}
\\\end{array}\right.
\end{eqnarray*}
in the category $\mathcal{R}_2$, where $P^{*n}$ denote the central product of $n$ copies of~$P$.
\end{enumerate}
\end{mth}
\noindent{\bf Proof~:}\ For Assertion~1, the assumptions imply that there is a unique isomorphism $\varphi:Z_P\to Z_Q$. Hence there is only one term in the summation of Proposition~\ref{edge of central products}.\par
Assertion~2 follows from Corollary~\ref{edge cyclic center}, which gives an isomorphism
$$\monast{i=1}{n}P_i\cong \partial \Big(\monast{i=1}{n}P_i\Big)\oplus \Big(\Big(\monast{i=1}{n}P_i\Big)/Y\Big)\;\;,$$
where $Y$ is the unique central subgroup of order $p$ in $\monast{i=1}{n}P_i$. Now an easy induction argument, using Assertion~1, shows that $\partial \Big(\monast{i=1}{n}P_i\Big)\cong\prod_{i=1}^n\limits(\partial P_i)$, and that $\Big(\monast{i=1}{n}P_i\Big)/Y\cong \prod_{i=1}^n\limits(P_i/Z_i)$.\par
Finally, when $P$ is one of the groups $D_{2^m}$, $SD_{2^m}$, or $SD_{2^m}$, then $P/Z\cong D_{2^{m-1}}$. Now Assertion~3 follows from Assertion~2 and from an easy induction argument using Corollary~\ref{edge product}. ~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\begin{rem}{Remark} It follows from Assertion~3 that when $n$ is even, the groups $(D_{2^m})^{*n}$ and $(Q_{2^m})^{*n}$ are isomorphic in the category $\mathcal{R}_2$~: it is actually easy to check that they are isomorphic as {\em groups}.
\end{rem}
\begin{rem}{Example} \label{edges}From Corollary~\ref{edge product} and Assertion~1, it follows that~:
\begin{itemize}
\item $\partial\big((D_8)^{*n}\big)\cong\partial C_2$.
\item $\partial\big((Q_8)^{*n}\big)\cong\left\{\begin{array}{cl}\partial C_2&\hbox{if $n$ is even}\\\partial Q_8&\hbox{if $n$ is odd}\end{array}\right.$
\item $\partial\big((SD_{2^m})^{*n}\big)\cong 2^{(n-1)(m-3)}\cdot\partial SD_{2^m}$, for $m\geq 4$.
\end{itemize}
More generally, if $P$ is any central product of groups isomorphic to $D_8$ or $Q_8$, that is, if $P$ is an extraspecial 2-group, then $\partial P\cong \partial C_2$, or $\partial P\cong\partial Q_8$. In particular (see Example~\ref{representation edge}), we recover the well known fact that $P$ has a unique faithful rational irreducible representation. But there is more~: let $Q$ be a non trivial 2-group. If the center of $Q$ is not cyclic, then the center of any central product $P*Q$ is not cyclic, hence $\partial Q=\partial (P*Q)=0$ in $\mathcal{R}_2$. If the center of $Q$ is cyclic, then there is a unique central product $P*Q$. By Theorem~\ref{Roquette category}, there is a finite sequence $\mathcal{S}$ of Roquette 2-groups such that
$$\partial Q\cong\dirsum{R\in\mathcal{S}}\partial R$$
in the category $\mathcal{R}_2$. By Assertion~1 of Corollary~\ref{edge of 2-groups}, it follows that
$$\partial(P*Q)\cong \dirsum{R\in\mathcal{S}}(\partial P\times\partial R)\;\;.$$
Now $\partial P\cong \partial C_2$ or $\partial P\cong\partial Q_8$. In both cases, by Proposition~\ref{edge C2} and Proposition~\ref{Q8 involution}, the multiplication by $\partial P$ is a permutation of the edges of the Roquette $2$-groups. It follows that there is a sequence $\mathcal{S}'$ of Roquette 2 groups, of the same length as $\mathcal{S}$, such that
$$\partial (P*Q)\cong\dirsum{R\in\mathcal{S}'}\partial R\;\;.$$
In particular, for any field $K$ of characteristic 0, the groups $R_K(\partial P)=\partial R_K(P)$ and $R_K\big(\partial(P*Q)\big)=\partial R_K(P*Q)$ are free of the same rank, equal to the length of $\mathcal{S}$ or $\mathcal{S}'$.\par
Hence in any case, the groups $Q$ and $P*Q$ have the same number (possibly~0 if the center of $Q$ is not cyclic) of faithful irreducible representations over $K$, up to isomorphism.\medskip\par
Similarly, the last example above means in particular that the group $(SD_{2^m})^{*n}$ admits $2^{(n-1)(m-3)}$ non isomorphic faithful rational irreducible representations.\par
\end{rem}
\begin{rem}{Example} \label{extraspecial} Let $p$ be an odd prime, and let $P=X^\epsilon$ (where $\epsilon\in\{\pm 1\}$) be one of the extraspecial groups of order $p^3$ considered in Examples~\ref{extra}. Then $\partial P\cong\partial C_p$. Moreover $Z(P)$ has order $p$, and any automorphism of $Z(P)=Z_P$ can be extended to an automorphism of $P$. It follows that $P*_\varphi Q$ is independent (up to a group isomorphism) of the choice of an embedding $\varphi:P\hookrightarrow Z(Q)$, for any non-trivial $p$-group $Q$ with cyclic center, so we can denote this group by $P*Q$. \par
Now if $Q$ is a non-trivial $p$-group, and $\mathcal{B}$ is a genetic basis of $Q$, it follows from Theorem~\ref{Roquette category} that
$$\partial Q\cong\oplusb{S\in\mathcal{B}}{S\cap Z(Q)={\bf 1}}\partial \sur{N}_Q(S)\;\;,$$
and the right hand side is a direct sum of {\em non-trivial} Roquette $p$-groups. By Equation~\ref{product cyclic edges}, for any non trivial Roquette $p$-group $R$
$$\partial C_p\times\partial R\cong (p-1)\partial R\;\;.$$
Hence for any non-trivial $p$-group $Q$
$$\partial C_p\times\partial Q\cong (p-1)\partial Q\;\;.$$
It follows that if $Q$ is a non-trivial $p$-group with cyclic center, and if $P\cong X^\epsilon$, then
$$\dirsum{\varphi:Z_P\stackrel{\cong}{\to}Z_Q}\partial(P*_\varphi Q)\cong(p-1)\partial(P*Q)\cong \partial C_p\times\partial Q\cong (p-1)\partial Q\;\;,$$
hence $\partial(X^\epsilon*Q)\cong\partial Q$, by Corollary~\ref{simplifiable}. Note that this is also true (for any central product of $X^\epsilon$ with $Q$) if the center of $Q$ is non-trivial, since in this case the center of $X^\epsilon*Q$ is also non trivial, and then $\partial(X^\epsilon*Q)\cong\partial Q\cong 0$ in $\mathcal{R}_p$, by Corollary~\ref{zero edge}. \par
It follows easily by induction that if $P$ is any central product of groups isomorphic to $X^+$ or $X^-$, i.e. if $P$ is an extraspecial $p$-group, then $\partial P\cong \partial C_p$. We recover this way the well known fact that for $p>2$ also, extraspecial $p$-groups have a unique faithful rational irreducible representation. The same argument shows more generally that $\partial(P*Q)\cong\partial Q$ in $\mathcal{R}_p$, for any non-trivial $p$-group $Q$. In particular $Q$ and $P*Q$ have the same number of faithful irreducible representations over a given field $K$ of characteristic 0.
\end{rem}
\begin{mth}{Theorem} \label{product of dihedral groups}\begin{enumerate}
\item Let $P$ be an arbitrary finite direct product of groups of order 2 and dihedral 2-groups. Then the Dade group of any factor group of $P$ is torsion free.
\item Let $4\leq m_1\leq m_2\leq\cdots\leq m_n$ be a non decreasing sequence of integers. Set $s=\sum_{i=1}^{n-1}\limits(m_i-3)$. Then the torsion part of the Dade group of $\monast{i=1}{n}SD_{2^{m_i}}$ is isomorphic to $(\mathbb{Z}/2\mathbb{Z})^{2^{s-1}}$ if $m_1<m_n$, and to $(\mathbb{Z}/2\mathbb{Z})^{2^{s}}$ if $m_1=m_n$.
\item In particular, for any integers $n\geq 1$ and $m\geq 4$, the torsion part of the Dade group of $(SD_{2^m})^{*n}$ is isomorphic to $(\mathbb{Z}/2\mathbb{Z})^{2^{(n-1)(m-3)}}$.
\end{enumerate}
\end{mth}
\noindent{\bf Proof~:}\ First by Example~\ref{D8Q8}
$$D_8\cong{\bf 1}\oplus 4\cdot \partial C_2$$
in $\mathcal{R}_2$. Now by Corollary~\ref{edge cyclic center}, since the center $Z$ of $D_{16}$ has order 2, and since $D_{16}/Z\cong D_8$
$$D_{16}\cong {\bf 1}\oplus 4\cdot \partial C_2\oplus \partial D_{16}\;\;.$$
Since for $n\geq 3$, the group $D_{2^n}$ has a center $Z$ of order 2, and since $D_{2^n}/Z\cong D_{2^{n-1}}$, it follows by induction that
\begin{equation}\label{d2n}
D_{2^n}\cong {\bf 1}\oplus 4\cdot \partial C_2\oplus\mathop{\bigoplus}_{l=4}^n\limits \partial D_{2^l}\;\;.
\end{equation}
Now by Corollary~\ref{edge product}, the product $\partial D_{2^l}\times\partial D_{2^m}$, for $l\leq m$, is isomorphic to $2^{l-3}\cdot\partial D_{2^{m}}$. Moreover $\partial C_2\times \partial P\cong\partial P$ for any 2-group $P$, since $\partial C_2\cong{\bf 1}$ by Proposition~\ref{edge C2}, it follows that any product of dihedral 2-groups and groups of order~2 is isomorphic to a direct sum of edges of the trivial group, of the edge of the group of order 2, and of edges of dihedral 2-groups.\par
It follows that if $P$ is a direct product of dihedral 2-groups and groups of order 2, then $P$ is isomorphic in $\mathcal{R}_2$ to the direct sum of the trivial group, and some copies of edges of the group of order~2 and the edge of dihedral 2-groups. In other words, if $S$ is a genetic subgroup of $P$, then $N_P(S)/S$ is either trivial, or of order~2, or dihedral. Now the Dade group of dihedral 2-groups is torsion free (by~\cite{cath} Theorem~10.3 (a)), and the Dade groups of the trivial group and the group of order 2 are trivial. It follows that the Dade group of $P$ is torsion free, as well as the Dade group of any quotient of $P$, as is it a direct summand of the Dade group of $P$. This proves Assertion~1.\par
For Assertion~2, set $P=\monast{i=1}{n}SD_{2^{m_i}}$. By Assertion~2 of Corollary~\ref{edge of 2-groups},\vspace{-2ex}
\begin{equation}\label{pisd}P\cong \prod_{i=1}^m\partial SD_{2^{m_i}}\oplus \prod_{i=1}^nD_{2^{m_i-1}}\;\;,
\end{equation}
in the category $\mathcal{R}_2$. Now an easy induction on $n$, using Corollary~\ref{edge product}, shows that
$$\prod_{i=1}^n\partial SD_{2^{m_i}}\cong\left\{\begin{array}{cl}
2^{s-1}\cdot\partial C_{2^{m_n-1}}&\hbox{ if }m_1<m_n\\
2^s\cdot \partial SD_{2^{m_n}}&\hbox{ if }m_1=m_n\end{array}\right.$$
where $s=\sum_{i=1}^{n-1}\limits(m_i-3)$.\par
By~\ref{pisd}, this means that in a genetic basis $\mathcal{B}$ of $P$, there are $2^{s-1}$ or $2^s$ subgroups $S$ such that $N_P(S)/S$ is semidihedral, depending on $m_1<m_n$ or $m_1=m_n$, and for the other $S\in\mathcal{B}$, the group $N_P(S)/S$ is trivial, of order~2, or dihedral. \par
The Dade group of a dihedral 2-group is torsion free, and the Dade groups of the trivial group and $C_2$ are trivial. Moreover, the faithful torsion part $\partial D^t(C_{2^m})$ of the Dade group of $C_{2^m}$ is isomorphic to $\mathbb{Z}/2\mathbb{Z}$, if $m\geq 2$ (see \cite{bisetfunctors} Theorem 12.10.3). Similarly, the faithful torsion part $\partial D^{t}(SD_{2^m})$, for $m\geq 4$, is isomorphic to $\mathbb{Z}/2\mathbb{Z}$. This completes the proof of Assertion~2. Assertion~3 is a particular case of Assertion~2.~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\begin{rem}{Remark} \label{units burnside}Let $P$ be a finite product of groups of order 2, and dihedral 2-groups, as in Assertion~1, and let $Q$ be a quotient of $P$. If $T$ is a genetic subgroup of $Q$, then $N_Q(T)/T$ is either trivial, of order 2, or dihedral~: indeed $T$ lifts to a genetic subgroup $S$ of $P$, such that $N_P(S)/S\cong N_Q(T)/T$. It follows in particular that the map
$$\sur{\epsilon}_Q:B^\times(Q)\to{\rm Hom}_\mathbb{Z}\big(R_\mathbb{Q}(Q),\mathbb{F}_2\big)$$
introduced in \cite{burnsideunits}, Notation~8.4, is a group isomorphism from the group of units of the Burnside ring of $Q$ to the $\mathbb{F}_2$-dual of $R_\mathbb{Q}(Q)$~: indeed, there are non-negative integers $a$ and $b_i$, for $i\in\{4,\ldots,m\}$ such that
$$Q\cong {\bf 1}\oplus a\cdot\partial C_2\oplus \dirsum{i=4}^mb_i\cdot \partial D_{2^i}$$
in the category $\mathcal{R}_2$. Then $B^\times(Q)\cong (\mathbb{F}_2)^r$, where $r=1+a+\sum_{i=4}^m\limits b_i$, by \cite{burnsideunits}, Theorem~8.5. Similarly $R_\mathbb{Q}(Q)\cong\mathbb{Z}^r$ (hence $r$ is equal to the number of conjugacy classes of cyclic subgroups of $Q$), so ${\rm Hom}_\mathbb{Z}\big(R_\mathbb{Q}(Q),\mathbb{F}_2\big)\cong(\mathbb{F}_2)^r$. As $\sur{\epsilon}$ is injective, it is an isomorphism.
\end{rem}
\begin{mth}{Proposition} Let $m\geq 3$ be an integer. Then for any integer $n$, there is an isomorphism
\begin{equation}\label{D2nk}(D_{2^m})^n\cong {\bf 1}\oplus (5^n-1)\cdot\partial C_2\oplus\mathop{\bigoplus}_{l=4}^m\limits\frac{(3+2^{l-2})^n-(3+2^{l-3})^n}{2^{l-3}}\cdot\partial D_{2^l}
\end{equation}
in the category $\mathcal{R}_2$.
\end{mth}
\noindent{\bf Proof~:}\ Let $\mathcal{S}_p$ denote the full subcategory of $\mathcal{R}_p$ consisting of finite direct sums of edges of Roquette $p$-groups, and let $\Gamma=K_0(\mathcal{S}_p)$ be the Grothendieck group of this category, for relations given by direct sum decomposition. Then Corollary~\ref{simplifiable} shows that $\Gamma$ is a free abelian group, and that two objects of $\mathcal{S}_p$ have the same image in $\Gamma$ if and only if they are isomorphic in~$\mathcal{R}_p$. Moreover, by Corollary~\ref{edge product}, the category $\mathcal{S}_p$ is a tensor subcategory of $\mathcal{R}_p$, and $\Gamma$ is actually a commutative ring.\par
It follows that $\Gamma$ identifies to a subring of the $\mathbb{Q}$-algebra $\mathbb{Q} \Gamma=\mathbb{Q}\otimes_\mathbb{Z}\Gamma$, and that, to prove the proposition, it suffices to check that the two sides of Equation~\ref{D2nk} have the same image in $\mathbb{Q}\Gamma$. Let $c$ denote the image of $\partial C_2$ in~$\Gamma$, and for $l\geq 4$, let $d_l$ denote the image of $\partial D_{2^l}$ in $\Gamma$. By Equation~\ref{d2n}, the image $i_m$ of $D_{2^m}$ in $\Gamma$ is equal to
$$i_m=1+4c+\sum_{l=4}^md_l\;\;.$$
By Corollary~\ref{edge product}, for $4\leq l\leq k$
$$d_l\times d_k = 2^{l-3}d_k\;\;.$$
It follows that the elements $e_l=\displaystyle\frac{1}{2^{l-3}}d_l$ of $\mathbb{Q}\Gamma$, for $l\geq 4$, are such that
$$\forall l,k,\;4\leq l\leq k,\;\;e_l\times e_k=e_k\;\;.$$
In particular $e_l$ is an idempotent, and the elements
$$f_l=e_l-e_{l+1},\;\hbox{for $4\leq l <m$, and }\;f_m=e_m$$
are orthogonal idempotents of $\mathbb{Q}\Gamma$. With this notation, for $l\geq 4$
$$e_l=f_l+f_{l+1}+\cdots+f_m\;\;,$$
and the element $i_r$ can be written as
\begin{eqnarray*}
i_m&=&1+4c+\sum_{l=4}^md_l\\
&=&1+4c+\sum_{l=4}^m2^{l-3}(f_l+f_{l+1}+\cdots+f_m)\\
&=&1+4c+\sum_{l=4}^m 2(2^{l-3}-1)f_l\;\;.
\end{eqnarray*}
Moreover, it follows from Proposition~\ref{edge C2} that $c\times f_l=f_l$, for $4\leq l\leq m$. Thus
\begin{eqnarray*}
(i_m)^n&=&(1+4c)^n+\sum_{j=1}^n\binom{n}{j}(1+4c)^{n-j}\left(\sum_{l=4}^m 2(2^{l-3}-1)f_l\right)^j\\
&=&(1+4c)^n+\sum_{j=1}^n\binom{n}{j}5^{n-j}\sum_{l=4}^m 2^j(2^{l-3}-1)^jf_l\\
&=&(1+4c)^n+\sum_{l=4}^m\left(\sum_{j=1}^n\binom{n}{j}5^{n-j}2^j(2^{l-3}-1)^j\right)f_l\\
\end{eqnarray*}
\begin{eqnarray*}
(i_m)^n&=&(1+4c)^n+\sum_{l=4}^m\Big(\big(5+2(2^{l-3}-1)\big)^n-5^n\Big)f_l\\
&=&(1+4c)^n+\sum_{l=4}^m\big((3+2^{l-2})^n-5^n\big)f_l\\
&=&(1+4c)^n+\sum_{l=4}^m\big((3+2^{l-2})^n-(3+2^{l-3})^n\big)e_l\;\;.
\end{eqnarray*}
The proposition follows, since $(1+4c)^n=1+(5^n-1)c$, by Equation~\ref{D2n}, and since $e_l=\displaystyle\frac{1}{2^{l-3}}d_l$.~\leaders\hbox to 1em{\hss\ \hss}\hfill~\raisebox{.5ex}{\framebox[1ex]{}}\smp
\begin{rem}{Remark} The isomorphism~\ref{D2nk} is equivalent to saying that a genetic basis of the group $P=(D_{2^m})^n$ consists of one subgroup $S$ such that $N_P(S)/S\cong {\bf 1}$ (namely $S=P$), of $5^n-1$ subgroups $S$ such that $N_P(S)/S\cong C_2$, and, for $4\leq l\leq m$, of $\displaystyle\frac{(3+2^{l-2})^n-(3+2^{l-3})^n}{2^{l-3}}$ subgroups $S$ such that $N_P(S)/S\cong D_{2^l}$.\par
Together with Assertion~3 of Corollary~\ref{edge of 2-groups}, this also gives the structure of genetic bases of the groups $(D_{2^m})^{*n}$, $(SD_{2^m})^{*n}$, $(Q_{2^m})^{*n}$~:
\begin{mth}{Corollary} Let $P$ be one of the groups $D_{2^m}$, $SD_{2^m}$, or $Q_{2^m}$, for $m\geq 4$. Then, for any positive integer $n$, any genetic basis of the group $Q=P^{*n}$ consists~:
\begin{itemize}
\item of one group $S$ such that $N_P(S)/S={\bf 1}$ (namely $S=Q$).
\item of $5^n-1$ subgroups $S$ such that $N_P(S)/S\cong C_2$.
\item for $4\leq l\leq m-1$, of $\displaystyle\frac{(3+2^{l-2})^n-(3+2^{l-3})^n}{2^{l-3}}$ subgroups $S$ such that $N_P(S)/S\cong D_{2^l}$.
\item of $2^{(n-1)(m-3)}$ subgroups $S$ such that $N_P(S)/S$ is isomorphic to $$\left\{\begin{array}{cl}D_{2^m}&\hbox{if $P=D_{2^m}$}\\SD_{2^m}&\hbox{if $P=SD_{2^m}$}\\D_{2^m}&\hbox{if $P=Q_{2^m}$ and $n$ is even}\\mathbb{Q}_{2^m}&\hbox{if $P=Q_{2^m}$ and $n$ is odd}\;\;.\end{array}\right.$$
\end{itemize}
\end{mth}
\end{rem}
|
\section{Introduction}
The Bianchi models, which describe homogeneous but anisotropic
spacetimes, have been extensively discussed in the literature,
motivated in part by attempts to explain small but significant
anisotropies in the cosmic microwave background (CMB) \cite{JD,EK}
and large structures \cite{MT}.
Among the anisotropic Bianchi models, the simplest ones are Bianchi
type-I (BI) models whose spatial sections are flat but the expansion
or contraction rates are direction dependent.
Primordial magnetic fields can have a significant impact on the CMB
anisotropy. Also the presence of strong magnetic fields raises
interesting problems like the formation of galaxies in the Universe.
The BI models are appropriate for the investigation of a Universe
which is permeated by a large scale, homogeneous magnetic field.
In the early stages of the evolution of the Universe it is expected
that topological defects could have formed naturally during phase
transitions followed by spontaneous broken symmetries. Cosmic
strings are linear topological defects, have very interesting
properties and might play an important role in structure formation
\cite{VS,HK}.
In the first part of the paper we shall investigate the evolution of a
BI model in presence of a cloud of strings and magnetic field. In order
to solve Einstein equations we resort to a tractable assumption
concerning a relation between the rest energy and tension
density of the system of strings \cite{PSL}.
Loop Quantum Gravity (LQG) \cite{CR,TT} represents one of the most
compelling attempt towards a complete non perturbative quantum theory
for the gravitational interaction. The cosmological application of LQG
was developed in terms of invariant connections \cite{MB1} and this
model was denoted by Loop Quantum Cosmology (LQC). LQC takes the
ingredients of LQG and applies them to expanding Universe or black
hole models. LQC permits the exploration of the effects of quantum
physics and quantum geometry in gravitation \cite{AA,MB2}.
In order to test the robustness of the LQC it is necessary to apply the
methodology to some concrete situations and one of the most favorable
model is represented by the simplest of anisotropic models, namely BI
cosmologies.
The detailed formulation for LQC in the BI models
\cite{DWC1,CV,DWC2,AE,MV} reveals the fact that gravity can behave
repulsively at Planckian energy densities leading to the replacement of
the big bang singularity with a big bounce.
The plan of the paper is as follows: In Section 2 we outline the
classical equations for a BI string cosmological model in the presence
of a magnetic field. In Section 3 we describe the effective loop
dynamics for the present BI cosmological model. In the next section we
present some numerical simulations and compare the classical and LQC
approaches. Some conclusions and open problems are discussed in the
last Section.
\section{Classical equations}
The line element of a BI Universe is
\begin{equation}
ds^2 = - dt^2 + a_1^2 dx^2 + a_2^2 dy^2 + a_3^2 dz^2\,, \label{BI}
\end{equation}
with three scale factors $a_i$ $(i=1,2,3)$ which are functions
of time $t$ only and consequently three expansion
rates. In principle all these scale factors could be different and it
is useful to express the mean expansion rate in terms of the average
Hubble rate:
\begin{equation}\label{Hubble}
H = \frac{1}{3}(H_1 + H_2 + H_3) = \frac{1}{3}\Bigl(\frac{\dot
a_1}{a_1}+\frac{\dot a_2}{a_2}+ \frac{\dot a_3}{a_3}\Bigr) =
\frac{1}{3} \frac{\dot V}{V}\,,
\end{equation}
where we have defined a new function
\begin{equation}
V = \sqrt{-g}= a_1 a_2 a_3 \,, \label{taudef}
\end{equation}
which is in fact the volume scale of the BI space-time. $H_i$ are
the so-called directional Hubble parameters:
\begin{equation}
H_i = \frac{\dot a_i}{a_i}. \label{DH}
\end{equation}
In \eqref{Hubble}, \eqref{DH} and further over-dot means
differentiation with respect to $t$.
The Einstein's gravitational field equation has the form
\begin{subequations}
\label{BID}
\begin{eqnarray}
\frac{\ddot a_2}{a_2} +\frac{\ddot a_3}{a_3} +
\frac{\dot a_2}{a_2}\frac{\dot
a_3}{a_3}&=& - \kappa T_{1}^{1}\,,\label{11}\\
\frac{\ddot a_3}{a_3} +\frac{\ddot a_1}{a_1} +
\frac{\dot a_3}{a_3}\frac{\dot
a_1}{a_1}&=& -\kappa T_{2}^{2}\,,\label{22}\\
\frac{\ddot a_1}{a_1} +\frac{\ddot a_2}{a_2} +
\frac{\dot a_1}{a_1}\frac{\dot
a_2}{a_2}&=& - \kappa T_{3}^{3}\,,\label{33}\\
\frac{\dot a_1}{a_1}\frac{\dot a_2}{a_2} +\frac{\dot
a_2}{a_2}\frac{\dot a_3}{a_3}+\frac{\dot a_3}{a_3}\frac{\dot
a_1}{a_1}&=& - \kappa T_{0}^{0}\,, \label{00}
\end{eqnarray}
\end{subequations}
where $\kappa$ is the gravitational constant.
The energy momentum tensor for a system of cosmic strings and
magnetic field in a comoving coordinate is given by
\begin{equation}
T_{\mu}^{\nu} = \rho_{string} u_\mu u^\nu - \lambda x_\mu x^\nu +
E_\mu^\nu\,, \label{imperfl}
\end{equation}
where $\rho_{string}$ is the rest energy density of strings with
massive particles attached to them and can be expressed as
$\rho_{string} = \rho_{p} + \lambda$, where $\rho_{p}$ is the rest
energy density of the particles attached to the strings and
$\lambda$ is the tension density of the system of strings
\cite{PSL,pradhan,tade} which may be positive or negative. Here
$u_i$ is the four velocity and $x_i$ is the direction of the string,
obeying the relations
\begin{equation}
u_iu^i = -x_ix^i = -1, \quad u_i x^i = 0\,. \label{velocity}
\end{equation}
In \eqref{imperfl} $E_{\mu\nu}$ is the electromagnetic field given by
Lichnerowich \cite{lich}. In our case the electromagnetic field tensor
$F_{\mu \nu}$ has only one non-vanishing component, namely
\begin{equation}
F_{23} = h\,,
\label{f23}
\end{equation}
where $h$ is assumed to be constant.
For the electromagnetic field $E_\mu^\nu$ one gets the following non-trivial
components
\begin{equation}
E_0^0 = E_1^1 = - E_2^2 = - E_3^3 = \frac{h^2}
{2 {\bar \mu} a_2^2 a_3^2}
\equiv \frac{1}{2}\frac{\beta^2}{(a_2 a_3)^2}\,,
\label{E}
\end{equation}
where $\bar \mu$ is a constant characteristic of the medium and called
the magnetic permeability. Typically $\bar \mu$ differs from unity
only by a few parts in $10^5$ ($\bar \mu > 1$ for paramagnetic
substances and $\bar \mu < 1$ for diamagnetic).
Choosing the string along $x^1$ direction and using comoving
coordinates we have the following components of energy momentum
tensor \cite{ass}:
\begin{eqnarray}
T_{0}^{0} + \rho_{string} = T_{1}^{1} + \lambda = - T_{2}^{2} = -
T_{3}^{3} = \frac{\beta^2}{2}\frac{a_1^2}{V^2}\,. \label{total}
\end{eqnarray}
Taking into account the conservation of the energy-momentum tensor,
i.e., $T_{\mu;\nu}^{\nu} = 0$, after a little manipulation of
\eqref{total} one obtains \cite{SV,SRV}:
\begin{equation}\label{rholambda}
\dot \rho_{string} + \frac{\dot V}{V}\rho_{string} - \frac{\dot
a_1}{a_1}\lambda = 0\,.
\end{equation}
Here we take into account that the conservation law for magnetic
field fulfills identically.
It is customary to assume a relation between $\rho_{string}$ and
$\lambda$ in accordance with the state equations for strings. The
simplest one is a proportionality relation \cite{PSL}:
\begin{equation}\label{rhoalphalambda}
\rho_{string} = \alpha \lambda \,.
\end{equation}
The most usual choices of the constant $\alpha$ are
\begin{equation}\label{alpha}
\alpha =\left \{
\begin{array}{ll}
1 & \quad {\rm geometric\,\,\,string}\\
1 + \omega & \quad \omega \ge 0, \quad p \,\,{\rm string\,\,\,or\,\,\,
Takabayasi\,\,\,string}\\
-1 & \quad {\rm Reddy\,\,\,string}\,.
\end{array}
\right.
\end{equation}
From eq. \eqref{rholambda} with \eqref{rhoalphalambda} we get
\begin{equation}\label{rhostring}
\rho_{string} = R a_1^{\frac{1-\alpha}{\alpha}} a_2^{-1} a_3^{-1}\,,
\end{equation}
with $R$ a constant of integration.
Let us also write the other features of BI metric such as expansion
and shear. The expansion for the BI metric takes the form
\begin{equation}
\vartheta = \frac{\dot a_1}{a_1}+\frac{\dot a_2}{a_2}+ \frac{\dot
a_3}{a_3} = \frac{\dot V}{V}, \label{expan}
\end{equation}
while the nonzero components for the shear tensor read
\begin{equation}
\sigma_i \equiv \sigma_i^i = \frac{\dot a_i}{a_i} -
\frac{1}{3}\vartheta.
\label{sigcom}
\end{equation}
In \eqref{sigcom} and henceforth there is no summation over repeated
index "$i$". The shear energy density in given by
\begin{eqnarray}\label{shed}
\Sigma^2 = \frac{1}{2}\sigma_{\mu\nu}\sigma^{\mu\nu} =
\frac{1}{6}\left( (H_1-H_2)^2 + (H_2-H_3)^2 + (H_3-H_1)^2 \right)\,.
\end{eqnarray}
\section{Effective loop quantum dynamics}
In the loop quantum cosmology approach we shall use a Hamiltonian
framework where the degrees of freedom of the Bianchi type-I model are
encoded in the triad components $p_i$ and momentum components $c_i$
as follows:
\begin{equation}\label{pc}
p_1 = a_2 a_3, \quad p_2 = a_1 a_3, \quad p_3 = a_1 a_2, \quad c_i
= \gamma \dot{a_i} \,.
\end{equation}
Here $\gamma$ is the Barbero-Immirzi parameter and represents a quantum
ambiguity of loop quantum gravity which is a non-negative real valued
parameter.
In terms of these variables, the total Hamiltonian of the model is
\begin{eqnarray}\label{htot}
\mathcal{H} &=& \mathcal{H}_{grav} + \mathcal{H}_{matter} \nonumber\\
&=& \frac{-1}{\kappa \gamma^2 \sqrt{p_1 p_2 p_3} }
(c_2 c_3 p_2 p_3 + c_1 c_3 p_1 p_3 + c_1 c_2 p_1 p_2)
+ \sqrt{p_1 p_2 p_3} \rho_M \,,
\end{eqnarray}
where $\rho_M$ is the matter energy density. In our model $\rho_M$
comprises the contribution of cosmological string density
$\rho_{string}$ given by \eqref{rhostring}:
\begin{equation}
\rho_{string} = R p_1^{- \frac{\alpha + 1}{2\alpha}} (p_2
p_3)^{\frac{1 - \alpha}{2 \alpha}}, \label{rhostringlqc}
\end{equation}
and the energy density of the magnetic field \eqref{E} \cite{MV}
\begin{equation}
\rho_{mag} = \frac{1}{2} \frac{\beta^2}{(a_2 a_3)^2} = \frac{1}{2}
\frac{\beta^2}{p_1^2}\,. \label{rhomaglqc}
\end{equation}
Einstein's equations are derived from Hamilton's equations:
\begin{equation}\label{ham}
\dot{p_i}= \{p_i,\mathcal{H}\} = -\kappa \gamma \frac{\partial
\mathcal{H}}{\partial c_i}, \quad \dot{c_i}= \{c_i,\mathcal{H}\} =
\kappa \gamma \frac{\partial \mathcal{H}}{\partial p_i}\,.
\end{equation}
On the other hand, the total Hamiltonian $\mathcal{H}$ is of constrained type whereby it
vanishes identically for solutions of Einstein's equations
\begin{equation}\label{h0}
\mathcal{H} = 0 \,.
\end{equation}
Using the explicit form of the Hamiltonian $\mathcal{H}$ we have for
$p_1$ the following equation
\begin{equation}\label{p1}
\frac{d p_1}{d t} = \frac{p_1}{\gamma \sqrt{p_1 p_2 p_3} }
(c_2 p_2 + c_3 p _3)\,,
\end{equation}
and similar equations for $p_2$ and $p_3$. For the evolution of $c_i$
we get:
\begin{subequations}
\label{ci}
\begin{eqnarray}
\frac{d c_1}{d t} =&-&\frac{c_1}{\gamma \sqrt{p_1 p_2 p_3} }
(c_2 p_2 + c_3 p _3)\nonumber\\
&+& \frac{1}{2 \gamma p_1 \sqrt{p_1 p_2 p_3} }
(c_2 c_3 p_2 p_3 + c_1 c_3 p_1 p_3 + c_1 c_2 p_1 p_2) \nonumber\\
&+& \frac{\kappa \gamma}{p_1} \left[-\frac{1}{2\alpha} R
\left(\frac{p_2 p_3}{p_1}\right)^{\frac{1}{2\alpha}}
-\frac{3}{4} \beta^2 \left(\frac{p_2 p_3}{p_1^3}\right)^{\frac{1}{2}}
\right] \,,\label{c1}
\end{eqnarray}
\begin{eqnarray}
\frac{d c_2}{d t} =&-&\frac{c_2}{\gamma \sqrt{p_1 p_2 p_3} }
(c_1 p_1 + c_3 p _3)\nonumber\\
&+& \frac{1}{2 \gamma p_2 \sqrt{p_1 p_2 p_3} }
(c_2 c_3 p_2 p_3 + c_1 c_3 p_1 p_3 + c_1 c_2 p_1 p_2) \nonumber\\
&+& \frac{\kappa \gamma}{p_2} \left[\frac{1}{2\alpha} R
\left(\frac{p_2 p_3}{p_1}\right)^{\frac{1}{2\alpha}} +
\frac{1}{4} \beta^2 \left(\frac{p_2 p_3}{p_1^3}\right)^{\frac{1}{2}}
\right] \,,\label{c2}
\end{eqnarray}
\begin{eqnarray}
\frac{d c_3}{d t} =&-&\frac{c_3}{\gamma \sqrt{p_1 p_2 p_3} }
(c_1 p_1 + c_2 p _2)\nonumber\\
&+& \frac{1}{2 \gamma p_3 \sqrt{p_1 p_2 p_3} }
(c_2 c_3 p_2 p_3 + c_1 c_3 p_1 p_3 + c_1 c_2 p_1 p_2) \nonumber\\
&+& \frac{\kappa \gamma}{p_3} \left[\frac{1}{2\alpha} R
\left(\frac{p_2 p_3}{p_1}\right)^{\frac{1}{2\alpha}} +
\frac{1}{4} \beta^2 \left(\frac{p_2 p_3}{p_1^3}\right)^{\frac{1}{2}}
\right] \,.\label{c3}
\end{eqnarray}
\end{subequations}
Let us observe that from equations for $p_i$ and $c_i$ we have
the following relation:
\begin{equation}
\frac{d}{d t} (p_i c_i) = \kappa \gamma \sqrt{p_1 p_2 p_3}
\left (\frac{1}{2} \rho_M + p_i \frac{\partial
\rho_M}{\partial p_i}\right )\,.
\end{equation}
The directional Hubble rates now reads
\begin{equation}
H_i = \frac{\dot{a_i}}{a_i} = \frac{\sqrt{p_i} c_i}{\gamma
\sqrt{p_j p_k}}\,, \quad i \ne j \ne k = 1,\,2,\,3. \label{DHlqc}
\end{equation}
Hamiltonian \eqref{htot} on account of the vanishing condition
\eqref{h0} and \eqref{DHlqc} leads to
\begin{equation}
H_1 H_2 + H_1 H_3 + H_2 H_3 = \kappa \rho_M,
\end{equation}
which is in fact the zero-zero component \eqref{00} of the Einstein
system of equations \eqref{BID}.
Taking into account the symmetry of the density $\rho_M$ with respect
to variables $p_2$ and $p_3$ we have
\begin{equation}\label{23}
\frac{d}{dt} (p_2 c_2 - p_3 c_3) = 0\,.
\end{equation}
This means that for the directional Hubble parameters $H_2\,,H_3$ we
have
\begin{equation}\label{H23}
H_2 - H_3 = \frac {\alpha_{23}}{a_1 a_2 a_3} = \frac {\alpha_{23}}
{\sqrt{p_1 p_2 p_3}}\,,
\end{equation}
with $\alpha_{23}$ a constant.
The quantum effects in loop quantum cosmology arise in the {\it
effective} Hamiltonian constructed from the classical one by replacing
the classical $c_i$ terms with sine functions:
\begin{equation}\label{csin}
c_i \longrightarrow \frac{\sin (\bar \mu_i' c_i)}{\bar \mu_i'}\,.
\end{equation}
where $\bar \mu_i$ are real valued functions of the triad coefficients
$p_i$.
The effective Hamiltonian is given by:
\begin{equation}\label{heff}
\mathcal{H}_{eff} =
\frac{-1}{\kappa \gamma^2 \sqrt{p_1 p_2 p_3}}
\left\{\frac{\sin ({\bar\mu_2'} c_2)
\sin ({\bar \mu_3'}c_3)}{{\bar\mu_2'}{\bar \mu_3'}} p_2 p_3 +
\text{cyclic~terms} \right \} + \sqrt{p_1 p_2 p_3} \rho_M \,,
\end{equation}
and the Hamilton's equations for $p_i$ and $c_i$ will be modified
accordingly.
It is quite evident that in the limit $\bar \mu_i \rightarrow 0$, the
classical Hamiltonian $\mathcal{H}$ \eqref{htot} is recovered.
The expression of the parameters $\bar \mu_i'$ as functions of the
triad components $p_i$ represent an ambiguity of the quantization. Two
most preferable constructions are discussed in \cite{CV,DWC2,AE}.
In what follows we shall adopt the $\bar \mu'$-scheme in which the
parameters $\bar \mu_i'$ are chosen as follows:
\begin{equation}
\bar \mu_1' = \sqrt{\frac{p_1 \Delta}{p_2 p_3}}\,,
\bar \mu_2' = \sqrt{\frac{p_2 \Delta}{p_1 p_3}}\,,
\bar \mu_3' = \sqrt{\frac{p_3 \Delta}{p_1 p_2}}\,,
\end{equation}
with $\Delta$ a constant related to the minimum area gap in LQG. For
the numerical simulations it is assumed that $\Delta = O(1)$ in Planck
units.
Let us remark that from the vanishing of the Hamiltonian \eqref{h0} we
have the bound:
\begin{equation}\label{bound}
p_1 p_2 p_3 \rho_M \leq \frac{1}{\kappa \gamma^2}
\left\{\frac{p_2 p_3}{\bar\mu_2 \bar\mu_3} +
\frac{p_1 p_3}{\bar\mu_1 \bar\mu_3}+
\frac{p_1 p_2}{\bar\mu_1 \bar\mu_2} \right\}\,.
\end{equation}
In particular, in the $\bar\mu'$ scheme the total density is bounded
by the critical value:
\begin{equation}\label{rhocrit}
\rho_{M\,crit} = 3(\kappa \gamma^2 \Delta)^{-1}\,,
\end{equation}
implying that the classical singularity is never approached. Indeed the
total energy of the matter must be below this value and the classical
collapse is replaced by a bounce.
\section{Numerical simulations and a comparison of the approaches}
The complexity of the equations does not allow for analytic solutions
and imposes numerical simulations. The classical equations of motion
given by \eqref{p1}-\eqref{ci} and the corresponding ones for quantum
effects in LQC approach with the replacement \eqref{csin} can be solved
once the initial values $p_i (t = t_0)$ and $c_i (t = t_0)$ are given.
In what follows we report some numerical studies on the behavior of
$\rho_M$, $V$ and the anisotropic
shear $\sigma_{\mu\nu}$. Taking into account that only diagonal
components of shear tensor are non-zero, in the new variables they
now read
\begin{equation}\label{shear}
\sigma_i =
\frac{\sqrt{p_i} c_i}{\gamma \sqrt{p_j p_k}} -
\frac{1}{3\gamma}\Bigl(\frac{\sqrt{p_1} c_1}{\sqrt{p_2 p_3}} +
\frac{\sqrt{p_2} c_2}{\sqrt{p_3 p_1}} + \frac{\sqrt{p_3}
c_3}{\sqrt{p_1 p_2}}\Bigr)\,, \quad i \ne j \ne k = 1,\,2,\,3\,,
\end{equation}
and the shear energy density is
\begin{eqnarray}\label{shedlqc}
\Sigma^2 = \frac{1}{6 \gamma^2 p_1 p_2 p_3}\bigl[(p_1c_1 - p_2c_2)^2
+ (p_2c_2 - p_3c_3)^2 + (p_3c_3 - p_1c_1)^2\bigr]\,.
\end{eqnarray}
In doing so we considered a number of cases that helps us to clarify
the role of various parameters. To begin with we examined both
positive and negative $\alpha$. In particular we considered the case
with $\alpha = 2$ and $\alpha = -2$ and it was found that the value or
sign of $\alpha$ leaves the overall picture qualitatively unchanged.
As a second consideration we set a large value of $R$ and small
value of $\beta$ and vice versa, for example, $R =
18.24,\,\beta = 1$ and $R = 0.90,\, \beta = 6.5$, respectively .
It was established that for both cases the overall picture remains
qualitatively unaltered. Finally we consider the case setting
different initial values for $c_i = c_0$.
In Figs. \ref{fig1}, \ref{fig3}, \ref{fig5}, \ref{fig7} we have
illustrated the evolution of volume scale $V$ (red solid line),
energy density $\rho_M$ (blue dash line) and shear energy $\Sigma^2$
(black dot line), whereas in Figs. \ref{fig2}, \ref{fig4},
\ref{fig6}, \ref{fig8} we have plotted the evolution of the
components of shear tenor $\sigma_1$ (red solid line), $\sigma_2$
(blue dash line) and $\sigma_3$ (black dot line).
In the classical case the initial condition $c_0 <0$ leads to a
collapsing Universe [cf. Figs. \ref{fig1}, \ref{fig2}], while
$c_0 > 0$ gives rise to an expanding one [cf. Figs. \ref{fig3},
\ref{fig4}].
In the LQC approach, even for $c_0 <0$ the energy density remains
bounded below the critical energy \eqref{rhocrit} as expected from
analytical considerations [cf. Fig. \ref{fig5}]. After the bounce, the
regime is an expanding one and the Universe isotropizes [cf. Fig.
\ref{fig6}]. On the other hand, for $c_0 > 0$ we have only expanding
phase of the Universe, the shear energy density $\Sigma^2$ remains
finite [cf. Fig. \ref{fig7}] and again $\sigma_i$ tend to zero
[cf. Fig. \ref{fig8}].
\myfigures{fig1}{0.4} {Classical: Evolution of $V$, $\rho_M$ and
$\Sigma^2$ for $\alpha=-2$, $R=18.24$, $\beta=1$, $c_0 = -1$. Here
and further the red solid line corresponds to $V$, blue dash line
to $\rho_M$ and black dot line to $\Sigma^2$.}{0.4}
{fig2}{0.4} {Classical: Evolution of $\sigma_i$'s
for $\alpha=-2$, $R=18.24$, $\beta=1$, $c_0 = -1$. Here and further
the red solid line corresponds to $\sigma_1$, blue dash line
to $\sigma_2$ and black dot line to $\sigma_3$.}{0.4}
\myfigures{fig3}{0.4} {Classical: Evolution of $V$, $\rho_M$ and
$\Sigma^2$ for $\alpha=-2$, $R=18.24$, $\beta=1$, $c_0 = 1$.}{0.4}
{fig4}{0.4} {Classical: Evolution of $\sigma_i$'s for $\alpha=-2$,
$R=18.24$, $\beta=1$, $c_0 = 1$.}{0.4}
\myfigures{fig5}{0.4} {LQC: Evolution of $V$, $\rho_M$ and
$\Sigma^2$ for $\alpha=-2$, $R=6.20$, $\beta=1$, $c_0 = -1$.}{0.4}
{fig6}{0.4} {LQC: Evolution of $\sigma_i$'s for $\alpha=-2$,
$R=6.20$, $\beta=1$, $c_0 = -1$.}{0.4}
\myfigures{fig7}{0.4} {LQC: Evolution of $V$, $\rho_M$ and
$\Sigma^2$ for $\alpha=-2$, $R=R=6.20$, $\beta=1$, $c_0 = 1$.}{0.4}
{fig8}{0.4} {LQC: Evolution of $\sigma_i$'s for $\alpha=-2$,
$R=6.20$, $\beta=1$, $c_0 = 1$.}{0.4}
\section{Conclusions}
In this report within the scope of Bianchi type-I cosmological model
we investigate the role of cosmic string and magnetic field on the
evolution of the Universe. In doing so we employ both classical and
LQC approaches. It is found that the qualitative picture of
evolution does not depend on the cosmic string ($\alpha$), though
the value of $\alpha$ leads to quantitative changes. On the other
hand magnetic field together with cosmic string (given by the pair
$\beta$ and $R$) also leave the qualitative feature unaltered. Only
initial value of $c_i$'s, i.e., initial rate of change of the metric
functions $a_i$'s play essential role.
In the classical approach the initial condition $c_0 < 0$ corresponds
to a classically collapsing Universe, while $c_0 > 0$ is associated
with an expansion. In the LQC approach for $c_0 < 0$ the singularity is
avoided via a bounce. After the bounce the Universe enters an expansion
phase with an asymptotic isotropization. For positive $c_0$ we get
always expansion and isotropization.
At the classical level other Bianchi models have richer phenomenology
than the cosmological BI model. From this point of view the extension of
the string cosmological model in the presence of electromagnetic fields
to other types of anisotropies and comparisons between the classical
approach and LQC one deserve further studies.
\subsection*{Acknowledgments}
This work is supported in part by a joint Romanian-LIT, JINR, Dubna
Research Project, theme no. 05-6-1060-2005/2013. M.V. is partially
supported by program PN-II-ID-PCE-2011-3-0137, Romania.
|
\section{A Brief History of Star Formation in M32}
M32 (NGC 221) is a compact, low-luminosity elliptical galaxy,
satellite of our neighbor M31. Due to its proximity, we can study M32
with great detail not only from its integrated light but also from its
individual, resolved stars in a way that is impossible for most of the
elliptical galaxies, given their greater distances and high
densities. Thus, M32 is a very important galaxy to understand the
formation and evolution of low-luminosity spheroidal star
systems. However, M32's SFH, and therefore its origin, is still
controversial. The different scenarios proposed to explain its origins
extend from a true elliptical galaxy at the lower extreme of the mass
sequence \citep[e.g.,][] {Faber73, Nieto_prugniel87, Kormendy_etal09}
to a threshed spiral galaxy \citep[e.g.,][]{Bekki_etal01,
Chilingarian_etal09}.
The only way to accurately determine the age, and thus the SFH, of a
galaxy is by directly observing its oldest main-sequence turnoff
(MSTO). With this goal in mind, we were awarded 64 orbits with HST
ACS/HRC to observe two fields near M32, F1 and F2
(Fig.~\ref{fig:location}), in order to detect the oldest MSTOs of this
galaxy.
\begin{figure} \centering
\includegraphics[width=80mm,clip]{fig1.jpg}
\caption{Location of our two HST ACS/HRC pointings: M32 (F1) field
and M31 background (F2) field, indicated as small black boxes.
Each field covers a region of $26 \times 29 \,\mathrm{arcsec^{2}}$
on the sky. The field F1 is located at 110\arcsec\ from the
nucleus of M32 and represents the best compromise between
minimizing image crowding and contamination from M31. The F2
field is at the same isophotal level in M31 as F1. At the distance
of M32, each field occupies an area of 11752 pc$^2$. Thirty-two
exposures in each of the $F435W$ ($B$) and $F555W$ ($V$) filters
were taken for each field. North is up and East is to the left.}
\label{fig:location}
\end{figure}
\subsection{The deepest HST CMD of M32}
In \citet[][hereafter Paper I]{Monachesi_etal11} we introduced our
observations and presented the deepest HST color-magnitude diagram
(CMD) of M32 yet obtained, reaching more than 2 mag fainter than the
RC and fully resolving the RGB and the AGB. Paper I significantly
improved our knowledge on the stellar populations of M32. We have
found that M32 is dominated by intermediate-age (2--8 Gyr old) and
old (8--10 Gyr old), metal-rich ($[\mathrm{Fe/H}] \sim -0.2$) stars
and it contains some ancient ($>10$ Gyr), metal-poor stars
($[\mathrm{Fe/H}] \sim -1.6$) as well as possible young populations
(0.5 -- 2 Gyr old stars).
These conclusions were provided by our qualitative analysis of the
CMD of M32, which shows a red clump (RC), a red giant branch (RGB), a
RGB bump (RGBb), an AGB bump (AGBb), and a blue plume (BP). Figure 12
of Paper I, reproduced here as Figure~\ref{fig:hess_boxes}, shows a
Hess representation of the CMD of M32 decontaminated of M31 stars,
where the different evolutionary features are highlighted. We
summarize here the main findings and conclusions of Paper I.
\begin{figure*}
\centering
\includegraphics[width=130mm, clip]{fig2.jpg}
\caption{Error-based Hess diagram for M32, corrected for
contamination from M31 background stars. The boxes indicate various
features that represent different stellar populations. MS: Main
Sequence; BP: Blue Plume; SGB: Subgiant branch; BHB: Blue
Horizontal Branch; BL: Blue Loop; RC: Red Clump; RGBb: Red Giant
Branch bump; R-RGB: Red-Red Giant Branch; B-RGB: Blue-Red Giant
Branch; TRGB: Tip of the Red Giant Branch; AGB: Asymptotic Giant
Branch; and AGBb: Asymptotic Giant Branch bump. The dotted-dashed
line indicates the 50\% completeness level of our data. Magnitudes
are calibrated onto the VEGAmag system. (This is Fig.~12 from Paper
I; we refer the reader to that paper for more details.)}
\label{fig:hess_boxes}
\end{figure*}
\begin{itemize}
\item The core-helium burning stars are concentrated in a RC and its
mean color and magnitude suggest a mean age of 8--10 Gyr for a
metallicity of $[\mathrm{M/H}]\sim -0.2$ in M32.
\item The first detection of the RGB bump and the AGB bump in M32
permits a constraint on the mean age and metallicity of the
population. This gives a mean metallicity of M32 higher than
$\mathrm{[M/H]} \sim -0.4$ dex and a mean age between 5 and 10 Gyr.
\item The metallicity distribution of M32 inferred from the CMD has a
peak at $[\mathrm{M/H}]\sim -0.2 \, \mathrm{dex}$. Overall, the
metallicity distribution function implies that there are more
metal-rich stars than metal-poor ones. Metal-poor stars with
$[\mathrm{M/H}] <-1.2$ contribute very little, \emph{at most} 6\% of
the total $V$-light or 4.5\% of the total mass, to M32 in F1,
implying that the enrichment process largely avoided the metal-poor
stage.
\item Bright AGB stars at $F555W < 24$, i.e. above the TRGB, confirm
the presence of an intermediate-age population in M32 (ages of 1--7
Gyr).
\item The observed blue plume is genuine, not an artifact of crowding,
and contains stars as young as $\sim0.5$ Gyr. The detected blue
loop, with stars having masses of $\sim 2$--$3 \,M_{\odot}$ and ages
between $\sim 0.3$ and $\sim 1$ Gyr, and the possible presence of a
bright SGB are different manifestations of the presence of a young
population. However, in Paper I we suggest that it is likely that
this young population belongs to the disk of M31 rather than to
M32. The fainter portion of the blue plume ($F555W > 26$)
\emph{does} belong to M32 and indicates the presence of stars with
ages 1--2 Gyr and/or the first direct evidence of blue straggler
stars (BSS) in M32.
\item The oldest MSTOs were out of reach, given the severe crowding in
F1, and there is no significant BHB observed in F1, so an ancient,
metal-poor population cannot be seen directly in our CMD. We have,
however, a hint of the presence of such a population from a
2--$\sigma$ detection of RR Lyrae stars found in F1 and associated
with M32 using our data \citep{Fiorentino_etal10}.
\item In general the CMDs of both fields F1 and F2 show an
unexpectedly similar morphology. By subtracting the normalized F1
CMD from the F2 one (see Figure 21 in Paper I), one can detect
subtle differences. M31 has a younger and more metal-poor population
than M32, and M32 has a more conspicuous intermediate-age population
(Fig.~\ref{fig:deconvolvedmcmd}).
\item The CMD of our M31 background field F2 exhibits a wide RGB,
indicative of a metallicity spread with its peak at
$[\mathrm{M/H}]\sim -0.4 \, \mathrm{dex}$. The presence of a blue
plume indicates the presence of stars as young as 0.3 Gyr. Bright
AGB stars in F2 reveal the presence of an intermediate-age
population in M31.
\end{itemize}
\subsection{Completing the picture of M32's SFH}
The analysis presented in Paper I provided initial constraints on the
ages and metallicities of the stellar populations of M32 at F1 and
M31 at F2. That work was based on traditional methods of isochrone
analysis, and was heavily based on the brighter evolved portions of
the CMDs, such as the RC, RGB, and bump (RGBb and AGBb) features.
The approach in this paper is independent. Here we use a
sophisticated method of CMD analysis and decomposition that digs into
the fainter, severely-crowded portions of the CMDs near the MSTO and
sub-giant branches. We recover information from the brighter MSTOs
present in the CMDs, thus providing quantitative information about
the younger populations of M32. In this paper, we derive the
detailed young and intermediate-age SFH of M32 at $\sim 2\arcmin$
from its center and of M31 at our background field's location, which
was not possible from the analysis in Paper I. We find that our field
in M32 has a substantial population of 2--5 Gyr old stars
contributing to $\sim 40\% \pm 17\%$ of its mass, an unexpectedly
large population of young stars at such a large distance from the
center of an elliptical galaxy.
The paper is organized as follows. In Section~\ref{sec:photo} we
briefly describe our observations and
photometry. Section~\ref{sec:method} describes the method used to
derive the SFH. We present the results of the SFH analysis obtained
for F1, F2 and M32 in Section~\ref{sec:results}. In
Section~\ref{sec:sfhm32} we provide a detailed and complete SFH of
M32 and discuss its implications on M32's origins, synthesizing a
complete picture based on both the present and Paper I analyses. In
Section~\ref{sec:m31} we discuss the SFH of the inner regions of
M31. Finally, we summarize our results and present our conclusions in
Section~\ref{sec:conclusions}.
\section{Observations and Photometry}
\label{sec:photo}
The field selection and observational strategy as well as the image
reduction are described in Paper I and we refer the reader to that
paper for details. Briefly, HST ACS/HRC images of two fields near M32
were observed during Cycle 14 (Program GO-10572, PI: Lauer). The M32
HRC field (F1) was centered on a location $110\arcsec$ south (the
anti-M31 direction) of the M32 nucleus. The background field (F2) was
located $327\arcsec$ from the M32 nucleus, roughly along its minor
axis, at the same isophotal level in M31 as F1. The field locations
are shown in Figure~\ref{fig:location}.
Each field was observed for 16 orbits in each of the $F435W$ ($\sim
B$) and $F555W$ ($\sim V$) filters. All of the images were combined in
an iterative procedure designed to detect and repair cosmic-ray
events, hot pixels, and other defects, with a Nyquist-sampled summed
image as the final product \citep{lauer99a}. Color images of F1 and F2
are shown in the left and right panels of Figure~\ref{fig:fields},
respectively, where the strong crowding in these fields is clearly
visible. There is however a difference between the stellar density in
F1 and F2: the crowding is more severe in F1 than in F2. This can
also be seen from the bottom panels of Figure~\ref{fig:fields}, where
zoomed-in images of the top panels are shown.
\begin{figure*
\vspace{-0.001cm}
\includegraphics[width=185mm,clip]{fig3a.jpg}
\includegraphics[width=185mm,clip]{fig3b.jpg}
\caption{Combined color images of the 32 exposures in the F1 (top left
panel) and F2 (top right panel) fields displayed with the same
logarithmic stretch. Each image has a size of 2048$\times$2048
pixels with a $0\farcs0125$ pixel scale. There is a clear difference
in stellar density between the images, indicating that crowding is
more severe in F1 than in F2. We also note a stellar density
gradient in the F1 image, becoming higher when approaching the
center of M32. The long black spot in the top center of each image
is the occulting finger of the ACS/HRC coronagraph. The bottom
panels are zoomed-in images of the centers of the top images for F1
(left) and F2 (right) fields, where we can better see individual
stars and the different crowding levels are also evident. Note the
blue stars in these images. Each zoomed-in image represents an area
of $\sim15\,\mathrm{arcsec}^2$ on the sky.}
\label{fig:fields}
\end{figure*}
Stellar photometry was performed on deconvolved combined images. A
detailed description of the deconvolution process is explained in
Paper I. In short, deconvolution was performed on the final images
using the Lucy-Richardson algorithm \citep{Lucy_74, Richardson_72} and
empirically constructed PSFs, one for each image. Stars were
identified in the deconvolved images and their fluxes were
measured. Change transfer efficiency (CTE) and aperture corrections
were applied to the magnitudes, which transform the instrumental
magnitudes into calibrated, apparent magnitudes.
Figure~\ref{fig:deconvolvedmcmd} shows the CMDs derived for F1 (left
panel) and F2 (right panel) from the deconvolved photometry,
calibrated onto the VEGAmag system. They contain 58143 and 27963
stars, respectively, as indicated in Table~\ref{table:photometry}. A
qualitative analysis of these CMDs allowed us to gain some insights
into its stellar populations. This was discussed in detail in Paper I
and we have summarized our conclusions above.
Note the difference between the CMD of F1 and F2 at magnitudes between
$F555W \sim 27$ and 28 (cyan boxes in
Figure~\ref{fig:deconvolvedmcmd}) . The number of stars in this
region, where the brighter MSTOs are located, is larger in F1 than
F2. This suggests that there is a bigger contribution of
intermediate-age stars in F1 than in F2. We can better appreciate this
difference in Figure 21 of Paper I, where we showed a Hess subtraction
of the normalized F1 CMD to the F2 CMD.
\begin{deluxetable}{lcccccc}
\tabletypesize{\scriptsize}
\tablecaption{Deconvolved photometry\label{table:photometry}}
\tablewidth{0pt}
\tablehead{\colhead{Field}&
\colhead{Detections\tablenotemark{a}}&
\colhead{$R^{F435W}_{\rm{PSF}}$\tablenotemark{b}}&
\colhead{$R^{F555W}_{\rm{PSF}}$\tablenotemark{b}}&
\colhead{AC$_{F435W}$\tablenotemark{c}}&
\colhead{AC$_{F555W}$\tablenotemark{c}}}
\startdata
F1&58,143&5&5&$-0.25$&$-0.22$ \\
F2&27,963&6&16&$-0.22$&$-0.10$
\enddata
\tablenotetext{a}{Final number of stars detected and used to derive
CMDs}
\tablenotetext{b}{PSF radius in HRC original pixels}
\tablenotetext{c}{Aperture correction}
\end{deluxetable}
\subsection{Crowding tests}\label{ast}
We performed artificial star tests (ASTs) to assess the completeness
level and quantify the photometric errors of our data. This is a
crucial step for the derivation of the SFH. The distribution of stars
in the observed CMD is modified from the actual distribution due to
the observational effects, particularly at the fainter magnitudes
where most of the information from the older star formation is
encoded. The ASTs are used to simulate the observational effects in
the synthetic CMDs that are then compared with the observed CMDs in
the analysis described below.
The procedure and results of the ASTs are presented in Paper I and we
refer to that paper for further details; we give a brief description
here in order to provide guidance for later sections of this paper. We
used IAC-STAR \citep{Aparicio_gallart04} to generate $5\times10^5$
artificial stars with realistic colors and magnitudes covering not
only the entire color and magnitude range of the observed stars
but also $\sim 2$ magnitudes fainter. We injected the artificial
stars into the real images after transforming their magnitudes into
instrumental ACS/HRC fluxes. The number of stars injected per
experiment is 2000, to avoid increasing the already severe crowding of
the real images. We performed 250 ASTs per field/filter combination
for a total of 1000 ASTs. The images containing real and artificial
stars are photometered exactly in the same way as the original images.
A comparison of the known injected magnitudes and colors of the
artificial stars to those obtained from their photometry allows us to
quantify the photometric errors. The completeness of our data at a
given color and magnitude is calculated as the ratio of
recovered-to-injected artificial stars on that color and magnitude
bin.
The results obtained from these ASTs indicate that the limiting
magnitudes of the F1 and F2 CMDs are $F555W \sim 28$ and $\sim 28.5$,
respectively, nearly independent of color. The CMD of F2 is therefore
slightly deeper than that of F1 (cf.\ Figs.\ 8 and 9 of Paper I). The
50\% completeness level as well as the photometric errors derived from
the ASTs for F1 and F2 are indicated in
Figure~\ref{fig:deconvolvedmcmd}.
\begin{figure*}\centering
\includegraphics[width=185mm, clip]{fig4.jpg}
\caption{($F435W - F555W$, $F555W$) CMDs of field F1 (left-hand
panel) and F2 (right-hand panel) obtained using deconvolved images.
These contain 58143 and 27963 stars respectively, and are
calibrated onto the VEGAmag HST system. Note the difference between
the CMDs in the region highlighted with dark blue boxes. The larger
number of stars in F1 indicate the presence of a more significant
intermediate-age population in this field compared to F2. The
region in the box is not an actual ``bundle'' used in the
derivation of the SFH but a similar one is used to obtain most of
the information about the SFH of both fields: see
Section~\ref{sec:method} for more details. The light blue line
indicates the 50\% completeness level of our data in each field and
the photometric errors from ASTs refer to a color of
$(F435W-F555W)=1$.}
\label{fig:deconvolvedmcmd}
\end{figure*}
\section{The IAC Method to resolve the SFH}
\label{sec:method}
To extract the detailed SFH of F1 and F2 we use the well-known method
of fitting synthetic CMDs to the data \citep[see e.g.,][]{Tosi_etal91,
Bertelli_etal92, Tolstoy_saha96, Aparicio_etal97}. There are
currently many approaches to derive detailed SFH of galaxies (e.g.,
StarFISH: \citealt{Harris_zaritsky01}, MATCH: \citealt{Dolphin02},
IAC-pop/MinnIAC: \citealt{Aparicio_hidalgo09, Hidalgo_etal11}) as well
as several different stellar libraries (e.g., BaSTI:
\citealt{Pietrinferni_etal04}, Padova/Girardi:
\citealt{Girardi_etal00, Marigo_etal08}) available to compute the
required synthetic CMDs. We use the IAC-pop/MinnIAC method and adopt
the BaSTI and Padova stellar libraries. The IAC-pop code
\citep{Aparicio_hidalgo09} uses a modified $\chi^2$ merit-function
\citep{Mighell99} to compare the observed and synthetic star counts in
different boxes (see below) of the CMDs. A genetic algorithm
\citep{Charbonneau95} is adopted to minimize $\chi^2$. An important
characteristic of the code is that it solves the SFH simultaneously
for age and metallicity distributions. It thus provides the SFH of a
stellar system as a linear combination of simple populations,
i.e. within small ranges of age and metallicity. We refer the reader
to \citet{Aparicio_hidalgo09} and \citet{Hidalgo_etal11} for more
details.
It is important to emphasize that, for the current analysis, we have
mainly used information from the extended MS, MSTO and SGB regions of
the CMDs, as we will see below. We have excluded the RC and most of
the RGB regions, which were the main features analyzed in Paper I and
from which we obtained estimates on the age and metallicity of
M32. This is because the physics governing these phases are more
uncertain than those on the MS and SGB, and differences between
stellar libraries are more severe \citep{Gallart_etal05}. For
instance, the morphology and number of stars occupying the HB/RC
evolutionary phases depend on unknown issues, like mass loss on the
RGB or He-core mass. Small differences in the adopted physics can
significantly alter the number of stars and morphology of these CMD
regions. The CMD regions that we probe in this paper allow us to
obtain detailed information about the young and intermediate-age
populations of M32, something that was not possible in Paper I, but
conversely, we cannot make a quantitative analysis of the older
populations and so we must rely on the qualitative results of Paper I.
\begin{figure*}
\includegraphics[width=90mm, clip]{fig5a.jpg}
\includegraphics[width=90mm, clip]{fig5b.jpg}
\caption{Left-hand panel: Hess representation with a logarithmic
stretch of the synthetic CMD generated using IAC-STAR for a range
of age between 0 and 14 Gyr and metallicities uniformly distributed
at all ages between 0.0001 and 0.04. It contains $5\times10^6$
stars. Right-hand panel: Hess representation of the model CMD,
i.e. the synthetic CMD after the observational effects have been
simulated. It contains $\sim 2\times10^6$ stars. The color bar
indicates the number of stars per color-magnitude bin in
logarithmic scale. This model CMD is the one to be compared with
the observed CMD to derive the SFH.}
\label{fig:scmd}
\end{figure*}
\subsection{Steps carried out to obtain the SFH }
1) \emph {Synthetic CMD}. We first generate a synthetic CMD using
IAC-STAR \citep{Aparicio_gallart04}. The bolometric corrections
applied to both libraries are those of \citet{Origlia_leitherer00}
which transform the theoretical tracks into the ACS/HRC photometric
system. We assume a constant star formation rate (SFR) from 0 to 14
Gyr, and metallicities from $Z = 0.0001$ ($[\mathrm{M/H}] = -2.3$) to
$Z = 0.04$ ($[\mathrm{M/H}] = 0.3$) uniformly distributed at all
ages. Note that there is no assumed age--metallicity relation as
input, and the selected age and $[\mathrm{M/H}]$ ranges are broader
than those expected for the solution. This allows the code to find the
SFH solution with minimum constraints and ensures no lost
information. We adopted a \citet{Kroupa02} initial mass function
(IMF)\footnote{Given that the range of masses we have in our observed
CMD is rather small and centered around $1\,M_\odot$, where the IMF
is not especially sensitive to changes, we do not expect the
effective SFH solution to significantly change if we assume another
IMF. However, since the large number of lower mass stars are not
well constrained by the data, different assumptions of the IMF will
affect the normalization of the SFH, i.e. the total mass of the
system.} from 0.1 to $100\,M_{\odot}$. The IMF has a slope of 1.3
for stars with masses lower than $0.5\,M_{\odot}$ and 2.3 for stars
with higher masses. We assume a 35\% binary fraction with a relative
mass ratio randomly distributed between 0.5 and 1 (the impact of
different binary fractions on the solution is discussed in the
Appendix). The synthesized CMD, shown in the left panel of
Figure~\ref{fig:scmd}, contains $5\times10^6$ stars and its faintest
magnitude is $\sim$ 2 magnitudes fainter than the 50\% completeness
level of our data. Observational effects (incompleteness and
photometric errors) are simulated using information obtained from the
ASTs described above \citep[see][and references therein for a detailed
description of this procedure]{Hidalgo_etal11}. The right panel of
Figure~\ref{fig:scmd} shows the synthetic CMD after observational
effects are simulated. We call it a ``model CMD'' following
\citet{Aparicio_etal97}'s notation. The model CMD is the one to be
compared with the observed CMD for the derivation of the SFH of our
fields.
2) \emph{Parametrization of the CMDs}. This is the main input of the
IAC-pop code and was performed using MinnIAC \citep{Hidalgo_etal11}, a
set of routines specially designed for this purpose. We first define
the ``simple populations'', the age and metallicity bins in which the
model CMD is to be divided. These simple populations represent the
bins in which the SFH is to be determined. The boundaries of the bins
that we used are [0, 0.5, 1, 2, 5, 14]$\times10^9$ years in age and
[0.02, 0.40, 0.80, 1.00, 2.00, 4.00]$\times10^{-2}$ in $Z$,
corresponding to $[\mathrm{M/H}]\approx[-2.0$, $-0.67$, $-0.35$,
$-0.25$, $0.02$, $0.32$], assuming $Z_{\odot}=0.019$. These
constitute $5\times5 = 25$ simple populations. The resolution in age
and metallicity was selected, after several experiments on mock
stellar populations, as the optimal choice for our data given the
observational uncertainties\footnote{Experiments on mock stellar
populations were conducted as follows: we generated synthetic CMDs
from arbitrary SFHs to which we apply the observational effects of
our observed CMDs obtained from the ASTs. We then solved for their
SFHs using the IAC-POP/MinnIAC method as if they were real data,
adopting different age and metallicity resolutions as first
reasonable guesses according to the limiting magnitude of our
data. The SFH solutions were compared with the input SFHs, and when
these agreed to within $1-\sigma$, we assumed the corresponding age
and metallicity resolutions were optimal.}. Note that the bin width
in age increases significantly for older populations. This is due to
the limits imposed by the crowding; we cannot extract more detailed
information about the oldest stars.
\begin{figure*}\centering
\includegraphics[width=80mm, clip]{fig6a.jpg}
\includegraphics[width=80mm, clip]{fig6b.jpg}
\caption{Left-hand panel: CMD of field F1 in absolute magnitudes,
assuming a distance $\mu_0=24.53$ and $E(B-V) = 0.08$, with the
location of the bundles superimposed. Right-hand panel: As in the
left-hand panel but for F2, assuming a distance $\mu_0=24.45$ and
$E(B-V) = 0.08$. Each bundle in both CMDs is subdivided into boxes
with sizes that vary from one bundle to another (see
Table~\ref{table:bundles}). This allows each CMD region used for
the analysis to have different weights on the extracted SFH. Note
that no bundles were added below the 50\% completeness level of
each CMD. Also note that most of the RGB and RC regions are
likewise excluded of our SFH analysis: uncertainties in the physics
governing these evolutionary phases are larger than those in the MS
and SGB region.}
\label{fig:bundles}
\end{figure*}
We then define five ``bundles'', macro-regions of the CMDs used for
the fitting. We show the CMDs of F1 and F2 with the selected bundles
superimposed in Figure~\ref{fig:bundles}. The bundles are subdivided
into boxes, whose sizes vary from bundle to bundle. The bundles and
boxes are equally sampled in the observed and model CMDs. Since the
number of stars in each box is the information provided to the IAC-pop
code, the different bundle subdivisions provide the weights that a
given CMD region has on the derived SFH. For instance, CMD regions
well-populated and/or where the input physics is better understood
(e.g., bundle 1) have smaller boxes than CMD regions where either the
number of stars is smaller or the uncertainties in the input physics
significantly impact stellar interior models (e.g., bundle 5). The
properties of the boxes for each bundle are specified in
Table~\ref{table:bundles} and both the observed and model CMDs with
one sample of the boxes superimposed are shown in
Figure~\ref{fig:boxes}. Note that the boxes inside the bundles are
shifted during the dithering process, as explained below, and only
stars inside a bundle are considered in the analysis, no matter how
big the box is. Also note that only stars brighter than the 50\%
completeness level were used to extract the SFH
(cf.~Fig.~\ref{fig:bundles}). Below this region, most of the
information is lost and results obtained from lower-completeness
regions are unreliable (see also the right panel of
Fig.~\ref{fig:scmd}). Also, as mentioned above, we did not use most of
the RGB and RC. Adding bundles in those regions significantly
increased $\chi^2$ from $\sim 2$ to $\sim 5$. Bundles 4 (bluer than
the observed MS) and 5 (redder than the observed RGB) were adopted to
mainly constrain the lowest and highest metallicity, respectively, in
the observed CMD. There are nearly no observed stars in these regions
(see Fig.~\ref{fig:bundles}), whereas there are stars in the model CMD
(see right panel of Fig.~\ref{fig:scmd}). The fact that stars of
certain ages and metallicities appear in the model CMD but not in the
observed one, indicates that those simple populations are not present
in M32.
\begin{figure*
\includegraphics[width=180mm]{fig7.jpg}
\caption{CMD of field F1 (left-hand panel) and model CMD (right-hand
panel) with the actual bundles and boxes used for the fitting
superimposed. We show here one sample of boxes in which the stars
are counted and compared between the observed and model CMDs. Note
that the boxes inside the bundles are shifted during the dithering
process, as explained in the text. Also note that only stars inside
a bundle are considered in the analysis, no matter how big the box
is.}
\label{fig:boxes}
\end{figure*}
3) \emph{Solution}. For a given parametrization, i.e., box sizes and
simple population boundaries, MinnIAC counts the stars in each of the
boxes for both the observed and model CMDs. The number of stars in
each box is the input information to run IAC-pop. IAC-pop compares the
observed and model star counts in each box using a modified $\chi^2$
merit-function \citep{Mighell99}, calculating which combination of
simple populations best reproduces the observed CMD. A SFH solution is
obtained as a linear combination of the simple populations. Thus,
IAC-pop solves the SFH considering the age and metallicity as
independent variables.
\begin{deluxetable}{lcccc}
\tabletypesize{\scriptsize}
\tablecaption{CMD regions used for the fitting\label{table:bundles}}
\tablewidth{0pt}
\tablehead{\colhead{Bundle}&
\colhead{\# of boxes}&
\colhead{Size of boxes (color, mag)} &
\colhead{CMD region sampled}}
\startdata
1 & 500 &(0.01, 0.20) &lower MS\\
2 & 150 &(0.03, 0.30) &upper MS\\
3 & ~~3 &(0.50, 0.40) &SGB\\
4 & ~~7 &(0.50, 0.50) &left of the MS\\
5 & ~~5 &(0.50, 0.90) &Right of the RGB
\enddata
\end{deluxetable}
4) \emph{Uncertainties and stability of the solution}. To minimize
biases in the solution due to the sampling of the CMD, MinnIAC allows
slight changes in the input parameters. The simple populations
(i.e. the age and metallicity bins) are shifted three times a 30\% of
their corresponding bin sizes for each of the following four different
configurations: (1) shifting the age bin toward increasing age at
fixed metallicity, (2) shifting the metallicity bin toward increasing
metallicity at fixed age, (3) shifting both age and metallicity bins
toward increasing values, and (4) shifting both age and metallicity
bins toward decreasing age and increasing metallicity. Furthermore,
for each of these 12 shifts, the boxes are shifted a fraction of their
[color, magnitude] sizes three times: [80\%, 20\%], [20\%, 80\%] and
[20\%, 0\%], respectively. These 36 sets of parameters are used to
generate 36 individual solutions. The final SFH solution is the
average of these. This ``dithering'' process significantly reduces
fluctuations in the solution associated with the sampling
\citep{Hidalgo_etal11}. The standard deviation of the ``dithers''
provides a measurement of the uncertainties on the solution
\citep[see][for further discussion of uncertainties in the
solution]{Aparicio_hidalgo09}.
To account for uncertainties in the distance modulus ($\pm 0.14$:
Paper I), reddening ($\pm 0.03$: \citealt{Burstein_heiles82}),
aperture corrections (Paper I), and other systematics possibly
affecting the zero points of our photometry, we allow the observed CMD
(not the model) to shift in both color and magnitude. The observed CMD
is shifted four times in magnitude and six times in color. The bundles
are correspondingly shifted. MinnIAC repeats the entire process of
generating the input information and averages the 36 individual
solutions generated by IAC-pop, for each of the positions in a
magnitude--color grid. The grid has 35 nodes, where the shifts in
magnitude are ($-0.14$, $-0.07$, $0$, $0.07$, $0.14$), and the shifts
in color are ($-0.12$, $-0.09$, $-0.06$, $-0.03$, $0$, $0.03$,
$0.06$). In total we generate $36\times 35 = 1260$ individual
solutions for each field (F1 and F2) and library (BaSTI and
Padova/Girardi) combination.
5) \emph{Final best solution}. After the observed CMD-shifting and
``dithering'' process, we have 35 averaged solutions, one for each
color-magnitude node. Among the 35 mean solutions, the one with the
lowest $\chi^2_{\nu}$ is chosen to be the final solution that best
reproduces our observed CMD.
\subsection{Best SFH solutions for F1 and F2}
\label{sec:bestsfh}
\begin{figure}
\includegraphics[width=40mm, clip]{fig8a.jpg}
\includegraphics[width=40mm, clip]{fig8b.jpg}
\caption{Left-hand panel: Grid of color and magnitude shifts applied
to the observed CMD of field F1. The color-magnitude shift nodes are
indicated with crosses. For each of these nodes, 36 individual
solutions were obtained, and the average of its corresponding
$\chi^2_{\nu}$ was calculated. We consider the solution at which the
minimum mean $\chi^2_{\nu}$ is reached as the best representation of
our data. The $\chi^2_{\nu, min}$ value obtained using BaSTI library
for a binary fraction of 35\% is indicated in the figure, at the
position where it was found. Contours around this position show the
1--, 2--, 3-- and 6--$\sigma$ confidence regions, where $\sigma$ is
defined as the standard deviations of the 36 $\chi^2_{\nu}$
individual solutions. Right-hand panel: As in the left-hand panel
but for F2. Note that the $1-\sigma$ confidence area is smaller in
this case.}
\label{fig:chiplot}
\end{figure}
As previously mentioned, for each shift in color and magnitude of the
observed CMD, we average the 36 individual solutions as well as its
corresponding $\chi^2_{\nu}$. For a 35\% binary fraction, the nodes at
which the mean minimum $\chi^2_{\nu}$, i.e. $\chi^2_{\nu, min}$ was
reached were found at $(\delta (F435W-F555W)_0, \delta M_{F555W}) =
(-0.09, 0.07)$ for F1 with $\chi^2_{\nu, min}=2.03$, and at $(\delta
(F435W-F555W)_0, \delta M_{F555W}) = (-0.06, 0.0)$ for F2 with
$\chi^2_{\nu, min}=2.23$. We consider the averaged solution
corresponding to $\chi^2_{\nu, min}$ as the one that best reproduces
our observations. Figure~\ref{fig:chiplot} shows how the mean
$\chi^2_{\nu}$ varies as a function of the color and magnitude shifts
applied to the observed CMD in F1 (left panel) and F2 (right panel),
using the BaSTI library and 35\% of binary fraction. The crosses
indicate the 35 nodes at which we calculated a mean $\chi^2_{\nu}$,
average of the 36 $\chi^2_{\nu}$ from the individual solutions. The
contours around the minimum $\chi^2_{\nu}$ (whose value is indicated
in the figure at the position where it was found) show the 1--, 2--,
3-- and 6--$\sigma$ confidence regions, with $\sigma$ defined as the
standard deviation of the mean $\chi^2_{\nu, min}$. We emphasize here
that the shifts in the observed CMD at which we obtained the best
solution do not necessarily represent corrections to the distance or
reddening estimates, since photometric corrections and model
systematics are also present.
The $\chi^2_{\nu, min}$ values suggest that the BaSTI library fits the
data better than the Padova/Girardi isochrones for both fields (see
Table~\ref{table:chivalues} in the Appendix). Nevertheless, the
solutions obtained with both libraries are very similar, with the
Padova/Girardi isochrones generating a best-fit mean solution slightly
more metal-rich than BaSTI. For simplicity, we consider the solutions
obtained using the BaSTI library for most of the following analysis.
\section{Results of the SFH Analysis}
\label{sec:results}
\subsection{The SFH of F1 and F2} \label{sec:sfh}
\begin{figure*}\centering
\subfigure
{\includegraphics[width=105mm, clip]{fig9a.jpg}}
\subfigure
{\includegraphics[width=105mm, clip]{fig9b.jpg}}
\caption{SFH(t,Z)$=\Psi(t,Z)$ of F1 (top panel) and F2 (bottom panel)
obtained using BaSTI models and assuming a 35\% binary
fraction. The blue and red lines are the two SFH projections:
metallicity distribution $\Psi(Z)$ and age distribution $\Psi(t)$,
respectively. Note that $\Psi(Z)$ does not represent metallicity
evolution, as it is integrated over age, and thus should not be
compared with panel (d) of Figures~\ref{fig:sfhf1}
and~\ref{fig:sfhf2}, which show Z as a function of age. Each
solution is calculated by averaging the 36 solutions at the
$\chi^2_{\nu, min}$ in the $\delta
\mathrm{mag}-\delta\mathrm{color}$ grid (Sec.~\ref{sec:bestsfh});
$\chi^2_{\nu,min}=2.03$ for F1 and 2.23 for F2. Recall that the
number of stars in F2 is $\sim 1/2$ of that in F1. Note the
prominent stellar population with ages 2--5 Gyr present in F1 but
nearly absent in F2. Although differences were expected (note the
different number of stars inside the blue box in
Figure~\ref{fig:deconvolvedmcmd} and the results in Paper I), the
significant different SFRs in the 2--5 Gyr bin between the two
fields is striking.}
\label{fig:sfhf1_f2}
\end{figure*}
In Figure~\ref{fig:sfhf1_f2} we show the best-fit mean SFH$=\Psi(t,Z)$
solution for F1 (top panel) and F2 (bottom panel) in a 3D-histogram
representation, as well as the two projections $\Psi(t)$ (red line)
and $\Psi(Z)$ (blue line). $\Psi(t)$ is the SFR as a function of time
or age distribution, i.e. $\Psi(t,Z)$ integrated over metallicity, and
$\Psi(Z)$ is the metallicity distribution function, i.e. $\Psi(t,Z)$
integrated over time. Both distributions are normalized by the area in
$\mathrm{pc}^2$. Recall that field F2 has $\sim 1/2$ the number of
stars as in F1.
The most striking feature of Figure~\ref{fig:sfhf1_f2} is the
significant star formation in F1 that occurred 2--5 Gyr ago. F2 is
predominantly old, with some contribution of young and
intermediate-age stars from 0.5 to 5 Gyr ago, but its 2--5 Gyr old
population is clearly not as prominent as that of F1. We emphasize
here that differences in the intermediate-age population between the
fields were expected (see Paper I and
Fig.~\ref{fig:deconvolvedmcmd}). However, the significant SFR in the
2--5 Gyr bin in F1 compared with F2 is rather surprising. As F1 has
contributions from both M32 and M31 stars and F2 is expected to have a
negligible contribution from M32, the derived SFHs suggest that the
2--5 Gyr old population in F1 is associated almost entirely with
M32. We discuss this further in the next section. To obtain the mean
age and metallicity of the system, we weight such quantities by the
mass of each bin in age or metallicity, respectively. We call them
hereafter mass-weighted mean age and mass-weighted mean metallicity.
\begin{deluxetable}{lcccc}
\tabletypesize{\scriptsize} \tablecaption{Integrated quantities
derived from the SFHs \label{table:sfhs}} \tablewidth{0pt}
\tablehead{\colhead{Field}&\colhead{$\langle\mathrm{Age}\rangle$
(Gyr)}&\colhead{$\langle[\mathrm{M/H}]\rangle$ (dex)} &\colhead{int(SFH)
($10^6\,M_{\odot}$)}}
\startdata
\sidehead{BaSTI library}
F1 & $7.95 ~\pm ~1.35$ & $-0.07 ~\pm ~0.10$ & $5.17 ~\pm ~0.50$\\
F2 & $9.12 ~\pm ~0.80$ &$-0.19 ~\pm ~0.10$ & $2.59 ~\pm ~0.24$ \\
M32 (F1-F2) & $6.80 ~\pm ~1.50$ & $-0.01 ~\pm ~0.08$ & $2.60 ~\pm ~0.50$\\
F2\tablenotemark{a} & $9.15~\pm~1.27$ & $-0.10~\pm~0.10$ & $2.50 ~\pm~ 0.18$\\
\sidehead{Padova/Girardi}
F1 & $7.99 ~\pm ~1.33$ & $0.01 ~\pm ~0.10$ & $5.88 ~\pm ~0.76$\\
F2 & $9.03 ~\pm ~0.85$ &$-0.07 ~\pm ~0.10$ & $2.81 ~\pm ~0.29$ \\
M32 (F1-F2) & $7.03 ~\pm ~1.50$ & $0.06 ~\pm ~0.10$ & $3.07 ~\pm ~0.75$
\enddata
\tablenotetext{a}{SFH of F2 was derived using BaSTI library with an
extra age bin, from 5--8 Gyr.}
\end{deluxetable}
\begin{figure*}\centering
\subfigure
{\includegraphics[width=85mm, clip]{fig10a.jpg}}
\subfigure
{\includegraphics[width=85mm, clip]{fig10c.jpg}}
\subfigure
{\includegraphics[width=85mm, clip]{fig10b.jpg}}
\subfigure
{\includegraphics[width=85mm, clip]{fig10d.jpg}}
\subfigure
{\includegraphics[width=150mm, clip]{fig10e.jpg}}
\caption{The SFH of F1. (a) SFR as a function of time; (b) cumulative
mass-weighted age distribution; (c) mass as a function of
metallicity; (d) age--metallicity relation; and (e) comparison
between the observed, calculated CMDs and a Hess representation of
the residuals. The calculated CMD is obtained by randomly extracting
stars from the model CMD in such a way that the final star
distribution represents the calculated SFH. Both the observed and
calculated CMDs were divided into the same $200\times200$ bins. The
Hess diagram of the residuals in panel (e) shows the subtraction of
the observed Hess diagram from the calculated one, in units of the
Poisson uncertainties. The vertical solid line in panel (b)
represents the mean age ($\sim 8$ Gyr) of the system, and the dashed
lines indicate the $1\sigma$ deviation of that value.}
\label{fig:sfhf1}
\end{figure*}
\begin{figure*}\centering
\subfigure
{\includegraphics[width=85mm, clip]{fig11a.jpg}}
\subfigure
{\includegraphics[width=85mm, clip]{fig11c.jpg}}
\subfigure
{\includegraphics[width=85mm, clip]{fig11b.jpg}}
\subfigure
{\includegraphics[width=85mm, clip]{fig11d.jpg}}
\subfigure
{\includegraphics[width=150mm, clip]{fig11e.jpg}}
\caption{As in Figure~\ref{fig:sfhf1} for F2. The mass in F2 is
roughly half that of F1.}
\label{fig:sfhf2}
\end{figure*}
Figures~\ref{fig:sfhf1} and \ref{fig:sfhf2} display the main results
projected from the extracted SFHs of F1 and F2, respectively. We find that:
\begin{itemize}
\item F1 acquired 75\% of its stellar mass between 5 and 14 Gyr ago.
Stars with ages of 2--5 Gyr contribute 23\% of the mass in F1. The
remaining 2\% of mass in F1 is constituted by stars younger than 2
Gyr.
\item F1 is metal-rich with an almost constant age--metallicity
relation, to the limits of the age resolution of the CMD.
\item F1's mass-weighted mean age is $7.95\pm1.35$ Gyr and its
mass-weighted mean metallicity is
$[\mathrm{M/H}]=-0.07\pm0.10~\mathrm{dex}$ (Table~\ref{table:sfhs}).
\item F2 is predominantly old, with 95\% of its mass already formed
5--14 Gyr ago. There is a small contribution of mass to the system
after that, and it stopped forming stars $\sim$ 0.5 Gyr ago.
\item F2 is also quite metal-rich, but is marginally more metal-poor
than F1, with a slight age--metallicity relation showing a small
increase in metallicity at younger ages.
\item F2's mass-weighted mean age is $9.12\pm0.80$ Gyr and its
mass-weighted mean metallicity is
$[\mathrm{M/H}]=-0.19\pm0.10~\mathrm{dex}$ (Table~\ref{table:sfhs}).
\end{itemize}
The integrated quantities derived for the SFHs of F1 and F2 using
Padova/Girardi Library are also indicated in Table~\ref{table:sfhs}.
Figures~\ref{fig:sfhf1}e and \ref{fig:sfhf2}e show comparisons between
the observed (left) and calculated CMD (middle) as well as the Hess
diagram of the residuals in units of the Poisson uncertainties
(right), for F1 and F2 respectively. The calculated CMDs have been
obtained by randomly extracting stars from the synthetic CMDs in such
a way that the resulting star distribution follows the best calculated
SFHs. For both F1 and F2, the model CMD shows reasonable agreement
with the observed CMD throughout most evolutionary phases, which is
also reflected in the residual Hess diagrams. The RC regions, however,
show significant discrepancies. This is not surprising; due to
uncertainties in, e.g., the mass loss during the RGB or the He content
of the stars, that particular evolutionary stage is not well-modeled
\citep{Gallart_etal05}---but we have not used this region in deriving
the solutions. There is also some discrepancies for magnitudes fainter
than the 50\% completeness level, but this region was also not used
for the derivation of the SFHs.
\subsection{The SFH of M32 as revealed by the IAC method}
To calculate the SFH of M32, we make use of the derived SFHs of F1 and
F2\footnote{We would ideally need a deep CMD composed solely of M32
stars to derive the SFH of M32, which we attempted to derive in
Paper I. Under the assumption that the M31 stellar populations in F1
and F2 are statistically the same, we subtracted the stars of the F2
CMD from the CMD of F1 taking into account the difference in
crowding of the fields. This produced the deepest CMD of M32 yet
obtained. However, the use of such CMD to extract the SFH of M32
would introduce uncertainties associated with the decontamination
process.}. Given the fact that both SFHs have been obtained using
the same stellar population sampling, and assuming that the SFH of M31
in F1 and in F2 is identical and that M32 is not present in F2,
calculating the SFH of M32 is straightforward: we simply subtract the
SFH of F2 from that of F1.
\begin{figure*}\centering
\includegraphics[width=150mm, clip]{fig12.jpg}
\caption{SFH of M32 obtained after subtracting the calculated SFH of
F2 from that of F1. We find two dominant populations contributing to
the SFH of M32. One is 2--5 Gyr old and contributes $\sim$ 40\% of
the total mass of M32 at F1. The population older than 5 Gyr
contributes $\sim$ 55\% of the total M32's mass at F1. Note that
some of the stars younger than 2 Gyr are quite metal-poor compared
to the nearly solar mean metallicity of M32. This suggests that
these are BSS and may be the first direct evidence of such a
population in M32.}
\label{fig:sfhm32}
\end{figure*}
\begin{figure*}\centering
\subfigure
{\includegraphics[width=85mm, clip]{fig13a.jpg}}
\subfigure
{\includegraphics[width=85mm, clip]{fig13c.jpg}}
\subfigure
{\includegraphics[width=85mm, clip]{fig13b.jpg}}
\subfigure
{\includegraphics[width=85mm, clip]{fig13d.jpg}}
\caption{The SFH of M32. (a) SFR as a function of time, clearly
indicating the two dominant populations: at $\sim 8$ Gyr and $\sim
4$ Gyr; (b) cumulative mass-weighted age distribution which shows
how much each population contributes to the total mass of M32 at F2;
(c) mass as a function of metallicity, indicates the mean
metallicity of the system, roughly solar; and (d) age--metallicity
relation, nearly constant. The vertical lines in panel (b) represent
the mean age ($\sim 6.8$ Gyr) of M32 in F1. The dashed lines
indicate the $1\sigma$ deviation of this value.}
\label{fig:sfhm32_bis}
\end{figure*}
Figure~\ref{fig:sfhm32} shows the inferred SFH of M32 for the first
time calculated from its resolved stellar population. We used the F1
and F2 SFHs shown in Figure~\ref{fig:sfhf1_f2}, inferred using the
BaSTI library and a 35\% binary fraction. We can see that a major
burst of star formation occurred in M32 2--5 Gyr ago, responsible for
$\sim$ 40\% of M32's current mass at F1's location. This can be seen
from the cumulative mass function, shown in panel (b) of
Figure~\ref{fig:sfhm32_bis}. Stars older than 5 Gyr contribute
$\sim$55\% of the total mass of M32 in this field. From this
CMD-fitting analysis, however, due to the limitations imposed by the
crowding of our fields, we cannot specify when the star formation
started, whether it was constant over the 5--14 Gyr period, or if it
peaked at some age. Integrated quantities derived from the calculated
M32 SFH are indicated in Table~\ref{table:sfhs}. Note that the
estimated mean age and metallicity of M32, $\sim 6.8$ Gyr and $\sim
-0.01~\mathrm{dex}$, respectively, are younger and more metal-rich
that the mean age and metallicity of F1, because M31's mean age and
metallicity in F2 is older and more metal-poor than M32 in F1.
The age--metallicity relation for M32 is nearly constant
(Fig.~\ref{fig:sfhm32_bis}d) although there is apparently a mild
increase at $\sim$ 5 Gyr followed by a small decrease at $\sim$ 2 Gyr.
We note that an almost constant age--metallicity relation appears to
suggest that M32 has not experienced any metal enrichment. However,
the lack of resolution in age means that we cannot extract detailed
information on stars older than 5 Gyr. Most of the chemical evolution
of the system has likely occurred during that 5--14 Gyr period. M32's
mass-weighted peak in metallicity is at $[\mathrm{M/H}] \sim 0.2$ dex
(Fig.~\ref{fig:sfhm32_bis}c).
\begin{figure*}\centering
\includegraphics[width=120mm, clip]{fig14a.jpg}
\includegraphics[width=140mm, clip]{fig14b.jpg}
\caption{Top panel: Calculated CMD of M32, obtained by randomly
extracting stars from the model CMD in such a way that they follow
the derived SFH of M32. The stars are color-coded according to age,
as indicated in the bottom panel, except for 2--5 Gyr old and stars
older than 5 Gyr, shown as yellow and black dots, respectively, in
the top panel whereas gray-scale Hess representations of their CMDs
are shown in the bottom panel. Note how the various ages fit
together and the age interval that populates each region of the
CMD. Bottom panel: Each CMD is composed by stars of a different age
interval, from only an extended main sequence (left panel, ages
$\sim 0.5$ Gyr) to a CMD with well populated RGB, RC and AGB
evolutionary phases (right panel, ages of 5--14 Gyr). Note the
differences in the MSTO region and fainter MS in the last two
CMDs. The MSTOs for the younger population (2--5 Gyr, yellow dots in
the top panel and a Hess representation of their CMD in the bottom
panel) are brighter and bluer than the ones for the 5--14 Gyr old
population (black dots in the top panel and a Hess representation of
their CMD in the bottom panel).}
\label{fig:scmd_m32}
\end{figure*}
We show in the top panel of Figure~\ref{fig:scmd_m32} the calculated
CMD of M32, with its stars color-coded according to age. The CMD was
obtained by randomly extracting stars from the model CMD, in such a
manner that their star distribution follows the calculated SFH. This
figure provides explicit information on the age interval that
populates each region of the CMD as well as showing how the various
ages combine. We see, for example, that stars of different ages
contribute to the RC. Younger stars populate the brighter bluer
portion of the RC while older stars populate the fainter, redder
portion of the RC. The BP is only populated by stars younger than 2
Gyr. The bottom panel shows the CMDs produced by each age interval
considered in the extraction of the SFH. We can appreciate in detail
the differences between each CMD as the ages vary, from only an
extended main sequence (bottom left panel, ages $\sim 0.5$ Gyr) to a
CMD with well-populated RGB, RC and AGB evolutionary phases (bottom
right panel, ages of 5--14 Gyr). Note the presence of only few BHB
stars in the bottom right panel, as expected for systems as metal-rich
as M32; in the composite CMD (top panel), these few BHB stars are
mixed with young, blue stars in the extended MS.
Finally, we can qualitatively compare these results with the SFHs
derived for some of the Local Group dwarf satellites, which have been
analyzed by the same or very similar methods\footnote{A more extensive
and quantitative comparison of our results with other the SFHs of
Local Group satellites is beyond the scope of this paper.}. It is
interesting to note that the mean age derived for M32 is comparable to
that of dwarf irregular galaxies, such as the Small and Large
Magellanic Clouds and Pegasus \citep{Noel_etal09, Harris_zaritsky09,
Dolphin_etal05}, whereas it is quite young compared with the mean
ages of dwarf spheroidal galaxies, such as Cetus, Tucana, and Ursa
Minor \citep{Monelli_etal10a, Monelli_etal10b, Dolphin_etal05}. We
also note that a synthetic CMD analysis was performed on archival
WFPC2 data of M32 by \citet{Dolphin_etal05}. Their results suggest an
older mean mass-weighted age for M32, $\sim 8.5$ Gyr. However, the
WFPC2 data not only are shallower than our data but also contain
significant contamination by M31 stars, which were not taken into
account in their SFH analysis.
\subsubsection{Exploring the SFH solution and its robustness}
We now address the robustness and uniqueness of the SFHs derived
here. First, due to the complex parameter space and thus multiple
local minima involved in the process to find a SFH, an algorithm that
guarantees finding the global minimum is strongly desirable. As
mentioned above, we have used a genetic algorithm to find the minimum
$\chi^2$ (see \citealt{Charbonneau95}). This type of algorithm, unlike
the so-called 'downhill' algorithms, is designed such that the
solution found is infinitesimally close to or at the global minimum
independently of the initial seed that started the process, provided
that a sufficiently large number of generations (i.e. variations of
individuals and mutations) is performed. \citet{Aparicio_hidalgo09}
have tested this and found that $\approx 10^5$ generations are enough
to guarantee that the solution will be at or as close as possible to
the global minimum. We have performed $2\times10^5$ generations per
solution, which assures us that we have reached convergence.
In addition, the IAC method does not introduce any systematic error to
the SFH solution, provided that the age and metallicity bins used to
extract the SFH are appropriate to the observed CMD. This has been
verified by several tests on mock stellar populations performed at
Instituto de Astrof\'isica de Canarias in which the SFH of mock
galaxies of known SFH have been recovered rather accurately
\citep[see][]{Aparicio_hidalgo09, Hidalgo_etal09, Hidalgo_etal11}. In
this work, we have also performed several experiments on mock data to
find the appropriate age and metallicity resolution at which,
according to the observational effects of our observed CMDs, the SFH
solutions of mock galaxies are within 1--$\sigma$ to the input ones
(see Sec.~3.1 above). Thus, each SFH solution obtained is ``unique,''
by which we mean that combinations of simple populations within the
error bars of the SFH will produce CMDs indistinguishable from the
best fit CMD. Any other SFH which is combination of simple populations
significantly different that those of the final SFH (i.e., not
possible within the error bars of our solution) will produce a CMD
significantly different than the best-fit CMD and, therefore, than the
observed one.
Finally, even though we only show our best solution, we have explored
other solutions that give similar good fits. We found that the SFH of
the solutions at 1--$\sigma$ confidence area (see
Fig.~\ref{fig:chiplot}) are not significantly different than the best
one, for both F1 and F2. The mass percentages per age range slightly
vary from one solution to the other, but the overall SFH remains the
same. Taking these nearby solutions into account, we find that M32 at
F1 has $\sim 40\% \pm 17\%$ of its mass in a 2--5 Gyr old, metal-rich
population and $\sim55 \% \pm 21\%$ of its mass in stars older than 5
Gyr, with slightly subsolar metallicities. The uncertainties represent
the 1--$\sigma$ error of our best solution and the variations of these
percentages when considering solutions of similar good quality fit.
\subsubsection{Young population (Ages $< 2$ Gyr) vs.\ Blue Stragglers}
In Paper I, we discussed the possibility that the fainter stars in the
BP of M32 could be old BSS rather than a young stellar population with
ages $<2$ Gyr. However, the analysis presented in Paper I did not
allow us to confirm or rule out either case. BSS are stars hotter,
bluer and brighter than the MSTOs in a CMD, thus generating a blue
plume. Given their locations on the CMD, they are burning hydrogen in
their cores with masses larger than the turn-off mass, which indicates
that some sort of mechanism rejuvenated their inner layers. Although
such a mechanism is still a matter of debate, there are currently two
theoretical possible scenarios to explain the BSS origin: they are the
result of either a collision between stars
\citep[e.g.,][]{Sigurdsson_etal94} or mass-transfer in a binary system
\citep[e.g.,][]{Mccrea64, Carney_etal01}.
We investigate the nature of these stars from the SFH presented
here. Stars younger than 2 Gyr constitute $\sim$ 4\% of the total mass
of M32 at F1. Figure~\ref{fig:sfhm32} shows that some of the young
stars, produced by a very low SFR event at look-back times $<2$ Gyr,
are rather metal poor ($[\mathrm{M/H}]\sim -0.7$) in comparison with
the mean metallicity of M32 ($[\mathrm{M/H}]\sim 0.0$). Given the
almost constant age--metallicity relation for M32 and the presence of
intermediate-age stars (2--5 Gyr old) of solar or even higher
metallicity, it is unlikely that M32 contains at the same time younger
stars with significantly sub-solar metallicities. The most plausible
explanation is that these stars are BSS belonging to an old metal-poor
population. BSS are found in open and globular clusters
\citep{Ferraro_etal04, Mapelli_etal04, Mapelli_etal06, Piotto_etal04,
Demarchi_etal06}, dwarf spheroidal galaxies
\citep{Hurleykeller_etal99, Carrera_etal02, Momany_etal07,
Mapelli_etal09, Monelli_etal10a}, and even in the Milky Way halo
field population \citep{Preston_sneden00}. Therefore, it seems natural
to consider that they can also be found in an elliptical galaxy. These
stars represent $\sim 2\%$ of the mass of M32 in F1 and might be the
first direct evidence of BSS in this galaxy. An alternative
explanation could be that these young and metal-poor stars were
generated by an episode of late infall of metal-poor gas. However, if
we assume that M32 is interacting with M31, we would not expect M32 to
accrete gas, but instead to lose gas to M31 through stripping.
The other $\sim 2$\% of stars with ages $<2$ Gyr that we find in the
SFH inferred for M32 may indeed represent a young metal-rich
population in M32 at F1.
\section{The Star formation history of M32}
\label{sec:sfhm32}
By combining the results in the present work with the analysis in
Paper I, we can finally provide a detailed and complete SFH of M32. We
conclude that M32 has had an extended SFH and is composed of two main
dominant populations at F1: $\sim 40\% \pm 17\%$ of the mass in a 2--5
Gyr old, metal-rich population and $\sim55 \% \pm 21\%$ of the mass in
stars older than 5 Gyr, with slightly subsolar metallicities. We
confirm the existence of the younger ($< 5$ Gyr) stars through the
presence of bright AGB stars observed in Paper I, with the appropriate
ages. From the RC, RGB bump and AGB bump analyzed in Paper I, the bulk
of the old population is 8--10 Gyr old. We therefore do not expect a
significant contribution from stars older than 10 Gyr in M32 at
F1. Nevertheless, there is a hint of the presence of a few ancient
metal-poor stars present in M32, as revealed by the 2--$\sigma$
detection of RR Lyrae belonging to M32 at F1. The remaining $\sim4$\%
of the mass is roughly equally divided between a young metal-rich
population and a young metal-poor population. We associate the latter
with blue straggler stars belonging to an old (likely metal-poor)
population.
The age--metallicity relation for M32 is nearly constant within our
age resolution, although there is a small increase in metallicity at
at $\sim$ 5 Gyr followed by a small decrease at ages younger than
$\sim$ 2 Gyr. The mass-weighted mean metallicity of M32 is
$[\mathrm{M/H}]\sim -0.01 \, \mathrm{dex}$ with a peak at
$[\mathrm{M/H}]\sim 0.02 \, \mathrm{dex}$. We emphasize here again
that an almost constant age--metallicity relation appears to suggest
that M32 has not experienced metal enrichment; but as in F1, this is
due to the poor age resolution and does not imply the lack of an
age--metallicity relation. Stars with metallicities lower than
$[\mathrm{M/H}]\la -1\, \mathrm{dex}$ only contribute $\sim 5$\% of
the total mass of M32 at $\sim2'$ from its center. This is consistent
with the photometric metallicity function (MDF) of M32 derived in
Paper I, which shows that the majority of the stars has a slightly
sub-solar metallicity at $[\mathrm{M/H}]\sim -0.2 \,
\mathrm{dex}$. The MDF also indicated that metal-poor stars with
$[\mathrm{M/H}] <-1.2$ contribute very little, \emph{at most} 6\% of
the total $V$-light to M32 or 4.5\% of the total mass in F1, implying
that the enrichment process largely avoided the metal poor stage.
\subsection{On the integrated light of M32}
The results obtained in this work are a fundamental step for
understanding the formation and evolution of other low-luminosity
spheroidal systems (elliptical galaxies or bulges). Since integrated
light spectra are, in general, the only means available to study the
stellar populations of these galaxies, we strongly rely on unresolved
stellar population models to learn about their SFHs. These models,
which have become very sophisticated in disentangling the non-trivial
age and metallicity degeneracy \citep{Worthey94, BC03, Thomas_etal03,
Vazdekis_etal10}, still suffer from several uncertainties: e.g., it
is difficult to distinguish between a young or hot old population
since the latter is not necessarily accounted for in the models
\citep[e.g.][]{Maraston_thomas00}. Calibration of these models, which
requires observations of individual stars in elliptical galaxies, is a
key ingredient that needs to be further developed. As briefly
mentioned in Section 1, M32 is located at a distance such that both
integrated spectroscopy and photometry of its individual stars can be
studied in great detail. A comparison of the stellar parameters
obtained using resolved stars and integrated luminosity is fundamental
to provide a calibration to the unresolved stellar models with an
actual elliptical.
Extensive spectroscopic studies of M32 have been performed, mostly in
the central regions and out to $\sim 1\,r_e$
\citep[e.g.,][]{Oconnell80,Gonzalez93, Worthey04, Rose_etal05}. All
studies agree that the central stellar population has an
SSP-equivalent age of 2.5--5 Gyr and roughly solar metallicity, with
an age gradient that increases the age at $1\,r_e$ by $\sim3$ Gyr and
a mild negative metallicity gradient. Various integrated-light studies
have suggested that M32 underwent a period of significant star
formation in the recent past, i.e., about 5--8 Gyr ago,
\citep[e.g.,][]{Oconnell80, Pickles85, Bica_etal90} based on the
presence of enhanced H$\beta$ absorption in the integrated spectrum of
M32, a signature of an intermediate-age population
\citep[e.g.,][]{Rose94, Trager_etal00a, Worthey04, Schiavon_etal04,
Rose_etal05, Coelho_etal09}. To date, only Coelho et al. (2009) have
attempted to probe the unresolved stellar populations as far from the
center of M32 as the ACS/HRC field presented in this paper lies,
using longslit observations with GMOS on Gemini. They propose that an
ancient and intermediate-age population are both present in M32 and
that the contribution from the intermediate-age population is larger
at the nuclear region. They claim that a young population is present
at all radii, and they further suggest that there is a strong
component of either very young ($<0.3$ Gyr) and/or very old ($>10$
Gyr), metal-poor stars even in their outermost field.
We can use the inferred SFH of M32 to compute line strength indices
using the models of \citet[][hereafter BC03]{BC03} and to calculate
single stellar population (SSP)-equivalent parameters that can then be
compared with the values obtained from the integrated light of this
galaxy.
Using the BC03 models, we obtain a $B$-band luminosity-weighted mean
age and metallicity of 4.9 Gyr and $[\mathrm{M/H}]=-0.12 \,
\mathrm{dex}$, respectively, for M32 at F1 from its resolved
SFH. \citet{Coelho_etal09} find an average luminosity-weighted age of
$5.7\pm1.5$ Gyr using BC03, which agrees with our result within the
uncertainties, but their inferred mean metallicity is much lower,
$[\mathrm{M/H}]=-0.6\pm 0.1$ (see their Table 3). Moreover, as
mentioned above, they suggest that there is a strong component of
either very young ($<0.3$ Gyr) or very old ($>10$ Gyr), metal-poor
stars in their field at a radius similar to our F1 field, which is
inconsistent with our data.
\begin{figure}
\centering
\includegraphics[width=84mm, clip]{fig15a.jpg}
\includegraphics[width=84mm, clip]{fig15b.jpg}
\caption{SSP-equivalent age (top panel) and metallicity (bottom panel)
values inferred from line-strength fits given by \citet{Worthey04}
as a function of radius in M32. The lines are linear fits to the
data. We can see the steep positive age gradient from M32's center
to $1\,r_e\sim 40''$. An extrapolation of these fits to $\log(110'')
\sim 2.04$ gives the SSP-equivalent age and metallicity values at
F1: 8.9 Gyr and $[\mathrm{M/H}]=-0.23\pm 0.03$ dex,
respectively. The SSP-equivalent parameters calculated from the
inferred SFH of M32 (blue dots) are in stark contrast with
predictions of spectral studies.}
\label{fig:extrapol}
\end{figure}
We have also calculated SSP-equivalent parameters that can be compared
with the values obtained from integrated spectra of this galaxy. Using
the BC03 models, the SSP-equivalent values of M32 obtained from its
inferred SFH at F1 are $2.9 \pm 0.2$ Gyr and $[\mathrm{M/H}]= 0.02 \pm
0.01\,\mathrm{dex}$, respectively. Given the radial line-strength
gradients present in M32 \citep[e.g.,][]{Rose_etal05}, we cannot
directly compare these SSP-equivalent values with those obtained from
central integrated spectra of M32 \citep[by,
e.g.,][]{Trager_etal00a}. We therefore do the following. We use the
values of the line strength indices from \citet{Worthey04} and compute
the SSP-equivalent parameters from polynomial fits to the
absorption-line strengths as a function of radius (his Table 1) using
the modified BC03 models described in \citet{Trager_etal08}. We then
fit straight-line gradients to the SSP-equivalent parameters as a
function of radius and extrapolate these fits to $110^{\prime\prime}$,
F1's position. The SSP-equivalent age and metallicity of M32 at F1
from this extrapolation are $8.9\pm0.5$ Gyr and
$[\mathrm{M/H}]=-0.23\pm 0.03$ dex, respectively. We note that
Worthey's values for Mg$b$ are low compared with \citet{Gonzalez93}
and \citet{Trager_etal98}, and thus the SSP-equivalent age we have
obtained may be slightly overestimated whereas the SSP-equivalent
metallicity may be underestimated. Taken this into account, we
obtained an SSP-equivalent age of $\sim 8.4$ Gyr and a SSP-equivalent
metallicity of $[\mathrm{M/H}]\sim-0.13$ dex.
Figure~\ref{fig:extrapol} shows the SSP-equivalent parameters from
\citet{Worthey04}. The linear fits to the $\log(\mathrm{age/Gyr})$ and
$[\mathrm{M/H}]$ as a function of $\log(r/'')$ are also shown. The
blue stars indicate the SSP-parameters obtained from the inferred SFH
of M32 from BC03 models.
The predicted age and metallicity at F1 from the extrapolation of the
absorption-line gradients are much older and more metal-poor,
respectively, than those obtained from the inferred SFH of M32. This
suggests that either the extrapolation of line strength indices or the
stellar population models, or both, may be in error, but we are
currently unable to discern which. Color profiles in many colors of
M32 are rather flat \citep[][]{Peletier93}. Since M32 does not contain
dust, integrated colours can be good population indicators and the
fact that there are no gradients in colors agree with the results from
the inferred SFH. \citet{Davidge_jensen07} have also challenged the
radial gradients in mean stellar parameters obtained from spectral
studies. They find no evidence for a radial age gradient in M32, based
on the properties of observed brightest AGB stars, in contrast to the
results by \citet{Worthey04} and \citet{Rose_etal05}, who found (as
described above) a significant radial gradient in the mean
luminosity-weighted age of M32.
To further investigate this apparent contradiction, integral-field
spectroscopic observations with VIRUS-P \citep{Hill_etal08} have been
taken at F1 and F2 and will be analyzed in a future paper. This will
provide spectra of the integrated stellar light of M32 for the
fundamental calibration for the study of stellar populations.
\subsection{On the formation of M32}
Certainly, the most striking result of this work is the substantial
contribution of 2--5 Gyr old metal-rich stars to the total mass of M32
at F1. How has an elliptical galaxy like M32 formed such a young
population of stars? What is the origin of this population? In this
section we attempt to address these questions and discuss, in
particular, the most popular proposed formation scenarios for M32.
A formation scenario for M32 has been proposed by \citet[][hereafter
K09]{Kormendy_etal09}, in which M32 is a normal, low-luminosity
elliptical galaxy. K09 find that both central and global parameter
correlations from recent accurate photometry of galaxies in the Virgo
cluster place M32 as a normal, low-luminosity elliptical galaxy in all
regards. K09 fit a \citet{Sersic68} profile to the SB of M32 with $n
= 2.8$, in agreement with S\'ersic indices of other low-luminosity
ellipticals studied by K09. They interpret the light at the center of
M32 that was not fit by their S\'ersic profile as a signature of
formation in dissipative mergers \citep[][]{Mihos_hernquist94}. Extra
central light is a general feature of coreless galaxies and is
observed in all the other low-luminosity ellipticals of K09's sample.
Within this scenario, the metal-rich 2--5 Gyr old stars contributing
to $\approx 40\% \pm 17\%$ of M32's mass at F1 could be the result of
such a dissipative merger event. Thus, the progenitors of M32 should
have been very gas-rich spiral galaxies, like M33 for
example. However, such progenitors should have stellar masses of the
order of $10^8\, M_{\odot}$, whereas M33's stellar mass is $\sim
3\times 10^9\, M_{\odot}$. There are in fact no gas-rich spiral
galaxies near M31 of the appropriate stellar mass.
An alternative scenario for the formation of M32 has been proposed by
\citet[][hereafter B01] {Bekki_etal01}, who assumed that M32 is the
result of a low-luminosity spiral galaxy, whose bulge, unlike most of
its outer disk, survived its dynamical interactions with
M31\footnote{The idea that M32, as well as other small high-surface
brightness galaxies, is a tidally truncated galaxy has been
discussed several decades before B01 models. In, for example,
\citet{Faber73}, the original truncated galaxy was a more massive
elliptical galaxy, from which only the tightly bound core of the
original elliptical remains after a strong tidal interaction.}. In
their N-body/smoothed particle hydrodynamics simulations, B01
considered a gas-rich low-mass disk galaxy with a bulge orbiting a
massive disk galaxy like M31. About 0.75 Gyr after the interactions
have started, the outer stellar disk (from 2 kpc to 5 kpc) of the
spiral galaxy is stripped away and only keeps $\approx40$\% of its
initial mass in stars initially located in the central regions, i.e.,
within 2 kpc of the center. New star formation is triggered by the
interaction of the gas-rich spiral with M31 but the outer new stellar
component is also tidally stripped away, and consequently only the
central starburst component survives. On the other hand, the bulge is
only weakly affected by tidal interactions with M31 due to its
compactness, and only $\approx 19$\% of its mass is lost. At the end
of their simulations, there is a fractional disk, bulge, and new stars
mass ratio of $\approx 49$\%, $\approx 42$\%, and $\approx 0.9$\%,
respectively, within 2 Kpc of the remnant compact galaxy. Our field F1
is located at 110$\arcsec$, i.e. $\sim 0.5$ kpc from the galactic
center at M32's distance and, assuming either an inside-out or
outside-in formation scenario for the disk \citep[see e.g.,][and
discussion in Section~\ref{sec:m31}]{Sommer-larsen_etal03} and
considering that we are looking at a $\sim0.5\,R_d$, where $R_d=0.9$
kpc is the scale length radius of B01's disk, the disk stars that we
should be observing there would have mean ages of 8--12 Gyr. Assuming
that the bulge is predominantly old, this scenario is difficult to
reconcile with our results, given the substantial 2--5 Gyr old
intermediate-age population detected in this work. However, there has
been some indications of small bulges which could have extended SFHs,
similar to that of M32, with 10--30\% of their total mass at look-back
times between 0.5 and 5 Gyr \citep{Thomas-davies08}.
Therefore it is unclear from our SFH results what the preferred model
for the origins of M32 is. While specific origin models differ in
detail, the general outlines overlap enough to make choosing a
specific model difficult with the age resolution of our current SFH.
More observations of M32-analog systems and simulations of spheroidal
systems with similar SFHs to that we have presented are needed to shed
light on M32's origins. Furthermore, finding evidence of a stellar
stream in the halo of M31 with the ages and metallicities obtained for
M32, which should be left if a major stripping of M32 by M31 has
occurred, would help to constrain the models and assess the validity
of the proposed scenarios.
\section{The disk and spheroid population of M31 in F2}
\label{sec:m31}
In Paper I, we compared our findings in F2, in particular its
metallicity distribution function (MDF), with several previous works
on the disk and bulge of M31 \citep[e.g.][]{Williams_02,
Worthey_etal05, Olsen_etal06, Brown_etal06}. We found a reasonably
good agreement with most studies. In this section we discuss our new,
quantitative results on the stellar populations at F2 and their
implications for the formation of the M31's disk. M31 seems to have
formed most of its stars between 5--14 Gyr ago at F2. As mentioned
above, we cannot precisely indicate when the star formation started in
either F1 nor F2 but we can see that M31 is older than M32 in F1.
\citet[][hereafter B06]{Brown_etal06} analyzed three CMDs of different
regions of M31: the spheroid, stream and outer disk. These CMDs
reached well below the oldest MSTOs, and B06 derived SFHs at each
field in great detail. Differences between these SFHs were mainly
found in the age and metallicity distributions of stars older than 5
Gyr. Within this age range (5--14 Gyr) we do not have the resolution
required to inspect different bursts of star formation in F2 in
detail, given the SFH extracted in Section~\ref{sec:results}. We can,
nevertheless study the SFH of F2 in more detail than what is presented
in Section~\ref{sec:results}. As we show in
Figure~\ref{fig:deconvolvedmcmd}, the CMD of F2 is $\sim 0.5$ mag
deeper than the one of F1, which allows us to obtain information of
fainter, i.e. older, MSTOs at F2\footnote{The previous selection of
age and metallicity bins to derive the SFHs was strictly based on
the resolution imposed by the CMD of F1. In order to subtract the
SFH of F2 from that of F1, we required the simple populations
considered be exactly the same.}. We therefore extracted again the
SFH of F2 following the steps indicated in Section~\ref{sec:method},
but with an extra bin in the age, from 5 to 8 Gyr, for the simple
populations considered. The boundaries of the bins in age are in this
case [0, 0.5, 1, 2, 5, 8, 14]$\times10^9$ years. The inferred best-fit
mean SFH of F2 with this new resolution in age was found at $(\delta
(F435W-F555W)_0, \delta M_{F555W}) = (-0.03, 0.00)$ with $\chi^2_{\nu,
min} = 2.12$. Figure~\ref{fig:sfhf2extrage} shows a 3D-histogram
representation of the new SFH solution for F2. We can now distinguish
two main populations that contribute substantially to our background
field F2, instead of only one old population: $\approx 30\% \pm 7.5\%$
of the total mass in F2 is composed of a 5--8 Gyr old metal-rich
population and $\approx 65\% \pm 9\%$ of the mass is composed of a
8--14 Gyr old, metal-poor population. An age--metallicity relation
shows a slightly steeper slope from an old metal-poorer population to
younger metal-richer ones than before, as shown in
Figure~\ref{fig:sfhf12-age}. We are still not able to answer when the
star formation started in F2. Nevertheless, our results for the mean
age and metallicity for F2, $9.12\pm 1.21$ Gyr and $-0.10\pm 0.10$ dex
respectively, are in good agreement with B06 results for their outer
disk field, $8.5$ Gyr and $-0.4$ dex, respectively\footnote{The cited
values correspond to the results obtained by B06 when a 40\% binary
fraction was assumed.}. In addition, young stars, with ages between
0.3 and 1 Gyr, that populate the BP in the CMD of F2 do not contribute
significantly to the total mass, which is also in agreement with B06's
results. Interestingly, kinematic data in our field imply that both
the disk and spheroid of M31 contribute to the populations in F2
(K. Howley, 2010, priv. commm). This was also the case for the outer
disk field of B06. B06, however, attempted to disentangle both
populations assuming that their spheroid field was representative of
the spheroid population present in their outer disk field. By
subtracting the spheroid population, they obtained a younger mean age
for the outer disk of M31---but still older than 5 Gyr.
\begin{figure*}\centering
\includegraphics[width=130mm, clip]{fig16.jpg}
\caption{A more detailed SFH of F2 in a 3d-histogram representation
than that shown in Figure~\ref{fig:sfhf1_f2}. This SFH of F2 was
constructed this time with an extra bin in age covering 5--8 Gyr. We
now find two dominant populations of M31 at F2: An old more
metal-poor population, older than 8 Gyr, and an intermediate-age
more metal-rich population, 5--8 Gyr old. Stars younger than 5 Gyr
old only contribute $\sim5$\% of the mass of M31 at F2.}
\label{fig:sfhf2extrage}
\end{figure*}
\begin{figure*}\centering
\subfigure
{\includegraphics[width=85mm, clip]{fig17a.jpg}}
\subfigure
{\includegraphics[width=85mm, clip]{fig17c.jpg}}
\subfigure
{\includegraphics[width=85mm, clip]{fig17b.jpg}}
\subfigure
{\includegraphics[width=85mm, clip]{fig17d.jpg}}
\caption{The SFH of F2 inferred using an extra bin, i.e. more
resolution, in age. (a) SFR as a function of time; (b) cumulative
mass-weighted age distribution; (c) mass as a function of
metallicity; (d) age--metallicity relation. The vertical solid and
dashed lines in panel (b) represent the mean age ($\sim 9.15$ Gyr)
and $1\sigma$ deviation of that value, respectively.}
\label{fig:sfhf12-age}
\end{figure*}
Given the resolution allowed by the depth of our data, the inner disk
and spheroid populations of M31 (at 5 kpc from its center) seem to be
indistinguishable from the outer disk and spheroid ones (at 25 kpc
from M31 galactic center, B06). Even though we are unable to subtract
the spheroid population that contributes to our field F2, most likely
the mean age of M31's disk at F2 is younger than 8.72 Gyr and older
than 5 Gyr, given the negligible contribution of stars younger than 5
Gyr. This result supports the inside-out disk formation models by
e.g., \citet[][]{Abadi_etal03a, Abadi_etal03b,
Sommer-larsen_etal03}. Abadi et al. find a mean age of 8--10 at 2
kpc, which radially decreases to 6--8 at 20 kpc. Sommer-Larsen et
al. simulated two spiral galaxies, with two different scenarios of
disk formation: inside-out and outside-in. Our expected mean age for
the disk of M31 at F2 agree with both scenarios within their
uncertainties, assuming a stellar disk scale length of $\approx5$ kpc
\citep[e.g.,][]{Walterbos_kennicutt88, Worthey_etal05}. They find
that, at 1 disk scale length, the mean ages of both simulated disks
are $\sim$6--8 Gyr. However, the significant fraction of stars younger
than 5 Gyr predicted by their outside-in model at F2 is not supported
by our data. Thus, we favor their inside-out model. Furthermore, the
inside-out formation model of \citet{Sommer-larsen_etal03} predicts
that the disk has almost no age gradient which, although surprising,
is also in agreement with the comparison of our and B06 results at
different disk locations. They explain that this prediction is a
consequence of the non linear dependence of the SFR on the cold gas
density, which makes the SFR rather low in the outer disk at late
times, thus the average outer disk stellar age is quite high. An
alternative scenario for the absence of an age gradient, found when
comparing our results with those of B06, is the radial migration of
stars seen in recent simulations of isolated disk formation and
evolution \citep{Roskar_etal08a, Minchev_etal11}. In these
simulations, inside-out disk growth yields a negative age gradient
within the break radius (2--3 disk scale length), after which there is
a positive age gradient due to the secular redistribution of stars,
given their interactions with transient spiral density waves. Of
course, we should keep in mind that what we presented here are the
results of one field in the inner regions of M31 and we need more
observations and statistics to either confirm or rule out what we
suggest. The multi-cycle Panchromatic Hubble Andromeda Treasury (PHAT)
project, which will cover 1/3 of M31 with HST WFC3 and ACS
observations, will resolve the SFH of the disk of M31: our
observations and analysis merely hint at what PHAT is likely to find.
\section{Summary and Conclusions}
\label{sec:conclusions}
We used deep HST ACS/HRC observations to derive the SFH of M32 for the
first time from a detailed modeling of its CMD. The two fields
observed, one closer to M32 (F1) and a background M31 field (F2), were
introduced and used in Paper I to construct deep CMDs of F1 and F2,
and the deepest optical CMD of M32 yet obtained. The IAC-POP/MinnIAC
method was used here to compare the distribution of stars in the
observed CMDs of F1 and F2 with that of a model CMD. We obtained the
SFH of M32 by linearly subtracting the SFHs of F2 from that of F1. The
use of different stellar evolutionary libraries (BaSTI and
Padova/Girardi) and assumptions of binary fractions (0, 0.35, 0.7, and
1) did not significantly modify the solutions obtained, indicating
that our results are robust.
Combining our present results with those of Paper I, we provide an
unprecedented census of the stellar content of M32. The derivation of
the SFH presented in this paper is independent of the analysis
performed in Paper I. In spite of using the same data, the CMD regions
that we have probed in this work are largely different from those used
in Paper I. Our analysis of these regions have allowed us to obtain
detailed information about the young and intermediate-age populations
of M32, whereas only the broadest sketch of these populations were
possible in Paper I. Conversely, detailed information about the older
populations cannot be obtained with our current approach, and
therefore we rely on the qualitative results of Paper I for those
populations.
The main finding of this work is that M32 is composed of two main
dominant populations at F1: $\sim 40\% \pm 17\%$ of the mass in a 2--5
Gyr old, metal-rich population and $\sim 55\% \pm 21\%$ of the total
mass in stars older than 5 Gyr, with slightly subsolar
metallicities. Its mass-weighted mean age and metallicity are
$\langle\mathrm{Age}\rangle=6.8\pm1.5\,\mathrm{Gyr}$ and
$\langle\mathrm{[M/H]}\rangle=-0.1\pm0.08\,\mathrm{dex}$,
respectively. Even though we are unable to specify when the star
formation started in M32 at F1, we make use of the analysis of Paper I
to constrain the older population. We know from the RC, RGb and AGB
bumps that the bulk of the old population is 8--10 Gyr old. Thus, we
do not expect a significant contribution from stars older than 10 Gyr
in M32. There has been, however, a marginal detection of RR Lyrae
belonging to M32 at F1, which reveal the presence of a few ancient
metal-poor stars in M32 \citep{Fiorentino_etal10}. The remaining
$\sim$ 4\% of the mass is distributed in genuine young metal-rich
stars ($\sim 2$\%) and young metal-poor stars ($\sim 2$\%) which we
associate with blue straggler stars belonging to an old metal-poor
population. In addition, we used the inferred SFH of M32 to calculate
the SSP-equivalent age and metallicity parameters from unresolved
stellar population models, which are $2.89 \pm 0.15$ Gyr and
$[\mathrm{M/H}]=0.02 \pm 0.01$ dex, respectively. These values,
however, contradict spectroscopic studies, which show a steep age
gradient from M32's center to $1\,r_e$.
Based on our present results, it is not currently possible to choose a
preferred model for M32's origins between two popular ones: a true
low-luminosity elliptical or a former spiral galaxy whose bulge
survived its dynamical interaction with M31. Future observations to
find M32-analog systems as well as simulations of spheroidal systems
with similar SFHs to M32 may shed light on this issue.
On the other hand, the inferred SFH for F2 shows that the stellar
populations of the inner regions of the disk and spheroidal components
of M31 are older and more metal-poor than M32. Its mass-weighted mean
age and metallicity are
$\langle\mathrm{Age}\rangle=9.15\pm1.2\,\mathrm{Gyr}$ and
$\langle\mathrm{[M/H]}\rangle=-0.10\pm0.10\,\mathrm{dex}$,
respectively. F2 has two main components: $65\% \pm 9\%$ of the mass
composed by a 8--14 Gyr old more metal-poor population and $30\% \pm
7.5\%$ of the mass in more metal-rich stars of 5--8 Gyr old. There is
a small contribution from stars younger to 5 Gyr to the total mass.
The inner disk and spheroidal stellar populations seem to be
indistinguishable from those of the outer disk and spheroid. Assuming
that M31's disk at F2 ($\sim$ 1 disk scale length) has a mean age
between $\sim$ 5 and 9 Gyr, our results are in agreement with
inside-out disk formation models. But of course, we need more
observations and statistics to confirm or rule out this suggestion.
Lastly, while this paper accounts for the SFH history of the bulk of
M32's mass, it does not offer strong constraints on small ``tracer''
populations that may testify to the present level of any very recent
star formation in M32, as well as fossil remnants that may date to
times well before the 10 Gyr mark, after which the bulk of M32 stars
were formed. Digging down to fainter stars in the M32 CMD to detect
any very old ( $>10$ Gyr) MSTO is not possible with HST, or any
instrument presently under development, given that the M31 background
prevents observation at lower surface brightness levels where the HST
angular resolution would be more effective. Instead, we believe the
best hope of detecting any ancient, metal-poor, remnant from the very
first stages in the M32 progenitor would come from unambiguously
detecting RR Lyrae stars at higher surface brightness levels than were
observed in F1, where the M31 contamination is relatively much weaker.
Our own image simulations show that RR Lyraes can be detected with HST
at substantially smaller radii than the F1 location. Likewise, better
constraints on a $< 2$ Gyr population native to M32 may be provided by
bright, if rare, tracers that can still be recognized at much higher
M32 surface brightnesses than were observed in the present work. As
with RR Lyraes, bright young blue MS stars or AGB stars should be
detectable throughout the body of M32. \\
\acknowledgments We thank Reynier Peletier, Eline Tolstoy and Antonio
Aparicio for their valuable comments on an early version of this
paper. A.M. wishes to thank the hospitality of the Instituto de
Astrof\'isica de Canarias and the Department of Physics and Astronomy
of Michigan State University, where part of this work was carried
out. We thank the anonymous referee for the very careful reading of
the manuscript and comments which helped to improve this paper. This
work has made use of the IAC-STAR synthetic CMD computation
code. IAC-STAR is supported and maintained by the computer division of
the Instituto de Astrof\'isica de Canarias. NOVA is acknowledged for
financial support. Support for program GO-10572 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,
Inc., under NASA contract NAS 5-26555.
\emph{Facility:} \facility{HST (ACS)}
\bibliographystyle{apj}
|
\section{Introduction}
\label{introduction}
Recently topological insulator has attracted much attention \cite{hasan10,qi10_QSH,qi10_topo,moore10,hasan11}.
Topological insulator has been firstly predicted
in a graphene with spin-orbit interaction
\cite{kane05QSH}, which consists of two copies of quantum Hall system, and shows a quantized spin Hall effect if $z$--component of spin of electrons is conserved.
Generally speaking, the present non--trivial system is characterized by $\mathbb Z_2$--index introduced by Fu, Kane, and Mele \cite{kane05Z2,fu06}, and has a gapless helical edge mode where spin current protected by time--reversal symmetry flows spontaneously.
However, quantum spin Hall (QSH) phase of graphene has not been observed experimentally since the spin--orbit interaction that is the driving force of topological insulator is much small.
After that,
HgTe/HgCdTe quantum well has been theoretically proposed as a candidate of two--dimensional topological insulator \cite{bernevig06}, and confirmed experimentally \cite{koenig07,roth09,brune11}.
Topological insulators have been realized in three dimensional systems \cite{fu07_3D,fu07_inversion,moore07}, {\it e.g.}, Bi$_{1-x}$Sb$_{x}$ alloy \cite{hsieh08}, the binary compounds Bi$_2$Se$_3$, Bi$_2$Te$_3$ \cite{zhang09,liu10_model,xia09,chen09}, and Tl--based ternary compound TlBiSe$_3$ \cite{yan10,lin10,sato10}.
Furthermore,
the quaternary compounds have also been theoretically predicted \cite{chen11,wang11arXiv}.
All of these systems have a single helical Dirac cone on the surface.
Nowadays, many exotic quantum phenomena are expected originating from surface states of three--dimensional topological insulators \cite{hasan10,qi10_QSH,TYN09,Linder10a,Linder10b2,Yokoyama}.
However, in the actual systems,
sufficient amount of carriers remain in the bulk due to
the difficulty of fabrication of samples \cite{hyde74}.
Then
the system becomes metallic and it is difficult to classify physical properties specific to surface Dirac cone \cite{chen09,taskin09,checkelsky09}.
To resolve this problem,
several approaches, \textit{e.g.,} chemical doping and surface adsorption \cite{hsieh09} have been performed.
The another new approach to control the carrier is
to fabricate high quality thin films, where carrier control by gating is possible
\cite{chen10PRL,chen10PRB}.
But
thin films may have the different electronic states from that of the bulk.
Especially, the surface states have an energy gap due to the hybridization between the Dirac cones on top and bottom on the film induced by finite--size effect.
Based on above backgrounds,
experimental studies of thin films of topological insulators have started \cite{zhang10, sakamoto10, hirahara10}.
Besides this,
there have been many theoretical studies based on
continuous models \cite{linder09, lu10, shan10}, first principle calculations \cite{liu10_oscillatory, park10, yazyev10, jin11, chang11}, and tight--binding model calculation \cite{liu10_oscillatory}.
Although
continuous model is simple,
it is
valid only for the long wavelength and low--energy limits.
First principle calculation gives detailed electronic states of thin film.
But
it is difficult to analyze
complicated phenomena, \textit{e.g.},
transport properties, disorder effects, and quantum many--body problems.
On the other hand,
tight--binding approach
is useful to calculate these interesting phenomena numerically, because many--body interaction and impurity effects are easily taken into account.
However, electronic properties of thin film of topological insulator have not been fully studied based on tight--binding model.
In the present paper, we study
electronic properties of thin film of
Bi$_2$Se$_3$ based on
a tight--binding model focusing on the
film--thickness dependencies of energy gap and surface states.
It is revealed that
the magnitude of energy gap is seriously influenced by material parameters.
The paper is organized as follows.
In section 2, we introduce a tight--binding model based on the Hamiltonian proposed by Refs. \cite{zhang09,liu10_model}.
In section 3, we calculate
energy spectrum of thin film for various number of quintuple layers
by changing material parameters.
In section 4, we conclude our results.
\section{Model}
\label{model}
We use the effective model derived in Refs. \cite{zhang09,liu10_model}
as
\begin{eqnarray}
H(\mathbf k) =
\mathcal{E}({\bf k}) + \left(
\begin{array}{cccc}
{\cal M}({\bf k}) & 0 & B_0 k_z & A_0 k_- \\
0 & {\cal M}({\bf k}) & A_0 k_+ & -B_0 k_z \\
B_0 k_z & A_0 k_-&-{\cal M}({\bf k})&0 \\
A_0 k_+ & -B_0 k_z & 0 & -{\cal M}({\bf k}) \\
\end{array}
\right),
\end{eqnarray}
with
\begin{eqnarray}
\mathcal{E}({\bf k}) &=& C_0+C_1k_z^2+C_2(k_x^2+k_y^2), \\
{\cal M}({\bf k}) &=& M_0+M_1k_z^2+M_2(k_x^2+k_y^2),
\end{eqnarray}
where $k_\pm=k_x\pm\im k_y$, and the base is taken as
$(|+,\uparrow\rangle, | +, \downarrow \rangle, |- , \uparrow \rangle,
| -, \downarrow \rangle)$, in which $\pm$ and $\uparrow(\downarrow)$
denote the parity eigenvalue and spin respectively.
Let us introduce the lattice model only with nearest neighbor hoppings in a tetragonal lattice with substitution as
\begin{eqnarray}
k_i a_i \to \sin k_i a_i, \quad (k_ia_i)^2 \to 2 (1-\cos k_i a_i).
\end{eqnarray}
As a result,
the bulk Hamiltonian is derived as
\begin{eqnarray}
H(\mathbf k) =
\tilde\mathcal{E}({\bf k}) + \left(
\begin{array}{cccc}
\tilde {\cal M}({\bf k}) & 0 & B_0 \sin k_zc & \bar A_- \\
0 & \tilde{\cal M}({\bf k}) & \bar A_+ & -B_0 \sin k_zc \\
B_0 \sin k_zc & \bar A_- &- \tilde{\cal M}({\bf k})&0 \\
\bar A_+ & -B_0 \sin k_zc & 0 & -\tilde{\cal M}({\bf k}) \\
\end{array}
\right),
\label{Hbulk}
\end{eqnarray}
where
\begin{eqnarray}
\tilde\mathcal E(k_x,k_y) &=& \bar C_0 + 2 \bar C_1(1- \cos k_z c ) + 2\bar C_2(2-\cos k_xa-\cos k_ya) ,\\
\tilde {\cal M}({\bf k}) &=& \bar M_0 + 2 \bar M_1 (1- \cos k_z c ) + 2 \bar M_2 (2 - \cos k_x a - \cos k_y a),
\\
\bar A_{\pm}(k_x,k_y) &=&
\bar A_0 (\sin k_xa \pm \mathrm i \sin k_ya),
\end{eqnarray}
with $a_i$ ($ a \equiv a_x=a_y, \, c \equiv a_z$) being the lattice constant along $i(=x,y,z)$--direction.
The relation between the original parameters
and
those in the present model
is
\begin{eqnarray}
\bar M_0 = M_0, \bar C_0 = C_0,
\bar M_1 = M_1/c^2, \bar C_1 = C_1/c^2,
\nonumber\\
\bar M_2 = M_2/a^2, \bar C_2 = C_2/a^2,
\bar A_0 = A_0/a, \bar B_0 = B_0/c.
\end{eqnarray}
In the following,
we express
the present Hamiltonian in real space along $z$--direction
perpendicular
to the quintuple layers
to focus on the surface states.
Here,
translational invariance is
satisfied for the direction parallel to the quintuple layers
i.e., $x$-- and $y$--directions.
Then $k_x$ and $k_y$ are good quantum numbers.
We apply open boundary condition only along $z$--direction, {\it i.e.}, the system is regarded as a one--dimensional chain for fixed $(k_x,k_y)$.
This condition corresponds to (111) cleavage surface of actual ${\rm Bi_2Se_3}$ which is easily cleaved.
The corresponding Hamiltonian is given as follows,
\begin{eqnarray}
H(k_x,k_y) &=& \sum_{n=1}^{N_z}
c^\dag_n(k_x,k_y) H_0(k_x,k_y) c_{n}(k_x,k_y)
\nonumber\\ && \quad +
\sum_{n=1}^{N_z-1}
\left[
c^\dag_n(k_x,k_y) H_1 c_{n+1}(k_x,k_y)
+
\mathrm{h.c.}
\right],
\label{tb}
\end{eqnarray}
where $N_z$ denotes number of quintuple layers.
It is noted that a lattice point $n$ in the above Hamiltonian
corresponds to position of a
quintuple layer in the actual crystal structure.
The on--site energy is given by
\begin{eqnarray}
H_0(k_x,k_y) = \left(
\begin{array}{cccc}
\bar{\cal M} + \bar{\cal E} & 0&0& \bar A_{-} \\
0 & \bar{\cal M} + \bar{\cal E} & \bar A_+ &0\\
0 & \bar A_- & -\bar{\cal M} + \bar{\cal E} &0 \\
\bar A_+ &0&0& -\bar{\cal M} + \bar{\cal E}
\end{array}
\right),
\label{H0}
\end{eqnarray}
with
\begin{eqnarray}
\bar\mathcal E(k_x,k_y) = \bar C_0 + 2 \bar C_1 + \bar C_2(2-\cos k_xa-\cos k_ya),\\
\bar\mathcal M(k_x,k_y) = \bar M_0 + 2 \bar M_1 + \bar M_2(2-\cos k_xa-\cos k_ya),
\end{eqnarray}
and the hopping between the nearest layers is as follows
\begin{eqnarray}
H_{1} = \left(
\begin{array}{cccc}
-\bar M_1 - \bar C_1 & 0 & \im \bar B_0/2 & 0 \\
0 & - \bar M_1 - \bar C_1 & 0 & - \im \bar B_0/2 \\
\im \bar B_0/2 & 0 & \bar M_1 - \bar C_1 & 0 \\
0 & - \im \bar B_0/2 & 0 & \bar M_1 - \bar C_1 \\
\end{array}
\right).
\end{eqnarray}
Since the Hamiltonian $H(k_x,k_y)$ has an inversion symmetry, it follows that $[H,P]=0$, or equivalently $PH(k_x,k_y)P^{-1}=H(-k_x,-k_y)$, where the parity operator $P$ is defined by
\begin{eqnarray}
P&=& \sum_{n=1}^{N_z} c_n^\dagger (k_x,k_y) {\rm diag}[1,1,-1,-1] c_{N_z+1-n}(k_x,k_y).
\end{eqnarray}
By using the parity operator $P$, we can derive the topological invariants $\nu$, which can be deduced from the parity of each pair of Kramers degenerate occupied energy band at the four time-reversal points at $\Gamma_\alpha$ ($\Gamma_1=(0,0)$, $\Gamma_2=(\pi,0)$, $\Gamma_3=(0,\pi)$, $\Gamma_4=(\pi,\pi)$,) in the Brillouin zone,
\begin{eqnarray}
(-1)^{\nu}&=&\prod_{\alpha=1}^{4} \prod_{m=1}^{N_z}\Braket{\phi_{2m}(\Gamma_\alpha)|P|\phi_{2m}(\Gamma_\alpha)},
\end{eqnarray}
where $\phi_m(\Gamma_\alpha)$ is the eigenvector of the Hamiltonian $H(\Gamma_\alpha)$, and $\Braket{\phi_{m}(\Gamma_\alpha)|P|\phi_{m}(\Gamma_\alpha)} (=\pm1)$ is the eigenvalue of parity operator $P$.
\section{Results and discussions}
We numerically obtain the eigenvalues and eigenvectors of bulk and surface states, diagonalizing the Hamiltonian given by eq. (\ref{tb}).
The value of parameters $\bar M_0, \bar M_2, \bar A_0, \bar C_0, \bar C_2$ are
the same as in Ref. \cite{liu10_model} with using $a=4.14 \AA$.
The values of $\bar M_1$ and $\bar C_1$ are determined
so that the eigen--energy at $Z$--point in Brillouin zone
coincides with that of first principle calculation in Ref. \cite{liu10_model}.
The value of $\bar{B_0}$ is chosen
in order to fit
the dispersion
along $\Gamma-Z$ line
as well as possible.
\begin{figure*}
\centering
\includegraphics{p2_EV_outline.eps}
\caption{Energy spectra of the bulk and surface states near $\bar{\Gamma}$-point in the slab geometry for $N_z= 3, 5,9$, and 16 quintuple layers.
The material parameters are set as $\bar M_0 = -0.28 \, \mathrm{eV}, \, \bar M_1=0.216 \, \mathrm{eV}, \, \bar M_2 = 2.60 \, \mathrm{eV}, \, \bar A_0 = 0.80 \, \mathrm{eV}, \, \bar B_0 = 0.32 \, \mathrm{eV}, \, \bar C_0 = -0.0083 \, \mathrm{eV}, \, \bar C_1 = 0.024 \, \mathrm{eV}, \, \bar C_2 = 1.77 \, \mathrm{eV}, \, a = 4.14 \, \AA, \,$ and $c= 9.55 \, \AA$.
}
\label{EV}
\end{figure*}
The indirect energy gap in the bulk Hamiltonian is located between $-0.071 \, {\rm eV}$ and $|\bar M_0 + \bar C_0| = 0.29 \, \rm eV$,
as derived from eq. (\ref{Hbulk}).
Figure \ref{EV} shows the energy spectrum for a slab geometry in the cases of $N_z = 3,5,9,$ and 16.
We can clearly see that eigenstates exist within the bulk energy gap.
These states can be regarded as surface states, which we can directly confirm from its density distribution localized in the vicinity of surface, as shown in Figure \ref{rho}.
The surface states have
a large magnitude of energy gap $E_{\rm g} \sim 0.033 \, \rm eV$
for $N_z=3$
since the two wave functions localized at the top and bottom surfaces overlap significantly.
(see (a) in Figure \ref{EV}).
The magnitude of the present energy gap
becomes small with the increase of $N_z$.
((b) and (c) in Fig. \ref{EV}.)
For $N_z=16$, the resulting $E_{\rm g}$ is significantly reduced
to be $0.15 \times 10^{-6} \, \rm eV$
((d) in Figure \ref{EV}).
Moreover,
it is noted
that the shape of valence subband depends on $N_z$.
For $N_z=3$ ((a) in Fig. \ref{EV}), there are two valence subbands in $-0.4 \, {\rm eV}< E < 0 {\rm eV}$.
The upper subband consists mainly of surface states
since it is located in the bulk energy gap.
The lower one, that is bulk energy band, is located at $\sim -0.3 \rm eV$.
The new subband appears between these two subbands at $\sim -0.2 \rm eV$ for $N_z=5$ ((b) in Fig. \ref{EV}).
Simultaneously,
valence subbands have
a local minimum at $k=0$,
and an indirect energy gap is generated.
For $N_z=9$ ((c) in Fig. \ref{EV})
there are much more subbands.
The energy bands
for $N_z=16$ as shown in (d) in Fig. \ref{EV} is almost similar to that of bulk three-dimensional topological insulator with $N_z \to \infty$.
\begin{figure}
\centering
\includegraphics[scale=1.]{p2_gap.eps}
\caption{The magnitude of energy gap $E_{\rm g}$ of the surface state as a function of $N_z$. The material parameters are the same as in Figure \ref{EV}.}
\label{gap}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=1.]{p2_rho.eps}
\caption{Density distribution $\rho(z)$ of surface state at $(k_xa, k_ya) = (\pi/32,\pi/32)$ for $N_z=16$. There are two surface states located at the top and bottom in the system. $z$ denotes the position of layer. The guided lines are drawn for view-ability. The material parameters are the same as in Figure \ref{EV}.}
\label{rho}
\end{figure}
In Figure \ref{gap},
$N_z$ dependence of $E_{\rm g}$ is plotted.
Hamiltonian of the monolayer system with $N_z=1$ is given by $H_0(k_x, k_y)$ (see eq. (\ref{H0})), which is equivalent to that of HgTe/HgCdTe quantum well \cite{bernevig06},
and $E_{\rm g}$ is given by
$2 |\bar M_0 + 2 \bar M_1| (= 0.30 \mathrm{eV})$.
$E_{\rm g}$ for $N_z=2$ is also derived analytically as $|2 \bar M_1 - [{\bar B_0^2 + 4(\bar M_0 + 2 \bar M_1)^2}]^{1/2} | (=0.0094 \mathrm{eV})$.
For $N_z = 2,3$, and $4$,
$E_{\rm g}$ decreases roughly exponentially as a function of $N_z$, and becomes $E_{\rm g} \sim 10^{-1} \, \rm eV$.
$E_{\rm g}$ becomes much smaller than room temperature for $N_z \geq 5$.
It is also noted that
$E_{\rm g}$ has an oscillatory behavior
as a function $N_z$
whose period is almost $3$.
The similar behavior has been obtained
based on continuous models \cite{linder09, lu10, shan10} and first principle calculations \cite{liu10_oscillatory, park10, yazyev10, jin11, chang11}.
Next, we investigate the relation between the magnitude of energy gap and wave functions.
Figure \ref{rho} shows
the density distribution of surface states for $N_z=16$, which is defined by
\begin{eqnarray}
\rho(z) = \left\langle c^\dag_z(k_x,k_y) c_z(k_x,k_y) \right\rangle,
\quad
z = 1, \cdots, N_z,
\end{eqnarray}
where the expectation value is evaluated for the surface state with momentum $(k_xa,k_ya) = (\pi/32, \pi/32)$.
The solid (dashed) line denotes the density distribution of surface state located on the top $z=1$ (bottom $z=16$).
The density distribution decays exponentially with oscillation whose period is nearly 3 quintuple layers.
This period is almost the same as that of $N_z$ dependence of $E_{\rm g}$.
Since
the two wave functions located at the top and bottom surfaces oscillate spatially,
the resulting $E_{\rm g}$ due to overlap between them also oscillates as a function of $N_z$.
In the following, we focus on the material parameters dependencies of $E_{\rm g}$.
For simplicity, we fix all parameters except for $\bar M_1$.
Here, we choose seven cases of $\bar M_1$ as shown in Figure \ref{souzu}.
In order to understand
electronic properties for the corresponding seven cases we have chosen,
we show the phase diagram of the system for $N_z=1$ and $N_z=\infty$ in Figure \ref{souzu}.
In the limit for $N_z = \infty$,
the system becomes weak topological insulator
(WTI) for $\bar M_1 < - \bar M_0/4 = 0.07 {\rm eV}$
while it becomes strong topological insulator (STI)
for $\bar M_1 > - \bar M_0/4 = 0.07 {\rm eV}$.
On the other hand,
in the limit for $N_z=1$
the present system is
QSH
for $\bar M_1 < - \bar M_0/2 = 0.14 {\rm eV}$,
while ordinary insulator (OI) for $\bar M_1 < - \bar M_0/2 = 0.14 {\rm eV}$.
\begin{figure}
\centering
\includegraphics[scale=0.5]{souzu.eps}
\caption{Phase diagrams of topological insulator for bulk limit $N_z = \infty$ and for monolayer $N_z=1$.
STI (WTI) denotes strong (weak) topological insulator
for $N_z=\infty$.
QSH and OI denotes quantum spin Hall insulator
where spin Hall conductance is quantized and ordinary insulator respectively for $N_z=1$.
${\bar M_1}=0.016, 0.070, 0.116, 0.140, 0.216, 0.316$, and $0.416 {\rm eV}$
for (a), (b), (c), (d), (e), (f), and (g), respectively.
These values
correspond to those used in Figure \ref{gap_all}.
}
\label{souzu}
\end{figure}
\begin{figure}
\centering
\includegraphics{gap_all.eps}
\caption{The magnitude of energy gap $E_{\rm g}$ as a function of $N_z$ for different values of ${\bar M_1}$.
The closed (open) circle describes a two--dimensional topological invariant $\nu=0 (\nu=1)$.
At the case (d), $E_g$ vanishes for odd numbers of layers: $N_z=1,3,\cdots, 15$.
}
\label{gap_all}
\end{figure}
Figure \ref{gap_all} shows $E_{\rm g}$ for various values of $\bar M_1$.
The curve in Figure \ref{EV} coincides with the curve (e)
obtained for $\bar M_1 = 0.216 {\rm eV}$ in Figure \ref{gap_all},
which is the same parameter as that of Bi$_2$Se$_3$.
The curve (e)
has a three-fold periodic damped oscillation.
For $\bar M_1 = 0.016 {\rm eV}$ (case (a) in Fig. \ref{gap_all}), where the system is WTI (QSH) with $N_z=\infty$ ($N_z=1$),
$E_{\rm g}$ does not decay with the increase of $N_z$
since there is no gapless surface Dirac cone at $\bar{\Gamma}$-point
for $N_z=\infty$.
For $\bar M_1 = 0.070 {\rm eV}$ (case (b) in Fig. \ref{gap_all}),
which is the transition point between WTI and STI,
where
closing of the bulk energy gap occurs at $\Gamma$--point,
$E_{\rm g}$ decreases monotonically as a function of $N_z$.
For $\bar M_1 = 0.116 {\rm eV}$ (case (c) in Fig. \ref{gap_all}),
$E_{\rm g}$ decays exponentially except for $N_z < 5$ as a function of $N_z$,
since surface Dirac cone is generated with $N_z = \infty$.
$E_{\rm g}$
has a strong oscillation for $\bar M_1 = 0.140 {\rm eV}$ (case (d) in Fig. \ref{gap_all}),
where transition between OI and QSH occurs for $N_z=1$.
$E_{\rm g}$
becomes exactly zero for odd numbers of $N_z$ (See Appendix).
The damped oscillation with four--fold periodicity
appears at $\bar M_1=0.316 {\rm eV}$ (case (f) in Fig. \ref{gap_all}) and at $\bar M_1=0.416 {\rm eV}$ (case (g) in Fig. \ref{gap_all}).
When the system is QSH for $N_z=1$,
$E_{\rm g}$ decreases monotonically in the wide parameter range of $N_z$.
On the other hand,
$E_{\rm g}$ shows a damped oscillation as a function of $N_z$
for
$\bar M_1 > 0.140 {\rm eV}$, {\it i.e.},
the system is OI
in the thin thickness limit.
As we have seen above,
the period of oscillation
depends on $\bar M_1$.
It can be concluded that
the
$N_z$ dependence of
$E_{\rm g}$ is sensitive to the material parameter $\bar M_1$.
We also show the topological invariant $\nu$ in Fig.\ref{gap_all}. The closed circle expresses non-topological phase with $\nu=0$, while the open circle expresses topological phase with $\nu=1$.
For cases with (a), (b) and (c), non-topological phase emerges for even number of layers and topological phase emerges for odd number of layers.
Topological phase and non-topological phase appear oscillatory also for cases with (e), (f) and (g).
However, the period of oscillation becomes three or four. These results are consistent with those by Liu $et$ $al.$ \cite{liu10_oscillatory}.
Since the magnitude of energy gap becomes zero at the boundary between topological and non-topological phases, the period of oscillation of topological number $\nu$ coincides with that of the energy gap.
\section{Conclusion}
We obtain the energy spectrum of
surface states in a topological insulator based on tight--binding model.
It is clarified that
there are various types of thickness dependencies of $E_{\rm g}$.
The origin of
the dumped oscillatory behavior of $E_{\rm g}$ is partitioned into two parts.
The dumped behavior
appears when the gapless surface Dirac cone is realized in the limit of
$N_z= \infty$.
The oscillatory behavior of $E_{\rm g}$
becomes prominent
when the system approaches to OI regime for $N_z=1$.
Based on these results,
we would expect various types of thin films by controlling material parameters, which could be controlled by external pressure along z-direction (c-axis) for $\rm{Bi_2Se_3}$.
Tuning of material parameters may be much more easier for optical lattices made from cold atoms \cite{Goldman_prl_neutral,Bermudez_arxiv_trap,Bermudez_arxiv_defect,Bermudez_arxiv_sweep,Goldman_prl_nonabelian,Bermudez_arxiv_honeycomb}.
If we can tune the corresponding material parameters with the case (d) in Fig.4, the strong even-odd effect is expected. In this case, various transport properties are sensitive to external fields.
There are several future unresolved problems.
The Anderson localizations of thin films have been recently studied \cite{hirahara10,liu11,wang11PRB} from various aspects.
It is interesting to study this problem with various types of thin films
with different electronic properties
of topological insulator with different material parameters.
\section{Acknowledgments}
This work was supported
in part by a Grant-in-Aid for Scientific Research from MEXT
of Japan, ``Topological Quantum Phenomena" No. 22103005
and No. 22340096.
\section*{Appendix}
As shown in Fig. \ref{gap_all}, the energy gap is exactly zero for the odd number of layers, and non-zero for the even number of layers at the case (d) in Fig. \ref{souzu}. In the following, we derive this behavior of the energy gap.
The Hamiltonian for $N_z$ at $\bar \Gamma$ point reads
\begin{eqnarray}
H_{N_z}(0,0) = F_{N_z}+G_{N_z},
\end{eqnarray}
with $4N_z\times 4N_z$ matrices $F_{N_z}$ and $G_{N_z}$ being
\begin{eqnarray}
F_{N_z}&=&\sum_{n=1}^{N_z} c^\dag_{n} {\rm diag}\left[ ( {\bar C_0}+2{\bar C_1}),\cdots,({\bar C_0}+2{\bar C_1})\right] c_{n}\\
G_{N_z}&=&\sum_{n=1}^{N_z-1}
\left[ c^\dag_{n} H_1 c_{n+1} + \mathrm{h.c.}
\right].
\end{eqnarray}
where ${\bar M_0}+2{\bar M_1}=0$ at the case (d). The eigenvalue of $ H_{N_z}(0,0) $ equals to ${\bar C_0}+2{\bar C_1}+E^G_n$ where $E^G_n$ is the eigenvalue of $G_{N_z}$.
If matrix $G_{N_z}$ has zero eigenvalue, the energy gap closes because $G_{N_z}$ has particle hole symmetry .
We show $|G_{N_z}| \equiv \det G_{N_z}=0$ for odd number of $N_z$ at $\bar \Gamma$ point, as follows.
\begin{eqnarray}\nonumber
|G_{N_z}|&=&\left|
\begin{array}{cccccc}
0 &H_1 &0& &&\\
H_1^\dagger&0&H_1&&&\\
0&H_1^\dagger&0&\ddots&&\\
&&\ddots&\ddots&\ddots&\\
&&&\ddots&0&H_1\\
&&&&H_1^\dagger&0
\end{array}
\right|=\left|
\begin{array}{cccccc}
H_1 &0&0& &&\\
0&H_1^\dagger&H_1&&&\\
H_1^\dagger&0&0&\ddots&&\\
&&H_1^\dagger&\ddots&\ddots&\\
&&&\ddots&0&H_1\\
&&&&H_1^\dagger&0
\end{array}
\right|\\\nonumber
&=&\left| H_1 \right| \left|
\begin{array}{ccccc}
H_1^\dagger&H_1&&&\\
0&0&\ddots&&\\
&H_1^\dagger&\ddots&\ddots&\\
&&\ddots&0&H_1\\
&&&H_1^\dagger&0
\end{array}
\right|=\left| H_1\right|^2\left|
\begin{array}{cccc}
0&H_1&&\\
H_1^\dagger&\ddots&\ddots&\\
&\ddots&0&H_1\\
&&H_1^\dagger&0
\end{array}
\right| \\
&=&\left| H_1\right|^2 |G_{N_z-2}|,
\end{eqnarray}
and we find
\begin{eqnarray}
\left|G_{N_z=1}\right|&=& 0,\\
\left|G_{N_z=2}\right|&=& \left|
\begin{array}{cc}
0&H_1\\
H_1^\dagger&0
\end{array}
\right|=\left| H_1\right| ^2 \neq 0.
\end{eqnarray}
Thus the energy gap at the case (d) is exactly zero for the odd number of layers, and non-zero for the even number of layers, and oscillates strongly as a function of $N_z$.
|
\section{Introduction}
The remarkable success of the mean-field Bardeen-Cooper-Schrieffer (BCS) theory in describing classic superconductors \cite{BCS57,Parks69}
may be attributed to the fact that an infinitesimal attraction between electrons suffices to form Cooper pairs responsible for superconductivity.\cite{Cooper56}
On the other hand, we now have an increasing number of superconductors and Fermi superfluids where relevant interactions are stronger beyond the applicability of the mean-field theory. They include superfluid $^{3}$He at high pressure, \cite{Leggett75,Vollhardt90} heavy-fermion superconductors with competing fluctuations,\cite{Flouquet05,Misra08,Pfleiderer09} high-$T_{c}$ cuprate superconductors,\cite{BK03} and the BCS-BEC (Bose-Einstein condensation) crossover in trapped atomic gases.\cite{GPS08}
The present paper is devoted to developing a convenient systematic method to describe those systems based on many-body quantum field theory.
To be specific, we will focus on a two-body interaction and extend the Luttinger-Ward self-consistent perturbation expansion for normal states\cite{LW60} to superconductors in such a way that, given a normal-state approximation, all the pair (i.e., ``anomalous'') processes derivable from it are automatically incorporated. Reproducing the Hartree-Fock theory as the lowest-order approximation, the Luttinger-Ward formalism enables us to include correlation effects systematically and microscopically up to a desired order. Its key ingredient is a functional $\Phi[G]$ of the single-particle Green's function $G$ described by closed skeleton Feynman diagrams, as properly identified and called ``$\Phi$-derivable approximations'' by Baym.\cite{Baym62} Indeed, choosing an approximate $\Phi$ enables us to calculate the whole thermodynamic series ranging from the thermodynamic potential and single-particle Green's function to the two-particle and higher-order Green's functions.\cite{Kita10} To put it another way, $\Phi$ uniquely determines how to resolve the Bogoliubov-Born-Green-Kirkwood-Yvon (BBGKY) hierarchy.\cite{Cercignani88,Kita10} An additional prominent advantage of the formalism is that it automatically satisfies conservation laws, as shown by Baym. \cite{Baym62} Hence, one can also study non-equilibrium phenomena by only changing the imaginary-time Matsubara contour into the real-time Schwinger-Keldysh contour. \cite{Schwinger61,Keldysh64,HJ08,Rammer07,Kita10}
It should be noted that a formal extension of the Luttinger-Ward theory to superconductors was performed by de Dominicis and Martin already in 1964.\cite{dDM64} However, they did not present any practical method of how to efficiently collect pair processes with anomalous Green's functions, whose number grows exponentially in comparison with normal diagrams as we proceed to higher orders. For example, (a)-(j) in Fig.\ \ref{Fig1} exhaust the third-order diagrams for $\Phi$ in terms of the symmetrized vertex $\Gamma^{(0)}$ of Abrikosov {\em et al}.,\cite{AGD63} where anomalous diagrams (c)-(j) are already four times as large in number as normal diagrams (a) and (b).
\begin{figure}[b]
\begin{center}
\includegraphics[width=0.9\linewidth]{Fig1.eps}
\end{center}
\caption{Third-order diagrams for $\Phi$. A square (a small circle) represents the bare vertex $\Gamma^{(0)}$ ($\underline{\Gamma}^{(0)}$) that is symmetric in the space-spin coordinates (both the space-spin coordinates and particle-hole indices). The number below each of diagrams (a)-(j) denotes its relative weight. \label{Fig1}}
\end{figure}
Here, we develop a perturbation expansion of gathering all the pair processes systematically and concisely without drawing anomalous diagrams for both the singlet and triplet pairings or their mixtures. This is made possible by introducing a bare interaction vertex $\underline{\Gamma}^{(0)}$ that is further symmetrized in the particle-hole (i.e., creation-annihilation or ``charge'') indices. Thus, processes (a)-(j) in Fig.\ \ref{Fig1} can be represented by a single diagram on the right-hand side with a definite analytic expression, i.e., eq.\ (\ref{Phi_3}) below, which reproduces their weights correctly. The key functional $\Phi$ is defined as a closed skeleton expansion in terms of the matrix Green's function $\hat{G}$ in the Nambu space, which may be approximated by some selected terms from the whole series in practical calculations. Besides the Dyson-Gor'kov equation for $\hat{G}$ and an expression of the thermodynamic potential, we also derive the Bethe-Salpeter equation for the two-particle Green's function that can be solved based solely on $\Phi$. The formalism enables us to study not only superconductivity itself but also competing effects of charge, spin, and superconducting fluctuations or orders simultaneously through the matrix structure of $\underline{\Gamma}^{(0)}$. This last point may be regarded as a definite advantage of the present approach over the one based on the Stratonovich-Hubbard transformation,\cite{Hubbard59} where only a single fluctuation relevant to the auxiliary field may be taken into account after some standard approximation.
Using the formalism, we also extend the fluctuation exchange (FLEX) approximation originally developed for normal states\cite{BS89,BSW89,BW91} so as to incorporate all the pair processes for the spin-singlet pairing. The FLEX approximation has been successful in describing anomalous normal-state transport phenomena of high-$T_{c}$ superconductors as well as some organic and heavy-fermion superconductors.\cite{Kontani08} Its extension to superconductivity as described above, which has not been performed yet and will be called ``FLEX-S approximation'' below, will be useful for understanding those superconductors quantitatively.
This will be carried out in a transparent and concise manner in terms of $\underline{\Gamma}^{(0)}$ in contrast to the original derivation for normal states.\cite{BS89,BSW89,BW91}
Naturally, considerable efforts have been made to go beyond the BCS theory.
Eliashberg incorporated the electron-phonon interaction explicitly as the source of the attraction to derive ``strong-coupling'' equations of superconductivity within the second-order perturbation in terms of the renormalized electron Green's function and bare phonon propagator.\cite{Eliashberg60} Leggett seriously considered interactions between particles other than the pairing part to develop a theory of superfluid Fermi liquids at low temperatures.\cite{Leggett65}
Anderson and Brinkman \cite{AB73} presented the idea of ``feedback effects,'' i.e., a change of the pairing interaction through the superfluid transition,
to understand the A phase of superfluid $^3$He that is realized against the mean-field Balian-Werthamer theory.\cite{BW63}
The idea was elaborated by Brinkman {\em et al}.\ \cite{BSA74} and Kuroda,\cite{Kuroda75} who both incorporated the first few anomalous processes into the normal-state pairing interaction composed of the particle-hole ``paramagnon'' diagrams\cite{LF71} and evaluated the extra free energy as a power series in the $p$-wave energy gaps. Subsequently, Tewordt \cite{Tewordt74} refined this approach into a self-consistent $\Phi$-derivable approximation, where he included those pair processes that can be obtained from the normal particle-hole diagrams by successively replacing a pair of normal Green's functions connecting adjacent vertices by a pair of anomalous Green's functions; relevant diagrams in the third order are (a), (c), (f), and (j) of Fig.\ \ref{Fig1}.
Tewordt and coworkers\cite{Tewordt93,Tewordt95} later applied the formalism to the two-dimensional Hubbard model to clarify quasiparticle and spin excitations of high-$T_c$ superconductors.
The same formalism was used by other groups for the two-dimensional Hubbard model\cite{PB94,MS94,GLSB96} and $d$-$p$ model.\cite{LB93,KFY97,TM98}
On the other hand, Eagles\cite{Eagles69} and Leggett\cite{Leggett80} discussed the BCS-BEC crossover of the isotropic spin-singlet pairing at zero temperature by combining the BCS gap equation with the equation for the chemical potential. The transition temperature of the BCS-BEC crossover problem was calculated subsequently by Nozi\`eres and Schmitt-Rink\cite{NSR85} based on the Thouless criterion.\cite{Thouless60}
Haussmann and coworkers\cite{Haussmann93,Haussmann07} studied this problem with an improved $\Phi$-derivable approximation to include interaction effects among unpaired and paired fermions. Specifically, their $\Phi$ consists of the normal particle-particle ladder diagrams appropriate for low-density systems, plus those anomalous processes obtained from the formers by successively replacing two pairs of normal Green's functions connecting adjacent vertices by two pairs of anomalous Green's functions;
relevant diagrams in the third order are (b) and (h) of Fig.\ 1, disregarding (e) and (g) which may be important even within the ladder approximation.
We should also mention an alternative approach on the issue where some effective pairing interaction is used to solve the Eliashberg equations.\cite{CPS03} However, there is an ambiguity as to how to construct the effective interaction.
Similarly, the FLEX approximation has sometimes been augmented rather phenomenologically to explain the pseudogap phenomena observed in underdoped cuprate superconductors.\cite{DMT97,Yanase01}
To be specific, the ``FLEX+T-matrix'' approximation\cite{Yanase01} consists of (i) identifying the pairing interaction obtained by the FLEX approximation as the ``bare'' interaction in the Thouless criterion
and (ii) incorporating the contribution of the two-particle ``superconducting fluctuations'' additionally into the single-particle self-energy.
However, the procedure may contain some double counting of elementary processes and has yet to be examined on a firm microscopic basis.
Note in this context that the Thouless criterion was originally derived by the particle-particle ladder approximation in the bare perturbation expansion for a two-particle Green's function,\cite{Thouless60} reproducing the same $T_c$ equation as the mean-field BCS theory.
Thus, one may be convinced that we still do not have practical methods of incorporating all the anomalous processes that are naturally present below $T_c$.
This paper is organized as follows. In \S 2, we develop a basic formalism to calculate the thermodynamic potential, single-particle Green's function, and two-particle Green's function of superconductivity self-consistently based on a single functional $\Phi[\hat{G}]$, whose expression may be obtained by a concise perturbation expansion in terms of $\underline{\Gamma}^{(0)}$. We carry it out in the coordinate representation so that general inhomogeneous systems
may be handled.
In \S 3, we focus on homogeneous systems and transform the results of \S 2 into the momentum-``energy'' representation.
Closed equations in the FLEX-S approximation are also derived. Section 4 presents a brief summary to indicate main results.
The whole contents of the present paper may be regarded as the Fermi-superfluid counterpart of the self-consistent perturbation expansion recently developed for BEC.\cite{Kita09,Kita10b, Kita11}
We set $\hbar=k_{\rm B}=1$ throughout with $k_{\rm B}$ the Boltzmann constant.
\section{Basic formalism}
The system we consider consists of identical particles with mass $m$ and spin $1/2$ described by the Hamiltonian
\begin{equation}
H=H_{0}+H_{\rm int},
\label{Hamil}
\end{equation}
with
\begin{subequations}
\label{Hamil2}
\begin{equation}
H_{0}=\int {\rm d}\xi_{1} \psi^{\dagger}(\xi_{1})K_{1}\psi(\xi_{1}),
\label{H_0}
\end{equation}
\begin{equation}
H_{\rm int}=\frac{1}{2}\int {\rm d}\xi_{1}\int {\rm d}\xi_{2}\,\psi^{\dagger}(\xi_{1})
\psi^{\dagger}(\xi_{2}) V({\bm r}_{1}-{\bm r}_{2})
\psi(\xi_{2})\psi(\xi_{1}) .
\label{H_int}
\end{equation}
\end{subequations}
Here
$\xi_{1}\equiv ({\bm r}_{1},\alpha_{1})$ with ${\bm r}_{1}$ and $\alpha_{1}$ denoting the space and spin coordinates, respectively,\cite{AGD63}
$\psi^{\dagger}$ and $\psi$ are the creation and annihilation operators of the fermion field, respectively,\cite{AGD63}
$K_{1}\equiv -{\hbar^{2}\nabla_{1}^{2}}/{2m}-\mu$
with $\mu$ the chemical potential,
and $V$ is the interaction potential.
Though disregarded here, the effect of a lattice or trap potential may be included easily in $K_{1}$.
The formulation in this section will be carried out in the coordinate representation so that general inhomogeneous systems can be handled.
It is also applicable to lattice models such as the Hubbard model, for which every integration over ${\bm r}$ above should be replaced by a summation
over the lattice sites.
Let us introduce the Heisenberg representations of the field operators in the Matsubara formalism by \cite{AGD63}
\begin{equation}
\left\{
\begin{array}{l}
\vspace{1mm}
\psi^{\rm H}_{1}(1) \equiv e^{\tau_{1}H}\psi(\xi_{1})e^{-\tau_{1}H} \\
\psi^{\rm H}_{2}(1) \equiv e^{\tau_{1}H}\psi^{\dagger}(\xi_{1})e^{-\tau_{1}H}
\end{array} \right. ,
\end{equation}
where argument $1$ in the round brackets is defined by $1\equiv (\xi_{1},\tau_{1})$ and variable
$\tau_{1}$ lies in $0 \leq \tau_{1} \leq \beta\equiv 1/T$ with
$T$ denoting the temperature.
The operators $\psi^{\rm H}_{1}(1)$ and
$\psi^{\rm H}_{2}(1)$ are denoted as $\tilde{\psi}(1)$ and $\tilde{\bar{\psi}}(1)$ by Abrikosov {\em et al}.\ \cite{AGD63}, respectively;
distinguishing them by the subscript $i=1,2$ (i.e., the particle-hole or ``charge'' index) as above enables us to simplify the notation and formulation considerably, as seen below.
\subsection{Green's function, Dyson-Gor'kov equation, and thermodynamic potential\label{subsec:TP}}
Using $\psi^{\rm H}_{i}(1)$, we define Green's functions by
\begin{equation}
G_{ij}(1,2)\equiv - \bigl< T_{\tau} \psi^{\rm H}_{i}(1)\psi^{\rm H}_{3-j}(2)\bigr> ,
\label{G_ij}
\end{equation}
where $T_{\tau}$ is the ``time''-ordering operator and $\langle\cdots\rangle$ denotes the grand-canonical average in terms of $H$.\cite{AGD63} The standard normal-state Green's function $G$ corresponds to $G_{11}$. The elements satisfy
\begin{equation}
G_{ij}(1,2)= -G_{3-j,3-i}(2,1)= G_{ji}^{*}(\xi_{2}\tau_{1},\xi_{1}\tau_{2}),
\label{G_ij-symm}
\end{equation}
with the superscript $^{*}$ denoting the complex conjugate.
The Nambu matrix $\hat{G}\equiv (G_{ij})$ is written explicitly as
\begin{equation}
\hat{G}(1,2)\equiv
\begin{bmatrix}
G_{11}(1,2) & G_{12}(1,2)
\\
G_{21}(1,2) & G_{22}(1,2)
\end{bmatrix} .
\label{hatG}
\end{equation}
It obeys the Dyson-Gor'kov equation
\begin{equation}
\bigl[\hat{G}_{0}^{-1}(1,\bar{3})-\hat{\Sigma}(1,\bar{3})\bigr] \hat{G}(\bar{3},2)=\hat{\sigma}_{0} \delta(1,2),
\label{DG}
\end{equation}
where $\hat{\sigma}_{0}$ is the $2\times 2$ unit matrix,
$\delta(1,2)\equiv \delta(\tau_{1}-\tau_{2})\delta(\xi_{1}-\xi_{2})$, and integrations over barred arguments are implied.
The quantity $\hat{G}_{0}^{-1}$ is defined by
\begin{equation}
\hat{G}_{0}^{-1}(1,2)\equiv \left(-\hat{\sigma}_{0}\frac{\partial}{\partial \tau_{1}}
-\hat{\sigma}_{3}K_{1}\right)\delta(1,2) ,
\label{G_0^-1(1,2)}
\end{equation}
where $\hat{\sigma}_{3}$ is the third Pauli matrix.
Finally, $\hat{\Sigma}$ in eq.\ (\ref{DG}) denotes the self-energy matrix
\begin{equation}
\hat{\Sigma}(1,2)\equiv
\begin{bmatrix}
\Sigma_{11}(1,2) & \Sigma_{12}(1,2)
\\
\Sigma_{21}(1,2) & \Sigma_{22}(1,2)
\end{bmatrix} .
\label{hatSigma}
\end{equation}
It follows from eqs.\ (\ref{G_ij-symm}), (\ref{DG}), and (\ref{G_0^-1(1,2)}) that $\hat{\Sigma}$ satisfies the same symmetry relations as $\hat{G}$ in eq.\ (\ref{G_ij-symm}).
As shown by de Dominicis and Martin,\cite{dDM64} we can express
the thermodynamic potential $\Omega\equiv -\beta^{-1}\ln {\rm Tr}\, {\rm e}^{-\beta H}$
as a functional of $\hat{G}$, i.e., $\Omega=\Omega[\hat{G}]$, so as to satisfy
\begin{equation}
\frac{\delta \Omega}{\delta G_{ji}(2,1)}=0.
\label{dOmega/dG=0}
\end{equation}
Let us write $\Omega[\hat{G}]$ as
\begin{equation}
\Omega = -\frac{1}{2\beta}{\rm Tr}\bigl[ \ln \bigl(-\hat{G}_{0}^{-1}+\hat{\Sigma}\bigr)+\hat{\Sigma}\hat{G}\bigr]+\Phi ,
\label{Omega}
\end{equation}
where $\hat{\Sigma}=\hat{\Sigma}[\hat{G}]$, $\Phi=\Phi[\hat{G}]$, and the operator ${\rm Tr}$ for the Nambu matrices is defined by
\begin{equation}
{\rm Tr}\,\hat{A}\equiv
A_{11}(\bar{1},\bar{1}_{+}) + A_{22}(\bar{1}_{+},\bar{1})
\label{Tr-def}
\end{equation}
with the subscript of $1_+$ denoting an extra infinitesimal positive constant in $\tau_{1}$ to
place creation operators to the left of annihilation ones for the equal-time average.
Using eqs.\ (\ref{DG}) and (\ref{Omega}), one may show easily that eq.\ (\ref{dOmega/dG=0}) is transformed into a relation between
$\hat{\Sigma}$ and $\Phi$ as
\begin{equation}
\Sigma_{ij}(1,2)=2\beta\frac{\delta \Phi}{\delta G_{ji}(2,1)} .
\label{Sigma-Phi}
\end{equation}
Equation (\ref{Omega}) with eqs.\ (\ref{DG}) and (\ref{Sigma-Phi}) forms an extension of the normal-state Luttinger-Ward functional\cite{LW60} to superconductors.
Note that we have given $\Omega$ as a functional of $\hat{G}$ following Baym\cite{Baym62} instead of $\hat{\Sigma}$ in the original treatment.\cite{LW60}
It follows from eqs.\ (\ref{DG}), (\ref{Omega}), and (\ref{Sigma-Phi}) that we can calculate $\hat{G}$ and $\Omega$ self-consistently once $\Phi=\Phi[\hat{G}]$ is given explicitly. The particle number $N$ is obtained by
\begin{equation}
N=\frac{1}{2}{\rm Tr}\,\hat{G} \hat{\sigma}_{3},
\end{equation}
which may be used to change an independent variable from $\mu$ to $N$ by the Legendre transformation $F\equiv \Omega+\mu N$.
\subsection{S-matrix and functional $\Phi$}
Perturbation expansions of $\Omega$ and $\hat{G}$ in terms of $H_{\rm int}$ in eq.\ (\ref{H_int})
can be carried out conveniently with the S-matrix\cite{AGD63}
\begin{equation}
{\cal S} \equiv T_{\tau} \exp\biggl[ -\frac{1}{4}
\Gamma^{(0)}(\bar{1}\bar{1}',\bar{2}\bar{2}')
\psi_{2}(\bar{1})\psi_{2}(\bar{2})\psi_{1}(\bar{2}')\psi_{1}(\bar{1}')\biggr] .
\label{S-matrix}
\end{equation}
Here $\psi_{i}(1)$ ($i \!=\! 1,2$) are the interaction representations of the field operators given explicitly by $\psi_{1}(1) \equiv e^{\tau_{1}H_{0}}\psi(\xi_{1})e^{-\tau_{1}H_{0}}$ and
$\psi_{2}(1) \equiv e^{\tau_{1}H_{0}}\psi^{\dagger}(\xi_{1}) e^{-\tau_{1}H_{0}}$,
and $\Gamma^{(0)}$ denotes the symmetrized bare vertex\cite{AGD63}
\begin{eqnarray}
&&\hspace{-14mm}
\Gamma^{(0)}(11',22')\equiv V({\bm r}_{1}-{\bm r}_{2})\delta(\tau_{1}-\tau_{2})
\nonumber \\
&&\hspace{10mm}\times
[\delta(1,1')\delta(2,2')-\delta(1,2')\delta(2,1')] ,
\label{Gamma^(0)}
\end{eqnarray}
satisfying
$\Gamma^{(0)}(11',22')\!=\!\Gamma^{(0)}(22',11')\!=\! \Gamma^{(0)}(1'1,2'2)\!=\! -\Gamma^{(0)}(12',21')$.
The interaction in eq.\ (\ref{S-matrix}) can be expressed graphically as Fig.\ \ref{Fig2}.
\begin{figure}[t]
\begin{center}
\includegraphics[height=10mm]{Fig2.eps}
\end{center}
\caption{Bare interaction vertex $\Gamma^{(0)}$ that is symmetric in a pair of outgoing (incoming) lines. \label{Fig2}}
\end{figure}
Now, the key functional $\Phi$ in eq.\ (\ref{Omega}) is given exactly in terms of eq.\ (\ref{S-matrix}) as
\begin{equation}
\Phi[\hat{G}] = -\frac{<{\cal S}>_{0{\rm c}}-1}{\beta}\biggr|_{{\rm skeleton},\hat{G}_{0}\rightarrow \hat{G}},
\label{Phi}
\end{equation}
where subscript $_{0}$ denotes the thermodynamic average with $H_{0}$ and another subscript $_{\rm c}$
implies retaining only connected Feynman diagrams.
The right-hand side may be expressed graphically by the skeleton diagrams (i.e., diagrams without self-energy insertions) in the perturbation expansion for $\Omega$ with $\hat{G}_{0}$ replaced by $\hat{G}$.\cite{LW60} Approximating eq.\ (\ref{Phi}) by a few terms or some partial series, we can construct a self-consistent
approximation, as already noted in the last paragraph of \S \ref{subsec:TP}.
Equation (\ref{Phi}) was proved by Luttinger and Ward for normal states.\cite{LW60,Kita10} However, it also holds true for superconductors by including all the pair processes.
The Feynman rules for the perturbation expansion of eq.\ (\ref{Phi}) with eq.\ (\ref{S-matrix}) was given by Abrikosov {\em et al}.\ \cite{AGD63} for normal states.
Following them, however, we need to calculate numerical factors from a separate expansion in terms of $V$.
This may be the reason why the expansion with $\Gamma^{(0)}$ has not been used widely.
It is shown in Appendix\ref{appendix:Feynman} that the rules can be simplified considerably so that one may perform it directly in terms of $\Gamma^{(0)}$,
even for superconductors.
For example, the first-order contribution to $\Phi$ is graphically given by Fig.\ \ref{Fig3},
which can be expressed analytically as
\begin{eqnarray}
&&\hspace{-10mm}
\Phi_{1}=\frac{1}{4\beta}\Gamma^{(0)}(\bar{1}\bar{1}',\bar{2}\bar{2}')\bigl[2G(\bar{1}',\bar{1})G(\bar{2}',\bar{2})
\nonumber \\
&&\hspace{-1mm}
+\bar{F}(\bar{1},\bar{2})F(\bar{1}',\bar{2}')\bigr],
\label{Phi_1}
\end{eqnarray}
where $G$, $F$, and $\bar{F}$ are defined together with $\bar{G}$ by
\begin{subequations}
\label{GF-def}
\begin{eqnarray}
&&\hspace{-10mm}
G(1,2)\equiv \frac{G_{11}(1,2)-G_{22}(2,1)}{2}\equiv \bar{G}(2,1),
\\
&&\hspace{-10mm}
F(1,2)\equiv \frac{G_{12}(1,2)-G_{12}(2,1)}{2},
\\
&&\hspace{-10mm}
\bar{F}(1,2)\equiv -\frac{G_{21}(1,2)-G_{21}(2,1)}{2}.
\end{eqnarray}
\end{subequations}
With eq.\ (\ref{GF-def}), we have incorporated the first symmetry of eq.\ (\ref{G_ij-symm}) manifestly into $\Phi_{1}$
for the purpose of deriving $\underline{\Gamma}^{(0)}$ below.
\begin{figure}[t]
\begin{center}
\includegraphics[height=17mm]{Fig3.eps}
\end{center}
\caption{First-order diagrams for $\Phi$. The number below each diagram denotes its relative weight.\label{Fig3}}
\end{figure}
\subsection{Concise perturbation expansion for $\Phi$}
The expansion with $\Gamma^{(0)}$ is still cumbersome for collecting anomalous diagrams
in superconductivity, however, as seen from Fig.\ \ref{Fig1}.
Hence, we further modify the expression of eq.\ (\ref{S-matrix}) so that the asymmetry between the incoming and outgoing lines
in Fig.\ \ref{Fig2} is removed.
To be specific, we introduce a bare vertex $\Gamma^{(0)}_{ii',jj'}(11',22')$ from eq.\ (\ref{Phi_1}) by
\begin{subequations}
\label{hatGamma^(0)}
\begin{equation}
\Gamma^{(0)}_{ii',jj'}(11',22')\equiv 2\beta\frac{\delta^{2}\Phi_{1}}{\delta G_{i'i}(1',1)G_{j'j}(2',2)}.
\label{hatGamma^(0)-1}
\end{equation}
A straightforward calculation using the symmetry of eq.\ (\ref{Gamma^(0)}) shows that results of all the above differentiations may be summarized as
\begin{eqnarray}
&&\hspace{-12mm}
\Gamma^{(0)}_{ii',jj'}(11',22')=\frac{\delta_{ij}}{2}\delta_{ii'}\delta_{jj'}\Gamma^{(0)}(11',22')
\nonumber \\
&&\hspace{16mm}
-\frac{\delta_{i,3-j}}{2}\delta_{ii'}\delta_{jj'}\Gamma^{(0)}(11',2'2)
\nonumber \\
&&\hspace{16mm}
-\frac{\delta_{i,3-j}}{2}\delta_{ij'}\delta_{ji'}\Gamma^{(0)}(12,1'2') ,
\label{hatGamma^(0)-2}
\end{eqnarray}
\end{subequations}
which satisfies
\begin{eqnarray}
&&\hspace{-4mm}
\Gamma^{(0)}_{ii',jj'}(11',22')=\Gamma^{(0)}_{jj',ii'}(22',11')=-\Gamma^{(0)}_{ij',ji'}(12',21')
\nonumber \\
&&\hspace{-8mm}
=-\Gamma^{(0)}_{3-i',3-i,jj'}(1'1,22')=-\Gamma^{(0)}_{3-j',i',j,3-i}(2'1',21).
\label{hatGamma-symm}
\end{eqnarray}
Now, we can express eq.\ (\ref{S-matrix}) alternatively as
\begin{eqnarray}
&&\hspace{-10mm}
{\cal S}= T_{\tau} \exp\biggl[-\frac{1}{12}\Gamma^{(0)}_{\bar{i}\bar{i}',\bar{j}\bar{j}'}(\bar{1}\bar{1}',\bar{2}\bar{2}')
\nonumber \\
&&\hspace{-3mm}\times
{\cal N}\psi_{\bar{i}'}(\bar{1}')\psi_{3-\bar{i}}(\bar{1})\psi_{\bar{j}'}(\bar{2}')\psi_{3-\bar{j}}(\bar{2})\biggr] ,
\label{S-matrix2}
\end{eqnarray}
where ${\cal N}$ denotes the normal-ordering operator of placing creation operators to the left of annihilation operators
with a sign change per every permutation of a pair of adjacent field operators.\cite{AGD63}
The equivalence between eqs.\ (\ref{S-matrix}) and (\ref{S-matrix2}) may be checked easily by substituting eq.\ (\ref{hatGamma^(0)-2}) into the latter
and writing the resultant expression without the ${\cal N}$ operator.
Note that ${\cal N}$ is only relevant to equal-time averages of the first order in the perturbation expansion.
\begin{figure}[t]
\begin{center}
\includegraphics[height=10mm]{Fig4.eps}
\end{center}
\caption{Bare interaction vertex $\Gamma^{(0)}_{ii',jj'}(11',22')$ that is symmetric in four external lines. \label{Fig4}}
\end{figure}
The interaction in eq.\ (\ref{S-matrix2}) may be expressed graphically as Fig.\ \ref{Fig4},
which is symmetric in the four external lines.
Using eq.\ (\ref{S-matrix2}), we can perform the perturbation expansion for $\Phi$ (and also for any other quantities) concisely in such a way that all the anomalous processes are incorporated automatically.
Figure \ref{Fig5} enumerates first- to forth-order diagrams for $\Phi$.
Thus, each of the first- to third-order contributions is exhausted by a single diagram, showing manifestly the advantage of using eq.\ (\ref{S-matrix2}).
Moreover, the symmetry of eq.\ (\ref{hatGamma-symm}) brings the simplification that every possible Wick decomposition for a distinct diagram
yields the same contribution.
Hence, we only need to consider a convenient decomposition (i.e., connection of vertices) for each distinct diagram and multiply the result by the number of possible connections,
which may be calculated easily based on a combinatorial consideration.
Now, the Feynman rules for the expansion of eq.\ (\ref{Phi}) with eq.\ (\ref{S-matrix2}) are summarized as follows.
\begin{enumerate}
\item[(a)] Draw all the $n$th-order closed skeleton diagrams that are topologically distinct.
For each such diagram, associate the factor $(-1)^{n+1}/ n!12^{n}\beta$.
\item[(b)] For each small circle, associate $\Gamma^{(0)}_{\bar{i}\bar{i}',\bar{j}\bar{j}'}(\bar{\lambda}\bar{\lambda}',\bar{\nu}\bar{\nu}')$.
\item[(c)] Identify the number $C_{n\alpha}$ of possible connections of vertices for diagram $\alpha$ under consideration.
\item[(d)] Consider a specific connection of vertices for diagram $\alpha$ where every primed argument is linked with an unprimed argument,
and associate $G_{\bar{i}'\bar{j}}(\bar{\lambda}',\bar{\eta})$ for each line connecting $\bar{\lambda}_{\bar{i}'}'$ and $\bar{\eta}_{\bar{j}}$.
\item[(e)] Identify the number $\ell_{n\alpha}$ of permutations to realize the connection.
With the choice in (d), we may find it by replacing every vertex as
$\Gamma^{(0)}_{ii',jj'}(11',22')\rightarrow \delta_{i1}\delta_{i'1}\delta_{j1}\delta_{j'1}\delta(1,1')\delta(2,2') V({\bm r}_{1}-{\bm r}_{2})\delta(\tau_{1}-\tau_{2})$
and equating $\ell_{n\alpha}$ with the number of closed particle loops in the resultant normal diagram with $V$.\cite{LW60}
\item[(f)] Multiply the expression by $(-1)^{\ell_{n\alpha}}C_{n\alpha}$.
\item[(g)] When calculating the two-particle irreducible vertex $\underline{\Gamma}^{({\rm ir})}$ by eq.\ (\ref{Gamma^(ir)}) below, replace every Green's function as
\begin{equation}
G_{ij}(1,2)\rightarrow \tilde{G}_{ij}(1,2)\equiv \frac{G_{ij}(1,2)-G_{3-j,3-i}(2,1)}{2},
\label{tG_ij}
\end{equation}
so as to incorporate the first symmetry of eq.\ (\ref{G_ij-symm}) manifestly in $\underline{\Gamma}^{({\rm ir})}$.
\end{enumerate}
\noindent
Rule (g) is relevant only to two-particle and higher-order Green's functions.
\begin{figure}[t]
\begin{center}
\includegraphics[height=14mm]{Fig5.eps}
\end{center}
\caption{First- to forth-order diagrams for $\Phi$. \label{Fig5}}
\end{figure}
As an example of rule (c) above, let us consider diagram (3) in Fig.\ \ref{Fig5}. Its Wick decomposition may proceed as follows:
Start from an arbitrary operator $\psi_{i}$ (labelled $a_{1}$ operator in $H_{\rm int}^{a}$) and find its partner (labelled $b_{1}$ operator in $H_{\rm int}^{b}$) from
$2\times 4$ possible candidates; to connect another line between $H_{\rm int}^{a}$ and $H_{\rm int}^{b}$, pick out a pair of operators $a_{2}$ and $b_{2}$ from
$3\times 3$ possible choices; to connect a pair of lines between $H_{\rm int}^{b}$ and $H_{\rm int}^{c}$, find the partners of $b_{3}$ and $b_{4}$ (labelled $c_{1}$ and $c_{2}$ operators, respectively) from $4\times 3$ possible choices in $H_{\rm int}^{c}$; finally, select one of the $2$ possibilities to
connect a pair of lines between ($c_{3}$,$c_{4}$) and ($a_{3}$,$a_{4}$). Thus, the number $C_{3}$ is obtained as
$$
C_{3}=(2\cdot 4\cdot 3^{2})\cdot (4\cdot 3)\cdot 2 =2!4^{2}3^{3}2.
$$
Now, rules (a)-(f) enable us to write down an analytic expression for $\Phi_{3}$ as
\begin{eqnarray}
&&\hspace{-10mm}
\Phi_{3} =
\frac{(-1)^{4+\ell_{3}}C_{3}}{3!12^{3}\beta}
\Gamma^{(0)}_{\bar{i}\bar{i}',\bar{j}\bar{j}'}(\bar{1}\bar{1}',\bar{2}\bar{2}')
G_{\bar{j}'\bar{k}}(\bar{2}',\bar{3})G_{\bar{k}'\bar{j}}(\bar{3}',\bar{2})
\nonumber \\
&& \hspace{-1mm} \times
\Gamma^{(0)}_{\bar{k}\bar{k}',\bar{l}\bar{l}'}(\bar{3}\bar{3}',\bar{4}\bar{4}')
G_{\bar{l}'\bar{m}}(\bar{4}',\bar{5})G_{\bar{m}'\bar{l}}(\bar{5}',\bar{4})
\nonumber \\
&& \hspace{-1mm} \times
\Gamma^{(0)}_{\bar{m}\bar{m}',\bar{n}\bar{n}'}(\bar{5}\bar{5}',\bar{6}\bar{6}')
G_{\bar{n}'\bar{i}}(\bar{6}',\bar{1})G_{\bar{i}'\bar{n}}(\bar{1}',\bar{6}),
\label{Phi_3-0}
\end{eqnarray}
with $\ell_{3}=3$ from rule (e) for this connection.
An elementary calculation using eq.\ (\ref{hatGamma^(0)-2}) shows that eq.\ (\ref{Phi_3-0}) correctly reproduces the weights of
diagrams (a)-(j) in Fig.\ \ref{Fig1}.
It is convenient at this stage to introduce matrices $\underline{\Gamma}^{(0)}$ and $\underline{\chi}^{(0)}$ by
\begin{subequations}
\label{uGamma-uchi}
\begin{eqnarray}
&&\hspace{-10mm}
\langle 11'_{ii'}|\underline{\Gamma}^{(0)}|22'_{jj'}\rangle \equiv
{\Gamma}^{(0)}_{ii',j'j}(11',2'2),
\label{uGamma^(0)}
\\
&&\hspace{-10mm}
\langle 11'_{ii'}|\underline{\chi}^{(0)}|22'_{jj'}\rangle \equiv
-G_{ij}(1,2)G_{j'i'}(2',1').
\label{chi^(0)}
\end{eqnarray}
\end{subequations}
Using them, we can express eq.\ (\ref{Phi_3-0}) concisely as
\begin{subequations}
\label{Phi_1-4}
\begin{equation}
\Phi_{3} = \frac{1}{6\beta}{\rm Tr}\,\bigl(\underline{\Gamma}^{(0)}\underline{\chi}^{(0)}\bigr)^{\! 3},
\label{Phi_3}
\end{equation}
where the basis of Tr is given by the bracket vectors in eq.\ (\ref{uGamma-uchi})
The above consideration may be extended to the other diagrams in Fig.\ \ref{Fig5},
where the numbers of rule (c) are identified as $C_{1}=3$, $C_{2}=4!$, $C_{4a}=3!4^{3}3^{4}2$, and
$C_{4b}=(_4 C_2)^{4}3\cdot 2^{2}2^{3}=3^{5}2^{9}=3!4^{4}3^{4}$.
We thereby obtain
\begin{equation}
\Phi_{1}=\frac{1}{4\beta}{\rm Tr}\,\underline{\Gamma}^{(0)}\underline{\chi}^{(0)},
\label{Phi_1-2}
\end{equation}
\begin{equation}
\Phi_{2}=-\frac{1}{12\beta}{\rm Tr}\,\bigl(\underline{\Gamma}^{(0)}\underline{\chi}^{(0)}\bigr)^{\! 2},
\end{equation}
\begin{equation}
\Phi_{4a}=-\frac{1}{8\beta}{\rm Tr}\,\bigl(\underline{\Gamma}^{(0)}\underline{\chi}^{(0)}\bigr)^{\! 4},
\label{Phi_4a}
\end{equation}
\begin{eqnarray}
&&\hspace{-15mm}
\Phi_{4b} =
\frac{1}{4\beta}
\Gamma^{(0)}_{\bar{i}\bar{i}',\bar{j}\bar{j}'}(\bar{1}\bar{1}',\bar{2}\bar{2}')
\Gamma^{(0)}_{\bar{k}\bar{k}',\bar{l}\bar{l}'}(\bar{3}\bar{3}',\bar{4}\bar{4}')
\nonumber \\
&& \hspace{-4.5mm} \times
\Gamma^{(0)}_{\bar{m}\bar{m}',\bar{n}\bar{n}'}(\bar{5}\bar{5}',\bar{6}\bar{6}')
\Gamma^{(0)}_{\bar{p}\bar{p}',\bar{q}\bar{q}'}(\bar{7}\bar{7}',\bar{8}\bar{8}')
\nonumber \\
&& \hspace{-4.5mm} \times
G_{\bar{j}'\bar{k}}(\bar{2}',\bar{3})G_{\bar{k}'\bar{j}}(\bar{3}',\bar{2})
G_{\bar{n}'\bar{p}}(\bar{6}',\bar{7})G_{\bar{p}'\bar{n}}(\bar{7}',\bar{6})
\nonumber \\
&& \hspace{-4.5mm} \times
G_{\bar{i}'\bar{q}}(\bar{1}',\bar{8})G_{\bar{q}'\bar{l}}(\bar{8}',\bar{4})G_{\bar{l}'\bar{m}}(\bar{4}',\bar{5})G_{\bar{m}'\bar{i}}(\bar{5}',\bar{1}).
\label{Phi_4b}
\end{eqnarray}
\end{subequations}
The equivalence between eqs.\ (\ref{Phi_1}) and (\ref{Phi_1-2}) may be checked easily by using eqs.\ (\ref{GF-def}), (\ref{hatGamma^(0)-2}),
and (\ref{uGamma-uchi}).
Now, one may be convinced that eq.\ (\ref{S-matrix2}) considerably simplifies the perturbation expansion of superconductivity.
\subsection{Self-consistent approximations}
As already mentioned, eq.\ (\ref{Phi}) forms a convenient basis for systematic approximations for superconductivity.
To be specific, Green's function and the self-energy are determined self-consistently by eqs.\ (\ref{DG}) and (\ref{Sigma-Phi})
based on some approximate $\Phi$.
The lowest-order approximation corresponds to the choice $\Phi\approx \Phi_{1}$ from the series of eq.\ (\ref{Phi_1-4}), which is exactly the mean-field BCS theory.\cite{BCS57,Parks69}
It may be improved by including the next-order term as $\Phi\approx \Phi_{1}+\Phi_{2}$, where two-body quasiparticle scatterings are included.
Hence, adopting this approximation on the Schwinger-Keldysh contour,\cite{Kita10,HJ08,Rammer07} we can describe thermalization of superconductors, for example.
We now introduce the FLEX-S approximation, which adds the sort of diagrams (3) and (4a) in Fig.\ 5 up to infinite order
besides $\Phi_{1}$ and $\Phi_{2}$.
As seen from Fig.\ 1, it contains the normal particle-hole and particle-particle diagrams, i.e., diagrams (a) and (b), respectively, where every exchange process is automatically incorporated through $\Gamma^{(0)}$, plus all the pair processes derivable from them
by successively changing directions of a pair of incoming and outgoing arrows at a vertex, i.e., diagrams (c)-(j).
Generalizing the consideration that led to eqs.\ (\ref{Phi_3}) and (\ref{Phi_4a}), we find the $n$th-order contribution $\Phi_{na}$ ($n\geq 4$) as
\begin{equation}
\Phi_{na}=\frac{(-1)^{n+1}}{2n\beta}{\rm Tr}\,\bigl(\underline{\Gamma}^{(0)}\underline{\chi}^{(0)}\bigr)^{\! n}.
\end{equation}
The functional $\Phi$ of our FLEX-S approximation is defined by
$\Phi_{\mbox{FLEX-S}}=\Phi_{1}+\Phi_{2}+\Phi_{3}+\sum_{n=4}^{\infty}\Phi_{na}$,
i.e.,
\begin{eqnarray}
&&\hspace{-13mm}
\Phi_{\mbox{FLEX-S}}=\frac{1}{4\beta}{\rm Tr}\, \underline{\Gamma}^{(0)}\underline{\chi}^{(0)}+
\frac{1}{6\beta}{\rm Tr}\bigl(\underline{\Gamma}^{(0)}\underline{\chi}^{(0)}\bigr)^{2}
\nonumber \\
&&\hspace{7mm}
+\frac{1}{2\beta}{\rm Tr} \! \left[\ln \bigl(\underline{1}+\underline{\Gamma}^{(0)}\underline{\chi}^{(0)}\bigr)
-\underline{\Gamma}^{(0)}\underline{\chi}^{(0)}\right] .
\label{Phi_FLEX}
\end{eqnarray}
The corresponding self-energy is obtained by eq.\ (\ref{Sigma-Phi}) with $\Phi\approx \Phi_{\mbox{FLEX-S}}$,
where the differentiation may also be carried out graphically by removing a single line from the diagrams for $\Phi$ in all possible ways.
Using eq.\ (\ref{hatGamma-symm}), we arrive at the expression
\begin{eqnarray}
&&\hspace{-8mm}
\Sigma_{ij}(1,2)=-\biggl[\langle 1\bar{3}_{i\bar{k}}|\underline{\Gamma}^{(0)} |2\bar{4}_{j\bar{l}}\rangle
\!+\! \frac{4}{3}\langle 1\bar{3}_{i\bar{k}}|\underline{\Gamma}^{(0)}\underline{\chi}^{(0)}\underline{\Gamma}^{(0)}|2\bar{4}_{j\bar{l}}\rangle
\nonumber \\
&&\hspace{9.5mm}
-2\langle 1\bar{3}_{i\bar{k}}|\underline{\Gamma}^{(0)}\underline{\chi}^{(0)}\underline{\Gamma}^{(1)}|2\bar{4}_{j\bar{l}}\rangle
\biggr] G_{\bar{k}\bar{l}}(\bar{3},\bar{4}) ,
\label{Sigma_FLEX}
\end{eqnarray}
where $\underline{\Gamma}^{(1)}$ is defined by
\begin{equation}
\underline{\Gamma}^{(1)}\equiv \underline{\Gamma}^{(0)}(\underline{1}+\underline{\chi}^{(0)}\underline{\Gamma}^{(0)})^{-1}.
\label{Gamma^(1)-def}
\end{equation}
\subsection{Two-particle Green's function}
Next, we consider the two-particle Green's function
\begin{eqnarray}
&&\hspace{-10mm}
{\cal K}_{ii',jj'}(11',22')
\equiv \langle T_{\tau}\psi_{i}^{\rm H}(1)\psi_{3-i'}^{\rm H}(1')\psi_{j}^{\rm H}(2)\psi_{3-j'}^{\rm H}(2')\rangle
\nonumber \\
&&\hspace{18mm}
-G_{ii'}(1,1')G_{jj'}(2,2') ,
\label{calK-def}
\end{eqnarray}
whose poles define collective modes such as density, spin, and superconducting fluctuations.
Functional $\Phi[\hat{G}]$ also enables us to calculate eq.\ (\ref{calK-def}) unambiguously.
To see this, let us apply an artificial nonlocal potential $\hat{\cal U}(1,2)$ described by the S-matrix
\begin{equation}
{\cal S}^{\cal U}\equiv T_{\tau}
\exp\!\left[ \,\frac{1}{2} \psi_{\bar{j}}^{\rm H}(\bar{2})\psi_{3-\bar{j}'}^{\rm H}(\bar{2}') {\cal U}_{\bar{j}'\bar{j}}(\bar{2}',\bar{2})\right]\! .
\label{calS_ex}
\end{equation}
Green's function in the presence of $\hat{\cal U}\equiv ({\cal U}_{ij})$ is defined by\cite{AGD63}
\begin{equation}
G^{\cal U}_{ii'}(1,1')\equiv -\frac{\langle T_{\tau}{\cal S}^{\cal U}\psi_{i}^{\rm H}(1)\psi_{3-i'}^{\rm H}(1')\rangle}
{\langle{\cal S}^{\cal U}\rangle} ,
\label{G_ij-U}
\end{equation}
which reduces to eq.\ (\ref{G_ij}) as $\hat{\cal U}\rightarrow \hat{0}$.
Now, one may see easily that eq.\ (\ref{calK-def})
is obtained from eq.\ (\ref{G_ij-U}) by
\begin{equation}
{\cal K}_{ii',jj'}(11',22')=-2\frac{\delta G^{\cal U}_{ii'}(1,1')}{\delta {\cal U}_{j'j}(2',2)}\biggr|_{\hat{\cal U}\rightarrow\hat{0}}.
\label{calK-deriv1}
\end{equation}
This expression tells us that we only need to know the linear response of $\hat{G}^{\cal U}$ to $\hat{\cal U}$ for obtaining the two-particle Green's function.
To find $\delta \hat{G}^{\cal U}$, we start from the Dyson-Gor'kov equation, which is modified from eq.\ (\ref{DG}) into\cite{MS59,BK61,Kita10b}
\begin{equation}
\bigl[ \hat{G}_{0}^{-1}(1,\bar{2})-\hat{\cal U}'(1,\bar{2})-\hat{\Sigma}^{\cal U\!}(1,\bar{2})\bigr]\hat{G}^{\cal U\!}(\bar{2},1')=\hat{\sigma}_{0}\delta(1,1') ,
\label{DG-U}
\end{equation}
with
\begin{equation}
{\cal U}_{ii'}'(1,1')\equiv \frac{{\cal U}_{ii'}(1,1')-{\cal U}_{3-i',3-i}(1',1)}{2} .
\label{U'}
\end{equation}
Let us change $\hat{\cal U}\rightarrow \hat{\cal U}+\delta \hat{\cal U}$ in eq.\ (\ref{DG-U}) and subsequently set $\hat{\cal U}=\hat{0}$.
We thereby obtain the first-order equation
\begin{equation}
\hat{G}^{-1}(1,\bar{2})\delta \hat{G}^{\cal U\!}(\bar{2},1')
=\bigl[\delta \hat{\cal U}'(1,\bar{2}) +\delta \hat{\Sigma}^{\cal U\!}(1,\bar{2})\bigr]\hat{G}(\bar{2},1').
\label{DG-dU}
\end{equation}
Using eq.\ (\ref{Sigma-Phi}) for the self-energy, we can express $\delta \hat{\Sigma}^{\cal U}$ above in terms of
$\delta \hat{G}^{\cal U}$ as
\begin{equation}
\delta \Sigma_{ii'}^{\cal U}(1,1')=\Gamma^{({\rm ir})}_{ii',\bar{j}\bar{j}'}(11',\bar{2}\bar{2}')\delta G_{\bar{j}'\bar{j}}^{\cal U}(\bar{2}',\bar{2}) ,
\label{dSigma}
\end{equation}
where $\Gamma^{({\rm ir})}_{ii',\bar{j}\bar{j}'}(11',\bar{2}\bar{2}')$ denotes the ``irreducible'' vertex defined by
\begin{equation}
\Gamma^{({\rm ir})}_{ii',\bar{j}\bar{j}'}(11',22')\equiv 2\beta \frac{\delta^{2}\Phi }{\delta G_{i'i}(1',1)\delta G_{j'j}(2',2)}.
\label{Gamma^(ir)}
\end{equation}
This differentiation should be performed after the replacement of eq.\ (\ref{tG_ij}) in $\Phi$, as already mentioned.
Let us substitute eq.\ (\ref{dSigma}) into eq.\ (\ref{DG-dU}), multiply the resultant equation by $\hat{G}(3,1)$ from the left, and integrate over $1$.
Changing the arguments appropriately, we obtain a closed equation for $\delta\hat{G}^{\cal U}$ as
\begin{eqnarray}
&&\hspace{-9mm}
\delta G^{\cal U}_{ii'}(1,1')
= G_{i\bar{j}}(1,\bar{2})G_{\bar{j}'i'}(\bar{2}',1')\delta {\cal U}_{\bar{j}\bar{j}'}'(\bar{2},\bar{2}')
+ G_{i\bar{j}}(1,\bar{2})
\nonumber \\
&&\hspace{12mm}
\times G_{\bar{j'}i'}(\bar{2}',1')
\Gamma^{({\rm ir})}_{\bar{j}\bar{j}',\bar{k}'\bar{k}}(\bar{2}\bar{2}',\bar{3}'\bar{3})\delta G^{\cal U}_{\bar{k}\bar{k}'}(\bar{3},\bar{3}').
\nonumber \\
\label{dG-eq}
\end{eqnarray}
At this stage, it is convenient to introduce the notations
\begin{equation}
\langle 11'_{ii'}|\delta\vec{G}^{\cal U}=\delta G_{ii'}^{\cal U}(1,1'),
\hspace{5mm}
\langle 11'_{ii'}|\delta\vec{\cal U}=\delta {\cal U}_{ii'}(1,1') ,
\label{Vec3}
\end{equation}
and
\begin{subequations}
\label{Mat}
\begin{eqnarray}
&&\hspace{-8mm}
\langle 11'_{ii'}|\underline{\cal K}|22'_{jj'}\rangle\equiv
{\cal K}_{ii',j'j}(11',2'2),
\label{calK-mat}
\\
&&\hspace{-11.5mm}
\langle 11'_{ii'}|\underline{\Gamma}^{({\rm ir})}|22'_{jj'}\rangle\equiv
\Gamma^{({\rm ir})}_{ii',j'j}(11',2'2),
\label{Gamma^(ir)-mat}
\\
&&\hspace{-12mm}
\langle 11'_{ii'}|\underline{\chi}^{(0{\rm e})}|22'_{jj'}\rangle\equiv -G_{ij}(1,2)G_{j'i'}(2',1')
\nonumber \\
&&\hspace{18.5mm} +G_{i,3-j'}(1,2')G_{3-j,i'}(2,1'),
\label{t-chi^(0)}
\\
&&\hspace{-6.5mm}
\langle 11'_{ii'}|\underline{1}|22'_{jj'}\rangle\equiv \delta_{ij}\delta_{i'j'}\delta(1,2)\delta(1',2').
\label{u1}
\end{eqnarray}
\end{subequations}
It follows from eqs.\ (\ref{G_ij-symm}), (\ref{calK-def}), (\ref{Gamma^(ir)}), and (\ref{Mat}) that matrices $\underline{\cal M}= \underline{\cal K}$,
$\underline{\Gamma}^{({\rm ir})}$, $\underline{\chi}^{(0{\rm e})}$ all satisfy
\begin{eqnarray}
&&\hspace{-10mm}
\langle 11'_{ii'}|\underline{\cal M}|22'_{jj'}\rangle=\langle 2'2_{j'j}|\underline{\cal M}|1'1_{i'i}\rangle
\nonumber \\
&&\hspace{14mm}
=-\langle 1'1_{3-i',3-i}|\underline{\cal M}|22'_{jj'}\rangle .
\label{GC-symm}
\end{eqnarray}
The first equality is also obeyed by $\underline{\chi}^{(0)}$ of eq.\ (\ref{chi^(0)}).
Using eqs.\ (\ref{chi^(0)}), (\ref{U'}), (\ref{Vec3}), and (\ref{Mat}), we can express eq.\ (\ref{dG-eq}) as
$\delta \vec{G}^{\cal U}= -\frac{1}{2}\underline{\chi}^{(0{\rm e})}\delta\vec{\cal U}
-\underline{\chi}^{(0)}\underline{\Gamma}^{({\rm ir})}\delta \vec{G}^{\cal U}$, or equivalently,
$$\delta\vec{G}^{\cal U} =- \frac{1}{2}\bigl(\underline{1}+\underline{\chi}^{(0)}\underline{\Gamma}^{({\rm ir})}\bigr)^{-1}\underline{\chi}^{(0{\rm e})} \delta\vec{\cal U}.$$
Finally, we obtain an expression of $\underline{\cal K}$ by eq.\ (\ref{calK-deriv1}) as
\begin{equation}
\underline{\cal K} = \bigl(\underline{1}+\underline{\chi}^{(0)}\underline{\Gamma}^{({\rm ir})}\bigr)^{-1}\underline{\chi}^{(0{\rm e})} .
\label{calK}
\end{equation}
It may be useful to introduce the full vertex
\begin{equation}
\underline{\Gamma} \equiv \underline{\Gamma}^{({\rm ir})}\bigl(\underline{1}+\underline{\chi}^{(0)}\underline{\Gamma}^{({\rm ir})}\bigr)^{-1} ,
\label{BS}
\end{equation}
which is nothing but the Bethe-Salpeter equation\cite{SB51} for superconductivity and given graphically by Fig.\ \ref{Fig6}. Using
$\underline{\Gamma}$, we can express eq.\ (\ref{calK}) alternatively as
\begin{equation}
\underline{\cal K} = \underline{\chi}^{(0{\rm e})} -\underline{\chi}^{(0)}\underline{\Gamma}\underline{\chi}^{(0{\rm e})} .
\label{calK2}
\end{equation}
Equation (\ref{calK}) or (\ref{calK2}) with eqs.\ (\ref{chi^(0)}), (\ref{Gamma^(ir)}), and (\ref{Mat}) tells us that we can also calculate the two-particle Green's function of
eq.\ (\ref{calK-def}) once $\Phi$ is given explicitly.
\begin{figure}[t]
\begin{center}
\includegraphics[height=10mm]{Fig6.eps}
\end{center}
\caption{Bethe-Salpeter equation for superconductivity. A black (shaded) circle denotes $\underline{\Gamma}$ ($\underline{\Gamma}^{({\rm ir})}$), and
an internal line represents $\hat{G}$. \label{Fig6}}
\end{figure}
It may be illuminating to write down $\underline{\Gamma}^{({\rm ir})}$ explicitly in the FLEX-S approximation.
Let us make the replacement of eq.\ (\ref{tG_ij}) in eq.\ (\ref{Phi_FLEX}), substitute the resultant $\Phi$ into eq.\ (\ref{Gamma^(ir)}),
and perform the differentiation.
Also using the symmetry noted around eq.\ (\ref{GC-symm}), we obtain
\begin{eqnarray}
&&\hspace{-10mm}
\langle 11'_{ii'}|\underline{\Gamma}^{({\rm ir})}|22'_{jj'}\rangle
\nonumber \\
&&\hspace{-14mm}
= \langle 11'_{ii'}|\underline{\Gamma}^{(0)}|22'_{jj'}\rangle
-2\bigl[ \langle 12_{ij}|\underline{\Gamma}^{(0)}\underline{\chi}^{(0)}\underline{\Gamma}^{(0)}|1'2'_{i'j'}\rangle
\nonumber \\
&&\hspace{-10mm}
-\langle 12'_{i,3-j'}|\underline{\Gamma}^{(0)}\underline{\chi}^{(0)}\underline{\Gamma}^{(0)}|1'2_{i',3-j}\rangle \bigr]
\nonumber \\
&&\hspace{-10mm}
+\langle 12_{ij}|\underline{\Gamma}^{(0)}\underline{\chi}^{(0)}\underline{\Gamma}^{(1)}|1'2'_{i'j'}\rangle
\nonumber \\
&&\hspace{-10mm}
-\langle 12'_{i,3-j'}|\underline{\Gamma}^{(0)}\underline{\chi}^{(0)}\underline{\Gamma}^{(1)}|1'2_{i',3-j}\rangle
\nonumber \\
&&\hspace{-10mm}
+2\bigl[\langle 1\bar{3}_{i\bar{k}}|\underline{\Gamma}^{(1)}|2\bar{4}'_{{j}\bar{l}'}\rangle
\langle2'\bar{4}_{j'\bar{l}}|\underline{\Gamma}^{(1)}|1'\bar{3}'_{i'\bar{k}'}\rangle
\nonumber \\
&&\hspace{-10mm}
-\langle 1\bar{3}_{i\bar{k}}|\underline{\Gamma}^{(1)}|2'\bar{4}'_{3-{j}',\bar{l}'}\rangle
\langle2\bar{4}_{3-j,\bar{l}}|\underline{\Gamma}^{(1)}|1'\bar{3}'_{i'\bar{k}'}\rangle \bigr]
\nonumber \\
&&\hspace{-10mm}
\times
\langle\bar{3}\bar{4}'_{\bar{k}\bar{l}'}|\underline{\chi}^{(0)}|\bar{3}'\bar{4}_{\bar{k}'\bar{l}}\rangle ,
\end{eqnarray}
which is different from $\underline{\Gamma}^{(1)}$ of eq.\ (\ref{Gamma^(1)-def}).
Thus, a clear distinction among $\underline{\Gamma}^{(1)}$, $\underline{\Gamma}^{({\rm ir})}$, and $\underline{\Gamma}$
is necessary.
\subsection{Alternative approach}
Before closing this section, we comment on an alternative self-consistent approach adopted by Haussmann {\em et al}.\ \cite{Haussmann07}
of expressing $\Omega$ in terms of a renormalized vertex $\Gamma^{\Omega}$ besides $\hat{G}$.
Its theoretical basis was given by de Dominicis and Martin;\ \cite{dDM64}
see also refs.\ \onlinecite{deDominicis63} and \onlinecite{Bloch65} on this point.
However, the vertices $\Gamma^{\Omega}$ obtained by de Dominicis \cite{deDominicis63,Bloch65} and used by Haussmann {\em et al}.\ \cite{Haussmann07}
have no internal degrees of freedom describing only density fluctuations, in contrast to our $\underline{\Gamma}^{(0)}$ of eq.\ (\ref{hatGamma^(0)}).
Indeed, constructing $\Omega=\Omega[\hat{G}, \Gamma^{\Omega}]$ with a matrix structure in $\Gamma^{\Omega}$ seems to be a non-trivial problem.
It will be practically more convenient to use $\Omega=\Omega[\hat{G}]$ that can easily be written down solely in terms of the smallest expansion unit $\hat{G}$.
Note in this context that: (i) the results obtained by using $\Omega=\Omega[\hat{G}, \Gamma^{\Omega}]$ should necessarily be reproduced by the approach based
on $\Omega=\Omega[\hat{G}]$, as in the case of the particle-particle ladder approximation by Haussmann {\em et al}.; \cite{Haussmann07}
(ii) the vertex $\Gamma^{\Omega}$ in $\Omega$ is generally different from both the irreducible vertex $\Gamma^{({\rm ir})}$ of eq.\ (\ref{Gamma^(ir)}) and the full vertex $\Gamma$
of eq.\ (\ref{BS}).
\section{Homogeneous systems}
We now simplify the formalism of \S 2 for homogeneous systems.
Lattice models may be handled similarly with modifications given in Appendix\ref{appendix:Hubbard}.
\subsection{Momentum-``energy'' representation}
Equations (\ref{G_ij}) and (\ref{hatSigma}) and the delta function $\delta(1,2)$ can be expanded for homogeneous systems as
\begin{subequations}
\label{GS-exp}
\begin{eqnarray}
&&\hspace{-10mm}
G_{ij}(1,2)=\sum_{\vec{p}}G_{i\alpha,j\beta}(\vec{p})\,{\rm e}^{i\vec{p}\cdot(\vec{r}_{1}-\vec{r}_{2})},
\label{G_ij-exp}
\\
&&\hspace{-9.5mm}
\Sigma_{ij}(1,2)=\sum_{\vec{p}}\Sigma_{i\alpha,j\beta}(\vec{p})\,{\rm e}^{i\vec{p}\cdot(\vec{r}_{1}-\vec{r}_{2})},
\label{Sigma_ij-exp}
\\
&&\hspace{-6.2mm}
\delta(1,2)=\delta_{\alpha\beta}\sum_{\vec{p}}{\rm e}^{i\vec{p}\cdot(\vec{r}_{1}-\vec{r}_{2})}.
\label{delta-fn}
\end{eqnarray}
\end{subequations}
Here $\vec{r}_{1}\!\equiv\! ({\bm r}_{1},i\tau_{1})$ and $\vec{p}\!\equiv\! ({\bm p},i\varepsilon_{n})$ with $\varepsilon_{n}\!\equiv\! (2n+1)\pi T$ the fermion Matsubara frequency ($n=0,\pm 1,\cdots$), subscripts $_\alpha, _\beta$ denote spin indices, and the summation over $\vec{p}$ is defined by
\begin{equation}
\sum_{\vec{p}}\equiv \frac{1}{\beta}\sum_{n}\int \frac{{\rm d}^{3}p}{(2\pi)^{3}}.
\label{p-sum}
\end{equation}
We call the imaginary quantity $i\varepsilon_{n}$ ``energy'' below.
Substituting eq.\ (\ref{GS-exp}) into eq.\ (\ref{DG}),
we can transform the Dyson-Gor'kov equation into the $4\times 4$ matrix equation
\begin{equation}
\bigl[\hat{G}_{0}^{-1}(\vec{p}\,)-\hat{\Sigma}(\vec{p}\,)\bigr]\hat{G}(\vec{p}\,)=\hat{1},
\label{DG(p)}
\end{equation}
where $\hat{G}_{0}^{-1}(\vec{p}\,)$ and $\hat{1}$ are defined by
\begin{equation}
\langle i\alpha|\hat{G}_{0}^{-1}(\vec{p}\,)|j\beta\rangle=\delta_{ij}\delta_{\alpha\beta}\bigl[i\varepsilon_{n}-(-1)^{i-1}\epsilon_{\bm p}\bigr]
\end{equation}
with $\epsilon_{\bm p}\equiv p^{2}/2m-\mu$ and $\langle i\alpha|\hat{1}|j\beta\rangle=\delta_{ij}\delta_{\alpha\beta}$, respectively.
The standard arrangement of the basis vectors for the direct product ${\rm C}\times {\rm S}$ of the particle-hole (i.e., charge) space C ($i=1,2$) and spin space S ($\alpha=\uparrow,\downarrow$) is given by $|{\rm cs}_{1}\rangle=|1\!\uparrow\rangle$, $|{\rm cs}_{2}\rangle=|1\!\downarrow\rangle$, $|{\rm cs}_{3}\rangle=|2\!\uparrow\rangle$, and $|{\rm cs}_{4}\rangle=|2\!\downarrow\rangle$.
Green's function in this representation reads
\begin{equation}
\hat{G}(\vec{p})=\begin{bmatrix}
\vspace{1mm}
\underline{G}(\vec{p}) & \underline{F}(\vec{p}) \\
- \underline{\bar{F}}(\vec{p}) & -\underline{\bar{G}}(\vec{p})
\end{bmatrix},
\label{hatG(p)}
\end{equation}
where each quantity with an underline is a $2\times 2$ matrix in spin space given as a Fourier coefficient of eq.\ (\ref{GF-def}).
It follows from eqs.\ (\ref{G_ij-symm}) and (\ref{GF-def}) that these submatrices satisfy
\begin{subequations}
\label{GF-symm}
\begin{equation}
G_{\alpha\beta}(\vec{p})=\bar{G}_{\beta\alpha}(-\vec{p})=G_{\beta\alpha}^{*}(\vec{p}^{\,*}),
\end{equation}
\begin{equation}
F_{\alpha\beta}(\vec{p})=-F_{\beta\alpha}(-\vec{p})=-\bar{F}_{\beta\alpha}^{*}(\vec{p}^{\,*}).
\end{equation}
\end{subequations}
The self-energy matrix may also be expressed as
\begin{equation}
\hat{\Sigma}(\vec{p})=\begin{bmatrix}
\vspace{1mm}
\underline{\Sigma}(\vec{p}) & \underline{\Delta}(\vec{p}) \\
- \underline{\bar{\Delta}}(\vec{p}) & -\underline{\bar{\Sigma}}(\vec{p})
\end{bmatrix},
\label{hatSigma(p)}
\end{equation}
whose submatrices clearly obey the relations of eq.\ (\ref{GF-symm}).
Next, we transform $\hat{\Sigma}=\hat{\Sigma}[\hat{G}]$ into the momentum-energy representation.
Since the expression depends on the approximation we adopt,
we specifically consider the FLEX-S approximation given by eq.\ (\ref{Sigma_FLEX}).
To begin with, we expand $V({\bm r}_{1}-{\bm r}_{2})\delta(\tau_{1}-\tau_{2})$ in eq.\ (\ref{Gamma^(0)}) as
\begin{equation}
V({\bm r}_{1}-{\bm r}_{2})\delta(\tau_{1}-\tau_{2})=\sum_{\vec{q}} V_{\bm q}\,{\rm e}^{i\vec{q}\cdot(\vec{r}_{1}-\vec{r}_{2})},
\label{V-exp}
\end{equation}
where $\vec{q}\equiv ({\bm q},i\omega_{\ell})$ with $\omega_{\ell}\equiv 2\ell \pi T$ the boson Matsubara frequency ($\ell=0,\pm 1, \cdots$),
and the summation over $\vec{q}$ is defined in the same way as eq.\ (\ref{p-sum}) with the replacement $p\rightarrow q$ and $n\rightarrow \ell$.
Let us substitute eq.\ (\ref{Gamma^(0)}) into eq.\ (\ref{hatGamma^(0)-2}), use eqs.\ (\ref{delta-fn}) and (\ref{V-exp}) subsequently, and make some changes of integration variables; see also Appendix\ref{appendix:Hubbard} for details. We thereby obtain an expansion of eq.\ (\ref{uGamma^(0)}) as
\begin{eqnarray}
&&\hspace{-10mm}
\langle 11'_{ii'}|\underline{\Gamma}^{(0)}|22'_{jj'}\rangle
= \sum_{\vec{p}\vec{p}^{\,\prime}\vec{q}}
\langle i\alpha, i'\alpha'|\underline{\Gamma}^{(0)}({\bm p},{\bm p}',{\bm q}) | j\beta, j'\beta'\rangle
\nonumber \\
&&\hspace{19mm}\times
{\rm e}^{i\vec{p}\cdot\vec{r}_{1}
-i\vec{p}_{-}\cdot\vec{r}_{1}^{\,\prime}
-i\vec{p}^{\,\prime}\cdot\vec{r}_{2}+i\vec{p}^{\,\prime}_{-}\cdot\vec{r}_{2}^{\,\prime}},
\label{hatGamma^(0)-exp}
\end{eqnarray}
where $\vec{p}_{-}$ is defined by
\begin{equation}
\vec{p}_{-}\equiv ({\bm p}-{\bm q},i\varepsilon_{n}-i\omega_{\ell}),
\end{equation}
and the Fourier coefficient is given by
\begin{eqnarray}
&&\hspace{-6mm}
\langle i\alpha, i'\alpha'|\underline{\Gamma}^{(0)}({\bm p},{\bm p}^{\prime},{\bm q}) | j\beta, j'\beta'\rangle
\nonumber \\
&&\hspace{-10mm}
=\frac{1}{2}\bigl[
(-1)^{i+j}\delta_{ii'}\delta_{jj'}\delta_{\alpha'\alpha}\delta_{\beta'\beta}V_{{\bm q}}
\nonumber \\
&&\hspace{-6mm}
-(-1)^{i+i'}\delta_{ij}\delta_{i'j'}\delta_{\alpha\beta}\delta_{\alpha'\beta'}V_{{\bm p}-{\bm p}'}
\nonumber \\
&&\hspace{-6mm}
+(-1)^{i+i'}\delta_{j',3-i}\delta_{j,3-i'}\delta_{\alpha\beta'}\delta_{\alpha'\beta}
V_{{\bm p}+{\bm p}'-{\bm q}}\bigr] .
\label{Gamma-p}
\end{eqnarray}
Note that the arguments $\tau_{1}$, $\tau_{1}'$, $\tau_{2}$, and $\tau_{2}'$ in eq.\ (\ref{hatGamma^(0)-exp}) are properly expanded
in fermion Matsubara frequencies;
this is why we have expressed eq.\ (\ref{V-exp}) in terms of boson Matsubara frequencies.
Using eq.\ (\ref{G_ij-exp}), we can also transform eq.\ (\ref{chi^(0)}) into
\begin{eqnarray}
&&\hspace{-10mm}
\langle 11'_{ii'}|\underline{\chi}^{(0)}|22'_{jj'}\rangle
= \sum_{\vec{p}\vec{p}^{\,\prime}\vec{q}}
\langle i\alpha, i'\alpha'|\underline{\chi}^{(0)}(\vec{p},\vec{q}) | j\beta, j'\beta'\rangle\delta_{\vec{p}\vec{p}^{\,\prime}}
\nonumber \\
&&\hspace{20mm}\times
{\rm e}^{i\vec{p}\cdot\vec{r}_{1}
-i\vec{p}_{-}\cdot\vec{r}_{1}^{\,\prime}
-i\vec{p}^{\,\prime}\cdot\vec{r}_{2}+i\vec{p}^{\,\prime}_{-}\cdot\vec{r}_{2}^{\,\prime}},
\label{hatChi^(0)-exp}
\end{eqnarray}
with
\begin{equation}
\langle i\alpha, i'\alpha'|\underline{\chi}^{(0)}(\vec{p},\vec{q}) | j\beta, j'\beta'\rangle
=-G_{i\alpha,j\beta}(\vec{p}\,)G_{j'\beta',i'\alpha'}(\vec{p}_{-}).
\label{chi^(0)(p,q)}
\end{equation}
Substituting eqs.\ (\ref{GS-exp}), (\ref{hatGamma^(0)-exp}), and (\ref{hatChi^(0)-exp}) into eq.\ (\ref{Sigma_FLEX}),
we obtain the self-energy for homogeneous systems in the FLEX-S approximation as
\begin{eqnarray}
&&\hspace{-9mm}
\Sigma_{i\alpha,j\beta}(\vec{p})
=\sum_{\vec{q}}{\rm Tr}\biggl[\underline{\Gamma}^{(0)}({\bm p},{\bm p},{\bm q})
\nonumber \\
&&\hspace{9mm}
+\frac{4}{3}\sum_{\vec{p}_{1}}\underline{\Gamma}^{(0)}({\bm p},{\bm p}_{1},{\bm q}) \underline{\chi}^{(0)}(\vec{p}_{1},\vec{q})
\underline{\Gamma}^{(0)}({\bm p}_{1},{\bm p},{\bm q})
\nonumber \\
&&\hspace{9mm}
-2\sum_{\vec{p}_{1}}\underline{\Gamma}^{(0)}({\bm p},{\bm p}_{1},{\bm q}) \underline{\chi}^{(0)}(\vec{p}_{1},\vec{q})
\underline{\Gamma}^{(1)}({\bm p}_{1},{\bm p},\vec{q})\biggr]
\nonumber \\
&&\hspace{9mm}
\times \underline{\delta\chi}^{(i\alpha,j\beta)}(\vec{p}-\vec{q}),
\label{Sigma_FLEX(p)}
\end{eqnarray}
where $\underline{\Gamma}^{(1)}$ is the momentum-energy representation of eq.\ (\ref{Gamma^(1)-def}) satisfying
\begin{eqnarray}
&&\hspace{-10mm}
\underline{\Gamma}^{(1)}({\bm p},{\bm p}',\vec{q})=\underline{\Gamma}^{(0)}({\bm p},{\bm p}',{\bm q})-
\sum_{\vec{p}_{1}}\underline{\Gamma}^{(0)}({\bm p},{\bm p}_{1},{\bm q})
\nonumber \\
&&\hspace{13mm}
\times \underline{\chi}^{(0)}(\vec{p}_{1},\vec{q})
\underline{\Gamma}^{(1)}({\bm p}_{1},{\bm p}',\vec{q}),
\label{Gamma^(1)(p)}
\end{eqnarray}
and $\underline{\delta\chi}^{(i\alpha,j\beta)}(\vec{q})$ is defined by
\begin{eqnarray}
&&\hspace{-6mm}
\langle j_{1}\beta_{1},j_{2}\beta_{2}| \underline{\delta\chi}^{(i\alpha,j\beta)}(\vec{p})| i_{1}\alpha_{1},i_{2}\alpha_{2}\rangle
\nonumber \\
&&\hspace{-10mm}
\equiv -\delta_{j_{1}j}\delta_{\beta_{1}\beta}\delta_{i_{1}i}\delta_{\alpha_{1}\alpha} G_{i_{2}\alpha_{2},j_{2}\beta_{2}}(\vec{p}) .
\label{dchi-def}
\end{eqnarray}
The basis vectors of Tr in eq.\ (\ref{Sigma_FLEX(p)}) are given by $|i\alpha,i'\alpha'\rangle$,
which belong to the direct product $({\rm C}\times {\rm S})\times ({\rm C}\times {\rm S})$
with dimension $2^{4}$.
It is convenient to arrange them as $|({\rm cs})^{2}_{\nu}\rangle$ $(\nu=1,\cdots,16)$ with
\begin{eqnarray}
\begin{array}{ll}
|({\rm cs})^{2}_{1}\rangle =|1\uparrow,1\uparrow\rangle , \,\,\,&\,\,\,
|({\rm cs})^{2}_{2}\rangle =|1\uparrow,1\downarrow\rangle ,\\
|({\rm cs})^{2}_{3}\rangle =|1\downarrow,1\uparrow\rangle , \,\,\,&\,\,\,
|({\rm cs})^{2}_{4}\rangle =|1\downarrow,1\downarrow\rangle ,\\
|({\rm cs})^{2}_{5}\rangle =|2\uparrow,2\uparrow\rangle , \,\,\,&\,\,\,
|({\rm cs})^{2}_{6}\rangle =|2\uparrow,2\downarrow\rangle ,\\
|({\rm cs})^{2}_{7}\rangle =|2\downarrow,2\uparrow\rangle , \,\,\,&\,\,\,
|({\rm cs})^{2}_{8}\rangle =|2\downarrow,2\downarrow\rangle ,\\
|({\rm cs})^{2}_{9}\rangle =|1\uparrow,2\uparrow\rangle , \,\,\,&\,\,\,
|({\rm cs})^{2}_{10}\rangle =|1\uparrow,2\downarrow\rangle ,\\
|({\rm cs})^{2}_{11}\rangle =|1\downarrow,2\uparrow\rangle , \,\,\,&\,\,\,
|({\rm cs})^{2}_{12}\rangle =|1\downarrow,2\downarrow\rangle ,\\
|({\rm cs})^{2}_{13}\rangle =|2\uparrow,1\uparrow\rangle , \,\,\,&\,\,\,
|({\rm cs})^{2}_{14}\rangle =|2\uparrow,1\downarrow\rangle ,\\
|({\rm cs})^{2}_{15}\rangle =|2\downarrow,1\uparrow\rangle , \,\,\,&\,\,\,
|({\rm cs})^{2}_{16}\rangle =|2\downarrow,1\downarrow\rangle .
\end{array}
\label{BV2}
\end{eqnarray}
Then, $\underline{\chi}^{(0)}(\vec{p},\vec{q})$ in eq.\ (\ref{chi^(0)(p,q)}) can be represented by a $16\times 16$ matrix as
\begin{equation}
\underline{\chi}^{(0)}=
\begin{bmatrix}
-\underline{GG} & \underline{F\bar{F}} & \underline{G\bar{F}} & -\underline{FG} \\
\underline{\bar{F}F} & -\underline{\bar{G}\bar{G}} & -\underline{\bar{F}\bar{G}} & \underline{\bar{G}F} \\
-\underline{GF} & \underline{F\bar{G}} & \underline{G\bar{G}} & -\underline{FF} \\
\underline{\bar{F}G} & -\underline{\bar{G}\bar{F}} & -\underline{\bar{F}\bar{F}} & \underline{\bar{G}G}
\end{bmatrix},
\label{chi-mat}
\end{equation}
where each element is a $4\times 4$ submatrix given in terms of the elements of eq.\ (\ref{hatG(p)});
for example, $\underline{F}\underline{\bar{F}}$ denotes
\begin{equation}
\underline{F}\underline{\bar{F}}=
\begin{bmatrix}
F_{\uparrow\uparrow}\bar{F}_{\uparrow\uparrow} & F_{\uparrow\uparrow}\bar{F}_{\downarrow\uparrow}
& F_{\uparrow\downarrow}\bar{F}_{\uparrow\uparrow} & F_{\uparrow\downarrow}\bar{F}_{\downarrow\uparrow} \\
F_{\uparrow\uparrow}\bar{F}_{\uparrow\downarrow} & F_{\uparrow\uparrow}\bar{F}_{\downarrow\downarrow}
& F_{\uparrow\downarrow}\bar{F}_{\uparrow\downarrow} & F_{\uparrow\downarrow}\bar{F}_{\downarrow\downarrow} \\
F_{\downarrow\uparrow}\bar{F}_{\uparrow\uparrow} & F_{\downarrow\uparrow}\bar{F}_{\downarrow\uparrow}
& F_{\downarrow\downarrow}\bar{F}_{\uparrow\uparrow} & F_{\downarrow\downarrow}\bar{F}_{\downarrow\uparrow} \\
F_{\downarrow\uparrow}\bar{F}_{\uparrow\downarrow} & F_{\downarrow\uparrow}\bar{F}_{\downarrow\downarrow}
& F_{\downarrow\downarrow}\bar{F}_{\uparrow\downarrow} & F_{\downarrow\downarrow}\bar{F}_{\downarrow\downarrow}
\end{bmatrix},
\label{FF-def}
\end{equation}
with $F_{\alpha\beta}\bar{F}_{\gamma\delta}=F_{\alpha\beta}(\vec{p})\bar{F}_{\gamma\delta}(\vec{p}_{-})$.
We can also express $\underline{\Gamma}^{(0)}({\bm p},{\bm p}',{\bm q})$ in eq.\ (\ref{hatGamma^(0)-exp}) as
\begin{equation}
\underline{\Gamma}^{(0)}=
\begin{bmatrix}
\underline{\Gamma}^{(0a)} & \underline{\Gamma}^{(0b)} & \underline{0} & \underline{0} \\
\underline{\Gamma}^{(0b)} & \underline{\Gamma}^{(0a)} & \underline{0} & \underline{0} \\
\underline{0} & \underline{0} & \underline{\Gamma}^{(0c)} & \underline{0} \\
\underline{0} & \underline{0} & \underline{0} & \underline{\Gamma}^{(0c)}
\end{bmatrix} ,
\label{Gamma-mat}
\end{equation}
where $\underline{\Gamma}^{(0a,0b,0c)}$ are given by
\begin{subequations}
\label{Gamma-mat^(a,b,c)}
\begin{eqnarray}
&&\hspace{-11mm}
\underline{\Gamma}^{(0a)}({\bm p},{\bm p}',{\bm q})
\equiv
\frac{V_{\bm q}}{2}\underline{\gamma}^a-\frac{V_{{\bm p}-{\bm p}'}}{2}\underline{1},
\\
&&\hspace{-10.4mm}
\underline{\Gamma}^{(0b)}({\bm p},{\bm p}',{\bm q})
\equiv
\frac{V_{{\bm p}+{\bm p}'-{\bm q}}}{2}\underline{\gamma}^b-\frac{V_{\bm q}}{2}\underline{\gamma}^a,
\\
&&\hspace{-10mm}
\underline{\Gamma}^{(0c)}({\bm p},{\bm p}',{\bm q})
\equiv
\frac{V_{{\bm p}-{\bm p}'}}{2}\underline{1}-\frac{V_{{\bm p}+{\bm p}'-{\bm q}}}{2}\underline{\gamma}^b,
\end{eqnarray}
\end{subequations}
with $\underline{1}$ denoting the $4\times 4$ unit matrix and
\begin{equation}
\underline{\gamma}^{a}\equiv
\begin{bmatrix}
1 & 0 & 0 & 1 \\
0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 \\
1 & 0 & 0 & 1
\end{bmatrix},\hspace{5mm}
\underline{\gamma}^{b}\equiv
\begin{bmatrix}
1 & 0 & 0 & 0 \\
0 & 0 & 1 & 0 \\
0 & 1 & 0 & 0 \\
0 & 0 & 0 & 1
\end{bmatrix}.
\end{equation}
Equation (\ref{dchi-def}) may also be transformed into a matrix representation.
Equations (\ref{DG(p)}) and (\ref{Sigma_FLEX(p)}) form a closed set of self-consistent equations
for $\hat{G}(\vec{p})$ in the FLEX-S approximation, which is applicable to an arbitrary pairing symmetry
including a mixture of singlet and triplet pairings.
By using the representation of eq.\ (\ref{BV2}), practical calculations may be performed based on the standard
matrix algebra plus treatment of integral equations.
Some unitary transformations will be helpful for simplifying those calculations.
For example, consider the orthogonal matrix
\begin{equation}
\underline{R}=
\begin{bmatrix}
\underline{R}^{a} & -\underline{R}^{b} & \underline{0} & \underline{0} \\
\underline{R}^{b} & \underline{R}^{a} & \underline{0} & \underline{0} \\
\underline{0} & \underline{0} & \underline{R}^{c} & \underline{0} \\
\underline{0} & \underline{0} & \underline{0} & \underline{R}^{c}
\end{bmatrix} ,
\label{R-def}
\end{equation}
with
\begin{subequations}
\begin{equation}
\underline{R}^{a}\equiv
\begin{bmatrix}
\frac{1}{2} & 0 & 0 & \frac{1}{2} \\
0 & \frac{1}{\sqrt{2}} & 0 & 0 \\
0 & 0 & \frac{1}{\sqrt{2}} & 0 \\
-\frac{1}{2} & 0 & 0 & \frac{1}{2}
\end{bmatrix},
\end{equation}
\begin{equation}
\underline{R}^{b}\equiv
\begin{bmatrix}
\frac{1}{2} & 0 & 0 & \frac{1}{2} \\
0 & 0 & \frac{1}{\sqrt{2}} & 0 \\
0 & \frac{1}{\sqrt{2}} & 0 & 0 \\
-\frac{1}{2} & 0 & 0 & \frac{1}{2}
\end{bmatrix},
\end{equation}
\begin{equation}
\underline{R}^{c}\equiv
\begin{bmatrix}
1 & 0 & 0 & 0 \\
0 & \frac{1}{\sqrt{2}} & -\frac{1}{\sqrt{2}} & 0 \\
0 & \frac{1}{\sqrt{2}} & \frac{1}{\sqrt{2}} & 0 \\
0 & 0 & 0 & 1
\end{bmatrix} .
\end{equation}
\end{subequations}
By using $\underline{R}$, eq.\ (\ref{Gamma-mat}) is transformed into the diagonal form
\begin{equation}
\underline{\tilde{\Gamma}}^{(0)}\equiv \underline{R}\,\underline{\Gamma}^{(0)}\underline{R}^{-1}=
\begin{bmatrix}
\underline{\tilde{\Gamma}}^{(0a)} & \underline{0} & \underline{0} & \underline{0} \\
\underline{0} & \underline{\tilde{\Gamma}}^{(0b)} & \underline{0} & \underline{0} \\
\underline{0} & \underline{0} & \underline{\tilde{\Gamma}}^{(0c)} & \underline{0} \\
\underline{0} & \underline{0} & \underline{0} & \underline{\tilde{\Gamma}}^{(0c)}
\end{bmatrix} ,
\label{tGamma-mat}
\end{equation}
where $\underline{\tilde{\Gamma}}^{(0a,0b,0c)}({\bm p},{\bm p}',{\bm q})$ are given by
\begin{subequations}
\label{tGamma-mat^(a,b,c)}
\begin{eqnarray}
&&\hspace{-10.5mm}
\underline{\tilde{\Gamma}}^{(0a)}=
\begin{bmatrix}
V^{a} \!&\! 0 \!&\! 0 \!&\! 0 \\
0 \!&\! -V^{c} \!&\! 0 \!&\! 0 \\
0 \!&\! 0 \!&\! -V^{c} \!&\! 0 \\
0 \!&\! 0 \!&\! 0 \!&\! -V^{c}
\end{bmatrix} ,
\label{tGamma^a}
\\
&&\hspace{-10mm}
\underline{\tilde{\Gamma}}^{(0b)}=V^{b}\underline{1} ,
\label{tGamma^b}
\\
&&\hspace{-10mm}
\underline{\tilde{\Gamma}}^{(0c)}=
\begin{bmatrix}
-V^{b} & 0 & 0 & 0 \\
0 & V^{c} & 0 & 0 \\
0 & 0 & -V^{b} & 0 \\
0 & 0 & 0 & -V^{b}
\end{bmatrix} ,
\label{tGamma^c}
\end{eqnarray}
\end{subequations}
with
\begin{subequations}
\label{V^eo}
\begin{eqnarray}
&&\hspace{-10.5mm}
V^{a}({\bm p},{\bm p}',{\bm q})\equiv 2V_{\bm q}-\frac{V_{{\bm p}+{\bm p}'-{\bm q}}+V_{{\bm p}-{\bm p}'}}{2},
\\
&&\hspace{-10mm}
V^{b}({\bm p},{\bm p}',{\bm q})\equiv \frac{V_{{\bm p}+{\bm p}'-{\bm q}}-V_{{\bm p}-{\bm p}'}}{2},
\\
&&\hspace{-10mm}
V^{c}({\bm p},{\bm p}',{\bm q})\equiv \frac{V_{{\bm p}+{\bm p}'-{\bm q}}+V_{{\bm p}-{\bm p}'}}{2}.
\end{eqnarray}
\end{subequations}
The other quantities in eq.\ (\ref{Sigma_FLEX(p)}) may also be expressed in some convenient forms after the transformation.
Finally, we briefly consider the two-particle Green's function defined by eq.\ (\ref{calK-def}), which obeys eq.\ (\ref{calK2}).
Let us expand the quantities in eq.\ (\ref{Mat}) as eq.\ (\ref{hatGamma^(0)-exp}), where
$\underline{\chi}^{(0{\rm e})}(\vec{p},\vec{p}^{\,\prime},\vec{q})$ is obtained as
\begin{eqnarray}
&&\hspace{-6.2mm}
\langle i\alpha, i'\alpha'|\underline{\chi}^{(0{\rm e})}(\vec{p},\vec{p}^{\,\prime},\vec{q}) | j\beta, j'\beta'\rangle
\nonumber \\
&&\hspace{-10mm}
=-G_{i\alpha,j\beta}(\vec{p})G_{j'\beta',i'\alpha'}(\vec{p}_{-})\delta_{\vec{p}^{\,\prime}\vec{p}}
\nonumber \\
&&\hspace{-6.2mm}
+G_{i\alpha,3-j',\beta'}(\vec{p})G_{3-j,\beta,i'\alpha'}(\vec{p}_{-})\delta_{\vec{p}^{\,\prime},-\vec{p}+\vec{q}}.
\label{uchi^(0)(p,q)}
\end{eqnarray}
Equation (\ref{calK2}) is thereby transformed into
\begin{eqnarray}
&&\hspace{-10mm}
\underline{\cal K}(\vec{p},\vec{p}^{\,\prime},\vec{q})=\underline{\chi}^{(0{\rm e})}(\vec{p},\vec{p}^{\,\prime},\vec{q})
-\sum_{\vec{p}_{1}}\underline{\chi}^{(0)}(\vec{p},\vec{q})\underline{\Gamma}(\vec{p},\vec{p}_{1},\vec{q})
\nonumber \\
&&\hspace{10mm}
\times \underline{\chi}^{(0{\rm e})}(\vec{p}_{1},\vec{p}^{\,\prime},\vec{q}),
\label{calK-p}
\end{eqnarray}
where $\underline{\Gamma}$ is the Fourier transform of eq.\ (\ref{BS}) obeying
\begin{eqnarray}
&&\hspace{-9mm}
\underline{\Gamma}(\vec{p},\vec{p}^{\,\prime},\vec{q})=\underline{\Gamma}^{({\rm ir})}(\vec{p},\vec{p}^{\,\prime},\vec{q})-
\sum_{\vec{p}_{1}}\underline{\Gamma}^{({\rm ir})}(\vec{p},\vec{p}_{1},\vec{q})\underline{\chi}^{(0)}(\vec{p}_{1},\vec{q})
\nonumber \\
&&\hspace{10mm}
\times\underline{\Gamma}(\vec{p}_{1},\vec{p}^{\,\prime},\vec{q}).
\label{BS-p}
\end{eqnarray}
Equations (\ref{calK-p}) and (\ref{BS-p}) correspond to eqs.\ (6) and (7) of Leggett for the spin-singlet pairing
in his theory of superfluid Fermi liquids,\cite{Leggett65} respectively.
Here, we can calculate $\underline{\cal K}$ microscopically based on $\Phi$ for both the singlet and triplet pairings.
\subsection{FLEX-S approximation for the contact interaction and spin-singlet pairing}
Now, we focus on the contact interaction
\begin{equation}
V_{\bm q}=U .
\label{U-def}
\end{equation}
In this case, we can simplify eq.\ (\ref{Sigma_FLEX(p)}) for the self-energy considerably.
Indeed, substituting eq.\ (\ref{U-def}) into eq.\ (\ref{tGamma-mat}),
we observe that only the 1st, 2nd, 3rd, 4th, 10th, and 14th diagonal elements are finite in the diagonal matrix $\underline{\tilde{\Gamma}}^{(0)}$;
the relevant $6\times 6$ submatrix is given by
\begin{equation}
\underline{\check{\Gamma}}^{(0)}=U
\begin{bmatrix}
1 & 0 & 0 & 0 & 0 & 0 \\
0 & -1 & 0 & 0 & 0 & 0 \\
0 & 0 & -1 & 0 & 0 & 0 \\
0 & 0 & 0 & -1 & 0 & 0 \\
0 & 0 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & 0 & 0 & 1
\end{bmatrix} .
\label{cGamma^(0)}
\end{equation}
Moreover, absence of momentum dependences in $\underline{\Gamma}^{(0)}$ enables us to perform the $\vec{p}_{1}$ summations
in eqs.\ (\ref{Sigma_FLEX(p)}) and (\ref{Gamma^(1)(p)}).
Using these properties in eq.\ (\ref{Sigma_FLEX(p)}),
we realize that we only need the 1st, 2nd, 3rd, 4th, 10th, and 14th rows and columns of the two matrices
\begin{subequations}
\label{tchi-tdchi}
\begin{equation}
\underline{\tilde{\chi}}^{(0)}(\vec{q}) \equiv \sum_{\vec{p}}
\underline{R}\underline{\chi}^{(0)}(\vec{p},\vec{q})\underline{R}^{-1},
\label{tchi-tdchi1}
\end{equation}
\begin{equation}
\underline{\delta\tilde{\chi}}^{(i\alpha,j\beta)}(\vec{p})\equiv \underline{R}\,\underline{\delta\chi}^{(i\alpha,j\beta)}(\vec{p})\underline{R}^{-1}.
\label{tchi-tdchi2}
\end{equation}
\end{subequations}
These statements on eq.\ (\ref{Sigma_FLEX(p)}) for the contact interaction hold true for an arbitrary pairing symmetry.
We further restrict ourselves to the spin-singlet pairing with no magnetic polarization for simplicity.
The condition implies $G_{\alpha\beta}(\vec{p})=\delta_{\alpha\beta}G_{\vec{p}}$,
$F_{\uparrow\downarrow}(\vec{p})=-F_{\downarrow\uparrow}(\vec{p})$, and
$F_{\alpha\alpha}(\vec{p})=0$ in eq.\ (\ref{hatG(p)}).
It may also be possible for this single-component superconductivity to choose the phase of $F_{\uparrow\downarrow}(\vec{p})$ such that
its imaginary part originates solely from the $i\varepsilon_{n}$ dependence as $F_{\uparrow\downarrow}(\vec{p})=F_{\uparrow\downarrow}^{*}(\vec{p}^{\,*})$.
These relations are summarized together with eq.\ (\ref{GF-symm}) as
\begin{subequations}
\label{GF-singlet}
\begin{equation}
\begin{array}{l}
\vspace{1mm}
G_{\alpha\beta}(\vec{p})=\bar{G}_{\beta\alpha}(-\vec{p})=\delta_{\alpha\beta}G_{\vec{p}}, \\
\vspace{1mm}
F_{\uparrow\downarrow}(\vec{p})=-F_{\downarrow\uparrow}(\vec{p})=\bar{F}_{\uparrow\downarrow}(\vec{p})=-\bar{F}_{\downarrow\uparrow}(\vec{p})
\equiv F_{\vec{p}}=F_{-\vec{p}},\\
F_{\alpha\alpha}(\vec{p})=0.
\end{array}
\end{equation}
Similarly, the elements of eq.\ (\ref{hatSigma(p)}) will obey
\begin{equation}
\hspace{-2mm}
\begin{array}{l}
\vspace{1mm}
\Sigma_{\alpha\beta}(\vec{p})=\bar{\Sigma}_{\beta\alpha}(-\vec{p})=\delta_{\alpha\beta}\Sigma_{\vec{p}}, \\
\vspace{1mm}
\Delta_{\uparrow\downarrow}(\vec{p})=-\Delta_{\downarrow\uparrow}(\vec{p})=\bar{\Delta}_{\uparrow\downarrow}(\vec{p})
=-\bar{\Delta}_{\downarrow\uparrow}(\vec{p})
\equiv \Delta_{\vec{p}}=\Delta_{-\vec{p}},\\
\Delta_{\alpha\alpha}(\vec{p})=0.
\end{array}
\end{equation}
\end{subequations}
Substituting eq.\ (\ref{GF-singlet}) into eq.\ (\ref{DG(p)}), we observe that the $4\times 4$ matrix equation is divided into a pair of $2\times 2$ equations that are equivalent to one another. The one composed of the 1st and 4th rows and columns are given by
\begin{equation}
\begin{bmatrix}
\vspace{1mm}
G_{\vec{p}} \!&\! F_{\vec{p}} \\
F_{\vec{p}} \!&\! -\bar{G}_{\vec{p}}
\end{bmatrix} =
\begin{bmatrix}
\vspace{1mm}
i\varepsilon_{n}-\epsilon_{\bm p}-\Sigma_{\vec{p}} \!&\! -\Delta_{\vec{p}} \\
-\Delta_{\vec{p}} \!&\! i\varepsilon_{n}+\epsilon_{\bm p}+\bar{\Sigma}_{\vec{p}}
\end{bmatrix}^{-1},
\label{DG-singlet}
\end{equation}
with $\bar{G}_{\vec{p}}\equiv G_{-\vec{p}}$ and $\bar{\Sigma}_{\vec{p}}\equiv \Sigma_{-\vec{p}}$.
The other finite submatrix from the 2nd and 3rd rows and columns are obtained from eq.\ (\ref{DG-singlet}) by changing the signs of
$F_{\vec{p}}$ and $\Delta_{\vec{p}}$ simultaneously.
Next, we set $(i\alpha,j\beta)=(1\!\uparrow,1\!\uparrow)$, $(1\!\uparrow,2\!\downarrow)$ in eq.\ (\ref{Sigma_FLEX(p)}) to express
$\Sigma_{\vec{p}}$ and $\Delta_{\vec{p}}$ in eq.\ (\ref{DG-singlet}) in terms of $G_{\vec{p}}$ and $F_{\vec{p}}$.
Elementary calculations with eqs.\ (\ref{dchi-def})-(\ref{FF-def}), (\ref{R-def}), and (\ref{GF-singlet}) show that the relevant $6\times 6$ submatrices of eqs.\ (\ref{tchi-tdchi1}) and (\ref{tchi-tdchi2}), which are denoted by
$\underline{\check{\chi}}^{(0)}(\vec{q})$ and $\underline{\delta\check{\chi}}^{(i\alpha,j\beta)}(\vec{p})$, respectively, are given by
\begin{subequations}
\label{check-mat}
\begin{eqnarray}
&&\hspace{-6mm}
\underline{\check{\chi}}^{(0)}
\nonumber \\
&&\hspace{-10mm}
=
\begin{bmatrix}
\vspace{1mm}
\chi^{(0)}_{-} \!&\! 0 \!&\! 0 \!&\! 0 \!&\! \sqrt{2}\chi^{(0)}_{GF} \!&\! -\sqrt{2}\chi^{(0)}_{\bar{G}F} \\
\vspace{1mm}
0 \!&\! \chi^{(0)}_{+} \!&\! 0 \!&\! 0 \!&\! 0 \!&\! 0 \\
\vspace{1mm}
0 \!&\! 0 \!&\! \chi^{(0)}_{+} \!&\! 0 \!&\! 0 \!&\! 0 \\
\vspace{1mm}
0 \!&\! 0 \!&\! 0 \!&\! \chi^{(0)}_{+} \!&\! 0 \!&\! 0 \\
\vspace{1mm}
\sqrt{2}\chi^{(0)}_{GF} \!&\! 0 \!&\! 0 \!&\! 0 \!&\! -\chi^{(0)}_{G\bar{G}} \!&\! -\chi^{(0)}_{FF} \\
-\sqrt{2}\chi^{(0)}_{\bar{G}F} \!&\! 0 \!&\! 0 \!&\! 0 \!&\! -\chi^{(0)}_{FF} \!&\! -\chi^{(0)}_{\bar{G}G}
\end{bmatrix}\! ,
\nonumber \\
\label{cchi^(0)}
\end{eqnarray}
\begin{equation}
\underline{\delta\check{\chi}}^{(1\uparrow,1\uparrow)}=-\frac{1}{4}
\begin{bmatrix}
G \!&\! 0 \!&\! 0 \!&\! -G \!&\! \sqrt{2}F \!&\! 0 \\
0 \!&\! 2G \!&\! 0 \!&\! 0 \!&\! 0 \!&\! 0 \\
0 \!&\! 0 \!&\! 0 \!&\! 0 \!&\! 0 \!&\! 0 \\
-G \!&\! 0 \!&\! 0 \!&\! G \!&\! -\sqrt{2}F \!&\! 0 \\
\sqrt{2}F \!&\! 0 \!&\! 0 \!&\! -\sqrt{2}F \!&\! -2\bar{G} \!&\! 0 \\
0 \!&\! 0 \!&\! 0 \!&\! 0 \!&\! 0 \!&\! 0
\end{bmatrix}\! ,
\end{equation}
\begin{equation}
\underline{\delta\check{\chi}}^{(1\uparrow,2\downarrow)}=\frac{1}{4}
\begin{bmatrix}
\vspace{1mm}
F \!&\! 0 \!&\! 0 \!&\! -F \!&\! -\sqrt{2}\bar{G} \!&\! 0 \\
\vspace{1mm}
0 \!&\! -2F \!&\! 0 \!&\! 0 \!&\! 0 \!&\! 0 \\
\vspace{1mm}
0 \!&\! 0 \!&\! 0 \!&\! 0 \!&\! 0 \!&\! 0 \\
\vspace{1mm}
F \!&\! 0 \!&\! 0 \!&\! -F \!&\! -\sqrt{2}\bar{G} \!&\! 0 \\
\vspace{1mm}
0 \!&\! 0 \!&\! 0 \!&\! 0 \!&\! 0 \!&\! 0 \\
\sqrt{2}G \!&\! 0 \!&\! 0 \!&\! -\sqrt{2}G \!&\! 2F \!&\! 0 \\
\end{bmatrix}\!,
\end{equation}
\end{subequations}
where the elements of eq.\ (\ref{cchi^(0)}) are defined by
\begin{subequations}
\label{chi_GG(q)}
\begin{eqnarray}
\chi^{(0)}_{GG}(\vec{q})\!\!\!\!&\equiv&\!\!\!\! -\sum_{\vec{p}}G_{\vec{p}}\,G_{\vec{p}_{-}}=\chi^{(0)}_{GG}(-\vec{q}),
\\
\chi^{(0)}_{FF}(\vec{q})\!\!\!\!&\equiv&\!\!\!\! -\sum_{\vec{p}} F_{\vec{p}}\,F_{\vec{p}_{-}}=\chi^{(0)}_{FF}(-\vec{q}),
\\
\chi^{(0)}_{G\bar{G}}(\vec{q})\!\!\!\!&\equiv&\!\!\!\! -\sum_{\vec{p}}G_{\vec{p}}\,\bar{G}_{\vec{p}_{-}}=\chi^{(0)}_{\bar{G}G}(-\vec{q}),
\\
\chi^{(0)}_{GF}(\vec{q})\!\!\!\!&\equiv&\!\!\!\! -\sum_{\vec{p}}G_{\vec{p}}\,F_{\vec{p}_{-}}=\chi^{(0)}_{\bar{G}F}(-\vec{q}),
\\
\chi^{(0)}_{\pm}(\vec{q})\!\!\!\!&\equiv&\!\!\!\! \chi^{(0)}_{GG}(\vec{q})\pm \chi^{(0)}_{FF}(\vec{q}).
\end{eqnarray}
\end{subequations}
Let us substitute eqs.\ (\ref{cGamma^(0)}) and (\ref{check-mat}) into eq.\ (\ref{Sigma_FLEX(p)}) for $(i\alpha,j\beta)=(1\!\uparrow,1\!\uparrow)$, $(1\!\uparrow,2\!\downarrow)$.
In both cases, the resultant Tr can be calculated separately for the two submatrices composed of the 2nd, 3rd, and 4th rows and columns and the 1st, 5th, and 6th rows and columns.
Also using $\sum_{\vec{q}}\chi^{(0)}_{G\bar{G}}(\vec{q})\bar{G}_{\vec{p}-\vec{q}}=
\sum_{\vec{q}}\chi^{(0)}_{GG}(\vec{q})G_{\vec{p}-\vec{q}}$, etc., in the second order, we obtain
\begin{subequations}
\label{SD-FLEX-S}
\begin{eqnarray}
& &\!\! \hspace{-12mm}
\Sigma_{\vec{p}}
=\frac{1}{2}Un-2U^{2}\sum_{\vec{q}}\chi^{(0)}_{+}(\vec{q})G_{\vec{p}-\vec{q}}
\nonumber \\
& &\!\!\! \hspace{-3mm}
+U^{2} \sum_{\vec{q}}\biggl( \frac{3}{2}\chi^{({\rm s})}_{\vec{q}}G_{\vec{p}-\vec{q}}
+\frac{1}{2}{\rm Tr}\,\underline{\chi}^{({\rm c})}_{\vec{q}}\underline{\cal G}_{\vec{p}-\vec{q}}\biggr) ,
\label{Sigma-FLEX-S}
\end{eqnarray}
\begin{eqnarray}
& &\!\! \hspace{-12mm}
\Delta_{\vec{p}}
=U\sum_{\vec{p}^{\,\prime}}F_{\vec{p}^{\,\prime}}
-2U^{2}\sum_{\vec{q}}\chi^{(0)}_{+}(\vec{q})F_{\vec{p}-\vec{q}}
\nonumber \\
& &\!\!\! \hspace{-3mm}
+U^{2} \sum_{\vec{q}}\biggl( \frac{3}{2}\chi^{({\rm s})}_{\vec{q}}F_{\vec{p}-\vec{q}}
-\frac{1}{2}{\rm Tr}\,\underline{\chi}^{({\rm c})}_{\vec{q}}\underline{\cal F}_{\vec{p}-\vec{q}}
\biggr) .
\label{Delta-FLEX-S}
\end{eqnarray}
\end{subequations}
Here $n$ is the particle density
\begin{equation}
n=2\sum_{\vec{p}}G_{\vec{p}}\,{\rm e}^{i\varepsilon_{n}0_+}
\label{n}
\end{equation}
with $0_+$ an infinitesimal positive constant,\cite{LW60}
the functions $\underline{\chi}^{(0{\rm c})}_{\vec{q}}$, $\chi^{({\rm s})}_{\vec{q}}$, and $\underline{\chi}^{({\rm c})}_{\vec{q}}$ are
defined by
\begin{subequations}
\label{chi^cs}
\begin{equation}
\underline{\chi}^{(0{\rm c})}_{\vec{q}}\equiv
\begin{bmatrix}
\vspace{1mm}
\chi^{(0)}_{-}(\vec{q}) & \sqrt{2}\chi_{GF}^{(0)}(\vec{q}) & -\sqrt{2}\chi_{\bar{G}F}^{(0)}(\vec{q})\\
\vspace{1mm}
\sqrt{2}\chi_{GF}^{(0)}(\vec{q}) & -\chi^{(0)}_{G\bar{G}}(\vec{q}) & -\chi^{(0)}_{FF}(\vec{q}) \\
-\sqrt{2}\chi_{\bar{G}F}^{(0)}(\vec{q}) & -\chi^{(0)}_{FF}(\vec{q}) & -\chi^{(0)}_{\bar{G}G}(\vec{q})
\end{bmatrix},
\label{chi^(0c)}
\end{equation}
\begin{equation}
\chi^{({\rm s})}_{\vec{q}}\equiv \frac{\chi^{(0)}_{+}(\vec{q})}{1-U\chi^{(0)}_{+}(\vec{q})},
\end{equation}
\begin{equation}
\underline{\chi}^{({\rm c})}_{\vec{q}}\equiv \underline{\chi}^{(0{\rm c})}_{\vec{q}}\bigl(\underline{1}+U\underline{\chi}^{(0{\rm c})}_{\vec{q}}\bigr)^{-1},
\end{equation}
with superscripts $^{\rm c}$ and $^{\rm s}$ denoting ``charge'' and ``spin,'' respectively,
and $\underline{\cal G}_{\vec{p}}$ and $\underline{\cal F}_{\vec{p}}$ are given by
\begin{equation}
\underline{\cal G}_{\vec{p}}\equiv
\begin{bmatrix}
G_{\vec{p}} & \sqrt{2}F_{\vec{p}} & 0 \\
\sqrt{2}F_{\vec{p}} & -2\bar{G}_{\vec{p}} & 0 \\
0 & 0 & 0
\end{bmatrix},
\end{equation}
\begin{equation}
\underline{\cal F}_{\vec{p}}=
\begin{bmatrix}
F_{\vec{p}} & -\sqrt{2}\bar{G}_{\vec{p}} & 0 \\
0 & 0 & 0 \\
\sqrt{2}G_{\vec{p}} & 2F_{\vec{p}} & 0
\end{bmatrix},
\end{equation}
\end{subequations}
respectively.
Setting $F=0$ in eq.\ (\ref{Sigma-FLEX-S}) yields the normal-state self-energy
\begin{eqnarray}
& & \hspace{-8.5mm}
\Sigma_{\vec{p}}^{({\rm n})}
\!=\frac{1}{2}Un+U^{2}\sum_{\vec{q}}\!\!\left[-2\chi^{(0)}_{GG}(\vec{q})G_{\vec{p}-\vec{q}}+\frac{3}{2}\frac{\chi^{(0)}_{GG}(\vec{q})G_{\vec{p}-\vec{q}}}{1-U\chi^{(0)}_{GG}(\vec{q})}\right.
\nonumber \\
& & \hspace{2.5mm} \left.
+\frac{1}{2}\frac{\chi^{(0)}_{GG}(\vec{q})G_{\vec{p}-\vec{q}}}{1+U\chi^{(0)}_{GG}(\vec{q})}
+\frac{\chi^{(0)}_{G\bar{G}}(\vec{q})\bar{G}_{\vec{p}-\vec{q}}}{1-U\chi^{(0)}_{G\bar{G}}(\vec{q})}\right] .
\label{Sigma^(n)}
\end{eqnarray}
The second term on the right-hand side completely agrees with eq.\ (2.31) by Bickers and Scalapino in the normal FLEX approximation,\cite{BS89}
as it should.
To find the equation for $T_{c}$, let us assume the second-order transition and linearize (i) the right-hand side of eq.\ (\ref{Delta-FLEX-S})
in terms of $F$ and (ii) $F$ itself in eq.\ (\ref{DG-singlet}) in terms of $\Delta$ as
\begin{equation}
F_{\vec{p}}\rightarrow F^{(1)}_{\vec{p}}\equiv \frac{\Delta_{\vec{p}}}{\bigl(i\varepsilon_{n}-\epsilon_{\bm p}-\Sigma^{({\rm n})}_{\vec{p}}\bigr)
\bigl(i\varepsilon_{n}+\epsilon_{\bm p}+\bar{\Sigma}^{({\rm n})}_{\vec{p}}\bigr)}.
\end{equation}
We thereby obtain the linearized self-consistency equation as
\begin{equation}
\Delta_{\vec{p}}=\sum_{\vec{p}^{\,\prime}}V^{\rm s}_{\vec{p}\vec{p}^{\,\prime}}F_{\vec{p}^{\,\prime}}^{(1)},
\end{equation}
where $V_{\vec{p}\vec{p}^{\,\prime}}^{\rm s}$ is given by
\begin{eqnarray}
&&\hspace{-10mm}
V_{\vec{p}\vec{p}^{\,\prime}}^{\rm s}
=U+U^{2}\biggl[-2\chi^{(0)}_{GG}(\vec{p}-\vec{p}^{\,\prime})
+\frac{3}{2}\frac{\chi^{(0)}_{GG}(\vec{p}-\vec{p}^{\,\prime})}{1-U\chi^{(0)}_{GG}(\vec{p}-\vec{p}^{\,\prime})}
\nonumber \\
&&\hspace{0.5mm}
-\frac{1}{2}\frac{\chi^{(0)}_{GG}(\vec{p}-\vec{p}^{\,\prime})}{1+U\chi^{(0)}_{GG}(\vec{p}-\vec{p}^{\,\prime})}
\biggr]
\nonumber \\
&&\hspace{0.5mm}
-U^{2}\sum_{\vec{q}}\frac{(G_{\vec{p}+\vec{q}}+G_{-\vec{p}+\vec{q}})G_{\vec{p}^{\,\prime}+\vec{q}}}{\bigl[1+U\chi^{(0)}_{GG}(\vec{q})\bigr]\bigl[1-U\chi^{(0)}_{G\bar{G}}(\vec{q})\bigr]}.
\label{Tc-FLEX-S-singlet}
\end{eqnarray}
This $V_{\vec{p}\vec{p}^{\,\prime}}^{\rm s}$ may be regarded as the pair potential for the spin-singlet pairing without feedback effects.\cite{AB73,Leggett75}
Note in this context that the feedback effects are included naturally in Eq.\ (\ref{Delta-FLEX-S}).
Equations (\ref{DG-singlet}), (\ref{SD-FLEX-S}), and (\ref{n}) form a closed set of self-consistent equations
for the spin-singlet pairing within the FLEX-S approximation.
The replacement
$$
\underline{\chi}^{(0{\rm c})}_{\vec{q}}\rightarrow
\begin{bmatrix}
\vspace{1mm}
\chi^{(0)}_{-}(\vec{q}) & 0 & 0 \\
\vspace{1mm}
0 & 0 & 0 \\
0 & 0 & 0
\end{bmatrix}
$$
in eq.\ (\ref{SD-FLEX-S}) yields the standard FLEX approximation for superconductivity
with only the particle-hole contribution;\cite{Tewordt74,Tewordt93,Tewordt95,PB94,MS94,GLSB96,TM98,KFY97}
the charge contribution is disregarded further in many cases.
This approximation may be appropriate for
the Hubbard model near half filling where spin fluctuations are dominant, for example.
On the other hand, neglecting the particle-hole contribution in eq.\ (\ref{SD-FLEX-S}) altogether by $\chi^{(0)}_{\pm}(\vec{q})\rightarrow 0$ and
$$
\underline{\chi}^{(0{\rm c})}_{\vec{q}}\rightarrow
\begin{bmatrix}
\vspace{1mm}
0 & 0 & 0 \\
\vspace{1mm}
0 & -\chi^{(0)}_{G\bar{G}}(\vec{q}) & -\chi^{(0)}_{FF}(\vec{q}) \\
0 & -\chi^{(0)}_{FF}(\vec{q}) & -\chi^{(0)}_{\bar{G}G}(\vec{q})
\end{bmatrix},
$$
we obtain the particle-particle ladder approximation
adopted by Haussmann {\em et al}.\ for the BCS-BEC crossover problem.\cite{Haussmann07}
Hence, eq.\ (\ref{SD-FLEX-S}) enables us to study the two cases of considerable physical interest on an equal footing,
including a crossover from the one limit to the other.
It is also clear that the FLEX+T-matrix approximation,\cite{DMT97,Yanase01} which incorporates a contribution of the two-particle Green's function into the single-particle self-energy, cannot be reproduced as a limit of the FLEX-S approximation, where both of the single-particle and two-particle Green's functions should be studied starting from the single functional of eq.\ (\ref{Phi_FLEX}). Hence, a microscopic foundation of the FLEX+T-matrix approximation remains to be examined critically by including more terms from eq.\ (\ref{Phi}) beyond the FLEX-S approximation.
\section{Summary}
We have developed a concise self-consistent perturbation expansion for superconductivity in terms of Green's function $\hat{G}$ and fully symmetrized bare vertex $\underline{\Gamma}^{(0)}$ defined by eqs.\ (\ref{hatG}) and (\ref{hatGamma^(0)}), respectively, so as to incorporate all the pair processes naturally and conveniently.
Indeed, distinct diagrams for the key functional $\Phi[\hat{G}]$ of eq.\ (\ref{Phi}) are exhausted within the fourth order by the five graphs of Fig.\ \ref{Fig5}.
Corresponding analytic expressions can be written down immediately as eq.\ (\ref{Phi_1-4}) based on the Feynman rules (a)-(f) given around eq.\ (\ref{tG_ij}).
Using an approximate $\Phi$ thereby constructed, we can calculate $\hat{G}$ self-consistently by eqs.\ (\ref{DG}) and (\ref{Sigma-Phi}), the thermodynamic potential subsequently by eq.\ (\ref{Omega}), and also the two-particle Green's function of eq.\ (\ref{calK-def}) by eq.\ (\ref{calK}) with
eqs.\ (\ref{chi^(0)}), (\ref{Gamma^(ir)}), and (\ref{Mat}). Moreover, entropy may be written down explicitly once $\hat{G}$ and $\hat{\Sigma}$ are known.\cite{Kita99}
Thus, a single procedure of choosing an approximate $\Phi$ enables us to calculate the whole thermodynamic hierarchy, and the results may be improved by starting from a better $\Phi$.
Various fluctuations and competing orders may be described simultaneously through the matrix structure of $\underline{\Gamma}^{(0)}$.
Since the conservation laws are obeyed automatically, the basic formalism of \S 2 in the coordinate representation can be applied to study non-equilibrium phenomena by changing the Matsubara contour into the Schwinger-Keldysh contour.\cite{Kita10}
The idea of symmetrizing the bare vertex in terms of the particle-hole indices will also be useful for the electron-phonon Hamiltonian to simplify its perturbation expansion.
This formalism has been transformed for homogeneous systems into the momentum-energy representation in \S 3.1. We have subsequently derived a closed set of self-consistent equations for the contact potential in the spin-singlet FLEX-S approximation as eqs.\ (\ref{DG-singlet}), (\ref{SD-FLEX-S}), and (\ref{n}), which adds all the pair processes characteristic of superconductivity to the normal FLEX approximation.\cite{BS89} As shown in the paragraph after eq.\ (\ref{Tc-FLEX-S-singlet}), it contains extra terms besides those in the standard FLEX approximation for superconductivity, and also embraces the particle-particle ladder approximation by Haussmann {\em et al}.\ \cite{Haussmann07} for the BCS-BEC crossover problem.
Thus, our FLEX-S approximation incorporates the particle-hole and particle-particle diagrams on an equal footing with their mixing effects, thereby enabling us to study the BCS-BEC problem from low- to high-density regions continuously.
\begin{acknowledgments}
This work is supported by a Grant-in-Aid for Scientific Research
from the Ministry of Education, Culture, Sports, Science and Technology of Japan.
\end{acknowledgments}
|
\section{Introduction}
Electromagnetic (EM) properties of particles are fundamental observables
for their internal structure. For hadrons it is still not possible
to reveal this structure analytically from QCD first principles. Two
prominent ways to obtain information on baryon EM properties are chiral
perturbation theories ($\chi$PT) and lattice QCD (lQCD) which both
allow to investigate the quark/pion mass dependence of observables.
At present, finite volume lQCD results are mainly compared to infinite
volume $\chi$PT ones
\cite{Alesandrou:NlQCD,Bratt:NlQCD,Collins:NlQCD,Syrisyn:NlCQD,Ledwig(2011):covChptNDelta}
where discrepancies in the small pion mass region are seen
\cite{Syrisyn:NlCQD,Collins:NlQCD,Ledwig(2011):covChptNDelta}.
EM finite volume effects on the $\chi$PT side are therefore of interest, however,
reveal certain subtleties. Special care has to be taken for the decomposition
of the vector-current matrix element in form factors as well as for their Lorentz
invariance \cite{Tiburzi3p}. An alternative approach, e.g. to investigate the
magnetic moment, is to use the particle's self-energy in an external EM
field. This was done in \cite{TiburziBFT} for pions and nucleons in the non-relativistic
heavy baryon $\chi$PT.
One motivation for this work is to derive the corresponding covariant infinite
volume $\chi$PT self-energies of the nucleon and $\Delta(1232)$ that would be needed
for a future finite volume study. Another one is to investigate further the non-analytic EM field dependence of
self-energies \cite{Ledwig(2010):nonAna}. We see that beside the found condition
for expanding a resonance self-energy in the field strength there exist two more.
For this, we modernize the EM background field technique (BFT) of \cite{Sommerfield}
and apply it to stable and unstable particles.
In 1958 C. M. Sommerfield used the BFT to obtain the electron's AMM
up to the fourth order, $\kappa_{e}=\frac{e^{2}}{8\pi^{2}}-0.328\frac{e^{4}}{4\pi^{3}}$,
correctly for the first time. However, it is the well known three
point function method, i.e. the one-photon approximation, that became
the preferred method to calculate EM moments in field theories. We
will use both techniques to check our formulas. It turns out that the
self-energies of particles, stable as well as unstable, generally
depend non-analytically on the magnetic field $B$. This is e.g.
seen for the nucleon and $\Delta\left(1232\right)$-isobar in the
covariant $SU(2)$ B$\chi$PT with $N-\pi$ loops where the $\tilde{MS}$
renormalized results with $\mu=m_{\pi}/M_{N}$ are:
\begin{eqnarray}
\Sigma_{p}\left(\mathcal{B}\right) & = & \frac{M_{N}C_{N}}{24\left(1
-\mathcal{B}\right)^{4}}\left[n_{1}+n_{2}\ln\mu+\frac{n_{3}}{\sqrt{4\mu^{2}
-\left(\mathcal{B}+\mu^{2}\right)^{2}}}\text{\ensuremath{\arccos}}
\frac{\mathcal{B}+\mu^{2}}{2\mu}\right]\,\,\,,\\
\Sigma_{\Delta^{+}}\left(\mathcal{B}\right) & = &
\frac{M_{\Delta}C_{\Delta}}{\left(1-\mathcal{B}\right)^{4}}\Big[d_{1}+d_{2}
\ln\mu+d_{3}\ln r+d_{4}\ln\left(1-\mathcal{B}\right)\nonumber\\
& & +\frac{\left(\mathcal{B}-1\right) d_{5}\mbox{arctanh
\ensuremath{\omega}}_{1}}{\sqrt{\lambda^{2}+4\mu^{2}\mathcal{B}}}
+\frac{\left(\mathcal{B}-1\right)d_{6}\mbox{arctanh
\ensuremath{\omega}}_{2}}{\sqrt{\lambda^{2}+\mathcal{B}\left(\mathcal{B}
-2+2r^{2}+2\mu^{2}\right)}}+d_{7}\ln\left(r^{2}-\mathcal{B}\right)\Big],
\end{eqnarray}
with certain polynomials $n_{i}=n_{i}\left(\mathcal{B},\mu\right)$,
$d_{i}=d_{i}\left(\mathcal{B},\mu,r\right)$, $\lambda^{2}=\lambda^{2}\left(\mu,r\right)$
and further non-analytic functions $\omega_{i}=\omega_{i}\left(\mathcal{B},\mu,r\right)$
and notations given later. One difference of the self-energies is
that the square root in $\Sigma_{p}\left(B\right)$ can be expanded
in a weak magnetic field $B$ for all pion masses $m_{\pi}>0$ whereas
for the $\Delta\left(1232\right)$-isobar only if the condition
\begin{equation}
\frac{eB}{2M_{\Delta}}\ll|M_{\Delta}-\left(M_{N}+m_{\pi}\right)|
\end{equation}
between the nucleon, $\Delta(1232)$ and pion masses is met. This
is a general situation for unstable particles and was discussed in
\cite{Ledwig(2010):nonAna} within a simpler field theory. In the
work \cite{Ledwig(2010):nonAna} only the one Feynman graph leading
to the above condition was investigated whereas we derive here the remaining two
conditions coming from the two more Feynman graphs.
In the next section we will give necessary notations for the BFT and
will use them in the third section to investigate the self-energies
of stable and unstable particles. In the forth and fifth section we
apply the BFT to the nucleon and $\Delta\left(1232\right)$-isobar
and derive the above expressions.
\section{Field equations in presence of an external em field}
We consider spin-1/2 and spin-3/2 fields moving in a constant electromagnetic
field given by the potential $A_{\mu}(x)=-\frac{1}{2}F_{\mu\nu}x^{\nu}$
with $F^{\mu\nu}=\partial^{[\mu}A^{\nu]}$ as the electromagnetic
field strength tensor. The Dirac equation for a particle $\Psi(x)$
of mass $M$ with the EM minimal substitution is:
\begin{equation}
\left[\s\Pi-M\right]\Psi_{A}(x)=0\,\,\,,\label{eq:dirac_eq_A}
\end{equation}
where we take $e>0$ and define the momentum $\Pi_{\mu}=i\partial_{\mu}-eA_{\mu}(x)$.
This operator is non-commutative with
\begin{equation}
\left[\Pi_{\mu},\Pi_{\nu}\right]=\frac{1}{i}F_{\mu\nu}\,\,\,.\label{eq:commutation_relation}
\end{equation}
The spin $3/2$ field $\psi_{\mu}(x)$ satisfies the equation
\begin{equation}
\left[\s\Pi-M\right]\Psi_{A}^{\mu}(x)=0,
\end{equation}
with the subsidiary conditions
\begin{eqnarray}
\gamma_{\mu}\Psi_{A}^{\mu}(x) & = & 0\,\,,\\
\Pi_{\mu}\Psi_{A}^{\mu}(x) & = & 0\,\,.
\end{eqnarray}
Because of the non-commutativity we have to symmetrize occurring expressions
and use for the propagators:
\begin{equation}
\frac{1}{\s\Pi-M}=\frac{1}{2}\left[\left(\s\Pi+M\right)\frac{1}{\Pi^{2}-M^{2}+\s
F}+\frac{1}{\Pi^{2}-M^{2}+\s F}\left(\s\Pi+M\right)\right]\,\,,
\end{equation}
where we introduce the notation $\s F=\frac{1}{2i}\gamma^{\alpha}F_{\alpha\beta}\gamma^{\beta}$
\cite{Sommerfield}. With this we calculate the particle's self-energy
$\Sigma\left(\s F\right)$ and obtain its anomalous magnetic moment
(AMM) $\kappa$ through the linear energy shift:
\begin{equation}
\langle\Psi_{A}|\Sigma\left(\s
F\right)|\Psi_{A}\rangle=\langle\Sigma\left(0\right)\rangle-\langle\s
F\rangle\frac{\kappa}{2M}+\mathcal{O}\left(\s F^{2}\right)\,\,\,.
\end{equation}
In Fig. \ref{fig:feynman_graphs} we list all Feynman graphs used
in this work for the BFT and three point function method. The upper
row shows all types of graphs that appear to the one-loop level in
the BFT, tadpole graphs not considered.
For better reading we use the notation $\vec{B}=M^{2}\,\mathcal{B}\,\vec{e}_{z}$,
i.e. $\mathcal{B}=|\vec{B}|/M^{2}$, with which the linear approximation
to the self-energy reads:
\begin{eqnarray}
\Sigma\left(B\right) & = &
\Sigma\left(0\right)-M\frac{1}{2}\kappa\mathcal{B}
+\mathcal{O}\left(B^{2}\right)\,\,\,,\label{eq:linear_approximation}\\
\Sigma\left(-B\right)-\Sigma\left(B\right) & = &
M\kappa\mathcal{B}+\mathcal{O}\left(B^{3}\right)\,\,\,,
\end{eqnarray}
where in the last equation the $B^{2}$ terms cancel out. The self-energy
formulas in this work omit some, but not all, $\mathcal{O}\left(B^{2}\right)$
terms. In particular non-analytic $B$ structures are preserved in
the\textbf{ }BFT\textbf{ }which are not in the one-photon approximation
Eq. (\ref{eq:linear_approximation}).
\begin{figure}
\includegraphics[scale=0.5]{SelfEnergy1_Psi+_PhiPsi}~~~\includegraphics[scale=0.5]{SelfEnergy2_Psi+_PhiPsi}
~~~\includegraphics[scale=0.5]{SelfEnergy3_Psi0_PhiPsi_BFT}\\
\includegraphics[scale=0.5]{N1}~~~\includegraphics[scale=0.5]{N2}\\
\includegraphics[scale=0.5]{D3}~~~\includegraphics[scale=0.5]{D4}
\caption{\label{fig:feynman_graphs} Upper row: Feynman graphs contributing
to the self-energy in the presence of an external electromagnetic
field. The blue lines indicate which loop-internal particle is affected
by the field. Middle and lower row: Feynman graphs contributing to
the three point function method to obtain the AMM. Single solid lines
correspond to stable particles (e.g. the nucleon), double solid lines
to unstable particles (e.g. the $\Delta\left(1232\right)$), dashed
lines to (pseudo-) scalar particles (e.g. the pion) and the blue cross
to the coupling of the photon.}
\end{figure}
\section{Scalar couplings and estimation of $\mathcal{O}(B^{2})$ effects}
We use the notations of the appendix and consider two spin 1/2 fields
$\Psi_{1}$, $\Psi_{2}$ interacting with a scalar field $\phi$:
\begin{eqnarray}
\mathcal{L} & = & \sum_{a=1}^{2}\,\overline{\Psi_{a}}\left(i\s
D_{a}-M_{a}\right)\Psi_{a}+\frac{1}{2}\left(D_{\mu}\phi\right)\left(D^{\mu}\phi\right)
-\frac{1}{2}m^{2}\phi^{2}+\sum_{a,b=1}^{2}g\overline{\Psi_{a}}\Gamma_{ab}\Psi_{b}\phi\,\,\,,\label{eq:ps_lagrangian}
\end{eqnarray}
with the covariant derivative $D_{\mu}=\partial_{\mu}+iqeA_{\mu}$,
$q$ as the charge of the field with $e>0$ and $\Gamma_{ab}$ either
$1$ or $\gamma_{5}$. In the upper row of Fig. \ref{fig:feynman_graphs}
we show all the types of self-energy graphs that can occur for a charged
or uncharged external particle $\Psi_{ex}$. The corresponding expressions
$\Sigma_{i}\left(B\right)$ with a loop-internal particle $\Psi_{in}$
are:
\begin{eqnarray}
\langle\Psi_{ex}|\Sigma_{1}\left(\s F\right)|\Psi_{ex}\rangle & = &
\frac{g^{2}}{i}
\langle\Psi_{ex}|\int\tilde{dl}\Gamma_{1}\frac{1}{\left[\s
l-M_{in}+i\varepsilon\right]}
\Gamma_{1}\frac{1}{\left[\left(\Pi-l\right)^{2}-m^{2}+i\varepsilon\right]}
|\Psi_{ex}\rangle\,\,\,,\\
\langle\Psi_{ex}|\Sigma_{2}\left(\s F\right)|\Psi_{ex}\rangle & = &
\frac{g^{2}}{i}\langle\Psi_{ex}|\int\tilde{dl}\Gamma_{2}\frac{1}{\left[\s\Pi-\s
l-M_{in}+i\varepsilon\right]}
\Gamma_{2}\frac{1}{\left[l^{2}-m^{2}+i\varepsilon\right]}|\Psi_{ex}\rangle\,\,\,,\\
\langle\Psi_{ex}|\Sigma_{3}\left(\s F\right)|\Psi_{ex}\rangle & = &
\frac{g^{2}}{i}\langle\Psi_{ex}|\int\tilde{dl}\Gamma_{3}\frac{1}{\left[\s\Pi-\s
l-M_{in}+i\varepsilon\right]}\Gamma_{3}\frac{1}{\left[\left(\Pi_{+}
+l\right)^{2}-m^{2}+i\varepsilon\right]}|\Psi_{ex}\rangle\,\,\,,
\end{eqnarray}
with $\Pi_{+}^{\mu}=i\partial^{\mu}+eA^{\mu}(x)$. For a third-spin
projection of $+1/2$ we can write these expressions in dimensional
regularization as:
\begin{eqnarray}
\Sigma_{i} & = &
\frac{-g^{2}}{\left(4\pi\right)^{2}}M_{ex}\int_{-\alpha_{i}}^{1-\alpha_{i}}dz\left[s_{5}
\left(z+\alpha_{i}\right)+r\right]\left[L+\ln\left(z^{2}-\lambda_{i}^{2}
-i\varepsilon\right)+\left(\delta_{i1}+\delta_{i2}\right)
\ln\left(1-\mathcal{B}\right)\right]
+\mathcal{O}\left(\mathcal{B}^{2}\right)\,\,\,,\label{eq:self_energy_ps}
\end{eqnarray}
with $\alpha_{i}$, $\lambda_{i}$ given below and $s_{5}=-1$ for
$\Gamma_{i}=\gamma_{5}$ and $+1$ for $\Gamma_{i}=1$. We see that
for $\vec{B}=0$ these expressions are the same and for $\vec{B}\neq0$
several different logarithms occur. The formula Eq. (\ref{eq:self_energy_ps})
omits some, but not all, $B^{2}$ terms and the integrated solution
for the real and imaginary parts are:
\begin{eqnarray}
\mbox{Re}\,\Sigma_{i} & = & \frac{-g^{2}}{\left(4\pi\right)^{2}}M_{ex}\,\,
\Big[+\left(s_{5}\alpha_{i}+r\right)\left(\beta_{i}\ln\left(\beta_{i}^{2}
-\lambda_{i}^{2}\right)+\alpha_{i}\ln\left(\alpha_{i}^{2}-\lambda_{i}^{2}\right)-2\right)\nonumber\\
& &
+s_{5}\frac{1}{2}\left(\alpha_{i}^{2}-\beta_{i}^{2}+\left(\beta_{i}^{2}
-\lambda_{i}^{2}\right)\ln\left(\beta_{i}^{2}-\lambda_{i}^{2}\right)
-\left(\alpha_{i}^{2}-\lambda_{i}^{2}\right)\ln\left(\alpha_{i}^{2}
-\lambda_{i}^{2}\right)\right)\nonumber\\
& &
+\left(\delta_{i1}+\delta_{i2}\right)\ln\left(1-\mathcal{B}\right)\left(s_{5}\alpha_{i}
+r+\frac{s_{5}}{2}\left(\beta_{i}^{2}-\alpha_{i}^{2}\right)\right)
+\left(s_{5}\alpha_{i}+r\right)\Omega_{i}\,\,\Big]\,\,\,,\\
\mbox{Im}\,\Sigma_{i} & = &
\frac{g^{2}\pi}{\left(4\pi\right)^{2}}M_{ex}\left(s_{5}\alpha_{i}+r\right)2\lambda_{i}
\end{eqnarray}
with $\beta_{i}=1-\alpha_{i}$ and
\begin{eqnarray}
\Omega_{i} & = & \begin{cases}
\begin{array}{c}
2\sqrt{-\lambda_{i}^{2}}\left(\arctan\frac{\beta_{i}}{\sqrt{-\lambda_{i}^{2}}}
+\arctan\frac{\alpha_{i}}{\sqrt{-\lambda_{i}^{2}}}\right)\\
2\sqrt{\lambda_{i}^{2}}\left(\mbox{arctanh}\frac{\beta_{i}}{\sqrt{\lambda_{i}^{2}}}
+\mbox{arctanh}\frac{\alpha_{i}}{\sqrt{\lambda_{i}^{2}}}\right)\end{array}
& \begin{array}{c}
\lambda_{i}^{2}<0\\
\lambda_{i}^{2}>0\end{array}\end{cases}\,\,\,,\label{eq:omega_contributions}
\end{eqnarray}
\begin{eqnarray}
\alpha_{0}=\frac{1}{2}\left(1+r^{2}-\mu^{2}\right) & \,\,,\,\, &
\lambda_{0}^{2}=\alpha_{0}^{2}-r^{2}\,\,,\\
\alpha_{1}=\frac{1}{2\left(1-\mathcal{B}\right)}\left(1+r^{2}-\mu^{2}-\mathcal{B}\right)
& \,\,,
\,\, & \lambda_{1}^{2}=\alpha_{1}^{2}-\frac{r^{2}}{1-\mathcal{B}}\,\,,\\
\alpha_{2}=\frac{1}{2\left(1-\mathcal{B}\right)}\left(1+r^{2}-\mu^{2}-2\mathcal{B}\right)
& \,\,
,\,\, & \lambda_{2}^{2}=\alpha_{2}^{2}-\frac{r^{2}-\mathcal{B}}{1-\mathcal{B}}\,\,,\\
\alpha_{3}=\frac{1}{2}\left(1+r^{2}-\mu^{2}-\mathcal{B}\right) & \,\,,\,\, &
\lambda_{3}^{2}=\alpha_{3}^{2}-r^{2}+\mathcal{B}\,\,.
\end{eqnarray}
The self-energies obtain imaginary parts if $\lambda_{i}^{2}$ becomes
positive together with $-\lambda_{i}<\beta_{i}$ and/or $\lambda_{i}>-\alpha_{i}$.
In addition, the non-analytic contributions $\Omega_{i}$ give constrains
on when the self-energies can be expanded for small $\mathcal{B}$.
These constrains can generically be written as:
\begin{eqnarray}
\sqrt{f_{i}(\mu,r)\mathcal{B}+\lambda_{0}^{2}} & \to &
|\mathcal{B}|<|\frac{\lambda_{0}^{2}}{f_{i}(\mu,r)}|\,\,\,,
\end{eqnarray}
for a certain function $f_{i}$ depending on the type of graph. One
of these constrains, coming from the first graph in Fig. \ref{fig:feynman_graphs},
was investigated in \cite{Ledwig(2010):nonAna} for the situation
of $M_{ex}>M_{in}$. In the following we investigate the remaining
two as well as further applications of Eq. (\ref{eq:self_energy_ps}).
\subsubsection{Nucleon-pion system}
The first example is the nucleon-pion ($N$, $\pi$) system with pseudo-scalar
couplings, $\Gamma_{a}^{NN\pi}=g_{a}\gamma_{5}$. For this we have
$s_{5}=-1$, $r=1$, $M_{ex}=M_{N}$, $m=m_{\pi}$ and write for the
proton and neutron energies:
\begin{eqnarray}
\Sigma_{p}\left(B\right) & = & 2\Sigma_{1}\left(B\right)+\Sigma_{2}\left(B\right)\,\,\,,\label{eq:ps_proton}\\
\Sigma_{n}\left(B\right) & = &
\Sigma_{3}\left(0\right)+2\Sigma_{3}\left(B\right)\,\,\,,\label{eq:ps_neutron_s3}
\end{eqnarray}
with
\begin{eqnarray}
\Sigma_{i} & = &
\frac{-g^{2}}{\left(4\pi\right)^{2}}M_{N}\int_{0}^{1}dz\,\left(1-z\right)\left[L+\ln\left(z\mu^{2}
+\left(1-z\right)^{2}+\mathcal{B}_{i}-i\varepsilon\right)\right]\,\,\,,
\end{eqnarray}
and $\mathcal{B}_{1}=z\left(1-z\right)\mathcal{B}$, $\mathcal{B}_{2}=-\left(1-z\right)^{2}\mathcal{B}$
and $\mathcal{B}_{3}=-\left(1-z\right)\mathcal{B}$. It is easy to
check that we get from this form the same AMM as obtained from the
usual three point function method:
\begin{eqnarray}
\kappa_{p}=2\kappa_{1}+\kappa_{2} & \,\,\,\,\,,\,\,\,\,\, & \kappa_{n}=-2\kappa_{1}+2\kappa_{2}\,\,\,,\\
\kappa_{1}=\frac{g^{2}}{\left(4\pi\right)^{2}}\int_{0}^{1}dz\frac{2z\left(1-z\right)^{2}}{z\mu^{2}+\left(1-z\right)^{2}}
& \,\,\,\,\,,\,\,\,\,\, &
\kappa_{2}=\frac{g^{2}}{\left(4\pi\right)^{2}}\int_{0}^{1}dz\frac{-2\left(1-z\right)^{3}}{z\mu^{2}+\left(1-z\right)^{2}}\,\,\,.
\end{eqnarray}
According to this, we can also write for the neutron and iso-vector
self-energies:
\begin{eqnarray}
\Sigma_{n}\left(B\right) & = & 3\Sigma_{1}\left(0\right)-2\Sigma_{1}\left(B\right)+2\Sigma_{2}\left(B\right)\,\,\,,\label{eq:ps_neutron_s12}\\
\Sigma_{v}\left(B\right) & = &
4\Sigma_{1}\left(B\right)-\Sigma_{2}\left(B\right)\,\,\,,
\end{eqnarray}
where the difference of Eq.(\ref{eq:ps_neutron_s12}) to Eq.(\ref{eq:ps_neutron_s3})
is of order $B^{2}$.
\begin{figure}
\caption{\label{fig:ps_self_energies}Energy shift of the proton (left) and
neutron (right) with pseudo-scalar coupling for $m_{\pi}=139$ MeV.
The upper row corresponds to the shift with $\Sigma\left(B\right)-\Sigma\left(0\right)$
and the lower row to $\Sigma\left(-B\right)-\Sigma\left(B\right)$.
The linear solid green lines corresponds to the linear approximation
Eq. (\ref{eq:linear_approximation}) while the curved solid red lines
to Eqs. (\ref{eq:ps_proton},\ref{eq:ps_neutron_s3}). The solid blue
line shows the result of Eq. (\ref{eq:ps_neutron_s12}). The dotted
lines correspond to the the imaginary parts.}
\includegraphics[scale=0.4]{PsProton1}~~~~~~~~\includegraphics[scale=0.4]{PsNeutron1}\\
\includegraphics[scale=0.4]{PsProton2}~~~~~~~~\includegraphics[scale=0.4]{PsNeutron2}
\end{figure}
In Fig. \ref{fig:ps_self_energies} we plot the nucleon energy shifts
together with the linear approximation with $M_{N}=939$ MeV and $m_{\pi}=139$
MeV. We see that the signs of the AMM are in agreement with phenomenology,
i.e. $\kappa_{p}>0$ and $\kappa_{n}<0$. To estimate $\mathcal{O}\left(B^{2}\right)$
effects, we use the difference of the red and blue lines in the lower
right neutron graph. The combination $\Sigma\left(-B\right)-\Sigma\left(B\right)$
does not have $\mathcal{O}\left(B^{2}\right)$ contributions whereas
the difference of the expressions Eq. (\ref{eq:ps_neutron_s3}) and
Eq. (\ref{eq:ps_neutron_s12}) is of $\mathcal{O}\left(B^{2}\right)$.
Plotting the same results for various pion masses shows that the $B^{2}$
effects are small for magnetic field strengths of $|B|<\frac{1}{5}M_{N}^{2}$
for pion masses larger than $m_{\pi}=100$ MeV and for $|B|<\frac{1}{2}M_{N}^{2}$
with pion masses around $m_{\pi}=600$ MeV. Within these region the
Eqs.(\ref{eq:linear_approximation},\ref{eq:ps_neutron_s3},\ref{eq:ps_neutron_s12})
give approximately the same results. We estimate therefore that the
self-energy formulas are applicable for magnetic field strengths of
$|B|<\frac{1}{5}M_{N}^{2}$ with pion masses between the physical
point $m_{\pi}=140$ MeV up to $m_{\pi}=600$ MeV. This is the applied
pion mass range in lattice QCD calculations.
In addition we have also for the nucleon the non-analytic $\mathcal{B}$
expressions, Eq. (\ref{eq:omega_contributions}), which would constrain
a definition of the magnetic moment by the linear energy shift. These
constrains read:
\begin{eqnarray}
\Sigma_{1,3}: & & |\mathcal{B}|<|\mu\left(\mu-2\right)|\,\,\,,\\
\Sigma_{1,3}: & & |\mathcal{B}|<|\mu\left(\mu+2\right)|\,\,\,,\\
\Sigma_{2}: & & |\mathcal{B}|<|\frac{1}{4}\mu^{2}-1|\,\,\,,
\end{eqnarray}
which can always be fulfilled for $0<\mu<2$.
Further, we obtain for the nucleon self-energy an imaginary part when
a stronger magnetic field is applied. The imaginary parts come only
from $\Sigma_{1}$ and $\Sigma_{3}$ and read:
\begin{eqnarray}
\mbox{Im}\Sigma_{i} & = &
M_{N}\frac{g^{2}\pi}{\left(4\pi\right)^{2}}\left(1-\alpha_{i}\right)2\lambda_{i}\,\,\,,
\end{eqnarray}
for $\mathcal{B}\leq-\mu\left(2+\mu\right)$ in $\Sigma_{1}$ and
$\mathcal{B}\ge\mu\left(2-\mu\right)$ in $\Sigma_{3}$.
The nucleon results of this section are obtained from the general
expression Eq. (\ref{eq:self_energy_ps}) for particles with Yukawa
couplings as in the Lagrangian Eq. (\ref{eq:ps_lagrangian}) and are
transcript-able for other stable particles.
\subsubsection{Nucleon-pion-resonance system}
For the second example we include a resonance ($R$) of mass $M_{R}$
with the coupling $\Gamma_{a}^{NR\pi}=ig_{a}$ and get with $s_{5}=1$
and $r=M_{N}/M_{R}$ the self-energies:
\begin{eqnarray}
\Sigma_{i} & = &
\frac{g^{2}}{\left(4\pi\right)^{2}}M_{R}\int_{0}^{1}dz\,\left(z+r\right)\left[L+\ln\left(z\mu^{2}
-z\left(1-z\right)+\left(1-z\right)r^{2}+\mathcal{B}_{i}-i\varepsilon\right)\right]
+\mathcal{O}\left(B^{2}\right)\,\,\,.\label{eq:resonance_self_energy}
\end{eqnarray}
From this, we also recover the results of the three point method:
\begin{equation}
\kappa_{1}=\frac{g^{2}}{\left(4\pi\right)^{2}}\int_{0}^{1}dz\,\frac{2z\left(1-z\right)
\left(-z-r\right)}{z\mu^{2}-z\left(1-z\right)+\left(1-z\right)r^{2}-i\varepsilon}\,\,\,\,\,\,\,\,\,,
\,\,\,\,\,\,\,\,\,\kappa_{2}=\frac{g^{2}}{\left(4\pi\right)^{2}}\int_{0}^{1}dz\,\frac{2\left(1-z\right)^{2}
\left(z+r\right)}{z\mu^{2}-z\left(1-z\right)+\left(1-z\right)r^{2}-i\varepsilon}\,\,\,.
\end{equation}
Since the resonance is unstable we have the following imaginary parts
for the self-energies:
\begin{eqnarray}
\mbox{Im}\Sigma_{i} & = &
-M_{R}\frac{g^{2}}{\left(4\pi\right)^{2}}\pi\left(r+\alpha_{i}\right)2\lambda_{i}\label{eq:imag_resonance}
\end{eqnarray}
for $\alpha_{i}>\lambda_{i}$. Explicitly, these parts are present
for magnetic fields of
\begin{eqnarray}
\Sigma_{1}: & & \mathcal{B}<-\sqrt{\left(1-r^{2}-\mu^{2}\right)^{2}-4\lambda_{0}^{2}}+1-r^{2}-\mu^{2}\,\,\,,\label{eq:imag_sigma1_resonance}\\
\Sigma_{2}: & & \mathcal{B}>-\frac{\lambda_{0}^{2}}{\mu^{2}}\,\,\,,\label{eq:imag_sigma2_resonance}\\
\Sigma_{3}: & &
\mathcal{B}>+\sqrt{\left(1-r^{2}+\mu^{2}\right)^{2}-4\lambda_{0}^{2}}-1+r^{2}+\mu^{2}\,\,\,.\label{eq:imag_sigma3_resonance}
\end{eqnarray}
In the case of $\Sigma_{2}$ and $\Sigma_{3}$ with a magnetic field
of $\mathcal{B}\geq r^{2}$ we have $\lambda_{i}>\alpha_{i}$, Eq.
(\ref{eq:imag_resonance}) has do be altered accordingly and an additional
cusp is present at $\mathcal{B}=r^{2}$.
In Fig. \ref{fig:sigma_resonance} we show all three possible self-energies
for the parameters $M_{N}=939$ MeV, $M_{R}=1232$ MeV and $m_{\pi}=139$
MeV together with the linear approximations. For the graph $\Sigma_{1}$
we see the linear behavior near $B=0$ and a cusp appearing according
to Eq. (\ref{eq:imag_sigma1_resonance}). This graph was investigated
in \cite{Ledwig(2010):nonAna}. In the case of $\Sigma_{2}$ we see
only one cusp at $\mathcal{B}=r^{2}\approx0.6$ since for the present
mass constellation Eq. (\ref{eq:imag_sigma2_resonance}) is fulfilled
for all $|\mathcal{B}|<1$ and an imaginary part is steadily present.
However, choosing $m_{\pi}$ closer to the mass-gap $M_{R}-M_{N}$
the second cusp appears on the $\mathcal{B}<0$ side. The occurrence
of such two cusps can be seen for $\Sigma_{3}$. The cusp for $\mathcal{B}<0$
is due to the imaginary part from Eq. (\ref{eq:imag_sigma3_resonance})
and the one for $\mathcal{B}>0$ due to $\mathcal{B}\geq r^{2}$.
As we approach with $m_{\pi}$ the mass-gap $M_{R}-M_{N}$, i.e. $\mu=1-r$,
all cusps corresponding to Eqs. (\ref{eq:imag_sigma1_resonance},\ref{eq:imag_sigma2_resonance},\ref{eq:imag_sigma3_resonance})
converge on $B=0$ where we have $1-r-\mu=0$. For a $m_{\pi}$ mass
larger than the mass gap the resonance will not be unstable anymore
and we get similar results as in the previous example.
The explicit conditions to expand the self-energies for small $\mathcal{B}$
due to the non-analytic contributions $\Omega_{i}$, Eq. (\ref{eq:omega_contributions}),
are:
\begin{eqnarray}
\Sigma_{1}: & & |\mathcal{B}|<2|1-r-\mu|\,\,\,,\\
\Sigma_{2}: & & |\mathcal{B}|<\frac{2r}{1-r}|1-r-\mu|\,\,\,,\\
\Sigma_{3}: & & |\mathcal{B}|<2r|1-r-\mu|\,\,\,,
\end{eqnarray}
for the parameters $r<1$, $\mu>0$ and $\mu<1+r$. The first condition
was found in \cite{Ledwig(2010):nonAna}. Which condition is the
most strict one on $\mathcal{B}$ depends on the charge of the external
particle and the actual values of the participating masses. For $\mu<r$
and $r>1/2$ it is the first one.
\begin{figure}
\caption{\label{fig:sigma_resonance}Self-energies of a resonance in an external
magnetic field $B$. The linear green lines correspond to the linear
approximation Eq. (\ref{eq:linear_approximation}) while the curved
red lines to Eqs. (\ref{eq:resonance_self_energy}). The solid lines
show the real parts and the dotted lines the imaginary parts.}
\includegraphics[scale=0.27]{PsR1}~~~~\includegraphics[scale=0.27]{PsR2}~~~~\includegraphics[scale=0.27]{PsR3}
\end{figure}
\section{Nucleon self-energy }
We apply now the BFT to a more involved situation, namely, the nucleon
self-energy in the SU(2) chiral perturbation theory of \cite{BChPTLagrangian}:
\begin{eqnarray}
\mathcal{L}_{N\pi} & = & \overline{N}(i\s
D-M_{N})N-\frac{g_{A}}{2f_{\pi}}\overline{N}\tau^{a}\left(\s
D^{ab}\pi^{b}\right)\gamma_{5}N+\frac{1}{2}(D_{\mu}^{ab}\pi^{b})(D_{ac}^{\mu}\pi^{c})
-\frac{1}{2}m_{\pi}^{2}\pi_{a}\pi^{a}\,\,\,.
\end{eqnarray}
The covariant derivatives are given in the appendix and the nucleon
axial-vector and the pion decay constants are $g_{A}=1.27$ and $f_{\pi}=92.4$
MeV. We consider the two graphs $\Sigma_{1}$ and $\Sigma_{2}$ in
Fig. \ref{fig:feynman_graphs} and obtain for the nucleon the following
unrenormalized self-energies in $d=4-2\varepsilon$ dimensions:
\begin{eqnarray}
\Sigma_{1}^{N}\left(B\right) & = &
i\left(\frac{g_{A}}{2f_{\pi}}\right)^{2}M_{N}
\int_{0}^{1}dz\Big[-M_{N}^{2}\left(1-z\right)^{3}J_{2}+\left(-6+3z\right)J_{1}
+\left(3-z\right)\varepsilon J_{1}+Bz^{2}\left(3-z\right)J_{2}\Big]\,\,\,,\\
\Sigma_{2}^{N}\left(B\right) & = &
i\left(\frac{g_{A}}{2f_{\pi}}\right)^{2}M_{N}
\int_{0}^{1}dz\Big[-M_{N}^{2}\left(1-z\right)^{3}J_{2}+\left(-6+3z\right)J_{1}+\left(3-z\right)\varepsilon J_{1}\nonumber\\
& &
\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,
+B\left(2\left(1-z\right)^{4}M_{N}^{2}J_{3}+\left(3-8z+6z^{2}-z^{3}\right)J_{2}-\left(3-4z+z^{2}\right)\varepsilon
J_{2}\right)\Big]\,\,\,.
\end{eqnarray}
The loop integrals $J_{i}=J_{i}\left(\mathcal{M}_{N}\right)$ with
$\mathcal{M}_{N}=zm_{\pi}^{2}+\left(1-z\right)^{2}M_{N}^{2}+z\left(1-z\right)B$
are listed in the appendix. The expressions $\Sigma_{1}$ and $\Sigma_{2}$
only differ by $B$ dependent terms and we write for the nucleon self-energies:
\begin{eqnarray}
\Sigma_{p}\left(B\right) & = & 2\Sigma_{1}\left(B\right)+\Sigma_{2}\left(B\right)\,\,\,,\\
\Sigma_{n}\left(B\right) & = &
3\Sigma_{1}\left(0\right)-2\Sigma_{1}\left(B\right)+2\Sigma_{2}\left(B\right)\,\,\,.
\end{eqnarray}
By integrating the Feynman parameter we get the $\tilde{MS}$ renormalized
proton self-energy
\begin{eqnarray}
\Sigma_{p}\left(\mathcal{B}\right) & = &
\frac{M_{N}C_{N}}{24\left(1-\mathcal{B}\right)^{4}}
\left[n_{1}+n_{2}\ln\mu+\frac{n_{3}}{\sqrt{4\mu^{2}-\left(\mathcal{B}+\mu^{2}\right)^{2}}}
\text{\ensuremath{\arccos}}\frac{\mathcal{B}+\mu^{2}}{2\mu}\right]\,\,\,,\label{eq:p_self_energy}\\
n_{1}\left(\mathcal{B},\mu\right) & = &
36-156\mathcal{B}+236\mathcal{B}^{2}-135\mathcal{B}^{3}
+12\mathcal{B}^{4}+7\mathcal{B}^{5}+\left(72-252\mathcal{B}\right)\mu^{2}
+\mathcal{O}\left(\mathcal{B}^{2}\right)\,\,\,,\\
n_{2}\left(\mathcal{B},\mu\right) & = &
-24\mathcal{B}^{3}+42\mathcal{B}^{4}-12\mathcal{B}^{5}
+96\mathcal{B}\mu^{2}+\left(-144+348\mathcal{B}\right)\mu^{4}+\left(36-72\mathcal{B}\right)\mu^{6}
+\mathcal{O}\left(\mathcal{B}^{2}\right)\,\,\,,\\
n_{3}\left(\mathcal{B},\mu\right) & = &
-24\mathcal{B}^{2}+42\mathcal{B}^{3}-12\mathcal{B}^{4}+120\mathcal{B}\mu^{2}
+\left(-36+72\mathcal{B}\right)\mu^{4}+\mathcal{O}\left(\mathcal{B}^{2}\right)\,\,\,,
\end{eqnarray}
with $C_{N}=\left(\frac{g_{A}M_{N}}{4\pi f_{\pi}}\right)^{2}$. We
give here only terms up to $\mathcal{O}\left(\mathcal{B}^{2}\right)$
or constant in $\mu$ and list in the appendix the full coefficients
$n_{i}\left(\mathcal{B},\mu\right)$. By setting $\mathcal{B}=0$
we recover the normal nucleon $\tilde{MS}$ renormalized B$\chi$PT
self-energy \cite{ChPTDelta:partI,ChPTmasses}:
\begin{eqnarray}
\Sigma_{p}\left(0\right) & = &
\frac{3}{2}M_{N}C_{N}\left[1+2\mu^{2}-2\mu^{3}\sqrt{1-\frac{\mu^{2}}{4}}
\text{\ensuremath{\arccos}}\frac{\mu}{2}-\mu^{4}\ln\mu\right]\,\,\,.\label{eq:BChPT_sigma}
\end{eqnarray}
The chiral expansion of the nucleon mass to order $p^{3}$ is
\begin{equation}
M_{N}\left(B\right)=\overset{\circ}{M_{N}}-4\overset{\circ}{c}_{1}m_{\pi}^{2}
+\Sigma_{N}^{(3)}\left(B\right)-\frac{\overset{\circ}{\kappa_{N}}}{2M_{N}}B\,\,\,,
\end{equation}
where $\overset{\circ}{M_{N}}$, $\overset{\circ}{c}_{1}$ and $\overset{\circ}{\kappa_{N}}$
are the low-energy constants for the nucleon mass and its AMM. The
$\mu^{0}$ and $\mu^{2}$ terms in Eq. (\ref{eq:BChPT_sigma}) break
the usual power counting scheme \cite{BChPTLagrangian} where e.g.
the EOMS renormalization scheme \cite{EOMS} is one way to deal with
this problem. The general behavior of the self-energy in this section
is similar to the nucleon self-energy of the last section for pseudo-scalar
$N-\pi$ couplings. Especially the non-analytic $\mathcal{B}$ term
is the same and we get again an imaginary part for $\mathcal{B}<-2\mu-\mu^{2}$.
\begin{figure}
\caption{\label{fig:proton_Sigma}Proton self-energy as function of a magnetic
field $B$ for $m_{\pi}=139$ MeV and $M_{N}=939$ MeV. The solid
red line is the result of Eq. (\ref{eq:p_self_energy}) and the solid
linear green line the linear approximation $\Sigma_{p}\left(B\right)=M_{N}-\frac{\kappa}{2M_{N}}B$
with $\kappa=1.73$. The dashed line is the imaginary part of Eq.
(\ref{eq:p_self_energy}).}
\includegraphics[scale=0.4]{PvProton}
\end{figure}
In Fig. \ref{fig:proton_Sigma} we show the proton self-energy as
function of the magnetic field $\mathcal{B}$ for the phenomenological
values. In case of the linear approximation we use the AMM obtained
from the three-point method, i.e. defined by $\kappa=F_{2}\left(0\right)$
via the matrix element
\begin{eqnarray}
\langle N(p^{\prime})|\overline{\Psi}(0)\gamma^{\mu}\Psi(0)|N(p)\rangle & = &
\bar{u}(p^{\prime})\left[\gamma^{\mu}F_{1}\left(q^{2}\right)+\frac{i\sigma^{\mu\nu}q_{\nu}}{2M_{N}}F_{2}\left(q^{2}\right)\right]u(p)\,\,\,,
\end{eqnarray}
with $q=p^{\prime}-p$ as the momentum transfer. The explicit results
for the nucleon graphs in the second row of Fig. \ref{fig:feynman_graphs}
are:
\begin{eqnarray}
\kappa_{1}^{N} & = &
\frac{1}{i}\left(\frac{g_{A}M_{N}}{2f_{\pi}}\right)^{2}\int_{0}^{1}dz2z
\left[\left(-6+12z-4z^{2}\right)J_{2}+\left(3-4z+z^{2}\right)\varepsilon J_{2}
-2M_{N}^{2}\left(1-z\right)^{4}J_{3}\right]\label{eq:pv_kappa1}\\
\kappa_{2}^{N} & = &
\frac{1}{i}\left(\frac{g_{A}M_{N}}{2f_{\pi}}\right)^{2}\int_{0}^{1}dz2
\left[\left(3-14z+15z^{2}-4z^{3}\right)J_{2}-\left(3-7z+5z^{2}-z^{3}\right)\varepsilon
J_{2}+2M_{N}^{2}\left(1-z\right)^{5}J_{3}\right]\label{eq:pv_kappa2}
\end{eqnarray}
with $J_{i}=J_{i}\left(\mathcal{M}\right)$ and $\mathcal{M}=zm_{\pi}^{2}+\left(1-z\right)^{2}M_{N}^{2}$
\cite{Ledwig(2011):covChptNDelta}. Extracting the AMM from the above
self-energy yields identical results where in the case of $\kappa_{1}^{N}$
this can be seen literally even before the Feynman parameter integration:
\begin{eqnarray}
\kappa_{1}^{N\left(BFT\right)} & = &
\frac{1}{i}\left(\frac{g_{A}M_{N}}{2f_{\pi}}\right)^{2}\int_{0}^{1}dz2z\left[\left(-6+9z-3z^{2}\right)J_{2}
+z\left(3-z\right)J_{2}+\left(3-4z+z^{2}\right)\varepsilon
J_{2}-2M_{N}^{2}\left(1-z\right)^{4}J_{3}\right].
\end{eqnarray}
The two $J_{2}$ integrals add up to the same expression in Eq. (\ref{eq:pv_kappa1}).
However, in the case of the three-point function method, e.g., the
$J_{2}$ contributions come purely from tensor integrals whereas in
the case of the BFT method it is a combination of tensor and scalar
loop-integrals. As we see for the infinite volume case this difference
is not important.
\section{Delta(1232) self-energy }
We consider now the $\Delta^{+}\left(1232\right)$-isobar and concentrate
on the graphs that give its decay width. We take the following Lagrangian
\cite{ChPTmasses}:
\begin{eqnarray}
\mathcal{L}_{\Delta\pi} & = &
\overline{\Delta}_{\mu}(i\gamma^{\mu\nu\alpha}D_{\alpha}-M_{\Delta}\gamma^{\mu\nu})\Delta_{\nu}
+i\frac{h_{A}}{2f_{\pi}M_{\Delta}}\overline{N}T^{a}\gamma^{\mu\nu\lambda}\left(D_{\mu}\Delta_{\nu}\right)
\left(D_{\lambda}^{ab}\pi^{b}\right)+\mbox{h.c.}\,\,\,,
\end{eqnarray}
and obtain the relevant self-energies as:
\begin{eqnarray}
\Sigma_{1}^{\Delta}\left(B\right) & = &
\frac{1}{i}M_{\Delta}\left(\frac{h_{A}}{2f_{\pi}}\right)^{2}\frac{1}{2}
\int_{0}^{1}dz\left[\left(z+r\right)J_{1}-B\left(z+r\right)z^{2}J_{2}\right]\,\,\,,\\
\Sigma_{2}^{\Delta}\left(B\right) & = &
\frac{1}{i}M_{\Delta}\left(\frac{h_{A}}{2f_{\pi}}\right)^{2}\frac{1}{2}
\int_{0}^{1}dz\left[\left(z+r\right)J_{1}-B\left(z+r\right)\left(1-z\right)^{2}J_{2}\right]\,\,\,,\\
\mathcal{M}_{\Delta} & = &
z\mu^{2}+\left(1-z\right)r^{2}-z\left(1-z\right)+\mathcal{B}_{i}\,\,\,,\label{eq:delta_self_energies}
\end{eqnarray}
with $\mathcal{B}_{1}=+z\left(1-z\right)\mathcal{B}$ and $\mathcal{B}_{2}=-\left(1-z\right)^{2}\mathcal{B}$
and all other definitions given in the appendix. Integrating these
expressions yield
\begin{eqnarray}
\Sigma_{1}^{\Delta}\left(B\right)\cdot\frac{144}{M_{\Delta}C_{\Delta}} & = &
\mathcal{A}_{1}\left(B\right)+48\left[2\left(r+\alpha_{1}\right)\lambda_{1}^{2}
+\mathcal{B}\left(3\left(r+\alpha_{1}\right)\alpha_{1}^{2}+\left(\alpha_{1}-r\right)
\lambda_{1}^{2}\right)\right]\,\,\Omega_{1}\left(B\right)\,\,\,,\\
\Sigma_{2}^{\Delta}\left(B\right)\cdot\frac{144}{M_{\Delta}C_{\Delta}} & = &
\mathcal{A}_{2}\left(B\right)+48\left[2\left(r+\alpha_{2}\right)\lambda_{2}^{2}
+\mathcal{B}\left(3r+3\alpha_{2}-6r\alpha_{2}-\left(3\alpha_{2}^{2}+1\right)\left(2
+r-\alpha_{2}\right)\lambda_{2}^{2}\right)\right]\,\,\Omega_{2}\left(B\right)\,\,,\\
\Omega_{i}\left(B\right) & = & \sqrt{\lambda_{i}^{2}}\left(\mbox{arctanh}
\frac{\beta_{i}}{\sqrt{\lambda_{i}^{2}}}+\mbox{arctanh}
\frac{\alpha_{i}}{\sqrt{\lambda_{i}^{2}}}\right)\,\,,\label{eq:Omega_Delta}
\end{eqnarray}
with $C_{\Delta}=\left(\frac{h_{A}M_{\Delta}}{8f_{\pi}\pi}\right)^{2}$
and the analytic parts $\mathcal{A}_{i}$ also listed in the appendix.
With these two expressions we can also write the self-energy of the
different iso-spin states as:
\begin{eqnarray}
\Sigma_{\Delta^{++}}\left(B\right) & = & -\Sigma_{1}^{\Delta}\left(0\right)
+\Sigma_{1}^{\Delta}\left(B\right)+\Sigma_{2}^{\Delta}\left(B\right)\,\,\,,\\
\Sigma_{\Delta^{+}}\left(B\right) & = &
\frac{1}{3}\Sigma_{1}^{\Delta}\left(B\right)
+\frac{2}{3}\Sigma_{2}^{\Delta}\left(B\right)\,\,\,,\\
\Sigma_{\Delta^{0}}\left(B\right) & = & \Sigma_{1}^{\Delta}\left(0\right)
-\frac{1}{3}\Sigma_{1}^{\Delta}\left(B\right)+\frac{1}{3}\Sigma_{2}^{\Delta}\left(B\right)\,\,\,,\\
\Sigma_{\Delta^{-}}\left(B\right) & = &
2\Sigma_{1}^{\Delta}\left(0\right)-\Sigma_{1}^{\Delta}\left(B\right)\,\,\,.
\end{eqnarray}
\begin{figure}
\caption{\label{fig:delta_sigma}The $\Delta^{+}\left(1232\right)$-isobar
self-energy loop results as function of a magnetic field $B$ for
$m_{\pi}=139$ MeV, $M_{N}=939$ MeV and $M_{\Delta}=1232$ MeV. The
left pictures shows the real part and the right one the imaginary
part. The curved red solid lines are from Eq. (\ref{eq:delta_self_energies})
while the linear green lines correspond to the linear approximation
$\Sigma_{\Delta}\left(B\right)=M_{\Delta}-\frac{\kappa_{\Delta}}{2M_{\Delta}}B$. }
\includegraphics[scale=0.4]{ReDelta}~~~~\includegraphics[scale=0.4]{ImDelta}
\end{figure}
In Fig. \ref{fig:delta_sigma} we show the $\Delta^{+}\left(1232\right)$
self-energy as function of the magnetic field $B$ for the phenomenological
parameters together with the linear approximation as obtained from
the three-point function method. The cusp comes from the $\Sigma_{1}$
while the second cusp of $\Sigma_{2}$ is not present for $m_{\pi}=139$
MeV, however it emerges on the $B<0$ side for larger pion masses
and both cusps fall on $B=0$ for $m_{\pi}=M_{\Delta}-M_{N}$. The
imaginary parts read
\begin{eqnarray}
\mbox{Im}\,\Sigma_{1}\left(B\right)\cdot\frac{3}{\pi M_{\Delta}C_{\Delta}} & =&
-2\left(1-\mathcal{B}\right)\left(r+\alpha_{1}\right)\lambda_{1}^{3}-\mathcal{B}
\lambda_{1}\left[3\alpha_{1}\left(\alpha_{1}^{2}+\lambda_{1}^{2}\right)
+r\left(3\alpha_{1}^{2}+\lambda_{1}^{2}\right)\right]\,\,\,,\\
\mbox{Im}\,\Sigma_{2}\left(B\right)\cdot\frac{3}{\pi M_{\Delta}C_{\Delta}} & =&
-2\left(1-\mathcal{B}\right)\left(r+\alpha_{2}\right)\lambda_{2}^{3}\nonumber\\
& &-\mathcal{B}\lambda_{2}\left[-6\alpha_{2}^{2}+3\alpha_{2}^{3}-2\lambda_{2}^{2}
+3\alpha_{2}\left(1+\lambda_{2}^{2}\right)+r\left(3-6\alpha_{2}+3\alpha_{2}^{2}+\lambda_{2}^{2}\right)\right]\,\,\,,
\end{eqnarray}
where the $\Delta^{+}\left(1232\right)$ decay width, experimentally
given by $\Gamma_{\Delta}\approx120$ MeV, is obtained from $\Gamma=-2\mbox{Im}\Sigma\left(B=0\right)$.
The slope of the imaginary part at $B=0$ in Fig. \ref{fig:delta_sigma}
is consistent with \cite{DeltaMagMom} and we obtain a vanishing
decay width at:
\begin{eqnarray}
\mathcal{B} & = &
-\frac{\left(1-2r+r^{2}-\mu^{2}\right)\left(1+2r+r^{2}-\mu^{2}\right)}{2\left(1+r-2r^{3}+3r^{4}
+2r\mu^{2}+3\mu^{4}-2r^{2}\left(1+3\mu^{2}\right)\right)}+\mathcal{O}\left(\mathcal{B}^{2}\right)\,\,\,.
\end{eqnarray}
The magnetic moment from the three-point function method is again
defined through the vector matrix element by $\kappa=F_{2}\left(0\right)$
and
\begin{eqnarray}
\langle\Delta(p^{\prime})|\overline{\Psi}(0)\gamma^{\mu}\Psi(0)|\Delta(p)\rangle
& = & -\bar{u}_{\alpha}(p^{\prime})\Big\{\left[F_{1}^{\Delta}\gamma^{\mu}
+\frac{i\sigma^{\mu\nu}q_{\nu}}{2M_{\Delta}}F_{2}^{\Delta}\right]g^{\alpha\beta}
+\left[F_{3}^{\Delta}\gamma^{\mu}+\frac{i\sigma^{\mu\nu}q_{\nu}}{2M_{\Delta}}F_{4}^{\Delta}\right]
\frac{q^{\alpha}q^{\beta}}{4M_{\Delta}^{2}}\Big\}u_{\beta}(p)\,\,\,,
\end{eqnarray}
with the results \cite{Ledwig(2011):covChptNDelta}:
\begin{eqnarray}
\kappa_{1}^{\Delta} & = &
\frac{1}{i}\left(\frac{h_{A}M_{\Delta}}{2f_{\pi}}\right)^{2}
\int_{0}^{1}dz\,\,2z\,\,\left[-\frac{1}{2}z+z^{2}-\frac{1}{2}r+zr\right]J_{2}\,\,\,,\\
\kappa_{2}^{\Delta} & = &
\frac{1}{i}\left(\frac{h_{A}M_{\Delta}}{2f_{\pi}}\right)^{2}\int_{0}^{1}dz\,2
\left(1-z\right)\left[\left(1+r\right)\left(1-z\right)-\left(1-z\right)^{2}\right]J_{2}\,\,\,,
\end{eqnarray}
with $J_{i}=J_{i}\left(\mathcal{M}_{\Delta}\right)$ and $\mathcal{M}_{\Delta}=z\mu^{2}+\left(1-z\right)r^{2}-z\left(1-z\right)$.
The results of the BFT method agree in this form literally. From $\sqrt{\lambda_{1}^{2}}$
in Eq. (\ref{eq:Omega_Delta}) we obtain the condition
\begin{equation}
\frac{eB}{2M_{\Delta}}\ll|M_{\Delta}-\left(M_{N}+m_{\pi}\right)|
\end{equation}
for expanding the $\Delta\left(1232\right)$-isobar self-energies
in small magnetic fields \cite{Ledwig(2010):nonAna}.
\section{Summary}
We investigated the self-energies of particles placed in a constant
electromagnetic (EM) field. Explicitly, we applied the EM background
field technique (BFT) once to the situation of stable and unstable
spin-1/2 particles coupling to (pseudo-) scalar fields and once to
the nucleon and $\Delta\left(1232\right)$-isobar baryons in the $SU\left(2\right)$
chiral perturbation theory (B$\chi$PT). We obtained the self-energies
of these particles as function of the external constant magnetic field
$B$ and calculated from these the anomalous magnetic moments (AMM)
by the linear energy shift. We summarize our findings as:
\begin{itemize}
\item Self-energies of Dirac particles coupling to (pseudo-) scalar fields:
We investigated all three types of Feynman graphs that can appear
in the one-loop BFT. The self-energies generally depend non-analytically
on $B$ where for stable particles the actual non-analytic points
are unproblematic for defining the AMM by the linear energy shift.
These points depend on the mass and charge constellations in the loop
and give three different conditions for resonances on when their self-energies
can be expanded for weak $B$. One of these conditions was reported
earlier. The self-energy formulas contain those $\mathcal{O}\left(B^{2}\right)$
terms that keep the non-analytical dependence intact, in contrast
to the one-photon approximation. However, they omit some $\mathcal{O}\left(B^{2}\right)$
contributions where we estimated for the nucleon-pion system with
pseudo-scalar couplings that these contributions are small for magnetic
fields of $|B|<\frac{1}{5}M_{N}^{2}$ with $M_{N}$ as the nucleon
mass.
\item Self-energies of the nucleon and $\Delta\left(1232\right)$-isobar
in B$\chi$PT: We derived the formulas for the nucleon and $\Delta\left(1232\right)$-isobar
self-energies depending on the pion mass and the magnetic field. We
recover the expressions as obtained from the three point function
method as well as the condition on when the $\Delta\left(1232\right)$-isobar
magnetic moment is well defined by the linear energy shift. Further,
we saw that the AMM expressions have different tensor- and scalar-loop
integral combinations depending on whether the AMM is derived by the
three point function method or from the self-energy. In the infinite
volume these combinations add up to the same AMM expressions.
\end{itemize}
\begin{acknowledgments}
The author is thankful to B. Tiburzi, V. Pascalutsa and M. Vanderhaeghen for valuable
discussions and critical comments.
\end{acknowledgments}
\begin{appendix}
\section{Notation}
We use the following notations: $\tilde{dl}=d^{n}l/\left(4\pi\right)^{n}$
with $n=4-2\epsilon$ dimensions,
\begin{eqnarray}
\mu=m/M_{ex}\,\,\, & \,\,\, r=M_{in}/M_{ex}\,\,\, &
\,\,\,\mathcal{B}=B/M_{ex}^{2}\,\,\,.
\end{eqnarray}
The covariant derivatives are:
\begin{eqnarray}
D_{\mu}^{ab}\pi^{b} & = & \delta^{ab}\partial_{\mu}\pi^{b}+ieQ_{\pi}^{ab}A_{\mu}\pi^{b}\,\,\,,\\
D_{\mu}N & = &
\partial_{\mu}N+ieQ_{N}A_{\mu}N+\frac{i}{4f_{\pi}^{2}}\epsilon^{abc}\tau^{a}\pi^{b}
\left(\mbox{\ensuremath{\partial}}_{\mu}\pi^{c}\right)\,\,\,,\\
D_{\mu}\Delta_{\nu} & = &
\partial_{\mu}\Delta_{\nu}+ieQ_{\Delta}A_{\mu}\Delta_{\nu}+\frac{i}{2f_{\pi}^{2}}\epsilon^{abc}
\mathcal{T}^{a}\pi^{b}\left(\mbox{\ensuremath{\partial}}_{\mu}\pi^{c}\right)\,\,\,,
\end{eqnarray}
with the operators $Q_{N}=\left(1+\tau^{3}\right)/2$ and $Q_{\pi}^{ab}=-i\varepsilon^{ab3}$.
The results for the loop integration in infinite volume and dimensional
regularization with $L=-\frac{1}{\varepsilon}+\gamma_{E}+\ln\frac{M_{sc}^{2}}{4\pi\Lambda^{2}}$
is:
\begin{eqnarray}
J_{1}\left(\mathcal{M}\right)=\frac{-i}{\left(4\pi\right)^{2}}m_{sc}^{2}
\tilde{\mathcal{M}}\left[L-1+\ln\tilde{\mathcal{M}}\right]\,\,\,,\,\,\,&
J_{2}\left(\mathcal{M}\right)=\frac{-i}{\left(4\pi\right)^{2}}
\left[L+\ln\tilde{\mathcal{M}}\right]\,\,\,,\,\,\,
&J_{3}\left(\mathcal{M}\right)=\frac{-i}{\left(4\pi\right)^{2}}
\frac{1}{2m_{sc}^{2}}\frac{1}{\tilde{\mathcal{M}}}\,\,\,.
\end{eqnarray}
The corresponding parts of the nucleon and $\Delta\left(1232\right)$
self-energies are:
\begin{eqnarray}
n_{1}\left(\mathcal{B},\mu\right) & = & 36-156\mathcal{B}+236\mathcal{B}^{2}
-135\mathcal{B}^{3}+12\mathcal{B}^{4}+7\mathcal{B}^{5} \nonumber\\
& & +\left(72-252\mathcal{B}+321\mathcal{B}^{2}-186\mathcal{B}^{3}
+45\mathcal{B}^{4}\right)\mu^{2}+\left(6\mathcal{B}^{2}-6\mathcal{B}^{3}\right)\mu^{4}\\
n_{2}\left(\mathcal{B},\mu\right) & = & -24\mathcal{B}^{3}+42\mathcal{B}^{4}
-12\mathcal{B}^{5}+\left(96\beta-180\beta^{2}+60\beta^{3}+12\beta^{4}
-12\beta^{5}\right)\mu^{2} \nonumber\\
& & +\left(-144+348\mathcal{B}-270\mathcal{B}^{2}+120\mathcal{B}^{3}
-18\mathcal{B}^{4}\right)\mu^{4}+\left(36-72\mathcal{B}
+12\mathcal{B}^{2}\right)\mu^{6}+6\mathcal{B}^{2}\mu^{8}\\
n_{3}\left(\mathcal{B},\mu\right) & = & -24\mathcal{B}^{2}+42\mathcal{B}^{3}
-12\mathcal{B}^{4}+\left(120\mathcal{B}-294\mathcal{B}^{2}+216\mathcal{B}^{3}
-60\mathcal{B}^{4}\right)\mu^{2} \nonumber\\
& &+\left(-36+72\mathcal{B}-24\mathcal{B}^{2}+6\mathcal{B}^{3}\right)\mu^{4}
-6\mathcal{B}^{2}\mu^{6}
\end{eqnarray}
\begin{eqnarray}
\mathcal{A}\left(B\right) & = & 27+40r-68\alpha-120r\alpha+42\alpha^{2}
+120r\alpha^{2}+12\alpha^{3}-54\lambda^{2}-168r\lambda^{2}-60\alpha\lambda^{2}\nonumber\\
& &
+\left(-18+48\alpha-36\alpha^{2}+36\lambda^{2}-24r\left(1-3\alpha+3\alpha^{2}
-3\lambda^{2}\right)\right)\ln\left(1-\mathcal{B}\right)\nonumber\\
& &
+\left(-24r\alpha^{3}-6\alpha^{4}+72r\alpha\lambda^{2}+36\alpha^{2}\lambda^{2}
+18\lambda^{4}\right)\ln\left(\alpha^{2}-\lambda^{2}\right)\nonumber\\
& &
+\left(-18+48\alpha-36\alpha^{2}+6\alpha^{4}+36\lambda^{2}-36\alpha^{2}\lambda^{2}
-18\lambda^{4}+24r(\alpha-1)\left((\alpha-1)^{2}-3\lambda^{2}\right)\right)
\ln\left(\beta^{2}-\lambda^{2}\right)\\
\mathcal{A}_{1}\left(B\right) & = & \Big[\mathcal{A}\left(B\right)+\mathcal{B}
\Big(-36-56r+56\alpha+96r\alpha-60\alpha^{2}-168r\alpha^{2}-48\alpha^{3}
+36\lambda^{2}+120r\lambda^{2}-48\alpha\lambda^{2}\nonumber\\
& &
+\sqrt{\lambda^{2}}\Omega\left(144r\alpha^{2}+144\alpha^{3}-48r\lambda^{2}
+48\alpha\lambda^{2}\right)\nonumber\\
& & +12\left(3-4\alpha+3\alpha^{2}-3\lambda^{2}+r\left(4-6\alpha+6\alpha^{2}
-6\lambda^{2}\right)\right)\ln\left(1-\mathcal{B}\right)\nonumber\\
& & +\left(48r\alpha^{3}+24\alpha^{4}+72\alpha^{2}\lambda^{2}\right)
\ln\left(\alpha^{2}-\lambda^{2}\right)-12\big(4\alpha+2\alpha^{4}
+3\left(\lambda^{2}-1\right)+\alpha^{2} \left(6\lambda^{2}-3\right) \nonumber\\
& &+r\left(6\left(\alpha-\alpha^{2}+\lambda^{2}\right)
-4+4\alpha^{3}\right)\big)\ln\left(\beta^{2}-\lambda^{2}\right)\Big)
\Big]_{\alpha,\beta,\lambda\to\alpha_{1},\beta_{1},\lambda_{1}}\\
\mathcal{A}_{2}\left(B\right) & = & \Big[\mathcal{A}\left(B\right)+\mathcal{B}
\Big(-40-128r+32\alpha+240r\alpha+36\alpha^{2}-168r\alpha^{2}-48\alpha^{3}
+132\lambda^{2}+120r\lambda^{2}-48\alpha\lambda^{2}\nonumber\\
& & +12\left(2-4\alpha+3\alpha^{2}-3\lambda^{2}+r\left(4-6\alpha+6\alpha^{2}
-6\lambda^{2}\right)\right)\ln\left(1-\mathcal{B}\right)\nonumber\\
& & +\left(72r\alpha+36\alpha^{2}-72r\alpha^{2}-48\alpha^{3}+48r\alpha^{3}
+24\alpha^{4}+36\lambda^{2}-72r\lambda^{2}-144\alpha\lambda^{2}+72\alpha^{2}
\lambda^{2}\right)\ln\left(\alpha^{2}-\lambda^{2}\right)\nonumber\\
& & -24(-1+\alpha)^{2}\left(-1+2r(-1+\alpha)+\alpha^{2}+3\lambda^{2}\right)
\ln\left(\beta^{2}-\lambda^{2}\right)\Big)
\Big]_{\alpha,\beta,\lambda\to\alpha_{2},\beta_{2},\lambda_{2}}
\end{eqnarray}
\end{appendix}
|
\section{\label{sec:Introduction}Introduction}
The investigation of rare $B$ decays induced by the
flavor-changing neutral current (FCNC) transitions $b\to s$ and
$b\to d$ represents an important test of the standard model (SM)
and its extensions (see \cite{Antonelli:2009} for a review).
Among the rare decays, the process $b\to s \ell^+\ell^-$, where
the virtual photon is converted to the lepton pair, is of
considerable interest. This decay proceeds through a loop
(penguin) diagram, to which high-mass particles introduced in
various extensions to the SM may contribute with sizable
amplitudes. In this decay the angular distributions and lepton
polarizations can probe the chiral structure of the matrix element
\cite{Grossman:2000, Melikhov:1998, Ali:2000, Kruger:2005,
Bobeth:2008, Altmannshofer:2009, Egede:2010} and thereby effects
of the new physics (NP) beyond the SM.
In order to unambiguously measure effects of NP in the observed
process ${\bar B}_d^0 \to {\bar K}^{*0} \, (\to K^{-}\, \pi^+) \,
\ell^+\, \ell^- $ ($l=e, \, \mu$), mediated by $b\to s
\ell^+\ell^-$ decay, one needs to calculate the SM predictions
with a high accuracy. The amplitude in the SM consists of the
short-distance (SD) and long-distance (LD) contributions. The
former are expressed in terms of the Wilson coefficients $C_i$
calculated in perturbative QCD up to a certain order in
$\alpha_s(\mu)$; they carry information on processes at energy
scales $ \sim m_W, \ m_t$. The LD effects describing the
hadronization process are expressed in terms of matrix elements of
several $b \to s$ operators between the initial $B$ and the $K^*$
final state. These hadronic matrix elements are parameterized in
terms of form factors \cite{Ali:2000} that are calculated in
various approaches (see, e.g., \cite{Ball:2005, Defazio:2006}).
The additional LD effects, originating from intermediate vector
resonances $\rho (770)$, $\omega(782)$, $\phi(1020)$,
$J/\psi(1S)$, $\psi(2S)$,$\ldots$, in general, may complicate
theoretical interpretation and make it more model dependent. The
vector resonances modify the amplitude and thus may induce, for
example, the right-handed currents which are absent in the SM.
Present experimental studies~\cite{Babar:2009, Belle:2009,
CDF:2011} of the $ B \to K^* \, \ell^+\, \ell^- $ decay aim at
the search of effects of the NP in the whole region of dilepton
invariant mass $m_{ee} \equiv \sqrt{q^2}$ GeV (here $q= q_+ +
q_-$). In these analyses certain cuts are applied in order to
exclude a rather big charmonia contribution.
Recently also the region of small dilepton invariant mass, $m_{ee}
\lesssim 1$ GeV, attracted attention \cite{Grossman:2000}, as
having a potential for searching signatures of the NP. The authors
of \cite{Lefrancois:2009} analyzed the azimuthal angular
distribution in the decay $\bar{B}^0 \to \bar{K}^{*0} \ell^+
\ell^-$ in this region, to test the possibility to measure this
distribution at the LHCb. They have shown the feasibility of
measurements with small systematic uncertainties. In
Ref.~\cite{Korchin:2010} the influence of the low-lying resonances
$\rho (770)$, $\omega(782)$ and $\phi(1020)$ on differential
branching ratio, polarization fraction of the $K^{0*}$ and
transverse asymmetry $A_{\rm T}^{(2)}$ has been studied.
In the present paper we extend calculations of \cite{Korchin:2010}
to the whole region of dilepton invariant mass up to $m_{ee}^{max}
= m_{B} -m_{K^{*}} =4.39 $ GeV. The effective SM Hamiltonian with
the Wilson coefficients in the next-to-next-to-leading order
(NNLO) approximation is applied. The LD effects mediated by the
resonances, {\it i.e.} $\bar{B}^0 \to \bar{K}^{*0} V \to
\bar{K}^{*0} e^+ e^-$ with $V=\rho(770), \ \omega(782), \
\phi(1020), \ J/\psi, \ \psi(2S), \ldots $, are included
explicitly in terms of the helicity amplitudes of the decays
$\bar{B}^0 \to \bar{K}^{*0} V $. The information on the latter is
taken from experiments if available; otherwise it is taken from
theoretical predictions.
The fully differential angular distribution over the three angles
and dilepton invariant mass for the four-body decay ${\bar B}_d^0
\to K^{-}\, \pi^+ \, e^+\,e^- $ is analyzed. We define a
convenient set of asymmetries which allows one to extract these
asymmetries from the angular distribution once sufficient
statistics is accumulated. These asymmetries may have sensitivity
to various effects of the NP, although in order to see signatures
of these effects, the resonance contribution should be accurately
evaluated.
One of the ingredients in calculation of the resonance
contribution is the transition vertex $V \, \gamma $. This vertex
is conventionally treated in the vector-meson-dominance (VMD)
model. In the present paper we apply two versions of the VMD model
(called subsequently VMD1 and VMD2) which result in rather
different $V \, \gamma $ vertices, in particular, far from the
vector-meson mass shell $q^2 = m_V^2$. Specifically, due to
explicit gauge-invariant construction of the VMD2 Lagrangian the
$V \, \gamma $ transition is suppressed in the region $q^2 \ll
m_V^2$ (for $V = J/\psi, \, \psi(2S), \ldots$). This observation
may be important for estimation of resonance contribution to those
asymmetries, which are small in the SM.
One should mention that the $c \bar{c}$ vector resonances $J/\psi,
\, \psi(2S), \ldots$ have been included earlier in
Refs.~\cite{Ligeti:1996} in the analysis of the $\bar{B}^0 \to
\bar{K}^{*0} \mu^+ \mu^-$ decay. This method of including
resonances has been originally suggested in \cite{Deshpande:1989}.
In order to see sensitivity of observables to the method of
including the $c \bar{c}$ resonances, we perform calculations
using the two methods, and compare the results.
Results of the present calculations are compared with the recent
data from Belle (KEKB) and CDF (Tevatron) experiments for the
differential branching, asymmetry $A_{\rm T}^{(2)}$, longitudinal
polarization fraction of $K^*$ and forward-backward asymmetry.
The paper is organized as follows. In Sec.~\ref{subsec:angle
distribution} the fully differential angular distribution is
discussed. In Section~\ref{subsec:asymmetries} one-dimensional
distributions and definition of asymmetries are defined.
Section~\ref{subsec:transversity} contains expressions for the
transversity amplitudes in framework of the SM. The models of
vector-meson dominance and contributions of vector resonances to
the amplitudes are discussed in Sec.~\ref{subsec:resonances}.
Results for the dependence of observables on the invariant mass
squared are presented in Sec.~\ref{subsec:observables}. In
Sec.~\ref{subsec:two_approaches} we compare two approaches to
inclusion of vector resonances in the amplitudes of the ${\bar
B}_d^0 \to K^{-}\, \pi^+ \, e^+\,e^- $ decay. In
Sec.~\ref{sec:conclusions} we draw conclusions. In
Appendix~\ref{sec:Appendix} some details of the calculation of the
matrix element and the model of the $B \to K^*$ transition form
factors are described. Appendix~\ref{subsec:vector mesons} deals
with calculation of the $\bar{B}^0 \to \bar{K}^{*0} V$ amplitudes
for the off-mass-shell vector meson $V$.
\section{\label{sec:formalism} Angular distributions and
amplitudes for the ${\bar B}_d^0\to {\bar K}^{*0}\,e^+\,e^-$ decay
}
\subsection{ \label{subsec:angle distribution}
Differential decay rate}
The decay ${\bar B}_d^0\to {\bar K}^{*0}\,e^+\,e^-$, with ${\bar K}^{*0} \to
K^- \pi^+$ on the mass shell~\footnote{This means the narrow-width
approximation for the ${\bar K}^{*0}$ propagator: \ $(k^2 - m_{K^*}^2 +
im_{K^*} \Gamma_{K^*})^{-1} \approx -i \pi \delta(k^2 - m_{K^*}^2) $.}, is
completely described by four independent kinematic variables: the
electron-positron pair invariant-mass squared, $q^2$, and the three angles
$\theta_l$, $\theta_K$, $\phi$. In the helicity frame (Fig.~\ref{fig1}), the
angle $\theta_l\,(\theta_K)$ is defined as the angle between the directions of
motion of $e^+\,(K^-)$ in the $\gamma^*\,({\bar K}^{*0})$ rest frame and the
$\gamma^*\,({\bar K}^{*0})$ in the ${\bar B}_d^0$ rest frame. The azimuthal
angle $\phi$ is defined as the angle between the decay planes of $\gamma^*\to
e^+\,e^-$ and ${\bar K}^{*0} \to K^- \pi^+$ in the ${\bar B}_d^0$ rest frame.
The fully differential angular distribution in these coordinates is given by
\begin{eqnarray}\label{eq:001}
{\cal W}(\hat{q}^2, \theta_l, \theta_K,
\phi)&\equiv&\frac{d^4\,\Gamma}{d\hat{q}^2d\cos\theta_l\,
d\cos\theta_K d\phi}/\frac{d \Gamma}{d\hat{q}^2}\nonumber \\ &=&
\frac{9}{64\,\pi}\sum_{k=1}^{9}\alpha_{k}(q^2)g_{k}(\theta_l,
\theta_K,\phi)\,,
\end{eqnarray}
where the angular terms $g_k$ are defined as
\begin{figure}
\centerline{\includegraphics[width=.45\textwidth]{fig1.eps}}
\caption{Definition of helicity angles $\theta_l$, $\theta_K$, and
$\phi$, for the decay ${\bar B}_d^0\to {\bar K}^{*0}\,e^+\,e^-$.}
\label{fig1}
\end{figure}
\begin{widetext}
\[g_1=4\sin^2\theta_l\cos^2\theta_K\,,\:
g_2=\left(1+\cos^2\theta_l \right)\sin^2\theta_K\,,\:
g_3=\sin^2\theta_l\sin^2\theta_K\cos2\phi\,,\]
\[g_4=-2\sin^2\theta_l\sin^2\theta_K\sin2\,\phi\,,\:g_5=-\sqrt{2}\sin2\,\theta_l\sin2\,\theta_K\cos\phi\,,
\: g_6=-\sqrt{2}\sin2\,\theta_l\sin2\,\theta_K\sin\phi\,,\]
\[g_7=4\cos\theta_l\sin^2\theta_K\,,\:g_8=-2\sqrt{2}\sin\theta_l\sin2\,\theta_K\cos\phi\,,
\:g_9=-2\sqrt{2}\sin\theta_l\sin2\,\theta_K\sin\phi\,,
\]
and the amplitude terms $\alpha_k$ as
\[\alpha_1=|a_{0}|^2=f_L\,, \: \alpha_2=|a_{\|}|^2+|a_{\perp }|^2=f_{\|}+f_{\perp }\,,\: \alpha_3=|a_{\perp }|^2-|a_{\|}|^2=f_{\perp }-f_{\|}\,,
\: \alpha_4={\rm Im}\left(a_{\|}a_{\perp }^*\right)\,, \:
\alpha_5={\rm Re}\left(a_{0}a_{\|}^*\right)\,,\]
\[\alpha_6={\rm
Im}\left(a_{0}a_{\perp }^*\right)\,,\:\alpha_7={\rm
Re}\left(a_{\|L}a_{\perp L}^*-a_{\|R}a_{\perp R}^*\right)\,,\:
\alpha_8={\rm Re}\left(a_{0L}a_{\perp L}^*-a_{0R}a_{\perp
R}^*\right)\, ,\:\alpha_9={\rm Im}\left(a_{0L}a_{\|
L}^*-a_{0R}a_{\| R}^*\right)\,, \]
\end{widetext} where $\hat{q}^2\equiv q^2/m_B^2$, $m_B$ is the mass of the $B^0_d$ meson, and
\begin{equation}
\frac{d\,\Gamma}{d\hat{q}^2}=m_B\,N^2\hat{q}^2\sqrt{\hat{\lambda}}\left(|A_0|^2+|A_{\|}|^2+|A_{\perp}|^2
\right) \,. \label{eq:002}
\end{equation}
\begin{equation}
a_i a^*_j \equiv a_{i L}(q^2) a^*_{jL}(q^2)+ a_{iR}(q^2)
a^*_{jR}(q^2) \, , \label{eq:003}
\end{equation}
\begin{equation}
a_{i L(R)}\equiv \frac{A_{i L(R)}}{\sqrt{\sum_j |A_j|^2}}
\label{eq:004} .
\end{equation}
Here $i,j = (0, \|, \perp )$, we have neglected the electron mass
$m_e$ and $A_{0L(R)}$, $A_{\|L(R)}$ and $A_{\perp L(R)}$ are the
complex decay amplitudes of the three helicity states in the
transversity basis, $f_L$, $f_{\|}$ and $f_\perp$ are polarization
parameters of the $K^*$ meson, $f_L+f_\|+f_\perp =1$,
$\hat{\lambda}\equiv\lambda(1,\hat{q}^2,
\hat{m}_{K^*}^2)=(1-\hat{q}^2)^2-2(1+\hat{q}^2)\hat{m}_{K^*}^2+\hat{m}_{K^*}^4$,\
$\hat{m}_{K^*}\equiv m_{K^*}/m_B$, where $m_{K^*}$ is the mass of
the $K^{*0}$ meson, and
\[N=|V_{tb}V_{ts}^*|\frac{G_F m_B^2 \alpha_{\rm em}}{32 \,\pi^2 \sqrt{3\,
\pi}}\,.\]
Here, $V_{ij}$ are the Cabibbo-Kobayashi-Maskawa (CKM) matrix
elements \cite{CKM}, $G_F$ is the Fermi coupling constant,
$\alpha_{\rm em}$ is the electromagnetic fine-structure constant.
With its rich multidimensional structure, the differential decay
rate in Eq.~(\ref{eq:001}) has sensitivity to various effects
modifying the SM, such as $CP$ violation beyond the CKM mechanism
and/or right-handed currents. Given sufficient data, all
$\alpha_k$ can, in principle, be completely measured from the full
angular distribution in all three angles $\theta_l$, $\theta_K$,
and $\phi$.
\subsection{ \label{subsec:asymmetries}
One-dimensional angular distributions and asymmetries}
The one-dimensional angular distributions in $\cos\theta_l$ and
$\cos\theta_K$ simply are
\begin{eqnarray}
{\cal W}_{\theta_l}(\hat{q}^2,\cos\theta_l) &\equiv&
\frac{d^2\,\Gamma}{d\hat{q}^2d\cos\theta_l}/\frac{d\,\Gamma}{d\hat{q}^2}=\frac{3}{4}f_L
(1-\cos^2\theta_l) \nonumber \\
&&+\frac{3}{8} (1-f_L)(1+\cos^2\theta_l)\nonumber \\
&&+\frac{d{\bar A}_{\rm FB}^{\rm (l)}}{d\hat{q}^2}\cos\theta_l
\label{eq:005}
\end{eqnarray}
and
\begin{eqnarray}
{\cal W}_{\theta_K}(\hat{q}^2,\cos\theta_K)&\equiv&
\frac{d^2\,\Gamma}{d\hat{q}^2d\cos\theta_K}/\frac{d\,\Gamma}{d\hat{q}^2}=\frac{3}{2}f_L
\cos^2\theta_K \nonumber \\
&&+\frac{3}{4} (1-f_L)(1-\cos^2\theta_K)
\label{eq:006},
\end{eqnarray}
where $d{\bar A}_{\rm FB}^{\rm (l)}/d\hat{q}^2$ is the normalized
lepton forward-backward asymmetry
\begin{eqnarray}
\frac{d{\bar A}_{\rm FB}^{\rm
(l)}}{d\hat{q}^2}&\equiv&\int\limits_{-1}^{1}{\rm
sgn}(\cos\theta_l){\cal W}_{\theta_l}(\hat{q}^2,\cos\theta_l)\:d\cos\theta_l\nonumber \\
&=&\frac{3}{2}{\rm Re}(a_{\parallel\,L}\,
a_{\perp\,L}^*-a_{\parallel\,R}\, a_{\perp\,R}^*) \equiv A_7 \, .
\label{eq:007}
\end{eqnarray}
While ${\cal W}_{\theta_K}(\hat{q}^2,\cos\theta_K)$ depends only
on $f_L$, ${\cal W}_{\theta_l}(\hat{q}^2,\cos\theta_l)$ depends
both on $f_L$ and $d{\bar A}_{\rm FB}^{\rm (l)}/d\hat{q}^2$. The
measurement of the lepton forward-backward asymmetry $d{\bar
A}_{\rm FB}^{\rm (l)}/d\hat{q}^2$ alone is not enough to fully
reconstruct the $\cos\theta_l$ distribution. One can then think
about other asymmetries. For any fixed $z$ in the interval
$[-1,1]$, one can define an asymmetry
\begin{equation} \label{eq:008}
{\cal A}_z\equiv \frac{{\cal E}_z-{\cal P}_z}{{\cal E}_z+{\cal
P}_z} ,
\end{equation}
where
\begin{equation}
{\cal E}_z\equiv\int\limits_{-z}^{z}{\cal
W}_{\theta}(\hat{q}^2,\cos\theta)\:d\cos\theta , \label{eq:009}
\end{equation}
and
\begin{equation}
{\cal P}_z\equiv\left(\int\limits_{-1}^{-z}d\cos\theta
+\int\limits_{z}^{1}d\cos\theta \right) {\cal
W}_{\theta}(\hat{q}^2,\cos\theta)\label{eq:010}.
\end{equation}
Measuring the asymmetry ${\cal A}_{z_l}$ for $z=z_l\approx0.596$
($\cos^3\theta_l+3\cos\theta_l-2=0$), we can find the fraction of
the longitudinal polarization of the $K^*$ meson
\begin{equation}
{\cal A}_{z_l}=3(2z_l-1)f_L\approx0.576 f_L \label{eq:011},
\end{equation}
Similarly, measuring the asymmetry ${\cal A}_{z_K}$ for
$z=z_K=2\cos\frac{4\pi}{9}\approx0.347$
($\cos^3\theta_K-3\cos\theta_K+1=0$), we can find the fraction of
the longitudinal polarization of the $K^*$ meson
\begin{equation}
{\cal A}_{z_K}=3(2z_K-1)f_L\approx -0.916 f_L \label{eq:012}.
\end{equation}
Finally, the one-dimensional angular distribution in the angle
$\phi$ between the lepton and meson planes takes the form
\begin{eqnarray} \label{eq:013}
{\cal W}_{\phi}(\hat{q}^2, \phi)
&\equiv&\frac{d^2\,\Gamma}{d\hat{q}^2d\phi}/\frac{d\,\Gamma}{d\hat{q}^2}=\frac{1}{2\pi}
\Bigl(1+\frac{1}{2}\bigl(1\nonumber \\&&-f_L\bigr)A^{(2)}_{\rm
T}\cos2\phi -A_{\rm Im}\sin2\phi\Bigr) ,
\end{eqnarray}
\begin{equation} \label{eq:014}
A^{(2)}_{\rm T}\equiv \frac{ f_\perp-f_\|}{f_\perp+f_\|}\,, \quad
A_{\rm Im}\equiv {\rm Im}(a_\|a^*_\perp) ,
\end{equation}
where the asymmetry $A^{(2)}_{\rm T}(q^2)$ is sensitive to new
physics from right-handed currents, and the amplitude $A_{\rm
Im}(q^2)$ is sensitive to complex phases in the hadronic matrix
elements. Sometimes $A^{(2)}_{\rm T}(q^2)$ is called transverse
asymmetry~\cite{Kruger:2005}.
Measurement of the angular distribution in the azimuthal angle $\phi$ allows
one to determine the quantities $(1-f_L)A^{(2)}_{\rm T}$ and $A_{\rm Im}$
\begin{eqnarray}
A_3&\equiv&
\Bigl(\int\limits_{0}^{\pi/4}d\,\phi-\int\limits_{\pi/4}^{3\pi/4}d\,\phi
+\int\limits_{3\pi/4}^{5\pi/4}d\,\phi \nonumber \\
&&-\int\limits_{5\pi/4}^{7\pi/4}d\,\phi
+\int\limits_{7\pi/4}^{2\pi}d\,\phi\Bigr) {\cal
W}_{\phi}(\hat{q}^2, \phi)\nonumber \\
&=& \frac{1}{\pi}(1-f_L)A^{(2)}_{\rm T}=\frac{f_\perp-f_\|}{\pi} ,
\label{eq:015}
\end{eqnarray}
\begin{eqnarray}
A_4&\equiv&\Bigl(\int\limits_{0}^{\pi/2}d\,\phi-\int\limits_{\pi/2}^{\pi}d\,\phi
+\int\limits_{\pi}^{3\pi/2}d\,\phi
\nonumber \\&&-\int\limits_{3\pi/2}^{2\pi}d\,\phi\Bigr){\cal
W}_{\phi}(\hat{q}^2, \phi) = -\frac{2}{\pi}A_{\rm Im} .
\label{eq:016}
\end{eqnarray}
Measurement of the azimuthal angle dependence of the
forward-backward asymmetry for positrons and $K^-$ mesons
\begin{eqnarray}
&&\frac{d^2{\bar A}_{\rm FB}^{\rm
(Kl)}}{d\hat{q}^2d\phi}\equiv\int\limits_{-1}^{1}{\rm
sgn}(\cos\theta_l)\,d\cos\theta_l \nonumber
\\&\times&\int\limits_{-1}^{1}{\rm
sgn}(\cos\theta_K)\,d\cos\theta_K \, {\cal W}(\hat{q}^2, \theta_l,
\theta_K, \phi)\nonumber \\
&=&-\frac{\sqrt{2}}{4\pi}\Bigl({\rm
Re}(a_{0}\, a_{\|}^*)\cos\phi +{\rm Im}(a_{0}\,
a_{\perp}^*)\sin\phi\Bigr) , \label{eq:017}
\end{eqnarray}
will allow one to find ${\rm Re}(a_{0}\, a_{\|}^*)$ and ${\rm Im}(a_{0}\,
a_{\perp}^*)$
\begin{eqnarray}
A_5&\equiv&\Bigl(\int\limits_{0}^{\pi/2}d\,\phi-\int\limits_{\pi/2}^{3\pi/2}d\,\phi+\int\limits_{3\pi/2}^{2\pi}d\,\phi
\Bigr)\frac{d^2{\bar A}_{\rm FB}^{\rm
(Kl)}}{d\hat{q}^2d\phi}\nonumber \\ &=& -\frac{\sqrt{2}}{\pi}{\rm
Re}(a_{0}\, a_{\|}^*) , \label{eq:018}
\end{eqnarray}
\begin{equation}
A_6\equiv\Bigl(\int\limits_{0}^{\pi}d\,\phi-\int\limits_{\pi}^{2\pi}d\,\phi
\Bigr)\frac{d^2{\bar A}_{\rm FB}^{\rm
(Kl)}}{d\hat{q}^2d\phi}= -\frac{\sqrt{2}}{\pi}{\rm Im}(a_{0}\,
a_{\perp}^*) . \label{eq:019}
\end{equation}
Measurement of the azimuthal angle dependence of the
forward-backward asymmetry for $K^-$ mesons
\begin{eqnarray}
&&\frac{d^2{\bar A}_{\rm FB}^{\rm (K)}}{d\hat{q}^2d\phi}\nonumber
\\&\equiv&\int\limits_{-1}^{1}d\cos\theta_l
\int\limits_{-1}^{1}{\rm sgn}(\cos\theta_K)\,d\cos\theta_K \,{\cal
W}(\hat{q}^2, \theta_l,
\theta_K, \phi) \nonumber \\
&=&-\frac{3\sqrt{2}}{16}\Bigl({\rm Re}(a_{0L}\, a_{\perp
L}^*-a_{0R}\, a_{\perp R}^*)\cos\phi \nonumber \\ &&+{\rm Im}(a_{0
L}\, a_{\| L}^*-a_{0 R}\, a_{\| R}^*)\sin\phi\Bigr)
\label{eq:020},
\end{eqnarray}
will allow one to find ${\rm Re}(a_{0L}\, a_{\perp L}^*-a_{0R}\,
a_{\perp R}^*)$ and ${\rm Im}(a_{0 L}\, a_{\| L}^*-a_{0 R}\, a_{\|
R}^*)$
\begin{eqnarray}
A_8&\equiv&\Bigl(\int\limits_{0}^{\pi/2}d\,\phi-\int\limits_{\pi/2}^{3\pi/2}d\,\phi+\int\limits_{3\pi/2}^{2\pi}d\,\phi
\Bigr)\frac{d^2{\bar A}_{\rm FB}^{\rm
(K)}}{d\hat{q}^2d\phi}\nonumber \\&=& -\frac{3\sqrt{2}}{4}{\rm
Re}(a_{0 L}\, a_{\perp L}^*-a_{0 R}\, a_{\perp R}^*)
,\label{eq:021}
\end{eqnarray}
\begin{eqnarray}
A_9&\equiv&\Bigl(\int\limits_{0}^{\pi}d\,\phi-\int\limits_{\pi}^{2\pi}d\,\phi
\Bigr)\frac{d^2{\bar A}_{\rm FB}^{\rm
(K)}}{d\hat{q}^2d\phi}\nonumber \\&=& -\frac{3\sqrt{2}}{4}{\rm
Im}(a_{0L}\, a_{\| L}^*-a_{0R}\, a_{\| R}^*) . \label{eq:022}
\end{eqnarray}
\subsection{ \label{subsec:transversity}
Transversity amplitudes}
The nonresonant amplitudes follow from the matrix element of the
${\bar B}_d^0 (p)\to {\bar
K}^{*0}(k,\epsilon)\,e^+(q_+)\,e^-(q_-)$ process in
Eq.~(\ref{eq:0A1}),
\begin{eqnarray}
A_{0L,R}^{\rm
NR}&=&\frac{C_0(q^2)}{2\,\hat{m}_{K^*}\sqrt{\hat{q}^2}}\,\Biggl(C_{9V}^{\rm
eff} \mp C_{10A} \nonumber \\ &&+2\hat{m}_b\left(C_{7\gamma}^{\rm
eff}-C_{7\gamma}^{\prime\,\rm
eff}\right)\kappa_0(q^2)\Biggr)\,,\label{eq:023}
\end{eqnarray}
\begin{eqnarray}
A_{\|L,R}^{\rm NR}&=&-\sqrt{2}\,C_{\|}(q^2)\,\Biggl(C_{9V}^{\rm
eff} \mp C_{10A}\nonumber
\\ &&+2\frac{\hat{m}_b}{\hat{q}^2}\left(C_{7\gamma}^{\rm
eff}-C_{7\gamma}^{\prime\,\rm
eff}\right)\kappa_{\|}(q^2)\Biggr)\,,\label{eq:024}
\end{eqnarray}
\begin{eqnarray}
A_{\perp L,R}^{\rm
NR}&=&\sqrt{2\hat{\lambda}}\,C_{\perp}(q^2)\,\Biggl(C_{9V}^{\rm
eff} \mp C_{10A}\nonumber
\\ &&+2\frac{\hat{m}_b}{\hat{q}^2}\left(C_{7\gamma}^{\rm
eff}+C_{7\gamma}^{\prime\,\rm
eff}\right)\kappa_{\perp}(q^2)\Biggr)\,,\label{eq:025}
\end{eqnarray}
where the form factors enter as
\begin{eqnarray}
C_0(q^2)&=&(1-\hat{q}^2-\hat{m}_{K^*}^2)(1+\hat{m}_{K^*})
A_1(q^2)\nonumber \\ &&-\hat{\lambda}
\frac{A_2(q^2)}{1+\hat{m}_{K^*}} \label{eq:026},
\end{eqnarray}
\begin{equation}
C_{\|}(q^2)=(1+\hat{m}_{K^*})A_1(q^2) \label{eq:027},
\end{equation}
\begin{equation}
C_{\perp}(q^2)=\frac{V(q^2)}{1+\hat{m}_{K^*}} \label{eq:028},
\end{equation}
\begin{eqnarray}
\kappa_0(q^2)&\equiv&
\Bigl((1-\hat{q}^2+3\hat{m}_{K^*}^2)(1+\hat{m}_{K^*})T_2(q^2)\nonumber
\\&&-\frac{
\hat{\lambda}}{1-\hat{m}_{K^*}}T_3(q^2)\Bigr)\Bigl((1-\hat{q}^2-\hat{m}_{K^*}^2)\nonumber
\\&&\times(1+\hat{m}_{K^*})^2 A_1(q^2)-\hat{\lambda}\,
A_2(q^2)\Bigr)^{-1} \label{eq:029},
\end{eqnarray}
\begin{equation}
\kappa_{\|}(q^2)\equiv \frac{T_2(q^2)}{A_1(q^2)}(1-\hat{m}_{K^*}),
\label{eq:030}
\end{equation}
\begin{equation}
\kappa_{\perp}(q^2)\equiv \frac{T_1(q^2)}{V(q^2)}(1+\hat{m}_{K^*})
. \label{eq:031}
\end{equation}
In the above formulas the definition $\hat{m}_b\equiv
\overline{m}_b(\mu)/m_B$, $\hat{m}_s\equiv
\overline{m}_s(\mu)/m_B$ are used, and $A_1(q^2), \, A_2(q^2), \,
V(q^2), \, T_1(q^2), \, T_2(q^2), \, T_3(q^2)$ are the $B \to K^*$
transition form factors, specified in Appendix~\ref{sec:Appendix}.
\subsection{\label{subsec:resonances}
Resonant contribution}
Next, we implement the effects of LD contributions from the decays ${\bar
B}_d^0\to {\bar K}^{*0}\,V$, where $V=\rho^0\,,\omega\,,
\phi\,,J/\psi(1S)\,,\psi(2S)\,,\ldots$ mesons, followed by $V\to e^+\,e^-$ in
the decay ${\bar B}_d^0\to {\bar K}^{*0}\,e^+\,e^-$~(see
Fig.~\ref{fig:resonances}).
\begin{figure}[tbh]
\centerline{\includegraphics[width=.45\textwidth]{Fig_resonances.eps}}
\caption{Nonresonant and resonant contributions to the decay
amplitude. } \label{fig:resonances}
\end{figure}
We apply vector-meson dominance (VMD) approach. In general, the $V
\, \gamma$ transition can be included into consideration using
various versions of VMD model. In the ``standard'' version (see,
e.g. \cite{Feynman}, chapter~6), the $V \, \gamma $ transition
vertex can be written as
\begin{equation} \langle \gamma (q); \, \mu | V (q); \, \nu\, \rangle = -e f_V Q_V
m_V \, g^{\mu \nu} , \label{eq:032}
\end{equation}
where $q$ is the virtual photon (vector meson) four-momentum,
$g^{\mu \nu}$ is the metric tensor, \ $Q_V$ is the effective
electric charge of the quarks in the vector meson:
\begin{eqnarray}
&&Q_\rho =\frac{1}{\sqrt{2}} \,, \quad Q_\omega=\frac{1}{3
\sqrt{2}}\,, \quad Q_\phi=- \frac{1}{3}\,,\nonumber \\ &&
Q_{J/\psi} = Q_{\psi(2S)}=\ldots = \frac{2}{3}\,. \label{eq:033}
\end{eqnarray}
The decay constants of neutral vector mesons $f_V$ can be
extracted from their electromagnetic decay width, using
\begin{equation}
\Gamma_{V \to e^+ e^-} = \frac{4 \pi \alpha_{em}^2 }{3\,m_V} f_V^2
Q_V^2\,. \label{eq:034}
\end{equation}
This version of VMD model will be called VMD1. The vertex
(\ref{eq:032}) comes from the transition Lagrangian
\begin{equation}
{\cal L}_{\gamma V} = -e A^\mu \sum_{V} f_V Q_V m_V \, V_\mu \,.
\label{eq:035}
\end{equation}
A more elaborate model (called hereafter VMD2) originates from
Lagrangian
\begin{equation}
{\cal L}_{\gamma V} = -\frac{e}{2} F^{\mu \nu} \sum_{V} \frac{f_V
Q_V}{m_V} \, V_{\mu \nu} \label{eq:036}
\end{equation}
where $V_{\mu \nu} \equiv \partial_\mu V_\nu -
\partial_\nu V_\mu$ and $F^{\mu \nu} \equiv \partial^\mu A^\nu -
\partial^\nu A^\mu$ is the electromagnetic field tensor.
Lagrangian (\ref{eq:036}) is explicitly gauge invariant, unlike
Eq.~(\ref{eq:035}), and gives rise to to the $ V \, \gamma^*$
vertex
\begin{equation}
\langle \,\gamma (q); \, \mu | V (q); \, \nu \,\rangle = -\frac{e
f_V Q_V }{m_V} \, (q^2 g^{\mu \nu} - q^\mu q^\nu )\,,
\label{eq:037}
\end{equation}
This transition vertex is suppressed at small invariant masses,
$q^2 \ll m_V^2$, i.e. in the region far from the vector-meson mass
shell~\footnote{The term $\propto q^\mu q^\nu /q^2 $ in
(\ref{eq:037}) does not contribute when contracted with the
leptonic current}.
Note that these two versions of the VMD model have been discussed
in Refs.~\cite{Klingl:1996by,O'Connell}. The VMD2 version
naturally follows from the Resonance Chiral Theory
\cite{EckerNP321}; in this context VMD2 coupling has been applied
in \cite{Eidelman:2010ta} for studying electron-positron
annihilation into $\pi^0 \pi^0 \gamma$ and $\pi^0 \eta \gamma$
final states.
Parameters of vector resonances are presented in
Table~\ref{tab:param1}.
\begin{table}[th]
\caption{Mass, total width, leptonic decay width and coupling
$f_V$ of vector mesons~\cite{PDG:2010} (experimental uncertainties
are not shown).}
\begin{center}
\begin{tabular}{c c c c c}
\hline \hline $V$ & $m_V{\rm (MeV)}$ &$\Gamma_V{\rm (MeV)}$&$\Gamma_{V\to e^+\,e^-}{\rm (keV)}$&$f_V{\rm (MeV)}$ \\
\hline
$\rho^0$ & $775.49$ & $149.1$ & $7.04$ & $221.2$ \\
$\omega$ &$782.65$ & $8.49$ & $0.60$ & $194.7$ \\
$\phi$ &$1019.455$ & $4.26$ & $1.27$ & $228.6$ \\
$J/\psi$ & $3096.916$ & $0.0929$ & $5.55$ & $416.4$ \\
$\psi(2S)$ & $3686.09$ & $0.304$ & $2.35$ & $295.6$ \\
$\psi(3770)$ & $3772.92$ & $27.3$ & $0.265$ & $100.4$ \\
$\psi(4040)$ & $4039$ & $80$ & $0.86$ & $187.2$ \\
$\psi(4160)$ & $4153$ & $103$ & $0.83$ & $186.5$ \\
$\psi(4415)$ & $4421$ & $62$ & $0.58$ & $160.8$ \\
\hline\hline
\end{tabular}
\end{center}
\label{tab:param1}
\end{table}
Based on VMD approach, we obtain the total amplitude including
nonresonant and resonant parts,
\begin{eqnarray} A_{0
L,R}& = &
\frac{1}{2\,\hat{m}_{K^*}\sqrt{\hat{q}^2}}\Biggl(C_0(q^2)\Bigl(C_{9V}^{\rm
eff} \mp C_{10A} \nonumber \\ &&+2\hat{m}_b\left(C_{7\gamma}^{\rm
eff}-C_{7\gamma}^{\prime\,\rm eff}\right)\kappa_0(q^2)\Bigr)
\nonumber \\
&& + 8\pi^2\sum_{V}C_V D_V^{-1}(\hat{q}^2)
\Bigl(\left(1-\hat{q}^2-\hat{m}_{K^*}^2\right)S_1^V\nonumber \\
&&+\hat{\lambda}\frac{S_2^V}{2}\Bigr)\Biggr)\,, \label{eq:038}
\end{eqnarray}
\begin{eqnarray}
A_{\| L,R}&=&-\sqrt{2}\Biggl(C_{\|}(q^2)\Bigl(C_{9V}^{\rm eff} \mp
C_{10A}\nonumber \\
&&+2\frac{\hat{m}_b}{\hat{q}^2}\left(C_{7\gamma}^{\rm
eff}-C_{7\gamma}^{\prime\,\rm
eff}\right)\kappa_{\|}(q^2)\Bigr)\nonumber \\
&&+8\pi^2\sum_{V}C_V
D_V^{-1}(\hat{q}^2)\,S_1^V\Biggr)\,,\label{eq:039}
\end{eqnarray}
\begin{eqnarray}
A_{\perp
L,R}&=&\sqrt{2\hat{\lambda}}\Biggl(C_{\perp}(q^2)\Bigl(C_{9V}^{\rm
eff} \mp
C_{10A}\nonumber \\
&&+2\frac{\hat{m}_b}{\hat{q}^2}\left(C_{7\gamma}^{\rm
eff}+C_{7\gamma}^{\prime\,\rm
eff}\right)\kappa_{\perp}(q^2)\Bigr)\nonumber \\
&&+4\pi^2\sum_{V}C_V
D_V^{-1}(\hat{q}^2)\,S_3^V\Biggr)\,,\label{eq:040}
\end{eqnarray}
where
\[ D_V(\hat{q}^2) = \hat{q}^2 - \hat{m}_V^2 +
i\hat{m}_V \hat{\Gamma}_V (\hat{q}^2) \] is the usual Breit-Wigner
function for the $V$ meson resonance shape with the
energy-dependent width ${\Gamma}_V ({q}^2)$ \ [$\hat{\Gamma}_V
(\hat{q}^2) = {\Gamma}_V ({q}^2) /m_B $], $\hat{m}_V\equiv
m_V/m_B$, $\hat{\Gamma}_V\equiv \Gamma_V/m_B$, $m_V(\Gamma_V)$ is
the mass (width) of a $V$ meson.
\begin{equation} C_V= \frac{Q_V m_V f_V}{q^2}\,\left(\rm{VMD1}\right)\,,
\quad C_V= \frac{Q_V f_V}{m_V}\,\left(\rm{VMD2}\right)\, .
\label{eq:041}
\end{equation}
In Eqs.~(\ref{eq:038})-(\ref{eq:040}), $S_i^V$ \ ($i=1,2,3$) are
the invariant amplitudes of the decay $B_d^0\to K^{*0}\,V$. These
amplitudes are calculated in Appendix~\ref{subsec:vector mesons}.
The energy-dependent widths of light vector resonances $\rho$,
$\omega$ and $\phi$ are chosen as in Ref.~\cite{Korchin:2010}. The
up-dated branching ratios for resonances decays to different
channels are taken from \cite{PDG:2010}. For the $c \bar{c}$
resonances $J/\psi$, \ $\psi (2S)$, $\ldots$ we take the constant
widths.
In order to calculate the resonant contribution to the amplitude
of the ${\bar B}_d^0\to {\bar K}^{*0}\,e^+\,e^-$ decay, one has to
know the amplitudes of the decays ${\bar B}_d^0\to {\bar K}^{*0}\,
\rho$, \ ${\bar B}_d^0\to {\bar K}^{*0}\,\omega, \, {\bar
B}_d^0\to {\bar K}^{*0}\,\phi, \, {\bar B}_d^0\to {\bar
K}^{*0}\,J/\psi, \, {\bar B}_d^0\to {\bar K}^{*0}\,\psi(2S),
\ldots$ At present the amplitudes of the ${\bar B}_d^0\to {\bar
K}^{*0}\,\phi, \, {\bar B}_d^0\to {\bar K}^{*0}\,J/\psi, \, {\bar
B}_d^0\to {\bar K}^{*0}\,\psi(2S) $ decays are known from
experiment~\cite{PDG:2010}, therefore, we use these amplitudes for
calculation of invariant amplitudes in
Appendix~~\ref{subsec:vector mesons} in Table~\ref{tab:ampl}. For
the light resonances $\rho$ and $ \omega$ we use the theoretical
prediction~\cite{Chen:2006} for the decay amplitudes. At the same
time, we are not aware of a similar prediction for the higher $c
\bar{c}$ resonances, such as $\psi(3770)$ an so on, therefore we
do not include contribution of these resonances to amplitudes.
The parameters of the model are indicated in
Table~\ref{tab:param2}. The SM Wilson coefficients have been
obtained in \cite{Altmannshofer:2009} at the scale $\mu=4.8$ GeV
to NNLO accuracy. In our notation (see Appendix~\ref{subsec:matrix
element}) these coefficients are given in Table~\ref{tab:Wilson}.
\begin{table}[t]
\caption{The numerical input used in our analysis.}
\label{tab:param2}
\begin{center}
\begin{tabular}{ll}
\hline \hline
$|V_{tb}V_{ts}^*|=0.04026$ & $G_F=1.16637\times 10^{-5}\, {\rm GeV^{-2}}$\\
$\mu=m_b=4.8\, {\rm GeV}$ & $\alpha_{\rm em}=1/137.036$\\
$m_c=1.4\, {\rm GeV}$ &$m_B=5.27950\, {\rm GeV}$ \\
$\overline{m}_b(\mu)=4.14\, {\rm GeV}$ &$\tau_B=1.525\, {\rm ps}$ \\
$\overline{m}_s(\mu)=0.079\, {\rm GeV}$ &$m_{K^*}=0.89594\, {\rm GeV}$\\
\hline \hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[th]
\caption{The SM Wilson coefficients at the scale $\mu=4.8$\,GeV,
to NNLO accuracy \cite{Altmannshofer:2009}.} \label{tab:Wilson}
\begin{center}
\begin{tabular}{ccccc}
\hline \hline
$\bar C_1(\mu)$ & ${\bar C}_2(\mu)$ &${\bar C}_3(\mu)$ &${\bar C}_4(\mu)$ &${\bar C}_5(\mu)$ \\
$-0.128$ & $1.052$ & $0.011$ & $-0.032$ &$0.009$ \\
${\bar C}_6(\mu)$ & $C_{7\gamma}^{\rm eff}(\mu)$ & $C_{8 g}^{\rm eff}(\mu)$ &$C_{9{\rm V}}(\mu)$&$C_{10{\rm A}}(\mu)$\\
$-0.037$ & $-0.304$ &$-0.167$& $4.211$&$-4.103$ \\
\hline \hline
\end{tabular}
\end{center}
\end{table}
In the numerical estimations, we use the form factors from the
light-cone sum rules (LCSR) calculation \cite{Ball:2005}.
\section{ \label{sec:results}
Results of the calculation for the ${\bar B}_d^0\to {\bar
K}^{*0}\,e^+\,e^-$ decay }
\subsection{\label{subsec:observables}
Dependence of observables on dilepton invariant mass squared}
\begin{figure*}
\begin{center}
\includegraphics[width=.40\textwidth]{dBr_10.eps}
\includegraphics[width=.40\textwidth]{fL_10.eps}
\includegraphics[width=.40\textwidth]{AFB_10.eps}
\includegraphics[width=.40\textwidth]{AT2_10.eps}
\caption{The upper row: differential branching ratio (left), longitudinal
polarization fraction of $K^*$ meson (right); the lower row: forward-backward
asymmetry $A_{\rm FB}$ (left), coefficient $A_{\rm T}^{(2)}$ (right) as
functions of $q^2$. Solid line corresponds to the SM calculation without
resonances taken into account. Dashed and dotted lines are calculated with
account of resonances in the VMD1 and VMD2 versions of VMD model respectively.
The form factors are taken from~\cite{Ball:2005}. The data from Belle
(KEKB)~\cite{Belle:2009} and CDF (Tevatron)~\cite{CDF:2011} are shown by the
circles and boxes respectively. Due to the choice of reference frame in
Fig.~\ref{fig1}, the forward-backward asymmetry $A_{\rm FB}$ in
Refs.~~\cite{Belle:2009, CDF:2011} is related to asymmetry in
Eq.~(\ref{eq:012}) via $A_{\rm FB} = - d{\bar A}_{\rm FB}^{(l)}/d\hat{q}^2$.
The dash-double-dot vertical lines in the figure for differential branching
ratio indicate ``charmonia veto'' (see the text). } \label{fig:Br}
\end{center}
\end{figure*}
In Figs.~\ref{fig:Br} we present results for the dependence of various
observables in the ${\bar B}_d^0\to {\bar K}^{*0}\,e^+\,e^-$ decay on the
dilepton invariant mass squared. The interval of $m_{ee}=\sqrt{q^2}$ is taken
from 10 MeV to $m_B- m_{K^*} \approx 4.384$ GeV. The phase $\delta_0^{V}$ is
chosen zero for all resonances except the $\phi$ meson, for which
$\delta_0^{\phi} =2.82\, {\rm rad}$ (see Table~\ref{tab:ampl}).
As is seen in Fig.~\ref{fig:Br}, predictions of VMD1 and VMD2
models differ for $f_L$, $A_{\rm FB}$ and $A_{\rm T}^{(2)}$ at
small $q^2 \lesssim 2$ GeV$^2$, while for the differential
branching at the bigger values, $q^2 \lesssim 8$ GeV$^2$. At the
bigger values of invariant mass, VMD1 and VMD2 yield close
results. Note that the difference between predictions of these two
models is especially large for the high-lying resonances $J/ \psi$
and $\psi(2S)$.
In addition, VMD1 and VMD2 models lead to a qualitatively different behavior of
longitudinal fraction $f_L$ at small $q^2$ (the upper right panel in
Fig.~\ref{fig:Br}). The data demonstrate that $f_L \to 0$ at very small $q^2$,
that is in agreement with calculation in the VMD2 model. In general, from
comparison with data the VMD2 version seems somewhat more preferable.
Let us comment on the coefficient $A_{\rm T}^{(2)}$ in the azimuthal
distribution (\ref{eq:013}), plotted in Fig.~\ref{fig:Br} (the lower right
panel). According to the definition (\ref{eq:014}) and due to the properties of
the $K^*$ polarization fractions \ $ 0 \leqslant f_{\|, \, \perp } \leqslant
1$, the coefficient $A_{\rm T}^{(2)}$ is constrained:
\begin{equation} -1 \leqslant A_{\rm T}^{(2)}
\leqslant 1 \,. \label{eq:041_c}
\end{equation}
The calculation in the SM with resonances yields the values of
$A_{\rm T}^{(2)}$ which are much smaller than the data (see
Fig.~\ref{fig:Br}). In this connection we note that the
experimental uncertainties are still big, and it is not clear if
the measured $q^2$-dependence of $A_{\rm T}^{(2)}$ indeed lies in
the limits (\ref{eq:041_c}).
\begin{table}[tbh]
\caption{Branching ratio for the decay ${\bar B}_d^0\to {\bar
K}^{*0}\,e^+\,e^-$ calculated within the limits: $m_{ee}^{min}
=30$ MeV, \ $m_{ee}^{max} = m_B-m_{K^*}$. The 2nd column: for the
whole interval of invariant mass, the 3rd column: with the
``charmonia veto'' (see the text).}
\begin{center}
\begin{tabular}{l c c }
\hline \hline & \multicolumn{2}{c}{BR ($10^{-6})$} \\
\cline{2-3}
model & no veto & with veto \\
\hline
SM, no res. & 1.32 & 1.01 \\
SM, res. VMD1 & 134.2 & 49.6 \\
SM, res. VMD2 & 85.6 & 1.03 \\
\hline \hline
\end{tabular}
\end{center}
\label{tab:Branching}
\end{table}
It is seen from Fig.~\ref{fig:Br} (the upper left panel) that the
charmonia resonances contribute to the differential branching far
beyond their pole positions, especially in the VMD1 model. In
order to investigate the role of the charmonia resonances we
calculate the total branching ratio in Table~\ref{tab:Branching}.
Calculation over the whole allowed interval of invariant masses,
shown in the 2nd column, demonstrates a very big resonance
contribution.
Usually in experimental analyses~\cite{Babar:2009,Belle:2009,CDF:2011} certain
cuts are applied in order to cut out the charmonia contributions (the so-called
charmonia veto). In the 3rd column of Table ~\ref{tab:Branching} we used the
integration region with the following cut out intervals taken as in the BaBar
analysis~\cite{Babar:2009} for the $e^+ e^-$ pairs: \ $ 8.11 \leqslant q^2
\leqslant 10.03$ GeV$^2$ and $ 12.15 \leqslant q^2 \leqslant 14.11$ GeV$^2$.
As is seen, the $c \bar{c}$-resonances contribution is completely
eliminated. The result obtained in the VMD2 model becomes close
(within a few percent) to the calculation without resonances. At
the same time the VMD1 calculation still yields a big value of
branching ratio, which is due to the steep rise of the
differential branching at small invariant masses in
Fig.~\ref{fig:Br}.
The calculations, presented in Table~\ref{tab:Branching}, are the
predictions of our model for the current experiments carried out
at LHCb~\cite{Lefrancois:2009}. We can also compare predictions
(with veto) in Table~\ref{tab:Branching} with experimental
measurements: \ $(1.07^{+0.11}_{-0.10} \pm{0.09}) \times 10^{-6}$
for the $B \to K^{*} \, \ell^+\,\ell^-$ decay ($\ell = e, \, \mu$)
at Belle~\cite{Belle:2009}, $(1.02 \pm 0.10 \pm 0.06) \times
10^{-6}$ for the $ B^0 \to K^{*0}\, \mu^+\,\mu^-$ decay at
CDF~\cite{CDF:2011}, and $(1.02^{+0.30}_{-0.28} \pm{0.06}) \times
10^{-6}$ for the $B^0 \to K^{*0}\, e^+\, e^-$ decay at
BaBar~\cite{Babar:2009}.
\begin{figure*}
\begin{center}
\includegraphics[width=.32\textwidth]{A3_10.eps}
\includegraphics[width=.32\textwidth]{A5_10.eps}
\includegraphics[width=.32\textwidth]{A8_10.eps}
\includegraphics[width=.32\textwidth]{A4_10.eps}
\includegraphics[width=.32\textwidth]{A6_10.eps}
\includegraphics[width=.32\textwidth]{A9_10.eps}
\caption{Asymmetries as functions of $q^2$. The upper row from
left to right: $A_3$, $A_5$ and $A_8$; the lower row from left ro
right: $A_4$, $A_6$ and $A_9$. Solid lines correspond to the SM
calculation without resonances, dotted (dashed) lines correspond
to calculation including resonances with the zero-helicity phase
$\delta_0^{J/\psi}=0$ ($\delta_0^{J/\psi}= 2.45$ rad). Calculation
is performed in the VMD2 version. } \label{fig:A3}
\end{center}
\end{figure*}
In Figs.~\ref{fig:A3} we present asymmetries calculated according
to Eqs.~(\ref{eq:014})--(\ref{eq:016}), (\ref{eq:018}),
(\ref{eq:019}), (\ref{eq:021}), (\ref{eq:022}).
Asymmetry $A_3 = \frac{1}{\pi}(1-f_L) A_{\rm T}^{(2)}$ takes
sizable values at large invariant masses, while in the wide region
of $m_{ee}$ this observable is small, of the order of $10^{-2}$
(see also Figs.~\ref{fig:Br} for $A_{\rm T}^{(2)}$). Account of
resonances changes it mainly in the vicinity of the resonance
positions, i.e. at $m_{ee} \approx m_V$.
As for $A_5$ and $A_8$, they take sizable values in the whole
region of $q^2$ (see Figs.~\ref{fig:A3}). The resonances give
considerable contribution, especially to $A_8$. One of features is
the point $q^2_0$, where these asymmetries cross zero. Calculation
shows that the zero point, $q^2_0 \sim 1.5$ GeV$^2$, is almost
insensitive to the presence of the resonances. This feature makes
these asymmetries convenient observables for experimental study,
similarly to the forward-backward asymmetry $A_{{\rm FB}}$ in
Fig.~\ref{fig:Br}.
Some of the asymmetries are very small in the SM without
resonances, in particular, $A_9 \equiv 0$, $A_4 \sim 10^{-4}$ and
$A_6 \sim 10^{-3}$ (Figs.~\ref{fig:A3}). Note that these
asymmetries are determined by the imaginary part of the bilinear
combinations of the amplitudes. The imaginary part of the
amplitudes in the SM (without resonances) is determined by the
$u,\, d,\, s$ and $c$ quark loop through the function $Y(q^2)$
\cite{BFS:2001}, and therefore the imaginary part of the
nonresonant amplitudes appears to be very small. Therefore in
framework of the naive factorization, applied in the present work,
it is not surprising that these asymmetries are small in the SM
without resonances. Inclusion of the resonances changes behavior
of these asymmetries in the wide region of invariant masses (see
Figs.~\ref{fig:A3}).
Recently the asymmetry $A_{{\rm Im}}= -\frac{\pi}{2}A_4$ in the
$B^0 \to K^{*0}\, \mu^+\,\mu^-$ decay has been measured for the
first time at CDF~\cite{CDF:2011}. Note that the average values of
this asymmetry over the $q^2$-ranges $[0.0 - 4.3)$ GeV$^2$ and
$[1.0 - 6.0)$ GeV$^2$ are consistent with our calculations in
framework of the SM.
One should note, that the amplitudes for the decay of $B$ meson to
two vector mesons are experimentally determined by the four
polarization parameters, branching fraction and one overall phase
$\delta_0^V $ (the phase of the amplitude with zero helicity for
decay $B \to K^* V$). As it is seen from
Eqs.~(\ref{eq:038})--(\ref{eq:040}) the contribution of resonances
depends on the invariant amplitudes $S^{V}_i$. Values of these
amplitudes are determined in Table~\ref{tab:ampl}, however their
phases are defined with respect to the phase $\delta_0^V$ , which
is experimentally known only for the decay $B \to K^* \phi$. For
other resonances, the phase $\delta_0^V$ is not known either
experimentally or theoretically.
As is seen in Figs.~\ref{fig:A3}, the asymmetries $A_4$ and $A_9$ essentially
depend on the choice of the $\delta_0^{J/\psi}$ phase. Thus, in order to
unambiguously determine the resonance contribution to the process ${\bar B}_d^0
\to {\bar K}^{*0} \, (\to K^{-}\, \pi^+) \, \ell^+\, \ell^- $, the phases
$\delta_0^V$ should be known for all vector resonances $\rho, \omega, \phi,
J/\psi, \ldots $. These phases can be found from experiments on $B$-meson
decays through the interference with other $B$ decays with the same final
states, for example, $B^0 \to \phi K^{*0}$ and $B^0 \to \phi (K \pi)_0^{*0}$,
where $(K \pi)_0^{*0}$ is the $J^P =0^+$ \ $K \pi$
component~\cite{Babar:2007}.
\subsection{\label{subsec:two_approaches}
Comparison of two approaches to including the resonances }
Earlier in the literature \cite{Deshpande:1989} it has been
suggested to combine the factorization assumption and VMD
approximation in estimating LD effects for the $B$ decays. This
can be accomplished in an approximate manner through the
substitution
\begin{equation}\label{eq:042}
C_{9V}^{\rm eff}\to C_{9V}^{\rm
eff}-\frac{3\,\pi}{\alpha_{em}^2}C^{(0)}\sum_{\psi} k_{\psi}
\frac{\hat{m}_{\psi}\,\hat{\Gamma}(\psi\to e^+e^-)}{\hat{q}^2 -
\hat{m}_{\psi}^2 + i\hat{m}_{\psi} \hat{\Gamma}_{\psi}}\,,
\end{equation}
where the properties of the vector mesons
$\psi=J/\psi(1S)\,,\psi(2S)\,,\ldots,\psi(4415)$ are summarized in
Table~\ref{tab:param1}, \ $\hat{\Gamma}_{\psi\to e^+e^-}\equiv
\Gamma_{\psi\to e^+e^-}/m_B$, \ $C^{(0)}=3\,\bar C_1+\bar
C_2+3\,\bar C_3+\bar C_4+3\,\bar C_5+\bar C_6$, and the Wilson
coefficients are presented in Table~\ref{tab:Wilson}.
The last term in Eq.~(\ref{eq:042}) describes the LD contribution
of the real intermediate $\bar c\,c$ states,
$k_\psi=|k_\psi|\,e^{i\delta_\psi}$ is the factor that the $B\to
K^* \psi$ amplitude, calculated using naive factorization, must be
multiplied by to get the measured $B\to K^* \psi$ rate. Under
naive factorization, the branching ratio for $B \to K^* \psi$ is
\begin{eqnarray}\label{eq:043}
{\rm BR}(B \to K^* \psi) &=&m_B\,\tau_B\,
\frac{\sqrt{\lambda(1,\hat{m}_{K^*}^2,\hat{m}_\psi^2)}}{16\pi}\nonumber \\
& \times & \left(\frac{G_F m_B^2}{\sqrt{2}}\right)^2\,
\left|V_{tb}V_{ts}^*\frac{C^{(0)}}{3}\right|^2 \nonumber \\
& \times &\left( |X_0^\psi|^2 + |X_\|^\psi|^2 + |X_\perp^\psi|^2
\right) \,,
\end{eqnarray}
where
\begin{eqnarray}\label{eq:044}
&& X_0^\psi=\frac{\hat{f_\psi}}{2\,\hat{m}_{K^*}(1+\hat{m}_{K^*})}
\Bigl((1+\hat{m}_{K^*})^2
(1-\hat{m}_{K^*}^2 \nonumber \\
&&-\hat{m}_\psi^2)\,A_1(m_\psi^2) -
\lambda(1,\hat{m}_{K^*}^2,\hat{m}_\psi^2)\,A_2(m_\psi^2) \Bigr)
\,,
\end{eqnarray}
\begin{equation}\label{eq:045}
X_\|^\psi=-\sqrt{2}\,\hat{m}_\psi\hat{f_\psi}(1+\hat{m}_{K^*})\,A_1(m_\psi^2)
\,,
\end{equation}
\begin{equation}\label{eq:046}
X_\perp^\psi=\sqrt{2\,\lambda(1,\hat{m}_{K^*}^2,\hat{m}_\psi^2)}\,\hat{m}_\psi\hat{f_\psi}/(1+\hat{m}_{K^*})\,V(m_\psi^2)
\,.
\end{equation}
Here, $\hat{f_\psi}\equiv f_\psi/m_B$. \
We calculate the values of $k_\psi$ for each $c \bar c$ resonance
using experimental information from Tables~\ref{tab:param1} and
\ref{tab:ampl} and equation
\begin{equation}
|k_\psi|^2 = \frac{{\rm BR}(B \to K^* \psi)_{exp} }{{\rm BR}(B \to
K^* \psi)_{theor}} \,. \label{eq:047}
\end{equation}
Using the form factors $A_1, \, A_2, \, V$ in the LCSR
model~\cite{Ball:2005} we find the values: \ $|k_{J/\psi}|=0.894$,
\ $|k_{\psi(2S)}|=0.841$ and for the higher resonances the average
of the $|k_{J/\psi}|$ and $|k_{\psi(2S)}|$ is used. The phase of
$k_\psi$ is chosen zero as in the factorization approach. Note,
that the above values of the parameters $k_\psi$ are considerably
smaller that the values used in the earlier
papers~\cite{Deshpande:1989,Ligeti:1996,Ali:2000}. Therefore we
can expect the smaller resonance contribution to the differential
branching and forward-backward asymmetry as compared with results
of these papers.
Using Eqs.~(\ref{eq:042}), (\ref{eq:023})--(\ref{eq:025}) we
calculate the observables for the ${\bar B}_d^0 \to {\bar K}^{*0}
e^+ e^-$ decay. These results are compared with results of
calculations performed in the previous sections, in which only
$J/\psi$ and $\psi(2S)$ resonances are included. This comparison
may show sensitivity of the branching, $K^*$ polarization
fractions and asymmetries to the method of including the vector
resonances.
\begin{figure*}
\begin{center}
\includegraphics[width=.32\textwidth]{dBr_11.eps}
\includegraphics[width=.32\textwidth]{AFB_11.eps}
\includegraphics[width=.32\textwidth]{A8_11.eps}
\end{center}
\caption{Differential branching (left), forward-backward asymmetry
(middle) and asymmetry $A_8$ (right) a function of $q^2$. Solid
lines are calculated without resonances, dotted lines are
calculated with resonances $J/\psi$, $\psi(2S)$ according
Eqs.~(\ref{eq:038})--(\ref{eq:040}), and dash-dotted lines with
resonances according to Eqs.~(\ref{eq:042}),
(\ref{eq:023})--(\ref{eq:025}). } \label{fig:01res}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=.32\textwidth]{fL_11.eps}
\includegraphics[width=.32\textwidth]{AT2_11.eps}
\includegraphics[width=.32\textwidth]{A5_11.eps}
\includegraphics[width=.32\textwidth]{A4_11.eps}
\includegraphics[width=.32\textwidth]{A6_11.eps}
\includegraphics[width=.32\textwidth]{A9_11.eps}
\end{center}
\caption{The upper row from left to right: longitudinal
polarization fraction $f_L$, asymmetries $A_{\rm T}^{(2)}$ and
$A_5$; the lower row from left ro right: asymmetries $A_4$, $A_6$
and $A_9$, as functions of $q^2$. Solid lines are calculated
without resonances, dotted lines are calculated with resonances
$J/\psi$, $\psi(2S)$ according
Eqs.~(\ref{eq:038})--(\ref{eq:040}), and dash-dotted lines with
resonances according to Eqs.~(\ref{eq:042}),
(\ref{eq:023})--(\ref{eq:025}). } \label{fig:02res}
\end{figure*}
Firstly, from Fig.~\ref{fig:01res} one sees that the two ways of
including the resonances give very close predictions for
differential branching, forward-backward asymmetry and asymmetry
$A_8$ in the region of $q^2$ up to $m^2_{\psi(2S)}$. Note, that
calculations of the branching and forward-backward asymmetry have
been performed in Ref.~\cite{Ali:2000} using formulas analogous to
Eqs.~(\ref{eq:042}). Our results in Figs.~\ref{fig:01res}
qualitatively agree with calculations in \cite{Ali:2000}.
Secondly, in Fig.~\ref{fig:02res} we present asymmetries, which strongly depend
on the method of including resonances. As for asymmetries $A_{\rm T}^{(2)}$ and
$A_5$, these two methods give essentially different predictions in the region
of resonances $J/\psi$ and $\psi(2S)$, while far from the resonance region, the
predictions coincide with the non-resonant calculations. In the other
asymmetries, $A_4$, $A_6$ and $A_9$, the two methods give different predictions
not only in the resonance region but also at $q^2$ away from resonances.
\section{ \label{sec:conclusions} Conclusions}
The rare FCNC decay ${\bar B}_d^0 \to {\bar K}^{*0} \, (\to K^{-}\, \pi^+) \,
e^+\,e^- $ has been studied in the whole region of electron-positron invariant
masses up to $m_B -m_{K^*}$. The fully differential angular distribution over
the three angles and dilepton invariant mass for the four-body decay ${\bar
B}_d^0 \to K^{-}\, \pi^+ \, e^+\,e^- $ is analyzed. We defined a convenient set
of asymmetries which allows one to extract them from measurement of the angular
distribution once sufficient statistics is accumulated. We performed
calculations of the differential branching ratio, polarization fractions of
$K^*$ meson and asymmetries. These asymmetries may have sensitivity to various
effects of NP, although in order to see signatures of these effects the
resonance contribution should be accurately estimated.
Contribution of the intermediate vector resonances in the process $\bar{B}_d^0
\to \bar{K}^{*0} \, (\to K^{-}\, \pi^+) \, V $ with $V = \rho(770), \,
\omega(782), \, \phi(1020), \, J/\psi, \, \psi(2S)$, decaying into the $e^+
e^-$ pair, has been taken into account. Various aspects of theoretical
treatment of this long-distance contribution have been studied.
The important aspect is the choice of the VMD model, describing
the $V \, \gamma$ transition. We used two variants of the VMD
model, called here VMD1 and VMD2 versions, in particular, the VMD2
one is explicitly gauge-invariant and yields a $V \, \gamma$
vertex which is suppressed at small values of the photon invariant
mass far from the vector-meson mass shell. This is especially
important in the treatment of the high-lying $J/\psi$ and other $c
\bar c$ resonances at values $q^2 \ll m_V^2$. Some of the
observables appeared to be rather sensitive to the choice of the
$V \, \gamma$ vertex. Based on comparison of calculation with the
recent data from Belle and CDF experiments we can conclude that
the VMD2 version is somewhat more preferable.
For the vertex $\bar{B}_d^0 \to \bar{K}^{*0} \, V $ we used an off-mass-shell
extension of the helicity amplitudes describing production of on-shell vector
mesons. For the latter the experimental information is used if available, and
otherwise theoretical predictions.
The total branching ratio of the decay ${\bar B}_d^0 \to {\bar
K}^{*0} \, e^+\,e^- $ in the interval $30 \, {\rm MeV} \leq m_{ee}
\leq m_B-m_{K^*}$ is calculated, and the resonance contribution is
evaluated. The latter appeared to be very big, as expected. The
calculated branching is ${\rm BR} = 1.32 \times 10^{-6} $ ($85.6
\times 10^{-6})$ without resonances (with resonances in VMD2
version). To cut out the $c \bar{c}$-resonance contribution, we
also applied the ``charmonia veto'' as is usually done in
experimental analyses~\cite{Babar:2009,Belle:2009,CDF:2011}. Then
our prediction for the total branching becomes ${\rm BR} = 1.01
\times 10^{-6} $ ($1.03 \times 10^{-6})$ for the SM calculation
without resonances (with resonances in VMD2 model).
All asymmetries are calculated in the whole region of invariant
masses. The polarization asymmetry $A_{\rm T}^{(2)}$ (and $A_3 =
\frac{1}{\pi}(1-f_L) A_{\rm T}^{(2)}$) takes sizable values only
at large $m_{ee}$. Account of resonances changes $A_{\rm T}^{(2)}$
mainly in the vicinity of the resonances, i.e. at $m_{ee} \approx
m_V$.
The asymmetries $A_5$ and $A_8$ take big values in the whole region of $q^2$,
and resonances noticeably contribute, especially to $A_8$. An interesting
feature of these asymmetries is their crossing zero at some $q^2_0$. This zero
point, $q^2_0 \sim 1.5$ GeV$^2$, turns out to be almost independent of the
presence of resonances, and this property makes these asymmetries convenient
observables for experimental study, similarly to the forward-backward asymmetry
$A_{\rm FB}$.
Some of the asymmetries are very small in the SM without
resonances, in particular, $A_9 \equiv 0$, \ $A_4 \sim 10^{-4}$
and $A_6 \sim 10^{-3}$; inclusion of the resonances changes
behavior of these asymmetries considerably.
Our calculations are compared with recent
data~\cite{Belle:2009,CDF:2011} for $q^2$-dependence of the
differential branching ratio, longitudinal polarization fraction
of $K^*$, forward-backward asymmetry $A_{{\rm FB}}$ and transverse
asymmetry $A_{{\rm T}}^{(2)}$. On the whole, results of
calculation are in reasonable agreement with the data.
We also compared predictions of our method with other existing in
the literature method of including the $c \bar c$ resonances
through a modification of the Wilson coefficient
$C_{9V}^{eff}$~\cite{Deshpande:1989}.
Our calculation shows that a few observables (differential
branching ratio, forward-backward asymmetry and asymmetry $A_8$)
are independent of the calculation method, if the parameters
$|k_\psi|$ in Eq.~(\ref{eq:042}) are equal to: $0.894$ for
$J/\psi$, \ 0.841 for $\psi(2S)$ and 0.8675 for the higher $c \bar
c$ resonances. The phase of $k_\psi$ is chosen zero. These values
are considerably smaller that the values used
in~\cite{Deshpande:1989,Ligeti:1996,Ali:2000}.
At the same time there exist asymmetries, predictions for which
are substantially different in these two methods, namely, $A_{\rm
T}^{(2)}$, \ $A_5$ are different in the vicinity of resonances
$J/\psi$ and $\psi(2S)$, while $A_4$, \ $A_6$, \ $A_9$ are
different in the whole region of invariant masses. Thus
measurement of the latter asymmetries may also be useful for
selecting a more adequate method of including the long-distance
resonance contribution to the ${\bar B}_d^0 \to {\bar K}^{*0} \,
(\to K^{-}\, \pi^+) \, \ell^+\,\ell^- $ decay.
Calculations performed in the present work may be useful for experiments aiming
at search of effects of the NP in the decay ${\bar B}_d^0 \to {\bar K}^{*0} \,
(\to K^{-}\, \pi^+) \, \ell^+\,\ell^- $.
|
\section{Introduction}
Microscopic theories for calculating nuclear masses and/or binding energies (see, e.g., \cite{boh58,BM69,BB05}), have been revived and further elaborated with the advance of computational
resources. These advances are now sufficient to perform global studies based on, e.g., self-consistent mean field theory, sometimes also denoted by density functional theory (DFT) \cite{Ben03,Stoi06}.
One particular aspect of the nuclear binding problem is a phenomenon of odd-even staggering (OES)
of the binding energy. Numerous microscopic calculations have been published that treat individual isotope
chains. However, it might be necessary to examine the whole body of OES data to draw
general conclusions \cite{Satula98}.
Theoretically, OES values are often inferred from the average HFBCS or Hartree-Fock-Bogoliubov gaps \cite{dug01,mar07,Yama08},
rather than directly calculated from the experimental binding energy differences between
even and odd nuclei.
It should be mentioned that the average HFB gaps are sometimes
substantially different from the odd-even mass differences calculated from experimental binding energies.
In this work, we compare directly the calculated OES with the ones extracted from experiment.
One should say that there are also several prescriptions to obtain the OES from experiments, such as
3-point, 4-point, and 5-point formulas \cite{Satula98}. We adopt the 3-point formula
$\Delta^{(3)}$ centered at an odd nucleus, i.e., odd-N nucleus for neutron gap and
odd-Z nucleus for proton gap \cite{BM69}:
\begin{eqnarray} \label{eq:oes}
\Delta^{(3)}(N,Z)\equiv\frac{\pi_{A+1}}{2} \Big[\mathrm{B}(N-1,Z)&-&2\mathrm{B}(N,Z) \\\nonumber
&+&\mathrm{B}(N+1,Z)\Big] \; , \end{eqnarray} where $B(N,Z)$ is
the binding energy of $(N,Z)$ nucleus and
$\pi_A=(-)^A$ is the number parity with $A=N+Z$.
For even nuclei, the OES is known to be sensitive not only to the
pairing gap, but also to mean field effects, i.e., shell effects
and deformations~\cite{Satula98,dug01}. Therefore, the comparison of a
theoretical pairing gap with OES should be done with some discretion.
One advantage of $\Delta^{(3)}_o$ ($N=$ odd in Eq. \eqref{eq:oes})
is the suppression of the contributions from the mean field to the
gap energy. Another advantage of $\Delta^{(3)}_o(N,Z)$ is that it can be applied to more
experimental mass data than the higher order OES formulas.
At a shell closure, the OES~(Eq. \eqref{eq:oes}) does not go to zero
as expected, but it increases substantially. This large gap is an
artifact due to the shell effect, which is totally independent of
the pairing gap itself.
Recently, an effective isospin dependent pairing interaction was proposed
from the study of nuclear matter pairing gaps calculated by realistic
nucleon-nucleon interactions.
In Ref. \cite{mar07},
the density$-$dependent pairing interaction was defined as \begin{eqnarray}
V_{pair}(1,2)= \mathrm{V}_0 \,\mathrm{g}_\tau[\rho,\beta\tau_z]
\,\delta(\r_1-\r_2),
\label{eq:pairing_interaction} \end{eqnarray}
where $\rho=\rho_n+\rho_p$ is the
nuclear density and $\beta$ is the asymmetry parameter
$\beta=(\rho_n -\rho_p)/\rho$.
The isovector dependence is introduced through the
density-dependent term $\mathrm{g}_\tau$. The function
$\mathrm{g}_\tau$ is determined by the pairing gaps in nuclear
matter and its functional form is given by
\begin{eqnarray}
\mathrm{g}_\tau[\rho,\beta\tau_z] = 1
-\mathrm{f}_\mathrm{s}(\beta\tau_z)\eta_\mathrm{s}
\left(\frac{\rho}{\rho_0}\right)^{\alpha_\mathrm{s}}
-\mathrm{f}_\mathrm{n}(\beta\tau_z)\eta_\mathrm{n}
\left(\frac{\rho}{\rho_0}\right)^{\alpha_\mathrm{n}},
\label{eq:g1t}
\end{eqnarray} where $\rho_0$=0.16~fm$^{-3}$ is the saturation
density of symmetric nuclear matter. We choose
$\mathrm{f}_\mathrm{s}(\beta\tau_z)=1-\mathrm{f}_\mathrm{n}(\beta\tau_z)$
and
$\mathrm{f}_\mathrm{n}(\beta\tau_z)=\beta\tau_z=\left[\rho_\mathrm{n}({\bf r})-\rho_\mathrm{p}({\bf r})
\right]\tau_z/\rho({\bf r})$. The parameters for $\mathrm{g}_\tau$ are obtained from the fit to the pairing gaps in
symmetric and neutron matter obtained by the microscopic
nucleon-nucleon interaction.
\begin{figure}[htb]
\begin{center}
\includegraphics[clip,scale=0.33,angle=-90]{isivp.ps}
\caption{(Color online) The mean square deviation $\sigma$ of OES between
experimental data and the HF+BCS calculations. The filled circles and squares correspond to
the results with IS pairing for neutron and proton gaps, respectively, while
the filled diamonds and triangles are those of IS+IV pairing for neutron and proton gaps.
Experimental data are taken from Ref. \cite{Audi2003}. See the text for details. }
\label{fig01}
\end{center}
\end{figure}
\begin{table*}[htb]
\begin{center}
\setlength{\tabcolsep}{.06in}
\renewcommand{\arraystretch}{1.5}
\caption{Parameters for the density-dependent function
$\mathrm{g}_\tau$ defined in Eqs.~(\ref{eq:pairing_interaction}) and (\ref{eq:g1t}) for the
IS+IV interaction (first row)
and $\mathrm{g}_s$ in Eq. \eqref{isoscalar} for the IS interaction.
The parameters for $\mathrm{g}_\tau$ are obtained from the fit to the pairing gaps in
symmetric and neutron matter obtained with the microscopic
nucleon-nucleon interaction. The paring strength $V_0$
is adjusted to give the best fit
to odd-even staggering of nuclear masses. The parameters for $\mathrm{g}_s$
correspond to a surface peaked pairing interaction with no isospin
dependence. \label{table1}}
\begin{tabular}{cccccccc}
\toprule
interaction & $V_0$ (MeVfm$^3$)& $\rho_0$ (fm$^{-3}$) & $\eta_\mathrm{s}$ & $\alpha_\mathrm{s}$ & $\eta_\mathrm{n}$ & $\alpha_\mathrm{n}$ \\
\colrule
$\mathrm{g}_\tau$ (isotopes) & 1040 & 0.16 & 0.677& 0.365 & 0.931 & 0.378 \\
$\mathrm{g}_\tau$ (isotones)& 1120 & 0.16 & 0.677& 0.365 & 0.931 & 0.378 \\
$\mathrm{g}_s$ (isotopes)& 1300 & 0.16 & 1. & 1. & --- & --- \\
$\mathrm{g}_s$ (isotones)& 1500 & 0.16 & 1. & 1. & --- & --- \\
\botrule
\end{tabular}
\end{center}
\end{table*}%
In the literature and in many mean field codes publicly available such as the original EV8 code \cite{EV8}, a pure contact interaction
is used without an isospin dependence. In our notation, this amounts replacing
the isospin dependent function $\mathrm{g}_\tau$ in Eq. \eqref{eq:pairing_interaction} by the isoscalar
function
\begin{equation}
\mathrm{g}_s=1-\eta_\mathrm{s}
\left(\frac{\rho}{\rho_0}\right)^{\alpha_\mathrm{s}}
\label{isoscalar}.
\end{equation}
The EV8 code has been modified, using the filling approximation, to account for mass calculations for odd-N and odd-Z nuclei and is publicly available as the EV8odd code \cite{BB08}. It has also been modified to include isospin dependent paring, by means of Eq. \eqref{eq:pairing_interaction}.
The parameters of the isoscalar interaction were adjusted with EV8 to a best global fit of nuclear masses \cite{bertsch09}. They correspond to
a surface peaked pairing interaction (Eq. \eqref{isoscalar} with $\eta_s$ not too far from the unity).
A recent publication has explored the isospin dependence of the pairing force for the OES effect for a few selected isotopic and isotonic chains \cite{BLS09}. Here we have made a more ambitious study by extending the calculation to the whole nuclear chart. We have also explored several other
observables such as neutron and proton radii systematically which may allow for
more solid conclusions on isospin dependent pairing interactions.
This paper is organized as follows.
In section II we discuss our numerical calculation strategy. Our results are presented in section III for the energies, separation energies, OES energies, and nuclear radii. Our conclusions are presented in section IV.
\section{Calculation strategy}
The HF+BCS calculations are performed by using SLy4 Skyrme interaction which was found to be the most accurate interaction for studying OES for a few selected ($N=50,82$) isotonic and (Sn and Pb)isotopic chains \cite{BLS09}.
Our iteration procedure used in connection to EV8odd
achieves an accuracy of about 100 keV, or less, with 500 Hartree-Fock iterations for each nuclear state. Our calculations were performed with the now decommissioned XT4 Jaguar supercomputer at ORNL, as part of the UNEDF-SciDAC-2 collaboration \cite{SciDAC}.
The HF+BCS calculations were first performed for even-even nuclei.
The variables in the theory are the orbital wave functions $\phi_i$
and the BCS amplitudes $v_i$ and $u_i = \sqrt{1 - v^2_i}$. By
solving the BCS equations for the amplitudes, one obtains the
pairing energy from
\begin{equation}
E_{pair} = \sum_{i\neq j} V_{ij} u_i v_i u_j
v_j + \sum_i V_{ii} v^2_i \label{Epair}
\end{equation}
where $V_{ij}$ are the matrix elements
of the pairing interaction, Eq. \eqref{eq:pairing_interaction}, namely
$$
V_{ij}=V_0\int d^3 r |\phi_i({\bf r})|^2 |\phi_j({\bf r})|^2
\mathrm{g}_\tau[\rho ({\bf r}),\beta({\bf r})\tau_z],
$$
where $\rho({\bf r})=\sum_i v_i^2 |\phi_i({\bf r})|^2$.
\begin{figure*}[t]
\begin{center}
\includegraphics[clip,scale=0.42]{dE.ps}
\caption{(Color online) Binding energy differences between experimental data and calculations
using the HF+BCS model with the IS and IS+IV pairing interactions. The left panels show the differences for even Z isotopes
varying neutron numbers including both odd and even numbers.
The right panel show those for even N isotones
varying proton numbers including both odd and even numbers. The thin lines show the closed
shells at $N(Z)$ = 20, 28, 40, 50, 64, 82 and 126.
Experimental data are taken from Ref. \cite{Audi2003}. See the text for details. }
\label{fig-dE}
\end{center}
\end{figure*}
\begin{figure*}[htb]
\begin{center}
\includegraphics[clip,scale=0.42]{dS.ps}
\caption{(Color online) The same as Fig. \ref{fig-dE} but for neutron and proton separation energies.
See the caption to Fig. \ref{fig-dE} and the text for details. }
\label{fig-dS}
\end{center}
\end{figure*}
After determining the single-particle energies of
even-even nuclei,
the odd-A nuclei are calculated with the so-called filling
approximation for the odd particle starting from the HF+BCS
solutions of neighboring even-even nuclei: ones selects a pair of
$i$ and $\widetilde{i}$ orbitals to be blocked, and changes the BCS parameters
$v^2_i$ and $v_{\widetilde{i}}^2$ for these orbitals.
The change is to set
$v^2_i = v_{\widetilde{i}}^2=1/2$ in
Eq. \eqref{Epair} for the pairing energy at an orbital near the Fermi
energy. Note that this approximation gives equal occupation
numbers to both time-reversed partners, and does not account for the
effects of time-odd fields. More details of the procedure are
presented in Ref. \cite{bertsch09}.
The effect of
time-odd HF fields on the mass were studied in Refs. \cite{Duguet-Jerome,Mar11}.
It was pointed out that the effect of the time-odd fields is of the order of 100 keV
for the binding energy
depending strongly on the configuration of the last particle, and does not show
any clear sign of isospin dependence. Thus the time-odd
field might not change
conclusions of the present study in the following, while quantitative
accuracy might need some fine tuning of the pairing parameters.
\begin{figure*}[htb]
\begin{center}
\vspace{-0.5cm}
\includegraphics[clip,scale=0.6,angle=-90]{Sn.ps}
\includegraphics[clip,scale=0.6,angle=-90]{Sp.ps}
\caption{(Color online) Neutron separation energies $S_n$
of three isotope chains with $Z=46$, 72 and 90 calculated
by IS and IS+IV interactions in HF+BCS model (upper panels).
The lower panels show the proton separation energies $S_p$ for $N=50$, 90 and 136 isotones.
Experimental data are taken from Ref. \cite{Audi2003}. See the text for details. }
\label{fig-S}
\end{center}
\end{figure*}
\begin{figure}[htb]
\begin{center}
\includegraphics[clip,scale=0.45,angle=-90]{Sp-N136.ps}
\caption{(Color online) Proton separation energies $S_p$
of $N=136$ isotones calculated
by IS and IS+IV interactions in HF+BCS model.
Experimental data are taken from Ref. \cite{Audi2003}. See the text for details. }
\label{fig-Sp}
\end{center}
\end{figure}
For the pairing channels we have taken the surface-type contact
interaction, Eq. \eqref{isoscalar}, and the isospin dependent
interaction, Eq. \eqref{eq:g1t}. The density dependence of the
latter one is essentially the mixed-type interaction between the
surface and the volume types. The pairing strength $V_0$ depends on
the energy window adopted for BCS calculations. The odd nucleus is
treated in the filling approximation, by blocking one of the
orbitals. The blocking candidates are chosen within an energy window
of 10 MeV around the Fermi energy.
This energy window is rather small, but it is the maximum allowed by
the program EV8odd. It is shown that the BCS model used in the EV8odd code gives
almost equivalent results
to the HF+Bogoliubov model with a larger energy window,
except for unstable nuclei very close to the neutron drip line \cite{bertsch09}.
The pairing strengths $V_0$ for IS and IS+IV pairing interactions are
adjusted to give the best fit
to odd-even staggering of nuclear masses in a wide region of the mass table.
\begin{figure*}[htb]
\begin{center}
\includegraphics[clip,scale=0.42]{dD.ps}
\caption{(Color online) The same as in Fig. \ref{fig-dE} but for OES for neutrons and protons.
See the caption to Fig. \ref{fig-dE} and the text for details. }
\label{figdD}
\end{center}
\end{figure*}
\section{Numerical results}
\subsection{Global data on odd-even staggering}
The results for the mean square deviation of our global mass table calculations are shown in Fig. \ref{fig01}.
Table \ref{table1} gives the values $V_0$ in Eq. \eqref{eq:pairing_interaction} and the parameters for
$\mathrm{g}_\tau$ and $\mathrm{g}_s$ is Eqs. \eqref{eq:g1t} and \eqref{isoscalar} used in the present work.
Optimal pairing strength values were found to be different for isotones (varying $Z$, constant $N$) and
for isotopes (varying $N$, constant $Z$).
Figure \ref{fig01} shows the mean square deviation $\sigma$ of OES between
experimental data and the HF+BCS calculations. The mean square deviation $\sigma$ is
defined as
\begin{eqnarray}
\sigma =\sqrt{\sum_{i=1}^{N_i}\left|\Delta_i^{(3)}(HF+BCS)-\Delta_i^{(3)}(exp)\right|^2/N_i}
\end{eqnarray}
where $N_i$ is the number of data points.
For the IS interaction, the results for neutrons show a shallow minimum at $V_0\sim(1100-1300)$
MeV$\cdot$fm$^3$. For protons, the minimum becomes at around $V_0\sim1500$ MeV$\cdot$fm$^3$.
This difference makes it difficult to determine a unique pairing strength common for both neutrons and
protons. The results of IS+IV pairing show a minimum at V$_0\sim1100$ MeV$\cdot$fm$^3$ for both
neutron and proton OES which makes it easier to determine the value for the pairing strength.
Adopted values for the following calculations are listed in Table \ref{table1}.
The systematic study of HF+BCS calculations are performed for various isotopes and isotones
for all available data sets with ($Z=8,\cdots,102$) and ($N=8,\cdots, 156$), respectively.
Binding energy differences between experimental data and HF+BCS
\begin{eqnarray}
\delta E=|E^{exp}-E^{cal}|
\label{Diff-E}
\end{eqnarray}
are shown for both IS and
IS+IV pairing interactions in Fig. \ref{fig-dE}. The left panels show the values $\delta E$
varying neutron numbers (including both odd and even numbers) for each even Z.
The right panels show the values $\delta E$
varying proton numbers (including both odd and even numbers) for each even N.
With IS pairing, we can see a rather large
deviation for $Z=50$ isotopes in the upper right panel. This difference disappears in the case
of IS+IV pairing shown in the lower right panel. On the other hand, for the $N=82$ nuclei,
the IS+IV interaction does not work that well. As far as the binding energies are concerned,
the best results with IS+IV interaction are obtained for nuclei with $N=60-78$ and $N=86-96$.
Separation energy differences between experimental data and HF+BCS for
protons and neutrons are plotted in Fig. \ref{fig-dS}.
The HF+BCS results of neutron separation energies $S_n$ are reasonable
for medium and heavy mass nuclei with $N=60-120$, except
near the closed shell $N=82$. For heavy nuclei with $N=126$, the calculated results are
poorer than in other mass regions. For proton separation energy $S_p$, the HF+BCS also gives
reasonable results, except in the $Z=50$ and 82 mass regions.
In order to see the different outcomes between IS and IS+IV pairing interactions, the HF+BCS model
calculations are shown together with empirical data in Fig. \ref{fig-S}.
In most of cases, the difference between the two pairing interactions are small.
However, we can see a clear improvement of the agreement of $S_p$
with empirical data of $N=136$ isotones
with IS+IV pairing in Fig. \ref{fig-Sp}.
\begin{figure*}[htb]
\begin{center}
\vspace{-0.5cm}
\includegraphics[clip,scale=0.6,angle=-90]{comparison-Dn.ps}
\includegraphics[clip,scale=0.6,angle=-90]{comparison-Dp.ps}
\caption{(Color online) The neutron and proton OES, $\Delta_n$ and $\Delta_p$, calculated
with the HF+BCS model with IS and IS+IV interactions.
Experimental data are taken from Ref. \cite{Audi2003}. See the text for details. }
\label{fig-Delta-np}
\end{center}
\end{figure*}
The differences of neutron OES $\Delta_n$ and proton OES $\Delta_p$
between HF+BCS and empirical data are shown in
Fig. \ref{figdD} for both IS and IS+IV pairing, respectively, in the upper and lower
panels.
The agreement between HF+BCS and the empirical data
are good in the overall mass region except for masses with $Z=50$ and
at a small mass region $A<60$. To clarify the difference between IS and IS+IV pairing,
the OES differences $\Delta_n$ are shown in the upper panel of Fig. \ref{fig-Delta-np}
for $Z=52$, 78 and 92 isotopes. The HF+BCS results
are compared with the experimental data and also the
phenomenological parameterization based on liquid drop model,
\begin{equation}
\bar{\Delta}=c/A^{\alpha}
\label{eq:gap-pheno}
\end{equation}
with $c=4.66(4.31)$ MeV for neutrons (protons)
and $\alpha=0.31$ which gives the rms residual of
0.25 MeV \cite{bertsch09}.
We can see clearly a better agreement of
IS+IV results with empirical data for all isotopes. In the lower panel of Fig. \ref{fig-Delta-np},
the OES differences $\Delta_n$ are shown for $N=76$, 102 and 112 isotones.
The results with IS+IV pairing
certainly improve systematically the agreement with empirical data, especially for $N=102$ isotones.
The large increase of the HF+BCS model results at $Z=81$ is an artifact due to the shell
closure at $Z=82$. It is interesting to notice that the liquid drop formula gives smooth
mass number dependence which reflects well that of very heavy isotones with $N=112$.
\begin{table*}
\caption{\label{table2}Average $\Delta^{(3)}$ for low isospin and high isospin nuclei
and its difference. See the text for details. }
\begin{ruledtabular}
\begin{tabular}{cccccc}
\multicolumn{3}{c}{Data set} & Low isospin & High isospin & Difference \\
\hline
Neutrons & $Z=52$ & Exp & 1.36 & 1.08 & -0.28 \\
& & IS & 1.52 & 1.41 & -0.11 \\
& & IS+IV & 1.40 & 1.19 & -0.21 \\
& $Z=78$ & Exp & 1.13 & 0.99 & -0.14 \\
& & IS & 0.96 & 1.16 & 0.20 \\
& & IS+IV & 0.87 & 0.91 & 0.04 \\
& $Z=92$ & Exp & 0.77 & 0.56 & -0.21 \\
& & IS & 0.90 & 0.80 & -0.10 \\
& & IS+IV & 0.70 & 0.55 & -0.15 \\
Protons & $N=76$ & Exp & 1.19 & 0.93 & -0.26 \\
& & IS & 1.13 & 0.87 & -0.26 \\
& & IS+IV & 1.13 & 0.98 & -0.15 \\
& $N=102$& Exp & 0.96 & 0.63 & -0.33 \\
& & IS & 0.79 & 0.39 & -0.40 \\
& & IS+IV & 0.92 & 0.59 & -0.33 \\
& $Z=112$& Exp & 0.87 & 0.66 & -0.21 \\
& & IS & 0.58 & 0.61 & 0.03 \\
& & IS+IV & 0.67 & 0.70 & 0.03 \\
\end{tabular}
\end{ruledtabular}
\end{table*}
\begin{figure*}[htb]
\begin{center}
\includegraphics[clip,scale=0.4,angle=-90]{Rn.ps}
\includegraphics[clip,scale=0.4,angle=-90]{Rp-m.ps}
\caption{(Color online) Neutron and proton radii of various isotopes and
isotones calculated
by means of the HF+BCS model with IS+IV interaction. The solid lines are empirical fits
used in Ref. \cite{Meng06}.
See the text for details. }
\label{RnRp}
\end{center}
\end{figure*}
\begin{figure*}[htb]
\begin{center}
\includegraphics[clip,scale=0.6]{rnrp-m.ps}
\caption{(Color online) Neutron skin neutron for isotopes and
isotones calculated
by means of the HF+BCS model with IS and IS+IV interactions. The thin lines indicate
the closed shells with N (or Z)=20, 28, 40, 50 ,64, 82 and 126.
See the text for details. }
\label{n-skin}
\end{center}
\end{figure*}
The average gaps $\Delta ^{(3)}$ are tabulated for high and low isospins in
Table \ref{table2}. Each isotope (isotone) in Fig. \ref{fig-Delta-np} is
divided into two subsets of almost equal numbers
of nuclei by a cut at some value of $I=(N-Z)/A$. Both the average proton and neutron
$\Delta ^{(3)}$ show smaller values for higher isospin so that the pairing
interaction is weaker for neutron-rich nuclei. The IS+IV interaction reproduces
properly the difference of the neutron $\Delta ^{(3)}$ between high and low isospin
nuclei. For proton $\Delta ^{(3)}$ also, the IS+IV pairing gives a good account of the
isospin effect than the IS pairing.
\subsection{Nuclear radii}
Nuclear radii provide basic and important information for various aspect of nuclear
structure problems. The proton radii, or equivalently the charge radii with the correction of
finite proton size, can be determined accurately by electron scattering and muon scattering
experiments. However it is difficult to determine the neutron radii of finite nuclei with the
same accuracy level as that of the proton radii while there were several experimental
attempts to determine the difference of the neutron to proton
radius ~\cite{Sn-rnrp-exp,Trz01,rnrp-exp}.
It should be noticed that the difference
of the neutron and proton radii, $\delta r_{np}=r_n-r_p$, is called the neutron skin. It is thought that $\delta r_{np}$ can give
important constrains on the effective interactions used in nuclear structure study \cite{Horwitz2001}.
The neutron and proton radii of various isotopes and isotones are calculated by using the HF+BCS
model with the two pairing interactions, IS and IS+IV.
The results of neutron radii are shown in the left panel of Fig. \ref{RnRp}.
Since we do not find any appreciable
differences between the two pairing interactions in the results,
the results of IS+IV interactions will be mainly discussed hereafter.
The results obtained by a
simple empirical formula $r_n=r_0 N^{1/3}$ with $r_0=1.139$ fm \cite{Meng06}
are also plotted in the figure.
In general, the simple formula for $r_n$ agrees well with the HF+BCS results. It is noticed that
the HF+BCS model gives larger neutron radii for nuclei with $N<40$ than the simple formula but
smaller for nuclei with $N>120$.
The proton radii for $N=20$, 28, 40, 50, 82 and 126 isotones are shown as a function of
proton number Z in the right panel. The simple Z$^{1/3}$ dependence is also plotted to follow the formula
$r_p=1.263$/Z$^{1/3}$. The simple formula in general gives a good account of the HF+BCS data
and could be
a good starting point for describing the isospin dependence of nuclear charge radii.
However we can see some deviation between the HF+BCS and the simple formula
especially heavy $N=50$ and $N=82$ isotones.
\begin{figure*}[htb]
\begin{center}
\includegraphics[clip,scale=0.5]{Sn-rnrp-m.ps}
\includegraphics[clip,scale=0.5]{c-rnrp-m.ps}
\caption{(Color online) {\it Left} - Neutron skin for Sn isotopes
obtained with the HF+BCS model with IS+IV interaction.
Experimental data are taken from Ref. \cite{Sn-rnrp-exp,Trz01}.
{\it Right} - Neutron skin as a function of isospin
parameter $I=(N-Z)/A$ calculated
by means of the HF+BCS model with IS+IV interaction.
Experimental data are taken from Ref. \cite{rnrp-exp}.
See the text for details. }
\label{Sn-rnrp}
\end{center}
\end{figure*}
The neutron skin $r_n-r_p$ calculated by HF+BCS model with the two pairing interactions
are shown in Fig. \ref{n-skin}. The neutron skin becomes as large as 0.4 fm
near the neutron drip line with $Z<28$. On the other hand, the neutron skin is at most 0.25 fm in
neutron-rich nuclei with $Z>50$. For proton-rich nuclei, the proton skin becomes 0.1 fm with $Z<56$
and smaller than 0.05 fm in heavier isotopes, larger than $Z=56$. The results of IS and IS+IV
pairings are shown in the left panel and right panel, respectively. In general, the two
pairing interactions give almost the same results as shown in Fig. \ref{n-skin}. However, it is noticed
that the IS+IV pairing gives somewhat smaller neutron skins than
the IS pairing in very neutron-rich nuclei such as $^{136}$Sn, $^{150}$Ba and $^{218}$Po.
The calculated values are
compared with empirical data of Sn isotopes obtained from studies of spin-dipole resonances ~\cite{Sn-rnrp-exp} and
antiprotonic atoms ~\cite{Trz01} in the left panel Fig. \ref{Sn-rnrp}.
The calculated values show reasonable agreement with the
empirical data within the experimental error. The isospin dependence of neutron skin is shown in
the right panel of Fig. \ref{Sn-rnrp}
together with empirical values obtained by antiprotonic atom experiments in a wide range of nuclei
from $^{40}$Ca to $^{238}$U. The slope of experimental data as a function of the isospin parameter
$I=(N-Z)/A$ is reproduced well by our calculations.
The neutron skin of $^{208}$Pb has been
discussed intensively in relation with neutron matter properties. The systematic studies of
scattering data yield the empirical value $r_n-r_p=0.17\pm 0.02$ fm which is close to another
empirical value $r_n-r_p=0.15\pm0.02 $ fm from the study of antiprotonic-atom systems.
The model independent determination of parity violation experiment at Jefferson Laboratory ~\cite{Horwitz2001}
has been proposed and performed recently to obtain the neutron skin of $^{208}$Pb. However
the statistics was poor and needs improvement by more data accumulation.
Our calculated value $r_n-r_p=0.157$ fm is close to the experimental values by the two systematic
studies.
\section{Summary and conclusions}
In summary, we studied the binding energies, separation energies
and OES by using HF+BCS model with SLy4
interactions together with the isospin dependence pairing (IS+IV
pairing)
and isoscalar (IS pairing) interactions. The calculations are performed with the EV8odd code for
even-even nuclei and also even-odd nuclei using the filling approximation.
For the
neutron pairing gaps, the IS+IV pairing strength
decreases gradually as a function
of the asymmetry parameter $(\rho_n(r)-\rho_p(r))/\rho(r)$. On the other hand,
the pairing strength for protons
increases for larger values of the asymmetry parameter
because of the isospin factor in Eq. \eqref{eq:g1t}.
The empirical isotope dependence of the neutron OES, $\Delta^{(3)}_n$, is well
reproduced by the present calculations with the isospin dependent pairing
compared with the IS pairing. We can also obtain a good agreement between
the experimental proton OES and the calculations with the
isospin dependent pairing for $N=50$ and $N=82$ isotones.
The neutron and proton radii were also studied by using the same HF+BCS model with the two
pairing interactions. The two pairing interactions give essentially the same results for the radii
except for a few very neutron-rich nuclei. We found systematically large neutron skins in very
neutron-rich nuclei with $|r_n-r_p|\sim0.4$ fm, while the proton skin is rather small even in nuclei
close to the proton drip line because of the Coulomb interaction. The calculated results of
neutron skin show reasonable agreement with the empirical data including the ($N-Z$) dependence of
the data.
We tested the IS+IV pairing for the Skyrme interaction SLy4
and found the results reproduce well
the systematical experimental data.
Thus, we confirm a clear manifestation of the isospin
dependence of the pairing interaction in the OES
in comparison with the experimental data
both for protons and neutrons.
\begin{acknowledgments}
This work was partially supported by the U.S. DOE
grants DE-FG02-08ER41533 and DE-FC02-07ER41457 (UNEDF, SciDAC-2),
the Research Corporation, and the JUSTIPEN/DOE grant DEFG02- 06ER41407, and
by the Japanese Ministry of Education, Culture, Sports,
Science and Technology by Grant-in-Aid for Scientific Research under
the Program number C(2) 20540277. Computations were carried out on
the XT4 Jaguar supercomputer at the Oak Ridge National Laboratory.
\end{acknowledgments}
|
\section{Introduction}\label{Intro}
We work over the field $\mathbb{C}$ of complex numbers.
A birational map $f: X \dashrightarrow X'$ of varieties
is a {\it pseudo-isomorphism} if it is an isomorphism outside
codimenion-two closed subsets of $X$ and $X'$. If we assume further $X = X'$, then $f$ is called a
{\it pseudo-automorphism}.
By the minimal model program (which we will not use at all), a variety of dimension $\ge 3$ may have more than two
minimal models, but all of them are pseudo-isomorphic to each other.
In dimension two, every pseudo-automorphism of a normal projective surface is an automorphism, and all the minimal models
of a given surface are isomorphic to each other.
The main result of the paper is the following:
\begin{theorem}\label{ThA}
Let $w = w_{p, q, r}$ be the Coxeter element $($unique up to conjugation$)$
of the Weyl group $W(T_{p, q, r})$ $($cf. Fig.~$1$ in $\ref{Weyl})$.
Suppose that $r \ge 3$ and $\frac{1}{p}+\frac{1}{q}+\frac{1}{r}<1$.
Then there exist a blowup $X = X_{p, q, r}$ of $(\mathbb{P}^{r-1})^{p-1} = \mathbb{P}^{r-1} \times \cdots \times \mathbb{P}^{r-1}$
at $q+r$ points $P_i$ lying on a multi-degree-$r$ cuspidal curve $C \subset (\mathbb{P}^{r-1})^{p-1}$
and a pseudo-automorphism $f_w : X \dashrightarrow X$
such that $(f_w)^* | H^2(X, \mathbb{Z})$ equals $w$. In particular, the first dynamical degree $d_1(f_w)$
of $f_w$ equals the spectral radius of $w$, hence the entropy $h(f_w)$ of $f_w$
satisfies $h(f_w) \ge \log d_1(f_w) > 0$ $($cf. \cite{DS}$)$.
\end{theorem}
Here $C$ is the cuspidal curve of arithmetic genus one embedded in $(\mathbb{P}^{r-1})^{p-1}$
by the product map $\Phi_{|D_1|} \times \cdots \times \Phi_{|D_{p-1}|}$
for some Cartier divisors $D_i$ of degree $r$ on $C$.
For instance, when $p = 2$, we can take $C = \{(1, z, z^2, ..., z^{r-2}, z^{r})\}$ in affine coordinates.
The construction of the pairs $(X_{2, q, r}, f_w)$, is given in \ref{constr}.
When $p = 2$ and $n = q+r$, we can take
$w = (12 \cdots n) r_{I,1}$, where the permutation is on the part $e_j$ of the
standard basis of the hyperbolic lattice
$\Lambda_n = h_1 \mathbb{Z} + \sum_{j=1}^{n} e_j \mathbb{Z}$
(naturally identified with $H^2(X, \mathbb{Z})$)
and $r_{I,1}$ is the reflection corresponding to the root $\alpha_{I, 1} = h_1 - \sum_{j=1}^r e_j$ (cf.~\ref{Weyl}).
As a consequence of Theorem \ref{ThA} we have:
\begin{corollary}\label{Cor1}
\begin{itemize}
\item [(1)]
When $\{p, q, r\} = \{2, 3, 7\}$ $($as unordered sets$)$ and $r \ge 3$,
$f_w$ is a pseudo-automorphism of the blowup of $(\mathbb{P}^{r-1})^{p-1}$
at $q+r$ points and $d_1(f_w) = 1.17628 \dots$ is the Lehmer number of
the Lehmer polynomial $x^{10} + x^9 - (x^7 + x^6 + x^5 + x^4 + x^3) + x + 1$.
\item [(2)]
When $\{p, q, r\} = \{2, 4, 5\}$ $($as unordered sets$)$ and $r \ge 3$,
$f_w$ is a pseudo-automorphism of the blowup of $(\mathbb{P}^{r-1})^{p-1}$
at $q+r$ points and $d_1(f_w) = 1.28064 \dots$ is
the largest root of the Salem polynomial $x^8 - x^5 - x^4 - x^3 + 1$.
\item [(3)]
When $\{p, q, r\} = \{3, 3, 4\}$ $($as unordered sets$)$,
$f_w$ is a pseudo-automorphism of the blowup of $(\mathbb{P}^{r-1})^{p-1}$
at $q+r$ points and $d_1(f_w) = 1.40127 \dots$ is
the largest root of the Salem polynomial $x^6 - x^4 - x^3 - x^2 + 1$.
\end{itemize}
\end{corollary}
The above three types of $T_{p, q, r}$ are the only $T$-shaped minimal hyperbolic Coxeter diagrams
(cf. \cite[Table 5]{Mc02}).
The three Salem numbers above are the smallest Salem numbers of
degrees $10$, $8$ and $6$, respectively. Especially, one can also realize
the Lehmer number as $d_1(f_w)$ of the pseudo-automorphism of $X$
which is a blowup of $\mathbb{P}^6$ at $10$ points.
\begin{remark}\label{rThA}
\begin{itemize}
\item[(1)]
$d_1(f_w) = 1.28064 \dots$ is the smallest known first dynamical degree $(> 1$)
of a pseudo-automorphism of a rational threefold which is not of product type.
\item[(2)]
In \cite{BK11}, the authors have constructed a pseudo-automorphism $f$ on the blowup of
$\mathbb{P}^3$ at $2$ points and $13$
curves with $d_1(f) = 1.28064 \dots$. $f$ is induced by a quadratic birational map on $\mathbb{P}^3$, while $f_w$ in Corollary \ref{Cor1}
with $(p, q, r) = (2, 5, 4)$
comes from a cubic map. Moreover, $f$ stabilizes a family of $K3$ surfaces, while $f_w$ in Corollary \ref{Cor1}
stabilizes the cuspidal curve $C$ and permutes
members $F_t'$ of the rational pencil $|{-}K_X/2|$ each of which is a strict transform
of an {\it irreducible} quadric hypersurface $F_t \subset \mathbb{P}^3$ with $F_t' \cap F_{t'}' = C$ ($t \ne t'$).
Every effective divisor $E$ with the class $[E]$ fixed by $f_w^*$ is a union of members in $|{-}K_X/2|$.
This $f_w$ stabilizes at least one member $F_0' \in |{-}K_X/2|$
but does not stabilize any cohomology class of a $K3$ surface on $X$ (if any).
\item[(3)]
When $(p, q, r) = (3, 4, 3)$, $X$ is the blowup of $\mathbb{P}^2 \times \mathbb{P}^2$ at $7$ points. $f_w$ permutes
members of the linear system $|{-}K_X/3|$ of dimension $\ge 2$. Every divisor $E$ with class $[E]$
fixed by $f_w^*$ satisfies $E \sim a(-K_X/3)$ ({\it linear equivalence}) for some $a \in \mathbb{Z}$.
\item[(4)]
When $(p, q, r) = (2, 7, 3)$, $f_w$ is an automorphism of the blow-up of the projective plane
at $10$ points. This automorphism coincides with the one constructed in \cite[Appendix]{BK} and \cite[Theorem 1.1]{Mc07}.
\item [(5)]
When $(p, q, r) = (2, 6, 4)$, $w$ (or its power) seems to have been geometrically realized early
by Coble and Cossec-Dolgachev (cf.~\cite[p.~39]{Do}).
\end{itemize}
\end{remark}
\par \vskip 1pc \noindent
{\bf Acknowledgement.}
The present work took place when the second author was visiting Bayreuth in November 2011
and in the realm of
the DFG Forschergruppe 790 �Classification of algebraic surfaces and
compact complex manifolds� and partly supported by an ARF of NUS.
We express our thanks to Professor
Fabrizio Catanese for his interest, warm encouragement and hospitality.
The second author would like to thank Max Planck Institute for the warm hospitality
and Professor C. McMullen
for the discussion on the Salem numbers.
\section{Preliminaries}\label{Pri}
\begin{setup}\label{Weyl}
{\rm Weyl groups and roots (cf. \cite{Hu})}
\end{setup}
Let $p \ge 2$, $q \ge 2$ and $r \ge 3$ be intergers.
Let $n := p + q + r - 2$.
We now define the {\it root system} $L_n$ of type
$T_{p, q, r}$.
Let
$$\Lambda = \Lambda_n = \mathbb{Z} h_1+ \cdots + \mathbb{Z} h_{p-1} + \mathbb{Z} e_1 + \cdots + \mathbb{Z} e_{q+r}$$
be the lattice of rank $n+1$ with basis
$$h_1, h_2, \dots, h_{p-1}, e_1, \dots , e_{q+r} .$$
Late on in Section \ref{p=2}, we treat the case $p = 2$ and set $e_0 = h_1$.
The following equations define an {\it inner product}
on $\Lambda$ (cf. \cite[\S 3]{Mu2}):
$$\begin{aligned}
h_i^2 &= h_i \cdot h_i = r-2 \,\, (1 \le i < p), \, \\
h_i \cdot h_j &= r-1 \,\, (i \ne j), \,\, h_i \cdot e_j = 0 , \\
e_i^2 &= e_i \cdot e_i = -1 \,\, (1 \le i \le q+r), \,
e_i \cdot e_j =0 \,\, (i \not= j).
\end{aligned}$$
Set
$$\kappa := r \, \sum_{i=1}^{p-1} h_i - ((p-1)(r-1) - 1) \, \sum_{j=1}^{q + r} e_j.$$
We will see that $\kappa$ corresponds to the anti-canonical divisor of some blowup $X$ of
$(\mathbb{P}^{r-1})^{p-1}$ at $q+r$ points, and $\Lambda_n$ is isomorphic to $H^2(X, \mathbb{Z})$.
The {\it root system} (of type $T_{p, q, r}$) is
$$
L_n := \kappa^{\perp} \cap \Lambda_n = \{\alpha \in \Lambda_n \, | \, \alpha \cdot \kappa = 0\} \, .
$$
The {\it simple roots}:
$$\begin{aligned}
\beta_1 &= -h_1 + h_2, \, \beta_2 = -h_2 + h_3, \, \dots, \beta_{p-2} = -h_{p-2} + h_{p-1}, \, \\
\alpha_0 &= h_1 - \sum_{i=1}^r e_i, \,\, \alpha_1 = e_1 - e_2, \,\, \alpha_2 = e_2 - e_3, \,\, \dots, \,\,
\alpha_{q+r-1} = e_{q+r-1} - e_{q+r}
\end{aligned}$$
form a basis of $L_n$.
The corresponding Dynkin diagram is shown in Figure 1.
\begin{center}
\setlength{\unitlength}{0.8cm}
\begin{picture}(8,3)
\thicklines
\put(-0.3,0.99){\makebox{$\dots$}}
\put(-1.2,1){\circle{0.2}}
\put(-1.3,1.5){\makebox{$\alpha_1$}}
\put(-1.1,1){\line(1,0){0.4}}
\put(-0.5,0.99){\makebox{$\dots$}}
\put(0.5,1){\line(1,0){0.4}}
\put(1,1){\circle{0.2}}
\put(1.1,1){\line(1,0){0.8}}
\put(2,1){\circle{0.2}}
\put(1.85,1.5){\makebox{$\alpha_r$}}
\put(2,0.1){\line(0,1){0.76}}
\put(2,0){\circle{0.2}}
\put(2.2,-0.1){\makebox{$\alpha_0$}}
\put(2,-0.1){\line(0,-1){0.8}}
\put(2,-1){\circle{0.2}}
\put(2.2,-1){\makebox{$\beta_1$}}
\put(2,-1.1){\line(0,-1){0.4}}
\put(1.94,-2){\makebox{$\vdots$}}
\put(2,-2.1){\line(0,-1){0.4}}
\put(2,-2.6){\circle{0.2}}
\put(2.2,-2.5){\makebox{$\beta_{p-2}$}}
\put(2.1,1){\line(1,0){0.8}}
\put(3,1){\circle{0.2}}
\put(3.1,1){\line(1,0){0.8}}
\put(4,1){\circle{0.2}}
\put(4.1,1){\line(1,0){0.4}}
\put(4.7,0.99){\makebox{$\dots$}}
\put(5.5,1){\line(1,0){0.4}}
\put(6,1){\circle{0.2}}
\put(5.7,1.5){\makebox{$\alpha_{q+r-1}$}}
\put(1,-3.5){\makebox{Figure 1. }}
\end{picture}
\end{center}
\vspace{3cm}
Any $\alpha \in L_n$ with $\alpha^2 = -2$ determines the {\it reflection}
$r_{\alpha}\in O(L_n)$ given by:
$$
r_{\alpha}(x) = x + (x \cdot \alpha) \alpha.
$$
For distinct $i, j \ge 1$,
$r_{e_i - e_j}$ (resp. $r_{h_i - h_j}$)
is the {\it transposition} interchanging the basis elements $e_i$ and $e_j$
(resp. $h_i$ and $h_j$) while fixing the other $e_k$'s and $h_{\ell}$'s.
For any $1 \le k < p$ and subset $I \subseteq \{1, 2, \dots, n\}$ with $|\, I\, | = r$,
we define the `root'
$$\alpha_{I, k} = h_k - \sum_{i \in I} e_i \in L_n$$
and the reflection (called {\it a Cremona involution}):
$$
r_{I, k} := r_{\alpha_{I, k}} \, .
$$
Its action on $\Lambda$ is given as follows:
$$\begin{aligned}
r_{I, k}(h_k) &= h_k + (h_k \cdot \alpha_{I, k}) \alpha_{I, k} = (r-1) h_k - (r-2) \sum_{i \in I} e_i, \\
r_{I, k}(h_i) &= h_i + (h_i \cdot \alpha_{I, k}) \alpha_{I, k} = (r-1)h_k + h_i - (r-1) \sum_{j \in I} e_j \,\,\, (i \ne k), \\
r_{I, k}(e_i) &= e_i + \alpha_I \,\,\, (i \in I), \\
r_{I, k}(e_j) &= e_j \,\,\, (j \not\in I).
\end{aligned}
$$
The {\it Weyl group}
$$W := W(p, q, r) = W(T_{p,q,r}) \subset O(L_n) \subset O(\Lambda_n)$$
is the subgroup of $O(L_n)$ generated by the reflections
$$
r_{\beta_i} \, (1 \le i \le p-2), \,\, r_{\alpha_j} \,\, (0 \le j < q+r) \, .
$$
Elements in the set below are called (real) {\it roots}
$$\Delta_n := \{w(\beta_i), \, w(\alpha_j) \, | \, w \in W, \, 1 \le i \le p-2, \, 0 \le j < q+r\} .$$
\begin{definition}\label{cox}
A Coxeter element $w$ of $W$ is the product $w = \prod_{i=1}^n r_{\gamma_i}$
where $\{\gamma_i\}_{i=1}^n = \{\beta_i\}_{i=1}^{p-2} \cup \{\alpha_j\}_{j=0}^{q+r-1}$ as sets.
When $p = 2$,
choose $(\gamma_1, \dots, \gamma_n) = (\alpha_1, \dots, \alpha_{q+r-1}, \alpha_0)$,
we get $w = (12 \dots n) r_{I, 1}$ with $I = \{1, 2, \dots, r\}$,
the product of a permutation (on $e_1, \dots, e_n$) and a Cremona involution.
\end{definition}
\begin{remark}
Coxeter elements are conjugate to each other, since the Dynkin diagram $T_{p, q, r}$ is a tree
(cf. \cite[3.16, 8.14]{Hu}).
\end{remark}
\begin{setup}\label{lead}
{\rm
Let $w\in W$ with {\it spectral radius} $\rho(w)>1$.
Then the root system $L_n$ is hyperbolic. Hence
$\rho(w)$ is a Salem number
and $\det(xI-w)=Q(x)\cdot R(x)$, where $Q(x)$ is a Salem polynomial (cf.~\cite[Proposition 7.1]{Mc02}).
We say that $\lambda\in \mathbb{C}$ is a {\it leading eigenvalue} if $Q(\lambda)=0$.
So $\rho(w)$ is a leading eigenvalue.
We say that $v \in L_n \otimes \mathbb{C}$ is a {\it leading eigenvector} if $w(v)=\lambda v$ with $\lambda$
a leading eigenvalue.
}
\end{setup}
\begin{proposition}\label{McTh}
Let $r \ge 3$.
Let $v \in L_n \otimes \mathbb{C} = (\Lambda \otimes \mathbb{C}) \cap \kappa^{\perp}$
be an eigenvector of some $w \in W$ with eigenvalue $\lambda$.
Then $0 \not\in \Delta_n \cdot v$, i.e. $(D^v, c^v) \in U_C$ in the sense of
Lemma $\ref{rhoint}$,
if either one of the following
two conditions is satisfied.
\begin{itemize}
\item [(1)]
$w$ is a Coxeter element and $v$ is a leading eigenvector.
\item[(2)]
$\lambda$ is not a root of unity; and $w$ has no periodic roots, i.e., no positive power of $w$
fixes a root in $\Delta_n$.
\end{itemize}
\end{proposition}
\begin{proof}
The results follow from the calculation in \cite[Theorems 2.6 and 2.7]{Mc07},
as our diagram is bipartite; see also \cite[Discussions before Theorem 1.3 and after Theorem 3.1]{Mc02}.
Indeed, in (1),
the root system $L_n$ is hyperbolic of signature $(1, n-1)$.
\end{proof}
\begin{remark}\label{pqr}
(1) happens exactly when $\frac{1}{p}+\frac{1}{q}+\frac{1}{r}<1$
(cf. \cite[Table 5]{Mc02}).
\end{remark}
\begin{setup}\label{marked curves}
{\rm Marked cuspidal curves}
\end{setup}
For simplicity, now (till the end of this section), we assume $p = 2$
and set $e_0 = h_1$.
Let $$C = \{YZ^2 = X^3\} \subset \mathbb{P}^2$$
be the plane {\it cuspidal curve} (of arithmetic genus $1$). Consider the subset
$$
\Lambda_C \, \subset \, \operatorname{Pic}^r(C) \times C^n, \,\, \text{or equivalently} \,\,
\Lambda_C \, \subset \, \operatorname{Pic}^r(C) \times (\operatorname{Pic}^1(C))^n \, , \, r\geq 3
$$
consisting of $(n+1)$-tuples $$(D; c) = (D; c_1, \dots, c_n)$$
with $c_i$ contained in the smooth locus $C \setminus \{(0, 1, 0)\}$ of $C$.
Given $(D; c) = (D; c_1, \dots, c_n) \in \Lambda_C$,
define a {\it marking} on $C$
$$\rho = \rho_{(D; c)} : \Lambda \to \operatorname{Pic}(C)$$
by setting
$$
\rho(e_0) = D, \,\,\rho(e_i) = [c_i].
$$
\begin{remark}\label{pair-mark}
The $(n+1)$-tuple $(D; c) = (D; c_1, \dots, c_n) \in \Lambda_C$ and the marking
$\rho = \rho_{(D; c)}$ on $C$ determine each other uniquely.
\end{remark}
As observed in \cite[Proposition 4.1, Theorem 4.3]{Mc07},
since $\operatorname{Aut}(C)$ acts transitively on
$$\operatorname{Pic}^0(C) \cong \mathbb{C}$$
and for $u = \sum_{i=0}^n u_i e_i$
$$\deg(\rho(u)) = \deg(u_0 D + \sum_{i=1}^n u_i c_i) = r u_0 + \sum_{i=1}^n u_i
= \frac{1}{r-2}(\kappa \cdot u),$$
we have:
\begin{lemma}\label{rho0}
$\rho$ is determined, up to isomorphism, by its restriction
$$\rho_0 : \operatorname{Ker}(\deg \circ \rho) = L_n \to \operatorname{Pic}^0(C) .$$
\end{lemma}
\par \noindent
Here two markings $\rho$ and $\rho'$ are {\it isomorphic} if there is an
$f\in \operatorname{Aut} (C)$ such that $f^*\circ \rho=\rho'$.
Set
$$U_C := \{(D; c_1, \dots, c_n) \, | \, \rho_{(D; c)}(\alpha) \not= 0, \, \forall \, \alpha \in \, \Delta_n\}.$$
As observed in \cite[Example 3]{Mu}, applying the defining condition of $U_C$ to the roots $\alpha = e_i - e_j$
and $\alpha_I$ with $|I| = r$, we have:
\begin{remark}\label{non-deg}
If $(D; c_1, \dots, c_n) \in U_C$, then
$c_i \ne c_j$ ($i \ne j$), and $\sum_{i \in I} c_i \not\in |D|$ ($\forall I$, $|\, I \, | =r$),
i.e., no $r$ points of $P_i := \Phi_{|D|}(c_i) \in \mathbb{P}^{r-1}$ are contained in a hyperplane of $\mathbb{P}^{r-1}$.
Here $ \Phi_{|D|}\colon C \to \mathbb{P}^{r-1}$ is the embedding determined by $D$ (cf. Lemma \ref{RR}).
\end{remark}
\begin{definition}\label{wpair}
For $w \in W$, via the marking $\rho = \rho_{(D; c)}$, $w$ acts on $\Lambda_C$ by:
$$w (D; c_1, \dots, c_n) = (w(D); w(c_1), \dots, w(c_n)) :=
((\rho \circ w)(e_0); (\rho \circ w)(c_1), \dots, (\rho \circ w)(c_n)).$$
\end{definition}
Thus $\rho_{w(D; c)} = \rho_{(D; c)} \circ w$,
so $W$ acts on $U_C$ because $w \Delta_n = \Delta_n$. Namely, we have:
\begin{lemma}
If $w \in W_n$ and $(D; c_1, \dots, c_n) \in U_C$ then $w (D; c_1, \dots, c_n) \in U_C$.
\end{lemma}
\begin{setup}\label{resp}
{\rm The correspondence between vectors of $\Lambda_n \otimes \mathbb{C}$ and markings on $C$}
\end{setup}
Let $v = \sum_{i=0}^{n} v_i e_i \in \Lambda \otimes \mathbb{C}$.
We will define an $(n+1)$-tuple $(D^v; c^v)$ in the following way.
Let $p(t) = (t, t^3, 1) \in C$ be a parametrization.
Define $t_i$ and $c_i^v \in C$ by
$$\begin{aligned}
r t_0 &= v \cdot e_0 = (r-2) v_0 , \\
t_i &= v \cdot e_i = -v_i \,\, (i \ge 1) , \\
c_i^v &= p(t_i - t_0), \\ D^v &= r \, [p(0)] .
\end{aligned}$$
Then $(D^v; c^v)$ determines a marking $\rho^v$ on $C$ by setting
$\rho(e_0) = D^v$, $\rho(e_i) = [c_i^v]$.
For the result below, we refer to \cite[Theorem 7.5]{Mc07} for a concise calculation.
\begin{lemma}\label{rhoint}
The restriction $\rho_0^v : L_n \to \operatorname{Pic}_0(C) \cong \mathbb{C}$ of $\rho^v$
satisfies:
$$\rho_0^v(u) = (u \cdot v) [p(1) - p(0)] .$$
Hence for a root $\alpha \in \Delta_n$,
we have $\rho^v(\alpha) = 0$ if and only if $\alpha \cdot v = 0$.
In particular, the $(n+1)$-tuple $(D^v; c^v) \in U_C$ if and only if
$0 \not\in \Delta_n \cdot v$.
\end{lemma}
\begin{remark}\label{D-v}
Conversely, for any $(n+1)$-tuple $(D; c)$, we can use the equations in \ref{resp}
to define a vector $v$
such that $(D; c) = (D^v, c^v)$.
\end{remark}
\begin{lemma}\label{w-v}
For any $w \in W$, we have $\rho^v \circ w^{-1} = \rho^{w(v)}$,
and $w^{-1}(D^v, c^v) = (D^{w(v)}; c^{w(v)})$.
\end{lemma}
\begin{proof}
The first part follows from Lemma \ref{rhoint} and Remark \ref{rho0},
since $w \in O(\Lambda_n)$.
The second follows from the first and Definition \ref{wpair}
(cf. Remark \ref{pair-mark}).
\end{proof}
\begin{lemma}\label{equiv} $($cf. \cite[Corollary 7.7]{Mc07}$)$
Let $u,v \in \Lambda \otimes \mathbb{C}$ with
$\Delta_n \cdot u \not\ni 0 \not\in \Delta_n \cdot v$.
Then
$$
u = a v + b \kappa \, \Longleftrightarrow \, (D^u; c^u) \cong (D^v; c^v)\, .
$$
\end{lemma}
\begin{proof}
The $(n+1)$-tuples are determined by their markings on $C$ or equivalently by their restrictions on $L_n$
(cf. Remarks \ref{pair-mark} and \ref{rho0}), while the latter is determined by the inner product
on $L_n = \Lambda \cap \kappa^{\perp}$ (cf. Lemma \ref{rhoint}). The lemma follows
since $\operatorname{Aut}(C)$ acts on $\operatorname{Pic}^0(C)$ by scalar multiplication.
\end{proof}
\section{Proof of Theorems when $p = 2$}\label{p=2}
We will prove the following more general theorem.
\begin{theorem}\label{ThB}
Let $w$ be an element in the Weyl group $W = W(p, q, r)$
of the root system of type $T_{p, q, r}$ with
$r \ge 3$.
Let $v \in L_n \otimes \mathbb{C} = (\Lambda \otimes \mathbb{C}) \cap \kappa^{\perp}$
be an eigenvector of some $w \in W_n$ with $w(v) = \lambda v$
such that $0 \not\in \Delta_n \cdot v$, i.e., $(D^v, c^v) \in U_C$
in the sense of Lemma $\ref{rhoint}$.
Then there exist a blowup $X = X_{p, q, r} \to (\mathbb{P}^{r-1})^{p-1} =
\mathbb{P}^{r-1} \times \cdots \times \mathbb{P}^{r-1}$
at $q+r$ points $P_i$ on the multi-degree-$r$ cuspidal curve $C \subset (\mathbb{P}^{r-1})^{p-1}$
and a pseudo-automorphism $f_w : X \dashrightarrow X$
such that $(f_w)^* | H^2(X, \mathbb{Z})$ equals $w$.
When $|\lambda| > 1$, $\lambda$ equals $|\lambda|$, the spectral radius $\rho(w)$ of $w$ and also
the first dynamical degree $d_1(f_w)$ of $f_w$.
\end{theorem}
We will frequently use the following result.
\begin{lemma}\label{RR}
Let $C$ be the cuspidal curve of arithmetic genus $1$
and $D$ a Cartier divisor of degree $r$.
\begin{itemize}
\item [(1)]
If $r = 1$, then there is a unique smooth point $P$ of $C$
such that $P \sim D$.
\item[(2)]
If $r = \deg(D) \ge 3$, then the complete linear system
$|D|$ is base point free and defines an
embedding $\Phi_{|D|} : C \to \mathbb{P}^{r-1}$.
\end{itemize}
\end{lemma}
\begin{proof}
By the Riemann-Roch theorem (true for all projective curves as in Hartshorne's book, Ch IV, Ex 1.9)) and Serre duality
for Cohen-Macaulay projective variety,
we have $h^0(C, D) = r$. The results follow.
Indeed, the second part of (1) is worked out in Hartshorne's book, Ch II, Example 6.11.4.
\end{proof}
We now prove Theorem \ref{ThB} when $p = 2$.
In the definition of the lattice $\Lambda_n$ and $L_n$, we set $p = 2$ and $e_0 = h_1$.
Let $w \in W$.
Let $(D; c) \in U_C$ be an $(n+1)$-tupe. Denote by $(D'; c') = w(D; c)$.
Let
$$\Phi_{|D|} : C \to \mathbb{P}^{r-1}$$
be the embedding given by the base-point free
complete linear system $|D|$. Set $P_i := \Phi_{|D|}(c_i)$.
Let
$$\pi_{(D; c)} : X = X_{(D; c)} \to \mathbb{P}^{r-1}$$ be the blowup of the $n$ points $P_i$
with $E_i = \pi_{(D; c)}^{-1}(P_i)$.
Similarly, we define $\Phi_{|D'|}$, $P_i'$, $\pi_{(D'; c')} : X' = X_{(D'; c')} \to \mathbb{P}^{r-1}$, $E_i'$.
The result below should be well known
but we work it out since we need to extend it
to the case $p > 2$ in Section \ref{p>2}.
Our statement also incorporates the marking on the curve $C$ embedded in $\mathbb{P}^{r-1}$.
\begin{proposition}\label{Cal}
Let $p = 2$.
Let $w\in W$ and $(D;c)\in U_C$. Define $(D';c'):=w(D;c)$. Consider the blowup
$$
\pi\colon X_{(D;c)}\to \mathbb{P}_{(D;c)} = \mathbb{P}^{r-1}\, , \quad \pi' \colon X_{(D';c')}\to \mathbb{P}_{(D';c')} = \mathbb{P}^{r-1}
$$
at the points $P_i=\Phi_{|D|}(c_i)$ $($resp. $P_i'=\Phi_{|D'|}(c_i'))$.
Then there exists a pseudo-isomorphism $f_w\colon X_{(D;c)} \dashrightarrow X_{(D';c')}$ such that the following diagram
commutes:
$$
\xymatrix{
H^2(X_{(D';c')},\mathbb{Z})\ar[r]^{{f_{w}}^*} & H^2(X_{(D;c)},\mathbb{Z}) \\
\Lambda \ar[u]^{\phi'} \ar[r]^w & \Lambda \ar[u]_{\phi}
}
$$
where $\phi$ $($resp. $\phi')$ is defined by $\phi(e_0)=\pi^*[H]$, $\phi(e_i)=[E_i]$ $(1\leq i \leq n)$
$($resp. $\phi'(e_0)=(\pi')^*[H']$, $\phi'(e_i)=[E_i']$ $(1\leq i \leq n))$. As usual $H$ $($resp. $H')$
is the hyperplane of $\mathbb{P}_{(D;c)}$ $($resp. $\mathbb{P}_{(D';c')})$,
and $E_i$ $($resp. $E_i')$ is the exceptional divisor over $P_i$ $($resp. $P_i')$.
\end{proposition}
\begin{proof}
Since $W$ is generated by the transpositions $r_{e_i - e_j}$ and the Cremona involution
$r_{I, 1}$, we need to prove the result only when $w$ is one of them.
Our proof is top down: first construct a pseudo-isomorphism $X = X_{(D; c)} \dashrightarrow X'$ and then show that
$X'$ equals $X_{(D'; c')}$, the blowup of $n$ points $\Phi_{|D'|}(c_i')$.
Consider first the case $w = r_{I, 1}$ with $I = \{1,2, \dots, r\}$.
Let $X_P = X_{P_1, \dots, P_r} \to \mathbb{P}^{r-1}$ be the blowup of the $r$ points $P_i$ ($1 \le i \le r$).
Since these $r$ points $P_i$ span the whole space (cf. Remark \ref{non-deg}), we can take
the standard Cremona involution $\Psi_P = \Psi_{P_1, \dots, P_r} : X_P \dashrightarrow X_P$.
$\Psi_P$ is given by the linear system $|\mathcal{O}_{\mathbb{P}^{r-1}}(r-1) - (r-2)\sum_{i=1}^r P_i|$.
This linear system has a basis: $\sum_{j\not=i}H_j$, $i\in \{ 1, \dots , r\}$, where
$H_i$ is the hyperplane passing through $r-1$ points $\{P_1, \dots, P_{i-1}, P_{i+1}, \dots, P_r\}$.
The base locus of the linear system (the place where $\Psi_P$ is not defined)
is the union of $H_i \cap H_j$ ($1 \le i < j \le r$).
If one uses new coordinate system so that $P_1 = [1: 0: \dots : 0]$, $P_2 = [0: 1: \dots : 0], \, \dots,$
$P_r = [0: \dots : 0 : 1]$, then $\Psi_P$ is given by
$$\Psi_P : [X_1: \dots : X_r] \,\, \to \,\, [\frac{1}{X_1}: \dots : \frac{1}{X_r}].$$
Let $E_i \subset X_P$ be the inverse image of $P_i$ and $E_0 \subset X_P$ the total transform of a hyperplane of $\mathbb{P}^{r-1}$.
Then it is known that $\Psi_P$ lifts to an involutive pseudo-automorphism $\widetilde{\Psi_P} : X_P \to X_P$
exchanging $E_i$ with the proper transform $H_i' \subset X_P$ of $H_i$ (cf. \cite{DO}).
This means that
$$\widetilde{\Psi_P}^*E_i = H_i' \sim E_0 - \sum_{j \ne i} E_j.$$
Denote by $e_i = [E_i] \in H^2(X_P, \mathbb{Z})$.
Then
$$\widetilde{\Psi_P}^*e_i = [\widetilde{\Psi_P}^*E_i] = [E_0 - \sum_{j \ne i} E_j] =
e_0 - \sum_{j \ne i} e_j = e_i + (e_0 - \sum_{i=1}^r e_i) = w(e_i).$$
By the definition of the Cremona involution in terms of the linear system,
$$\widetilde{\Psi_P}^* E_0 = (r-1)E_0 - (r-2)\sum_{i=1}^r E_i$$
and hence
$\widetilde{\Psi_P}^*e_0 = w(e_0)$.
The blowup $X_P \to \mathbb{P}^{r-1}$ is centred at $r$ smooth points $P_i = \Phi_{|D|}(c_i)$,
and hence gives an isomorphism between the proper transform $C_{X} \subset X_P$ of $C$ and $C$.
Since $C = \Phi_{|D|}(C) \subset \mathbb{P}^{r-1}$ is a non-degenerate curve,
it is not contained in any hyperplane $H_i$.
Hence $C_X$ is not contained in $H_i'$.
Now $\deg(H_i' | C_{X}) = \deg(E_0 - \sum_{j \ne i} E_j) | C_{X}) = \deg(\mathcal{O}_{\mathbb{P}^{r-1}}(1) | C) - (r-1) = 1$
since $C = \Phi_{|D|}(C)$ is a curve of degree $r$ in $\mathbb{P}^{r-1}$.
Thus $C_X$ meets $H_i'$ only at one point and transversally.
Since the Cremona involution $\Psi_P : \mathbb{P}^{r-1} \dasharrow \mathbb{P}^{r-1}$
blows up $r$ smooth points $P_i$ on $C$ and collapses $H_i'$ to a point called $P_i'$ in the codomain $\mathbb{P}^{r-1}$,
it maps $C \subset \mathbb{P}^{r-1}$ isomorphically to a curve $C'$ in the codomain $\mathbb{P}^{r-1}$.
As sets, we have $\{P_i'\} = \{P_i\}$. This $C'$ is also the isomorphic image of $C_X \subset X_P$
via the map $X_P \overset{\widetilde{\Psi_P}}{\dashrightarrow} X_P \to \mathbb{P}^{r-1}$. This isomorphism of curves
factors as $C_X \overset{\widetilde{\Psi_P}}{\to} C'_X \to C'$.
Let us calculate the very ample divisor $D' = \mathcal{O}_{\mathbb{P}^{r-1}}(1) | C'$ giving rise to the
embedding $C' \subset \mathbb{P}^{r-1}$. By the above identification $C_X = C'_X = C'$ and further the identification
$C_X = \Phi_{|D|}(C) = C$, we have
$$D' = E_0 | C'_X = \widetilde{\Psi_P}^* E_0 | C_X =
((r-1)E_0 - (r-2)\sum_{i=1}^r E_i) | C_X = (r-1)D - (r-2) \sum_{i=1}^r c_i = w(D)$$
(cf. Definition \ref{wpair}).
Let $c_i' \in C'$ be the preimage of the point $P_i' \in \mathbb{P}^{r-1}$ via the embedding $\Phi_{|D'|} : C' \to \mathbb{P}^{r-1}$.
Under the same identification $C = \Phi_{|D|}(C) = C_X = C'_X = C'$,
we have (cf. Lemma \ref{RR}):
$$C = C' \ni c_i' = P_i' = H_i' \, | \, C_X \sim (E_0 - \sum_{j \ne i} E_j) | C_X =
D - \sum_{j \ne i}c_j = w(c_i).$$
For $r+1 \le j \le n$, the point
$P_j$ is not contained in the indeterminacy set: the union of $H_i \cap H_j$,
otherwise, the $r$ points $P_1, \dots, P_{i-1}, P_j, P_{i+1}, \dots, P_r$ are contained in the hyperplane
$H_i$, contradicting Remark \ref{non-deg}.
Let $Q_j$ ($r+1 \le j \le n$) be the $\Psi_P$-image of $P_j$.
For $1 \le i \le r$, set $Q_i = P_i$.
Let $\pi_{(D; c)} : X = X_{(D; c)} \to \mathbb{P}^{r-1}$ be the blowup of the $n$ points $P_i$,
$E_0 \subset X$ the pullback of the hyperplane of $\mathbb{P}^{r-1}$,
$E_i = \pi_{(D; c)}^{-1}(P_i)$ ($i \ge 1$), and $e_i$ ($i \ge 0$) the cohomology class of $E_i$ in $H^2(X, \mathbb{Z})$.
Let $\pi': X' \to \mathbb{P}^{r-1}$ be the blowup of the $n$ points $Q_i$,
$E_0' \subset X'$ the pullback of the hyperplane of $\mathbb{P}^{r-1}$,
$E_i' = (\pi')^{-1}(Q_i)$ ($i \ge 1$), and $e_i'$ ($i \ge 0$) the cohomology class of $E_i'$ in $H^2(X', \mathbb{Z})$.
Then $\widetilde{\Psi_P}$ lifts to a pseudo-isomorphism $f_w : X \to X'$.
Identify $H^2(X_P, \mathbb{Z})$ with its embedded image (via pullback) in $H^2(X, \mathbb{Z})$.
By the calculation above and the construction, we have $f_w^* e_i' = w(e_i)$ for all $i \le r$
and $f_w^*(E_j') = E_j$ ($j > r$) (so $f_w^*(e_j') = e_j = w(e_j)$), if we identify $H^2(X', \mathbb{Z}) = H^2(X, \mathbb{Z})$
by letting $e_i = e_i'$ ($i \ge 0$); thus $f_w^* = w$.
By the argument above, if we set $(D'; c') = w(D; c)$, then the above $\pi' : X' \to \mathbb{P}^{r-1}$ is just the blowup
of $n$ points $P_i' = \Phi_{|D'|}(c_i')$ on the curve $C' = \Phi_{|D'|}(C') \subset \mathbb{P}^{r-1}$,
i.e., it is $\pi_{(D'; c')}$.
This proves Propositoin \ref{Cal} when $w$ is a Cremona involution.
\par \vskip 1pc
Next we consider the case where $w = r_{e_a - e_b}$ is a transposition of the basis elements $e_a$ and $e_b$
and fixing the others.
Take an automorphism $\sigma$ of $\mathbb{P}^{r-1}$ interchanging two points $P_a$ and $P_b$.
Let $C' = \sigma(C) \subset \sigma(\mathbb{P}^{r-1}) = \mathbb{P}^{r-1}$.
Set $P_a' = P_a$, $P_b' = P_b$ and $P_j' = \sigma(P_j)$ ($j \ne a, b$).
Let $X' \to \mathbb{P}^{r-1}$ be the blowup of the $n$ points $P_i'$
with $E_i'$ the inverse of $P_i'$.
Then $\sigma$ lifts to an isomorphism $f_w: X \to X'$.
We see that $f_w^* = w$ if we identify $H^2(X', \mathbb{Z}) = H^2(X, \mathbb{Z})$ by letting $[E_i'] = e_i = [E_i]$ as above.
Define $(D'; c')$ so that $D' = D$, $c_a' = c_a, c_b' = c_b$ and $c_j' = \sigma(c_j)$ ($j \ne a, b$).
Using the identification $C = C_X = C_X' = C'$ as above, we obtain
$(D'; c') = w(D; c)$. This implies Proposition \ref{Cal} as in the previous case.
\end{proof}
\begin{setup}
{\bf Proof of Theorem \ref{ThB} when $p = 2$}
\end{setup}
Given $v$ as in Theorem \ref{ThB}, we define $(D^v; c^v)$ as in \ref{resp} (cf.~Lemma \ref{rhoint}).
Set $(D; c) = (D^v; c^v)$.
Then we get the pseudo-isomorphism $f_w : X = X_{(D; c)} \to X' = X_{(D'; c')}$ as in Proposition \ref{Cal}
with $(D'; c') = w(D; c)$ and $f_w^* = w$ on $H^2(X', \mathbb{Z})$ identified with $H^2(X, \mathbb{Z})$
by letting $[E_i'] = [E_i]$ and $[(\pi')^*H'] = [\pi^*H]$.
By Lemmas \ref{w-v} and \ref{equiv},
$$(D'; c') = w(D^v; c^v) = (D^{w^{-1}(v)}; c^{w^{-1}(v)}) = (D^{\lambda^{-1}v}; c^{\lambda^{-1}v})
= (D^v; c^v) = (D; c)$$
(up to the action of $\operatorname{Aut}(C)$).
Thus we get an isomorphism between $\pi_{(D'; c')} : X_{(D'; c')} \to \mathbb{P}^{r-1}$ in Proposition \ref{Cal}
and $\pi_{(D; c)} : X = X_{(D; c)} \to \mathbb{P}^{r-1}$ so that $f_w$ is a pseudo-automorphism.
This proves Theorem \ref{ThB}.
Indeed, for the final part (when $|\lambda| > 1$),
the Coxeter system is hyperbolic, so $\lambda$ is the largest root of a Salem polynomial
and hence also the spectral radius $\rho(w)$ of $w$ (cf. \cite[Proposition 7.1]{Mc02}). Thus
$d_1(f_w) = \rho(f_w^* | H^2(X, \mathbb{Z})) = \rho(w) = \lambda$ by the definition of $d_1(f_w)$ (cf. \cite{DS}).
\begin{setup}
{\bf Proof of Theorem \ref{ThA} when $p = 2$}
\end{setup}
Theorem \ref{ThA} follows from Proposition \ref{McTh}, Theorem \ref{ThB} and its proof,
by taking $\lambda$ in Proposition \ref{McTh} to be the spectral radius of $w$;
see also \ref{lead} and Remark \ref{pqr}
(cf.~\cite{DS} for the relation between $d_1(f_w)$ and $h(f_w)$).
\begin{setup}\label{constr}
{\rm Concrete construction of $f_w$ on $X_{2, q, r}$ as in Theorem \ref{ThA}}
\end{setup}
We first construct a pseudo-automorphism $f$ such that $f_* = w$ where $w = (12 \cdots n) r_{I, 1}$
is a Coxeter element of the root system $L_n$ of type $T_{2, n-r, r}$ (cf. Definition \ref{cox}).
Then $f_w = f^{-1}$ meets the requirement. To do so,
take an eigenvector $v$ of $w$ such that $w(v) = \lambda v$ and $\lambda$
is the spectral radius of $w \in O(L_n)$ (which turns out
to be $d_1(f_w)$, since $f_{w}^* = w$).
Define the $(n+1)$-tuple $(D; c) = (D^v; c^v)$ as in \ref{resp}.
Let $P_i = \Phi_{|D|}(c_i) \in \Phi_{|D|}(C) \subset \mathbb{P}^{r-1}$.
Choose new coordinate system of $\mathbb{P}^{r-1}$ such that $P_1 = [1 : 0 : \dots : 0], \dots, P_r = [0 : \dots : 0 : 1]$.
Write $P_{r+1} = [a_1 : \dots : a_r]$.
Let
$$\gamma : \mathbb{P}^{r-1} \to \mathbb{P}^{r-1}, \,\, [X_1 : \dots : X_r] \mapsto [\frac{1}{X_1} : \dots : \frac{1}{X_r}]$$
be the standard Cremona involution.
Let $g$ be a projective
automorphism of $\mathbb{P}^{r-1}$
taking $P_i$ to $P_{i+1}$ ($1 \le i \le r+1$).
Let $\pi = \pi_{(D; c)} : X = X_{(D; c)} \to \mathbb{P}^{r-1}$ be the blowup at the $n$ points $P_i$
and let $E_i = \pi^{-1}(P_i)$ and $E_0 \subset X$ the total transform of a hyperplane of $\mathbb{P}^{r-1}$.
Then by the construction and as in the proof of Theorem \ref{ThB} or Proposition \ref{Cal} (twisted by the permutation $g$),
$g \circ \gamma$ lifts to a pseudo-isomorphism $f : X \to X'$ with
$f_* = g_* \circ \gamma_* = w$ on $H^2(X, \mathbb{Z})$ identified with $H^2(X', \mathbb{Z})$ by letting $[E_i] = e_i = [E_i']$,
where
$$X' = X_{(D'; c')} = X_{w(D^v; c^v)} = X_{(D^{w^{-1}(v)}; c^{w^{-1}(v)})} = X_{(D^{\lambda^{-1}v}; c^{\lambda^{-1}v})}
= X_{(D^v; c^v)} = X$$
(up to isomorphism) and is the blowup of $n$ points $P_i' = \Phi_{|D'|}(c_i') \in \mathbb{P}^{r-1}$.
Now the identification $(D'; c') = w(D; c)$ with $(D; c)$ and the fact that
$w(c_n) = c_1$ forces $(g \circ \gamma)(P_n) = P_{1}$.
Conversely, if we can find $g$ as above
such that $(g \circ \gamma)(P_n) = P_{1}$,
then we can forget about the eigenvector $v$ or so, and straightaway say that $(g \circ \gamma)^{-1}$ lifts to
a pseudo-automorphism $f_w$ on the blowup $X \to \mathbb{P}^{r-1}$ at the $n$ points $P_i$
which satisfies the conclusion of Theorem \ref{ThA}.
\begin{setup}
{\rm Proof of Remark \ref{rThA} (2) (3)}
\end{setup}
We show (2), since (3) is similar.
Let $C_X\subset X$ be the proper transform of $C_D:=\Phi_{|D|}(C)\subset \mathbb{P}^3$ in Corollary \ref{Cor1}
with $(p, q, r) = (2, 5, 4)$.
By the proof, $f_w(C_X) = C_X$ holds in Theorem \ref{ThA} for any $(p, q, r)$.
Since $C_D$ has arithmetic genus $1$, it is contained in a linear system
$|\mathcal{I}(2)|$ of quadrics of dimension $\geq 1$.
This follows from the long cohomology sequence associated to
$$
0\to \mathcal{I}(2) \to \mathcal{O}_{\mathbb{P}^3}(2) \to \mathcal{O}_{C_D}(2)\to 0\, .
$$
Let $\pi : X \to \mathbb{P}^3$ be the blowup at the $9$ points $P_i$ as in Corollary \ref{Cor1} (2)
with $E_i = \pi^{-1}(P_i)$ and $E_0 \subset X$ the total transform of a hyperplane of $\mathbb{P}^3$.
For $F\in |\mathcal{I}(2)|$, the proper transform $F'$ of $F$
satisfies $F' \sim 2E_0 - \sum_{i=1}^9 E_i$ , so $-K_X = -(\pi^*K_{\mathbb{P}^3} + 2\sum_{i=1}^9 E_i) \sim 2F'$.
Since $-K_X$ is preserved by $f_w$, we have $2(f_w^*F' - F') \sim 0$, so
$f_w^*F' - F' \sim 0$, because the rational manifold $X$ is simply connected and hence
cohomologous divisors are just linear equivalent divisors.
If $E$ is a divisor whose class $[E]$ is fixed by $f_w^*$ (e.g., $E = a F'$), then
either $\dim |E| \le 0$ or $|E|$ is composed of a pencil, otherwise, $f_w$
would descend to a surface or threefold automorphism of null entropy
via a fibration with general fibre a curve or a point, contradicting the positivity of the entropy of $f_w$.
In particular, $|aF'|$ ($a > 0$) is composed of a pencil (necessarily parametrized by a curve $B \cong \mathbb{P}^1$
because the irregularity $q(B) \le q(X) = 0$) stabilized by $f_w$.
The induced action of $f_w$ on $B \cong \mathbb{P}^1$ has at least one fixed point, so
at least one $F'_0 \in |F'|$ is $f_w$-stable.
$F'$ or equivalent $F = \pi(F')$ is irreducible, otherwise,
$F = L_1 \cup L_2$ with two hyperplanes $L_i$. Since all $P_i \in C_D$ belong to $F$,
we may assume that $L_1$ contains $5$ points $P_1, \dots, P_5$. This contradicts Remark \ref{non-deg}
(cf. Proposition \ref{McTh}). Also, for two distinct such $F$, say $F_i$,
the intersection $F_1 \cap F_2$ includes $C_D$ and hence equals $C_D$ by comparing the degree.
Since the characteristic polynomial of $f_w^* | H^2(X, \mathbb{Z})$ is of the form $\phi_8(x) (x+1)(x-1)$
(cf. \cite[Table 5]{Mc02}) where $x-1$ corresponds to the $f_w$-invariant class $\kappa = [-K_X] = 2[F']$,
if $E$ is an integral divisor with $f_w^*[E] = [E]$ then $E \sim_{\mathbb{Q}} (a/b) F'$ for some coprime integers $a, b$.
Since $[F'] \cdot [F'] = (\kappa)^2/4 = -1$, we get $b = \pm 1$. In particular, every effective (resp. irreducible)
divisor $E$
with $[E]$ fixed by $f_w^*$ is a member of the pencil $|aF'|$ and hence equal to a union of $F_i' \in |F'|$
(resp. equal to some $F_1' \in |F'|$) by the Stein factorization.
\section{Proof of Theorems for all $p \ge 2$}\label{p>2}
We now prove Theorem \ref{ThB} for $p \ge 3$.
Let $p(t) = (t, t^3) \in C$ be the parametrization of the
cuspidal curve $C$ of arithemetic genus $1$.
Let $v = \sum_{i=1}^{p-1} \xi_i h_i + \sum_{j=1}^{q+r} \eta_j e_j \in L_n \otimes \mathbb{C}$ as in Theorem \ref{ThB}.
Define $t_j$, $c^v_j$ and $D^v_i$ ($1 \le i < p$), by
\begin{eqnarray}\label{1}
\begin{aligned}
r \, t_0 &= v \cdot h_1 = (r-2) \xi_1 + (r-1) \sum_{i =2}^{p-1} \xi_{i} , \\
t_j &= v \cdot e_j = -\eta_j \,\, (1 \le j \le q+r) , \\
c^v_j &= p(t_j - t_0) \in C, \,\, D^v_i = [rp(0) + \, p(\xi_1) -p( \xi_i)] \in \operatorname{Pic}^{r}(C) .
\end{aligned}
\end{eqnarray}
Then we get a $(p+q+r-1)$-{\it tuple}
$$(D^v; c^v) := (D^v_1, \dots , D^v_{p-1}; \, c^v_1, \dots , c^v_{q+r}) \, \in \,
(\operatorname{Pic}^r(C))^{p-1} \, \times \, C^{q+r} .$$
A {\it marking} of $C$ is a group homomorphism
$$
\rho \colon \Lambda = \mathbb{Z} h_1 +\cdots +\mathbb{Z} h_{p-1}+\mathbb{Z} e_1 +\dots + \mathbb{Z} e_{q+r} \, \to \, \operatorname{Pic}(C)
$$
such that $\rho (h_i)\in \operatorname{Pic}^r (C)$ and $\rho(e_j) = [p_{j}]$, where $p_j\in C$ is a smooth point.
There is an obvious bijection between markings $\rho$ and $(p+q+r-1)-tuples$
$$
(D; c):=(D_1, \dots , D_{p-1}; c_1, \dots , c_{q+r})\, \in \, (\operatorname{Pic}^r(C))^{p-1} \, \times \, C^{q+r}\, ,
$$
where $c_j$ are smooth points: $\rho(h_i) = D_i$, $\rho(e_j) = [c_j]$.
Denote by $\Lambda_C$ the set of such tuples.
Denote by $\rho_{(D;c)}$ the marking associated to the $(p+q+r-1)-tuple$ $(D;c)$.
Using \eqref{1} we associate a marking $\rho^v := \rho_{(D^v; c^v)}$ to any $v \in L_n \otimes \mathbb{C}$.
A simple calculation shows:
$$u \cdot \kappa = ((r-1)(p-1) - 1) \deg (\rho(u)) .$$
{\it We also get Lemma \ref{rhoint}.}
Indeed, it is enough to check the formula there for the elements of the basis
$\beta_i$ ($1 \le i \le p-2$), $\alpha_j$ ($0 \le j < q+r$) defined in \ref{Weyl}.
Thus $0 \not\in \Delta_n \cdot v$ if and only if $(D^v; c^v) \in U_C$, where, as before,
$$U_C = \{(D; c) \in \Lambda_C \, | \, \rho_{(D; c)}(\alpha) \not= 0,
\, \forall \, \alpha \in \, \Delta_n\}.$$
{\it Similarly, \ref{D-v} $\sim$ \ref{equiv} are also true.}
Using the marking $\rho_{(D; c)}$, the Weyl group $W$ acts naturally on the set of $(p+q+r-1)$-tuples
$(D; c)$ as in Definition \ref{wpair}.
Let $w \in W$.
Let $(D; c) \in U_C$. Denote by $(D'; c') = w(D; c)$.
Consider the embedding:
$$\Phi_{(D; c)} : C \to (\mathbb{P}^{r-1})^{p-1}, \,\,\,
(x \mapsto (\Phi_{|D_1|}(x), \dots, \Phi_{|D_{p-1}|}(x))) .$$
Set $P_j = (\Phi_{|D_1|}(c_j), \dots, \Phi_{|D_{p-1}|}(c_j))$.
Let
$$\pi_{(D; c)} : X = X_{(D; c)} \to (\mathbb{P}^{r-1})^{p-1}$$ be the blowup at the $q+r$ points $P_j$
with $E_j = \pi_{(D; c)}^{-1}(P_j)$.
Similarly, we define $\Phi_{(D'; c')}$, $P_j'$, $\pi_{(D'; c')} : X' = X_{(D'; c')} \to (\mathbb{P}^{r-1})^{p-1}$, $E_j'$.
For the result below, see \cite[Theorem 1]{Mu2}. Our statement also incorporates the marking on the curve $C$
embedded in $(\mathbb{P}^{r-1})^{p-1}$.
\begin{proposition}\label{Cal2}
Suppose that $w \in W$ and $(D; c) \in U_C$. Then there is a pseudo-isomorphism $f_w : X \to X' = X_{(D', c')}$
such that $f_w^* = w$ if we identify $H^2(X, \mathbb{Z}) = \sum_{i=1}^{p-1} \mathbb{Z} h_i + \sum_{j=1}^{q+r} \mathbb{Z} e_j = H^2(X', \mathbb{Z})$
by letting $[E_j] = e_j = [E_j']$ $(j \ge 1)$ and
$[\pi_{(D; c)}^*\mathcal{O}_{\mathbb{P}_i^{r-1}}(1)] = h_i = [\pi_{(D'; c')}^*\mathcal{O}_{\mathbb{P}_i^{r-1}}(1)]$
where $\mathbb{P}_i^{r-1}$ is the $i$-th factor of the product $(\mathbb{P}^{r-1})^{p-1}$.
\end{proposition}
\begin{proof}
The proof is similar to Proposition \ref{Cal}.
Since the Weyl group is generated by the reflections
$r_{h_i - h_j}$ (resp. $r_{e_i - e_j}$) corresponding to the exchange of the factors $\mathbb{P}_i^{r-1}$ and $\mathbb{P}_j^{r-1}$
(resp. $P_i$ and $P_j$ of the blowup),
and the Cremona involution $r_{\alpha_0}$ with $\alpha_0 = h_1 - \sum_{i=1}^r e_i$, we have only to
consider the case $w = r_{\alpha_0}$.
This $w$ is realized by the lifting $f_w : X \to X'$ of the
following standard (geometric) Cremona involution (cf. \cite[Lemma in \S 3]{Mu2}):
$$\begin{aligned}
\Psi \, : \, & \, (\mathbb{P}^{r-1})^{p-1} \, \to \, (\mathbb{P}^{r-1})^{p-1}, \\
& \, ([X_1 : \dots : X_r], [Y_1 : \dots : Y_r], \dots, [Z_1 : \dots : Z_r]) \, \mapsto \, \\
& \, ([\frac{1}{X_1} : \dots : \frac{1}{X_r}],
[\frac{Y_1}{X_1} : \dots : \frac{Y_r}{X_r}], \dots, [\frac{Z_1}{X_1} : \dots : \frac{Z_r}{X_r}]) .
\end{aligned}$$
Here, with new coordinates, we may assume that $P_1, \dots, P_r$ are images of the
standard vertices $[1 : 0 : \dots : 0], \dots, [0 : \dots : 0 : 1]$ in $\mathbb{P}^{r-1}$
via the diagonal embedding $\mathbb{P}^{r-1} \to (\mathbb{P}^{r-1})^{p-1}$ ($P \mapsto (P, \dots, P)$),
and
$X \to (\mathbb{P}^{r-1})^{p-1}$ is the blowup of $q+r$ points $P_i$
and $X' \to (\mathbb{P}^{r-1})^{p-1}$ is the blowup of $Q_1 := P_1, \dots, Q_r := P_r$ and $Q_j := \Psi(P_j)$ ($r < j \le q+r$).
By the form of the map,
$$f_w^* h_1 = (r-1) h_1 - (r-2) \sum_{i=1}^{r} [E_i] = w(h_1)$$
if we identify $H^2(X, \mathbb{Z}) = H^2(X', \mathbb{Z})$ (here and below) by letting $h_i = h_i', [E_j] = e_j = [E_j']$.
Here and below $E_i \subset X$ (resp. $E_i' \subset X'$) is the inverse of $P_i$ (resp. $Q_i$),
$h_i$ (resp. $h_i'$) is the (cohomology class of) total transform of the hyperplane
$\mathcal{O}_{\mathbb{P}_i^{r-1}}(1)$ of the $i$-th factor of the domain
(resp. codomain) of $\Psi$.
From the form of $\Psi$, we have also
$$f_w^* h_i = (r-1) h_1 - (r-2+1) \sum_{i=1}^{r} [E_i] + h_i = w(h_i) \,\, (1 \le i < p)$$
where the $h_i$'s in the middle of the display and the extra $1$ in $r-2+1$ correspond to the numerators $Y_1, \dots, Z_r$ in
the defining rational functions of $\Psi$.
As observed in \cite[Lemma in \S 3]{Mu2},
using the affine coordinates $((x_2, \dots, x_r), (y_2, \dots, y_r), \dots, (z_2, \dots, z_r))$
of $(\mathbb{P}^{r-1})^{p-1}$
around the point $P_1$ (the diagonal image of the point $[1 : 0 : \dots : 0] \in \mathbb{P}^{r-1}$),
the map $X \overset{f_w}\to X' \to (\mathbb{P}^{r-1})^{p-1}$ takes the following form around $E_1$:
$$\begin{aligned}
E_1 \, \ni \, & \, ((x_2, \dots, x_r), (y_2, \dots, y_r), \dots, (z_2, \dots, z_r) \\
\, \mapsto \,
& \, ([0 : \frac{1}{x_2} : \dots : \frac{1}{x_r}],
[1 : \frac{y_2}{x_2} : \dots : \frac{y_r}{x_r}], \dots, [1 : \frac{z_2}{x_2} : \dots : \frac{z_r}{x_r}]) .
\end{aligned}$$
Hence for the hyperplane $H_{1i} \subset (\mathbb{P}^{r-1})^{p-1}$ defined by $X_i = 0$, its
proper transform $H_{1i}' \subset X'$ satisfies
(when $i = 1$)
$f_w^*H_{1i}' = E_i$.
Since $\Psi$ is an involution and by a similar observation, for all $1 \le i \le r$, we have
(noting that $[H_{1i}] = h_1$):
$$[f_w^* E_i'] = [H_{1i}'] = h_1 - \sum_{i \ne j=1}^r [E_j] = w(e_i)$$
if we identify $H^2(X, \mathbb{Z}) = H^2(X', \mathbb{Z})$ as above.
The equality $f_w^* e_j' = e_j$ ($r < j \le q+r$) is by the definition of $Q_j$.
Thus we have $f_w^* = w$ on $H^2(X', \mathbb{Z})$ (identified with $H^2(X, \mathbb{Z})$).
To check that $X' \to (\mathbb{P}^{r-1})^{p-1}$ is just the blowup of points $P_i'$
determined by the $(p+q+r-1)$-tuple $w(D; c)$,
we can argue as in Proposition \ref{Cal}.
Indeed, let $C_X \subset X$ be the proper transform of $C = \Phi_{(D; c)}(C) \subset (\mathbb{P}^{r-1})^{p-1}$
(which is isomorphic to $C$ since we blow up only smooth points on $C$).
Then for $1 \le i \le r$,
we have $\deg(H_{1i}' | C_X) = \deg(H_{1i} | \Phi_{(D; c)}(C)) - \deg(\sum_{i \ne j=1}^r E_j) | C_X = r - (r-1) = 1$.
Hence $C_X$ meets $H_{1i}'$ at only one point and transversally.
So the map $X \overset{f_w}\to X' \to (\mathbb{P}^{r-1})^{p-1}$ collapses $H_{1i}'$ to a smooth point $Q_i$ on the image $C'$
of $C$ which is
contained in the codomain $(\mathbb{P}^{r-1})^{p-1}$ of the Cremona involution $\Psi$.
With the identification $C' = C_X = \Phi_{(D; c)}(C) = C$, we have
$$[\mathcal{O}_{\mathbb{P}_i^{r-1}}(1) | C'] = ((r-1) h_1 - (r-1) \sum_{i=1}^{r} [E_i] + h_i) | C_X = w(h_i) | C = w(D)_i = D'_i$$
which is a degree $r \ge 3$ (very ample) divisor and embeds $C'$ onto $C_i'$ ($\subset \mathbb{P}_i^{r-1}$,
the $i$-th factor of the codomain of $\Psi$).
With the identification $C' = C_X = \Phi_{(D; c)}(C) = C = C_1
:= \Phi_{|D_1|}(C)$ ($\subset \mathbb{P}_1^{r-1}$,
the first factor of the domain of $\Psi$),
the point $Q_i \in C'$ is given by
$[H_{1i}' | C_X] = [H_{1i} | C_1] - \sum_{i \ne j=1}^r E_j | C_X
= D_1 - \sum_{i \ne j=1}^r c_j = w(c_i) \in C$. Hence $Q_i$ is one of $P_i'$ ($1 \le i \le r$)
defined before Proposition \ref{Cal2}. By the construction, $Q_j = \Psi(P_j) = P_j'$ for $r < j \le q+r$.
Thus $X' = X_{(D'; c')}$.
This proves Proposition \ref{Cal2}.
\end{proof}
\begin{setup}
{\bf Proof of Theorems \ref{ThB} and \ref{ThA}}
\end{setup}
The same argument for $p = 2$ now works for all $p \ge 2$, but with Proposition \ref{Cal}
replaced by Proposition \ref{Cal2}.
|
\section{#1} \setcounter{equation}{0}}
\newcommand{\beq}{\begin{equation}}
\newcommand{\eeq}[1]{\label{#1}\end{equation}}
\newcommand{\ber}{\begin{eqnarray}}
\newcommand{\eer}[1]{\label{#1}\end{eqnarray}}
\newcommand{\re}[1]{(\ref{#1})}
\newcommand{\bar{u}}{\bar{u}}
\newcommand{\bar{z}}{\bar{z}}
\newcommand{\bar{x}}{\bar{x}}
\newcommand{\bar{y}}{\bar{y}}
\newcommand{\bar{\zeta}}{\bar{\zeta}}
\newcommand{\tilde{\Upsilon}}{\tilde{\Upsilon}}
\newcommand{\bar{\tilde{\Upsilon}}}{\bar{\tilde{\Upsilon}}}
\newcommand{\bar{\Upsilon}}{\bar{\Upsilon}}
\newcommand{\Upsilon}{\Upsilon}
\newcommand{\bar{\Upsilon}}{\bar{\Upsilon}}
\newcommand{\bar{\tilde{\Upsilon}}}{\bar{\tilde{\Upsilon}}}
\begin{document}
\renewcommand{\theequation}{\thesection.\arabic{equation}}
\setcounter{page}{0}
\thispagestyle{empty}
\bigskip
\begin{center} \LARGE
{\bf A twistor sphere of generalized K\"ahler potentials on hyperk\"ahler manifolds}
\\[12mm] \normalsize
{\bf Malte~Dyckmanns\footnote{E-mail: <EMAIL>}} \\[8mm]
{\small\it ~\\
C.N.Yang Institute for Theoretical Physics, Stony Brook University, \\
Stony Brook, NY 11794-3840,USA\\
~\\~\\}
\end{center}
\vspace{10mm}
\centerline{\bfseries Abstract}
\bigskip
\noindent
We consider the generalized K\"ahler structures $(g,J_+,J_-)$ that arise on a hyperk\"ahler mani-fold $(M,g,I,J,K)$ when we choose $J_+$ and $J_-$ from the twistor space of $M$. We find a relation between semichiral and arctic superfields which can be used to determine the generalized K\"ahler potential for hyperk\"ahler manifolds whose description in projective superspace is fully understood. We use this relation to determine an $S^2$-family of generalized K\"ahler potentials for Euclidean space and for the Eguchi-Hanson geometry. Cotangent bundles of Hermitian symmetric spaces constitute a class of hyperk\"ahler manifolds where our method can be applied immediately since the necessary results from projective superspace are already available. As a non-trivial higher-dimensional example, we determine the generalized potential for $T^\ast\mathbb{C}P^n$, which generalizes the Eguchi-Hanson result.
\bigskip
\bigskip ~
\eject
\normalsize
\eject
\section{Introduction}
Hyperk\"ahler manifolds admit various generalized K\"ahler structures. The corresponding generalized K\"ahler potentials can be used to reconstruct the hyperk\"ahler geometry. These generalized potentials are in general quite different from the ordinary K\"ahler potential and thus provide a new way of studying hyperk\"ahler geometry and finding hyperk\"ahler metrics. To gain insight into this new way of looking at hyperk\"ahler geometry, one first needs to study some examples. In this paper, we make use of the twistor space of hyperk\"ahler manifolds to develop a general method for determining their generalized K\"ahler potentials and we also explicitly work out some examples.
In the next section, we review the relevant features of generalized K\"ahler geometry in its bihermitian formulation. This geometry involves two complex structures $J_+$, $J_-$ on a Riemannian manifold $(M,g)$ and can locally be described by a generalized K\"ahler potential. In this paper, we consider the case where the kernel of $[J_+,J_-]$ is trivial. Then the potential is defined as the generating function for a symplectomorphism between coordinates $(x_L,y_L)$ and $(x_R,y_R)$ that are holomorphic w.r.t. $J_+$ and $J_-$ respectively. Generalized K\"ahler geometry was initially found as the target space geometry of $2D$ $\mathcal{N}=(2,2)$ supersymmetric sigma models, where the potential is the superspace Lagrangian and the coordinates $x_L,x_R$ describe semichiral superfields.
In section \ref{sectionhyper}, we review aspects of hyperk\"ahler geometry and its twistor space. We parametrize the twistor sphere of complex structures by a complex coordinate $\zeta$ and introduce holomorphic Darboux coordinates $\Upsilon(\zeta),\tilde{\Upsilon}(\zeta)$ for a certain holomorphic symplectic form. This construction is relevant for the projective superspace description of $2D$ $\mathcal{N}=(4,4)$ sigma models, where $\Upsilon,\tilde{\Upsilon}$ are arctic superfields.
Using its twistor space, a hyperk\"ahler manifold can be seen as a generalized K\"ahler manifold in various ways while keeping the metric fixed. In section \ref{sectionstructures}, we consider a two-sphere of generalized K\"ahler structures on a hyperk\"ahler manifold and express the coordinates $x_L,x_R,y_L,y_R$ in terms of $\Upsilon,\tilde{\Upsilon}$. This enables us to determine the generalized K\"ahler potential on a hyperk\"ahler manifold if we can find the decomposition of the arctic superfields $\Upsilon,\tilde{\Upsilon}$ in terms of their $\mathcal{N}=(2,2)$ components, i.e. in terms of coordinates on $M$.
In section \ref{sectionfour}, we consider four-dimensional hyperk\"ahler manifolds and explicitly determine the partial differential equations that the coordinates describing those arctic superfields have to fulfill. In section \ref{sectioneuclid}, we determine the potential for Euclidean space, where the differential equations for $\Upsilon,\tilde{\Upsilon}$ are easy to solve. In section \ref{sectioneguchi}, we look at the Eguchi-Hanson metric, where the relevant coordinates $\Upsilon,\tilde{\Upsilon}$ have been found previously in \cite{properties}. We give an explicit expression for the $S^2$-family of generalized K\"ahler potentials for this geometry, which belongs to the family of gravitational instantons and is thus of interest to physicists.
The Eguchi-Hanson geometry lives on the cotangent bundle of $\mathbb{C}P^1$. For hyperk\"ahler structures on cotangent bundles over arbitrary K\"ahler manifolds, projective superspace can be used to determine the coordinates $\Upsilon,\tilde{\Upsilon}$. This has been done in particular for all Hermitian symmetric spaces. In section \ref{sectionproj}, we review this procedure and as a non-trivial higher-dimensional example, we use the results for $T^\ast\mathbb{C}P^n$ to determine its generalized K\"ahler potential, which generalizes the Eguchi-Hanson result.
In an appendix, we extend our results from section \ref{sectionstructures} and consider the full $S^2\times S^2$-family of generalized K\"ahler structures on a hyperk\"ahler manifold, i.e. we let both $J_+$ and $J_-$ be an arbitrary point on the twistor sphere of complex structures. We also give the explicit potential depending on two complex parameters $\zeta_+$,$\zeta_-$ for the simplest hyperk\"ahler manifold, namely for Euclidean space.
\section{The generalized K\"ahler potential}
Generalized K\"ahler geometry first appeared in the study of 2D $\mathcal{N}=(2,2)$ nonlinear $\sigma$-models \cite{gates} and was later rediscovered by mathematicians as a special case of generalized complex geometry \cite{gualtieri}. In its bihermitian formulation, generalized K\"ahler geometry consists of two (integrable) complex structures $J_+$, $J_-$ on a Riemannian manifold $(M,g)$, where the metric is hermitian with respect to $J_+$ and $J_-$. Furthermore, the forms $\omega_\pm:=gJ_\pm$ have to fulfill \cite{linearizing}
\beq
d^c_+\omega_++d^c_-\omega_-=0,\quad dd^c_+\omega_+=0,
\eeq{integrability}
where $d^c_\pm=i(\bar{\partial}_\pm-\partial_\pm)$. This allows us to define the closed three-form $H:=d^c_+\omega_+=-d^c_-\omega_-$, whose local two-form potential we denote by $B$ ($H=dB$).
In general, $\omega_\pm$ is not closed and thus $(M,g,J_\pm)$ is not K\"ahler. In this paper, we will however consider the case where $H=0$. Then $\partial_\pm\omega_\pm$, $\bar{\partial}_\pm\omega_\pm$ have to vanish separately, so $d\omega_\pm=0$, i.e. $(M,g,J_\pm)$ is K\"ahler.
In \cite{offshell} it was shown that like ordinary K\"ahler geometry, generalized K\"ahler geometry is locally described by a single function, the generalized K\"ahler potential. On a generalized K\"ahler manifold, one can define the Poisson structure $\sigma:=[J_+,J_-]g^{-1}$ \cite{hitchinP}. Here, we consider the case where $[J_+,J_-]$ is invertible and recall how the generalized K\"ahler potential is defined in this case \cite{potential}:
Inverting $\sigma$ gives
\beq \Omega_G:=\sigma^{-1}=g[J_+,J_-]^{-1}, \eeq{omegag}
which is a real, closed and non-degenerate two-form that fulfills $J_\pm^T\Omega_GJ_\pm=-\Omega_G$ \cite{hitchinP}, i.e. it is a real holomorphic symplectic form both w.r.t. $J_+$ and w.r.t. $J_-$. This means that $\Omega_G$ can be split into the sum of a $(2,0)$- and a $(0,2)$-form both w.r.t. $J_+$ and w.r.t. $J_-$:
\beq \Omega_G=\Omega_+^{(2,0)}+\Omega_+^{(0,2)}=\Omega_-^{(2,0)}+\Omega_-^{(0,2)}, \eeq{plusminus}
where $\bar{\partial}_\pm\Omega_\pm^{(2,0)}=0$ and $\Omega_\pm^{(0,2)}=\overline{\Omega_\pm^{(2,0)}}$ (here the complex conjugate is taken w.r.t. $J_+$ and $J_-$ respectively).
One then introduces Darboux coordinates $x^p_L$ and ${y_L}_p$, holomorphic w.r.t. $J_+$, for $\Omega_+^{(2,0)}$; and $x^p_R$ and ${y_R}_p$, holomorphic w.r.t. $J_-$, for $\Omega_-^{(2,0)}$ ($p=1,...,n$, where $dim_\mathbb{R}M=4n$) \cite{potential}. Then
\ber \Omega_G&=&\Omega_+^{(2,0)}+\Omega_+^{(0,2)}=dx_L^p\wedge {dy_L}_p+d\bar{x}_L^p\wedge {d\bar{y}_L}_p, \nonumber \\
\Omega_G&=&\Omega_-^{(2,0)}+\Omega_-^{(0,2)}=dx_R^p\wedge {dy_R}_p+d\bar{x}_R^p\wedge {d\bar{y}_R}_p, \eer{asdf}
i.e. the coordinate transformation from $\{x_L,\bar{x}_L,y_L,\bar{y}_L\}$ to $\{x_R,\bar{x}_R,y_R,\bar{y}_R\}$ is a symplectomorphism (canonical transformation) preserving $\Omega_G$. It is thus described by a generating function $P(x_L,x_R,\bar{x}_L,\bar{x}_R)$ such that (omitting indices from now on)
\beq \frac{\partial P}{\partial x_L}=y_L,\quad \frac{\partial P}{\partial x_R}=-y_R,\quad \frac{\partial P}{\partial \bar{x}_L}=\bar{y}_L,\quad \frac{\partial P}{\partial \bar{x}_R}=-\bar{y}_R. \eeq{lrRelation}
This generating function is the generalized K\"ahler potential\footnote{We can choose to let the potential depend on other combinations of old and new coordinates as well. The potentials corresponding to the four different choices of variables are then related via Legendre transforms. In previous papers, the roles of $x_R$ and $y_R$ were interchanged. However, formulas in previous papers for reconstructing $g$, $J_+$, $J_-$ and $B$ from the potential remain unchanged when using our convention.} and can be used to locally reconstruct all the geometric data of generalized K\"ahler geometry \cite{potential}, i.e. the two complex structures $J_+$, $J_-$, the metric $g$ and the $B$-field. It also turns out to be the superspace Lagrangian for the $\mathcal{N}=(2,2)$ $\sigma$-models that led to the discovery of generalized K\"ahler geometry \cite{offshell}.
\section{Hyperk\"ahler manifolds and their twistor spaces}\label{sectionhyper}
In this paper, we consider generalized K\"ahler structures $(g,J_+,J_-)$ on a hyperk\"ahler mani-fold $M$ and investigate their generalized K\"ahler potentials. For the choice of the two complex structures $J_+$ and $J_-$, we will make use of the twistor space $\mathcal{Z}=M\times S^2$ of $M$.
Hyperk\"ahler manifolds appear for instance as the target spaces for hypermultiplet scalars in four-dimensional nonlinear $\sigma$-Models with rigid $\mathcal{N}=2$ supersymmetry on the base space \cite{freedman}. In geometric terms, they are described by the data $(M,g,I,J,K)$, where $g$ is a Riemannian metric on $M$ that is K\"ahler with respect to the three complex structures $I,J,K$, which fulfill the quaternion algebra (i.e. $IJ=K=-JI$). In fact, there exists a whole two-sphere of complex structures on $M$ with respect to which $g$ is a K\"ahler metric, namely $(M,g,\mathcal{J}=v_1I+v_2J+v_3K)$ is K\"ahler for each $(v_1,v_2,v_3)\in S^2$. Using (the inverse of) the stereographic projection, we parametrize this family of complex structures on $M$ in a chart of $S^2$ including the north-pole by a complex coordinate $\zeta$:
\beq \mathcal{J}(\zeta):=v_1(\zeta)I+v_2(\zeta)J+v_3(\zeta)K:=\frac{1}{1+\zeta\bar{\zeta}}\left[(1-\zeta\bar{\zeta})I+(\zeta+\bar{\zeta})J+i\,(\bar{\zeta}-\zeta)K\right]. \eeq{J}
We define the complex two-forms
\beq \omega^{(2,0)}:=\omega_2+i\omega_3,\quad\quad\omega^{(0,2)}:=\omega_2-i\omega_3; \eeq{omegapm}
where $\omega_1=gI$, $\omega_2=gJ$, $\omega_3=gK$ are the three K\"ahler forms. Then for each $\zeta\in \mathbb{C}$
\beq \Omega_H(\zeta):=\omega^{(2,0)}-2\zeta\omega_1-\zeta^2\omega^{(0,2)} \eeq{OmegaH}
turns out to be a holomorphic symplectic form with respect to the complex structure $\mathcal{J}(\zeta)$ \cite{hitchin}. In particular, $\omega^{(2,0)}=\Omega_H(\zeta=0)$ is a $(2,0)$-form w.r.t. $I=\mathcal{J}(\zeta=0)$.
Starting from $\zeta=0$, we can locally find holomorphic Darboux coordinates $\Upsilon^p(\zeta)$ and $\tilde{\Upsilon}_p(\zeta)$ ($p=1,...,n$, where $dim_\mathbb{R}M=4n$) for $\Omega_H(\zeta)$ that are analytic in $\zeta$ such that \cite{properties}
\beq \Omega_H(\zeta)=i\,d\Upsilon^p(\zeta)\wedge d\tilde{\Upsilon}_p(\zeta). \eeq{Darboux}
These canonical coordinates $\Upsilon,\tilde{\Upsilon}(\zeta)$ for $\Omega_H$ are crucial for the projective superspace formulation of $\sigma$-models with eight real supercharges\footnote{If we define $\breve{\Upsilon}(\zeta):=\bar{\Upsilon}(-\frac{1}{\zeta})$, $\breve{\tilde{\Upsilon}}(\zeta):=\bar{\tilde{\Upsilon}}(-\frac{1}{\zeta})$, then $(\Upsilon,\tilde{\Upsilon})$ and $(\breve{\Upsilon}$,$\breve{\tilde{\Upsilon}})$ are related by a $\zeta^2$-twisted symplectomorphism whose generating function $f(\Upsilon,\breve{\Upsilon};\zeta)$ can be interpreted as the projective superspace Lagrangian \cite{properties}. \label{Footnote}}, where they describe "arctic" superfields. They have been determined for instance in \cite{properties} for the Eguchi-Hanson metric and we will use them in this paper to determine the generalized K\"ahler potential for hyperk\"ahler manifolds.
\section{Gen. K\"ahler structures on hyperk\"ahler manifolds}\label{sectionstructures}
We want to transport the idea of a twistor space from hyperk\"ahler to generalized K\"ahler geometry, namely we interpret a hyperk\"ahler manifold $(M,g,I,J,K)$ as a generalized K\"ahler manifold $(M,g,J_+,J_-)$, where we fix the left complex structure $J_+=I$ and let the right complex structure depend on $\zeta$: $J_-=\mathcal{J}(\zeta)$ (see eq. \eqref{J}). So for a given hyperk\"ahler manifold, we consider an $S^2$-family of generalized K\"ahler structures whose generalized K\"ahler potentials we now try to determine.
First, we need an explicit expression for the symplectic form $\Omega_G$ (eq. \eqref{omegag}), which now depends on $\zeta$. The anticommutator of two complex structures on a locally irreducible hyperk\"ahler manifold is equal to a constant times the identity, $\{J_+,J_-\}=c\mathbbm{1}$ (see, e.g., \cite{lectures}). If $J_+\neq\pm J_-$, then $|c|<2$ and $\frac{1}{\sqrt{4-c^2}}[J_+,J_-]$ is another complex structure, so in particular it squares to $-\mathbbm{1}$. Using this, we have
\beq \Omega_G=g[J_+,J_-]^{-1}=-\frac{1}{4-c^2}g[J_+,J_-], \eeq{OmGen}
which in our case, where we have $c=-2v_1=-2\frac{1-\zeta\bar{\zeta}}{1+\zeta\bar{\zeta}}$ and $[J_+,J_-]=2v_2K-2v_3J$, gives
\beq \Omega_G(\zeta)=-\frac{1}{2-2v_1^2}\left(v_2\omega_3-v_3\omega_2\right)=-\frac{1+\zeta\bar{\zeta}}{8\zeta\bar{\zeta}}\left[(\zeta+\bar{\zeta})\omega_3-i(\bar{\zeta}-\zeta)\omega_2\right]. \eeq{OG}
This can be split into the sum of the holomorphic form $\Omega_+^{(2,0)}=i\bar{\zeta}\frac{1+\zeta\bar{\zeta}}{8\zeta\bar{\zeta}}\omega^{(2,0)}$ and the antiholomorphic form $\Omega_+^{(0,2)}=-i\zeta\frac{1+\zeta\bar{\zeta}}{8\zeta\bar{\zeta}}\omega^{(0,2)}$ with respect to $J_+$ (see equation \eqref{omegapm}). Combining equations \eqref{OmegaH} and \eqref{Darboux}, we can choose the following Darboux coordinates for $\Omega_G(\zeta)$:
\ber x_L^p&=&\Upsilon^p(\zeta=0), \quad {y_L}_p=-\bar{\zeta}\frac{1+\zeta\bar{\zeta}}{8\zeta\bar{\zeta}}\tilde{\Upsilon}_p(\zeta=0); \nonumber \\
\bar{x}_L^p&=&\bar{\Upsilon}^p(\zeta=0), \quad \bar{y}_{L_p}=-\zeta\frac{1+\zeta\bar{\zeta}}{8\zeta\bar{\zeta}}\bar{\tilde{\Upsilon}}_p(\zeta=0). \eer{xyL}
With respect to $J_-=\mathcal{J}(\zeta)$, $\Omega_G$ splits into the sum of $\Omega_-^{(2,0)}=i\bar{\zeta}\frac{1}{8\zeta\bar{\zeta}}\Omega_H(\zeta)$ and $\Omega_-^{(0,2)}=-i\zeta\frac{1}{8\zeta\bar{\zeta}}\overline{\Omega_H(\zeta)}$. Consequently, we can choose\footnote{We denote the complex conjugate of $\Upsilon(\zeta)$ by $\bar{\Upsilon}\equiv\bar{\Upsilon}(\bar{\zeta})\equiv\overline{\Upsilon(\zeta)}$ which is not to be confused with the notation in \cite{properties}, where $\bar{\Upsilon}$ is shorthand for $\breve{\Upsilon}(\zeta)=\bar{\Upsilon}(-\zeta^{-1})$.}
\ber x_R^p&=&\Upsilon^p(\zeta),\quad {y_R}_p=-\bar{\zeta}\frac{1}{8\zeta\bar{\zeta}}\tilde{\Upsilon}_p(\zeta); \nonumber \\
\bar{x}_R^p&=&\overline{\Upsilon^p(\zeta)},\quad {\bar{y}}_{R_p}=-\zeta\frac{1}{8\zeta\bar{\zeta}}\overline{\tilde{\Upsilon}_p(\zeta)}. \eer{xyR}
We are thus able to express the coordinates $x_{L,R}$ and $y_{L,R}$ that describe semichiral superfields in $\mathcal{N}=(2,2)$ models in terms of the coordinates $\Upsilon(\zeta)$, $\tilde{\Upsilon}(\zeta)$ describing arctic superfields in the projective superspace formulation of $\mathcal{N}=(4,4)$ supersymmetric sigma models. This will enable us to determine the $\zeta$-dependent generalized K\"ahler potential for hyperk\"ahler manifolds whose projective superspace description is known.
\section{The four-dimensional case \label{sectionfour}}
In this section, we consider the four-dimensional case and explicitly determine the partial differential equations for $\Upsilon(\zeta)$ and $\tilde{\Upsilon}(\zeta)$ in order to be holomorphic w.r.t. $\mathcal{J}(\zeta)$ and to fulfill equation \eqref{Darboux}. A four-dimensional K\"ahler manifold $(M,g,I)$ is hyperk\"ahler if and only if around each point there are holomorphic coordinates $(z,u)$ on $M$ such that the K\"ahler potential $K(z,u)$ fulfills the following Monge-Amp\`ere equation \cite{newman}:
\beq K_{u\bar{u}}K_{z\bar{z}}-K_{u\bar{z}}K_{z\bar{u}}=1. \eeq{Monge}
From a K\"ahler potential fulfilling this equation, we can construct the three K\"ahler forms:
\ber
\omega_1&=&-\frac{i}{2}\partial\bar{\partial} K, \nonumber \\
\omega_2&=&\frac{i}{2}(dz\wedge du-d\bar{z}\wedge d\bar{u}), \nonumber \\
\omega_3&=&\frac{1}{2}(dz\wedge du+d\bar{z}\wedge d\bar{u}).
\eer{kForms}
Together with the metric $g$ whose line element is
\beq ds^2=K_{u\bar{u}}\,du d\bar{u}+K_{u\bar{z}}\,du d\bar{z}+K_{z\bar{u}}\,dz d\bar{u}+K_{z\bar{z}}\,dz d\bar{z}, \eeq{metric}
we get the three complex structures, where equation \eqref{Monge} ensures that $J=g^{-1}\omega_2$ and $K=g^{-1}\omega_3$ indeed square to $-\mathbbm{1}$.
We find the following basis for the $(1,0)$ forms w.r.t. $\mathcal{J}(\zeta)$:
\beq \theta^1=dz-\zeta K_{u\bar{u}} d\bar{u}-\zeta K_{u\bar{z}} d\bar{z},\quad \theta^2=du+\zeta K_{z\bar{u}} d\bar{u}+\zeta K_{z\bar{z}} d\bar{z}. \eeq{basisForms}
For $\Upsilon,\tilde{\Upsilon}$ to be holomorphic w.r.t. $\mathcal{J}(\zeta)$, $d\Upsilon(\zeta)$ and $d\tilde{\Upsilon}(\zeta)$ must be linear combinations of $\theta^1$ and $\theta^2$ (here the differential does not act on $\zeta$):
\beq d\Upsilon=\frac{\partial\Upsilon}{\partial z} \theta^1+\frac{\partial \Upsilon}{\partial u} \theta^2,\quad d\tilde{\Upsilon}=\frac{\partial\tilde{\Upsilon}}{\partial z} \theta^1+\frac{\partial \tilde{\Upsilon}}{\partial u}\theta^2. \eeq{sdgdfklj2}
Here the coefficients have been determined by comparing the $dz$- and $du$-terms on both sides. From equations \eqref{basisForms} and \eqref{sdgdfklj2}, we get the requirement that both $\Upsilon$ and $\tilde{\Upsilon}$ have to fulfill the following two PDEs:
\beq \frac{\partial \Psi}{\partial \bar{z}}=\zeta\left(K_{z\bar{z}} \frac{\partial \Psi}{\partial u}- K_{u\bar{z}} \frac{\partial \Psi}{\partial z}\right),\quad \frac{\partial \Psi}{\partial \bar{u}}=\zeta\left(K_{z\bar{u}}\frac{\partial \Psi}{\partial u}-K_{u\bar{u}} \frac{\partial \Psi}{\partial z}\right)\quad (\Psi=\Upsilon,\tilde{\Upsilon}).\eeq{requirement1}
Furthermore, we find that
\beq \Omega_H(\zeta)= idz\wedge du+i\zeta \partial \bar{\partial} K+i\zeta^2 d\bar{z}\wedge d\bar{u}=i\theta^1\wedge\theta^2, \eeq{afsdf}
so using \eqref{sdgdfklj2}, we obtain that equation \eqref{Darboux} corresponds to the requirement
\beq \frac{\partial \Upsilon}{\partial z}\frac{\partial \tilde{\Upsilon}}{\partial u}-\frac{\partial \Upsilon}{\partial u}\frac{\partial \tilde{\Upsilon}}{\partial z}=1.\eeq{requirement2}
\section{Euclidean space}\label{sectioneuclid}
We now use the relation between $x,y$ and $\Upsilon,\tilde{\Upsilon}$ derived in section \ref{sectionstructures} to determine the generalized K\"ahler potential for Euclidean space. Here the K\"ahler potential is given by
\beq K=u\bar{u}+z\bar{z}, \eeq{kEuclid}
which clearly fulfills equation \eqref{Monge}. Assuming that $(z,u)$ are holomorphic coordinates w.r.t. $I$ and setting $\omega^{(2,0)}=idz\wedge du$, we get the complex structures as described in section \ref{sectionfour}. They are the differentials (pushforwards) of the left action of the imaginary basis quaternions\footnote{We stick to the convention from previous papers and include the $i$-factor in the choice of $\omega^{(2,0)}$. This interchanges the complex structures $J$ and $K$, s.t. $J$ corresponds to left multiplication by $k$ and $K$ corresponds to left multiplication by $-j$.} $i,j,k$ on $\mathbb{H}\approx \mathbb{C}^2$, where we make the identification $(z,u)=(x_0+ix_1,x_2+ix_3)\mapsto x_0+ix_1+jx_2+kx_3$.
\beq \Upsilon(\zeta)=z-\zeta \bar{u},\quad \tilde{\Upsilon}(\zeta)=u+\zeta \bar{z} \eeq{uFlat}
fulfill equations \eqref{requirement1} and \eqref{requirement2}, i.e. they are holomorphic w.r.t. $\mathcal{J}(\zeta)$ and satisfy equation \eqref{Darboux}.
Using equations \eqref{xyL} and \eqref{xyR}, we make the identifications
\beq x_L=\Upsilon(\zeta=0)=z,\,\,x_R=\Upsilon(\zeta)\equiv \Upsilon\text{ and }y_L=-\frac{1+\zeta\bar{\zeta}}{8\zeta}u,\,\,y_R=-\frac{1}{8\zeta}\tilde{\Upsilon}. \eeq{identification4d}
Solving for $y_L$, $y_R$ in terms of $x_L$, $x_R$, we get
\beq y_L=-\frac{1+\zeta\bar{\zeta}}{8\zeta\bar{\zeta}}(\bar{x}_L-\bar{x}_R),\quad y_R=-\frac{1}{8\zeta\bar{\zeta}}\left((1+\zeta\bar{\zeta})\bar{x}_L-\bar{x}_R\right),\eeq{sdfkjshgkjfdgc}
which leads (up to an additive constant) to the generating function (see equation \eqref{lrRelation})
\beq P=-\frac{1}{8\zeta\bar{\zeta}}\left[x_R\bar{x}_R+(1+\zeta\bar{\zeta})\cdot(x_L\bar{x}_L-x_L\bar{x}_R-\bar{x}_Lx_R)\right]. \eeq{potflat}
This is the generalized K\"ahler potential for Euclidean space, where $J_+=I$ and $J_-$ is an arbitrary point on the twistor-sphere of complex structures, $J_-\neq \pm I$. However, we notice that $P$ only involves the combination $\zeta\bar{\zeta}$, i.e. it only depends on the angle between $J_+$ and $J_-$ in the space spanned by the three complex structures $(I,J,K)$. Also $P$ turns out to be asymmetric between left- and right-coordinates. This can be resolved however, as there are various ambiguities in the generalized K\"ahler potential. For instance, we could distribute factors differently in \eqref{identification4d} or even perform a more complicated symplectomorphism, going to new coordinates $x'_{L/R}$, $y'_{L/R}$. If we make the identifications
\beq x'_L=i\sqrt{\frac{1+\zeta\bar{\zeta}}{8\zeta}}z,\,\,x'_R=i\sqrt{\frac{1}{8\zeta}}\Upsilon\text{ and }y'_L=i\sqrt{\frac{1+\zeta\bar{\zeta}}{8\zeta}}u,\,\,y'_R=i\sqrt{\frac{1}{8\zeta}}\tilde{\Upsilon}, \eeq{identification24d}
the potential is left-right-symmetric. Furthermore, we can perform Legendre transforms and express the potential in terms of a different set of variables. If we use for instance \eqref{identification24d} and in addition exchange the roles of $x'_R$ and $y'_R$, we arrive at the potential
\beq P'=\sqrt{1+\zeta\bar{\zeta}}\cdot(x'_Ly'_R+\bar{x}'_L\bar{y}'_R)+\sqrt{\zeta\bar{\zeta}}\cdot(x'_L\bar{x}'_L+y'_R\bar{y}'_R). \eeq{potentialprime}
\section{Example: Eguchi-Hanson geometry}\label{sectioneguchi}
The real function
\beq K=\sqrt{1+4u\bar{u}(1+z\bar{z})^2}+\frac{1}{2}\text{log}\left[\frac{4u\bar{u}(1+z\bar{z})^2}{\left(1+\sqrt{1+4u\bar{u}(1+z\bar{z})^2}\right)^2}\right] \eeq{Kpot}
in the two complex variables $z,u$ fulfills the Monge-Amp\`ere equation \eqref{Monge}. It thus defines a hyperk\"ahler metric, where the K\"ahler forms are given by equation \eqref{kForms}. The first K\"ahler form takes the form
\ber \omega_1&=&-\frac{i}{2}\frac{1+z\bar{z}}{\sqrt{1+4u\bar{u}(1+z\bar{z})^2}}\Bigg[(1+z\bar{z})\,du\wedge d\bar{u}+2u\bar{z} \,dz\wedge d\bar{u} \nonumber \\
&&\quad\quad\quad\quad\quad+2z\bar{u} \,du\wedge d\bar{z}+\left(\frac{1}{(1+z\bar{z})^3}+4u\bar{u}\right)dz\wedge d\bar{z}\Bigg],
\eer{formHanson}
from which the metric can be read off. This is the well-known Eguchi-Hanson geometry\footnote{Setting $u=\frac{1}{2}u'^2$, $z=\frac{z'}{u'}$ and $r:=\sqrt{u'\bar{u}'+z'\bar{z}'}$ gives the familiar K\"ahler potential $K=\sqrt{1+r^4}+\text{log} \frac{r^2}{1+\sqrt{1+r^4}}$ for the Eguchi-Hanson metric \cite{properties}.}. The holomorphic Darboux coordinates for $\Omega_H(\zeta)$ (fulfilling equations \eqref{requirement1} and \eqref{requirement2}) can be chosen as \cite{properties}
\ber \tilde{\Upsilon}&=&u+\zeta^2\bar{z}^2\bar{u}+\frac{\bar{z}\zeta}{1+z\bar{z}}\sqrt{1+4u\bar{u}(1+z\bar{z})^2},\nonumber \\
\Upsilon&=&z-\frac{2\bar{u}\zeta (1+z\bar{z})^2}{1+\sqrt{1+4u\bar{u}(1+z\bar{z})^2}+2\bar{u}\bar{z}\zeta(1+z\bar{z})}. \eer{Upsilon}
We solve $\Upsilon,\bar{\Upsilon}(z,\bar{z},u,\bar{u})$ for $u$ and $\bar{u}$ to get
\beq u(z,\bar{z},\Upsilon,\bar{\Upsilon})=\frac{\zeta}{1+z\bar{z}}\cdot\frac{(\bar{z}-\bar{\Upsilon})(1+\Upsilon\bar{z})}{\zeta\bar{\zeta}(1+\bar{\Upsilon} z)(1+\Upsilon\bar{z})-(z-\Upsilon)(\bar{z}-\bar{\Upsilon})} \eeq{U}
and its complex conjugate. Using this and the identifications derived in section \ref{sectionstructures} (equation \eqref{identification4d}), we get $y_L(x_L,x_R)$. We then integrate $y_L(x_L,x_R)$ w.r.t. $x_L$ to get the generalized K\"ahler potential up to a possible additive term that is independent of $x_L$:
\ber P&=&\int\,y_L(x_L,x_R)\,dx_L=-\bar{\zeta}\frac{1+\zeta\bar{\zeta}}{8\zeta\bar{\zeta}}\int\,u(z,\Upsilon)\,dz \nonumber \\
&=&-\frac{1}{8}\cdot \text{log}\frac{1+x_L\bar{x}_L}{\zeta\bar{\zeta}(1+x_L\bar{x}_R )(1+\bar{x}_Lx_R)-(x_L-x_R)(\bar{x}_L-\bar{x}_R)} \eer{potentialEH}
Plugging $u(z=x_L,\Upsilon=x_R)$ into $\tilde{\Upsilon}(z=x_L,u)$ (equation \eqref{Upsilon}) gives
\beq y_R(x_L,x_R)=-\bar{\zeta}\frac{1}{8\zeta\bar{\zeta}}\tilde{\Upsilon}(x_L,x_R)=-\frac{1}{8}\cdot\frac{(1+\zeta\bar{\zeta})\cdot \bar{x}_L-\left(1-\zeta\bar{\zeta}\cdot x_L\bar{x}_L\right)\bar{x}_R}{\zeta\bar{\zeta}(1+x_L\bar{x}_R )(1+\bar{x}_Lx_R)-(x_L-x_R)(\bar{x}_L-\bar{x}_R)} \eeq{yr}
which is indeed equal to $-\frac{\partial P}{\partial x_R}$. $P$ is real, so $\frac{\partial P}{\partial \bar{x}_L}=\bar{y}_L$ and $\frac{\partial P}{\partial \bar{x}_R}=-\bar{y}_R$ are also fulfilled and thus equation \eqref{potentialEH} gives indeed the $\zeta$-dependent generalized K\"ahler potential for the Eguchi-Hanson geometry:
\beq
P(x_L,\bar{x}_L,x_R,\bar{x}_R)=-\frac{1}{8}\cdot \text{log}\frac{1+|x_L|^2}{\zeta\bar{\zeta}\cdot|1+x_L\bar{x}_R |^2-|x_L-x_R|^2}. \eeq{potentialEH2}
Again, the generalized K\"ahler potential turns out to depend only on the combination $\zeta\bar{\zeta}$, i.e. on the angle between $J_+$ and $J_-$. Of course, there are again many ambiguities in the potential, but \eqref{potentialEH2} seems to be already in its simplest form.
\section{Hyperk\"ahler structures on cotangent bundles of K\"ahler manifolds and projective superspace}\label{sectionproj}
The target space of 4D $\mathcal{N}=2$ sigma models is constrained to be a hyperk\"ahler manifold \cite{freedman}. This corresponds to $2D$ $\mathcal{N}=(4,4)$ sigma models without $B$-field\footnote{For vanishing $B$-field, all the results from $4D$ $\mathcal{N}=2$ projective superspace can be immediately transferred to $2D$ $\mathcal{N}=(4,4)$ projective superspace. Actually the results that we are using only depend on the target space geometry, not on the number of space-time dimensions.}. Projective superspace provides methods to construct such models and for a large class of examples it can be used to extract the arctic superfields $\Upsilon$ and $\tilde{\Upsilon}$ \cite{GatesKuzenko1},\cite{Arai1},\cite{Arai2} (see \cite{ulfProjective} for a review) that we need in order to determine the generalized K\"ahler potential of the hyperk\"ahler target space using the method derived in chapter \ref{sectionstructures}. This has been done in particular for the hyperk\"ahler structure on cotangent bundles of Hermitian symmetric spaces (\cite{GatesKuzenko1}-\cite{Arai3}). As an example, we use the results from \cite{GatesKuzenko1},\cite{Arai1},\cite{Arai2} to determine the generalized K\"ahler potential for $T^\ast\mathbb{C}P^n=T^\ast(SU(n+1)/U(n))$. The special case $n=1$ then corresponds to the Eguchi-Hanson geometry that we considered in the last section.
It is a well-known fact that a hyperk\"ahler metric exists on (some open subset of) the cotangent bundle of every K\"ahler manifold $M$ \cite{Feix},\cite{Kaledin},\cite{GatesKuzenko2}. In projective superspace, this corresponds to models where the projective superspace Lagrangian $f(\Upsilon,\breve{\Upsilon};\zeta)$ (see footnote \ref{Footnote}) does not explicitly depend on $\zeta$ \cite{GatesKuzenko1}. To obtain the coordinates $\Upsilon$ and $\tilde{\Upsilon}$ for $T^\ast M$, one takes the $\mathcal{N}=(4,4)$ projective superspace Lagrangian $f(\Upsilon,\breve{\Upsilon})$ to be the K\"ahler potential $K(\phi,\bar{\phi})$ of the base space $M$, where the arctic and antarctic superfields $\Upsilon,\breve{\Upsilon}$ replace the chiral and antichiral superfields $\phi,\bar{\phi}$ of the $\mathcal{N}=(2,2)$ description, i.e. the holomorphic coordinates on $M$. In $\mathcal{N}=(2,2)$ components, $\Upsilon$ decomposes as
\beq \Upsilon=\phi+\zeta \Sigma+\sum_{j=2}^\infty \zeta^j X_j,\eeq{sdfcvfwe}
where $\phi$ is a chiral superfield describing the coordinates on $M$, $\Sigma$ is a complex linear superfield describing coordinates in the fiber of the tangent bundle $TM$ and all higher order terms are unconstrained auxiliary superfields \cite{GatesKuzenko2},\cite{Arai3}. Solving the algebraic equations of motion for the auxiliary superfields yields $\Upsilon$, and thus the Lagrangian in terms of the $\mathcal{N}=(2,2)$ superfields $(\phi,\Sigma)$ and the auxiliary complex variable $\zeta$. Integrating out $\zeta$ and dualizing\footnote{The duality between chiral $\psi$ and complex linear superfields $\Sigma$ is just an ordinary coordinate transformation that does not change the target space geometry.} the action, i.e. performing a Legendre transform replacing the complex linear superfields $\Sigma$ by chiral superfields $\psi$, gives the transformation $\Sigma(\psi)$ from which we obtain $\Upsilon(\phi,\psi)$. Here, $\psi$ describes coordinates in the fiber of the cotangent bundle $T^\ast M$. $\tilde{\Upsilon}(\phi,\psi)$ can then be obtained from $f$ and $\Upsilon(\phi,\psi)$ via $\tilde{\Upsilon}=\zeta\frac{\partial f}{\partial \Upsilon}$ \cite{properties}. One can also read off the ordinary K\"ahler potential of $T^\ast M$ from the dualized action \cite{GatesKuzenko2},\cite{Arai3}.
\subsection{Generalized K\"ahler potential for $T^\ast\mathbb{C}P^n$}
The crucial step in the above procedure is to eleminate the infinite tower of unconstrained auxiliary $\mathcal{N}=(2,2)$ superfields using their algebraic equations of motion. This has been done for instance in \cite{Arai1} for $T^\ast\mathbb{C}P^n$.
For the projective sigma model with target space $T^\ast\mathbb{C}P^n$, we take the projective superspace Lagrangian $f$ to be the K\"ahler potential of the Fubini-Study metric on $\mathbb{C}P^n$:
\beq f(\Upsilon^i(\zeta),\breve{\Upsilon}^{i}(\zeta))=a^2 \text{log}\left(1+\frac{\Upsilon^j\breve{\Upsilon}^{j}}{a^2}\right). \eeq{kahlerpotentialcpn}
Here, $a$ is a real parameter. The equations of motion for the auxiliary superfields have been solved in \cite{Arai1}. This gives $\Upsilon$ in terms of chiral and complex linear $\mathcal{N}=(2,2)$ superfields\footnote{This result was already obtained in the preparation of \cite{GatesKuzenko2} and later independently derived and first published in \cite{Arai1}.}:
\beq \Upsilon^i=z^i+\zeta\frac{\Sigma^i}{1-\zeta\frac{\bar{z}^{\bar{k}}\Sigma^k}{a^2+ z^l\bar{z}^{\bar{l}}}}. \eeq{UInTermsOfPhiSigma}
Here, we change notation and let $z\equiv\phi$ parametrize the base space and $u\equiv\psi$ the fibers of the cotangent bundle.
Dualizing the action of the sigma model to go from complex linear coordinates $\Sigma$ to chiral coordinates $u$ gives the equations\footnote{Repeated indices are always summed over $1,...,n$ and the metric is always written out explicitly, i.e. we never use it to raise or lower indices.} \cite{Arai1}
\beq u_i=-\frac{g_{i\bar{j}}\bar{\Sigma}^{\bar{j}}}{1-\frac{g_{k\bar{l}}\Sigma^k\bar{\Sigma}^{\bar{l}}}{a^2}},\quad
\bar{u}_{\bar{i}}=-\frac{g_{\bar{i}j}\Sigma^j}{1-\frac{g_{k\bar{l}}\Sigma^k\bar{\Sigma}^{\bar{l}}}{a^2}};
\eeq{kjfdg}
which have to be solved for the old coordinates $\Sigma$, $\bar{\Sigma}$ in terms of $u$, $\bar{u}$. Here $g_{i\bar{j}}$ is
\beq g_{i\bar{j}}=\frac{a^2\delta_{ij}}{a^2+z^k\bar{z}^{\bar{k}}}-\frac{a^2\bar{z}^{\bar{i}}z^j}{(a^2+z^l\bar{z}^{\bar{l}})^2}, \eeq{FubiniStudyMetric}
the Fubini-Study metric on $\mathbb{C}P^n$, and $g^{i\bar{j}}$ is its inverse. We find the following solution:
\beq
\Sigma^i=-\frac{2g^{i\bar{j}}\bar{u}_{\bar{j}}}{1+\sqrt{1+4\frac{g^{k\bar{l}}u_k\bar{u}_{\bar{l}}}{a^2}}},\quad
\bar{\Sigma}^{\bar{i}}=-\frac{2g^{\bar{i}j}u_j}{1+\sqrt{1+4\frac{g^{k\bar{l}}u_k\bar{u}_{\bar{l}}}{a^2}}}.
\eeq{sdfsdg}
Plugging this into \eqref{UInTermsOfPhiSigma} gives the arctic superfields $\Upsilon$ in terms of chiral $\mathcal{N}=(2,2)$ superfields $z$ and $u$:
\beq \Upsilon^i=z^i-\zeta\frac{2\bar{u}_{\bar{j}}g^{i\bar{j}}}{1+\sqrt{1+4\frac{g^{k\bar{l}}u_k\bar{u}_{\bar{l}}}{a^2}}+2\zeta\frac{g^{p\bar{q}}\bar{z}^p\bar{u}_{\bar{q}}}{a^2+z^m\bar{z}^{\bar{m}}}}.\eeq{NeededUInTermsOfUZ}
Together with
\beq \tilde{\Upsilon}^i=\zeta\frac{\partial f}{\partial \Upsilon^i}=\frac{\zeta\breve{\Upsilon}^i}{1+\frac{\Upsilon^j\breve{\Upsilon}^{j}}{a^2}}, \eeq{UTilde}
this is all the information we need to determine the generalized K\"ahler potential for $T^\ast \mathbb{C}P^n$ using the identifications found in section \ref{sectionstructures}.
Solving \eqref{NeededUInTermsOfUZ} and its complex conjugate for $u$ and $\bar{u}$ gives $u(z,\Upsilon)$:
\beq
u_i=\zeta \cdot \frac{a^2(a^2+z^{k}\bar{z}^{\bar{k}})(a^2+\bar{z}^{\bar{k}}\Upsilon^{k})\cdot g_{i\bar{j}}(\bar{z}^{\bar{j}}-\bar{\Upsilon}^{\bar{j}})}{a^2\zeta\bar{\zeta}(a^2+\bar{z}^{\bar{k}} \Upsilon^{k})(a^2+z^{k}\bar{\Upsilon}^{\bar{k}})-(z^l-\Upsilon^l)g_{l\bar{m}}(\bar{z}^{\bar{m}}-\bar{\Upsilon}^{\bar{m}})(a^2+z^{k}\bar{z}^{\bar{k}})^2}.
\eeq{uOfZU}
Integrating this with respect to $z^i$ gives:
\beq \int u_i\,dz^i= \frac{\zeta a^2}{1+\zeta\bar{\zeta}}\text{log}\frac{a^2+\mathbf{z}^T\mathbf{\bar{z}}}{a^2\zeta\bar{\zeta}(a^2+\mathbf{\bar{z}}^T\mathbf{\Upsilon})(a^2+\mathbf{z}^T\mathbf{\bar{\Upsilon}})-(\mathbf{z}-\mathbf{\Upsilon})^T\mathbf{g}(\mathbf{\bar{z}}-\mathbf{\bar{\Upsilon}})(a^2+\mathbf{z}^T\mathbf{\bar{z}})^2}. \eeq{sdgfdh}
Here, no sum is implied on the left-hand side and on the right-hand side we use vector notation ($\mathbf{z}:=(z^1,...,z^n)^T$, etc.) and $\mathbf{g}:=(g_{i\bar{j}})_{1\leq i,j\leq n}$. So, up to an additive term $c(x_R,\bar{x}_R)$, the generalized K\"ahler potential for $T^*\mathbb{C}P^n$ is
\beq P=-\frac{a^2}{8}\text{log}\frac{a^2+\mathbf{x_L}^T\mathbf{\bar{x}_L}}{a^2\zeta\bar{\zeta}(a^2+\mathbf{\bar{x}_L}^T\mathbf{x_R})(a^2+\mathbf{x_L}^T\mathbf{\bar{x}_R})-(\mathbf{x_L}-\mathbf{x_R})^T\mathbf{g}(\mathbf{\bar{x}_L}-\mathbf{\bar{x}_R})(a^2+\mathbf{x_L}^T\mathbf{\bar{x}_L})^2}. \eeq{finalpotential}
For $n=1$ and $a=1$, we have $g_{z\bar{z}}=\frac{1}{(1+z\bar{z})^2}$ and all the results from section \ref{sectioneguchi} are reproduced. Therefore, we assume that $c(x_R,\bar{x}_R)$ can be set to zero.
\section{Discussion}
There is an increasing number of examples, most notably among Hermitian symmetric spaces, where the decomposition of the $\mathcal{N}=(4,4)$ arctic superfields $\Upsilon$, $\tilde{\Upsilon}$ in terms of their $\mathcal{N}=(2,2)$ components ($z,u$) has been determined. In these cases, one can apply the methods developed in this paper to determine more examples of generalized K\"ahler potentials on hyperk\"ahler manifolds. Having whole classes of manifolds available for our analysis, one could try to find more general statements about the generalized K\"ahler potential in the case of hyperk\"ahler manifolds.
The Eguchi-Hanson geometry is one of the hyperk\"ahler manifolds that can be obtained from the generalized Legendre transform construction in \cite{hitchin} (generalized T-duality). The manifolds stemming from that construction are $4n$-dimensional hyperk\"ahler manifolds admitting $n$ commuting tri-holomorphic killing vectors. They are called toric hyperk\"ahler manifolds and have been classified in \cite{Bielawski}. It should be possible to determine the relevant coordinates $\Upsilon(z,u;\zeta)$ and $\tilde{\Upsilon}(z,u;\zeta)$ for toric hyperk\"ahler manifolds. For four-dimensional toric hyperk\"ahler manifolds, \cite{sevrin} gives a formula for the generalized K\"ahler potential as a certain threefold Legendre transform in the special case $\zeta\bar{\zeta}=1$. One could compare this construction with our results at least for the examples given in this paper or try to relate the two methods in general for four-dimensional toric hyperk\"ahler manifolds. As a further explicit example, one could for instance consider the Taub-NUT geometry and determine its generalized K\"ahler potential.
The generalized K\"ahler potential for the Eguchi-Hanson geometry can also be obtained from a generalized quotient of Euclidean $8$-dimensional space by a $U(1)$-isometry and in this setting turns out to be exactly \eqref{potentialEH2} as well \cite{marcos}.
$2D$ $\mathcal{N}=(2,2)$ sigma models have a target space that is a generalized K\"ahler manifold and in general, they are described by chiral, twisted chiral and semichiral superfields. The models with hyperk\"ahler target space and $J_+\neq \pm J_-$ are described purely in terms of semichiral superfields. These models do not in general admit off-shell $\mathcal{N}=(4,4)$ supersymmetry, since it was shown in \cite{Malin1} and \cite{Malin2} that a sigma model parametrized by semichiral fields can only be extended to off-shell $\mathcal{N}=(4,4)$ supersymmetry if the target space is $4n$-dimensional with $n>1$. However, they are always dual to models with $\mathcal{N}=(4,4)$ supersymmetry that are parametrized by chiral and twisted chiral superfields \cite{sevrin}. The exact relation between the $\mathcal{N}=(4,4)$ sigma models described by chiral and twisted chiral superfields, and their dual semichiral models will be described in \cite{Malin3}.
The relation between the coordinates $x_{L/R}$, $y_{L/R}$ and $\Upsilon$, $\tilde{\Upsilon}$ has been obtained in this paper from a purely differential geometric approach. $x_{L/R}$, $y_{L/R}$ describe left- and right-semichiral superfields in $2D$ $\mathcal{N}=(2,2)$ sigma models. For a target space that is hyperk\"ahler, these models are dual to models with $\mathcal{N}=(4,4)$ supersymmetry that are parametrized by chiral and twisted chiral superfields. The coordinates $\Upsilon(\zeta)$, $\tilde{\Upsilon}(\zeta)$ however describe arctic superfields in $\mathcal{N}=(4,4)$ sigma models in projective superspace. The field theoretical interpretation and understanding of this relation between arctic $\mathcal{N}=(4,4)$ models and the semichiral models that are dual to $\mathcal{N}=(4,4)$ models remains an open problem. The complex coordinate $\zeta$ is an auxiliary variable that gets integrated out in projective superspace, but for ordinary superspace it is just a constant parameter. Thus in our relation, an arctic model corresponds to a two-sphere (or more precisely to a cylinder) of presumably equivalent semichiral models.
In this paper, we mainly focused on the special case, where the bihermitian structure $(J_+,J_-)$ only depends on one complex parameter $\zeta$ and established a relation to projective superspace. In the appendix, we show that the results from section \ref{sectionstructures} can be generalized to the full $S^2\times S^2$-family of generalized complex structures on a hyperk\"ahler manifold parametrized by two complex parameters $\zeta_+$ and $\zeta_-$. Many hints point towards an intimate relation of this formulation to doubly-projective superspace \cite{doublyProjective},\cite{doublyProjective2}.
\bigskip\bigskip
\noindent{\bf\Large Acknowledgements}:
\bigskip
\noindent
The author owes a lot to Martin Ro\v cek, who initiated the project and kept it alive, providing motivation and knowledge. The author would like to thank P. Marcos Crichigno for valuable discussions during the creation of this paper. Ulf Lindstr\"om and Malin G\"oteman have provided comments helping to put the paper into its final form.
|
\section{Introduction}
Counting the number of solutions for random constraint satisfaction
problems is a very important and nontrivial problem which belongs to
\#P-complete class in computational complexity~\cite{Valiant-1979}
and is much harder than determining whether a random formula has any
solutions. In practice, we can only sample a very small part of a
huge solution space which contains an exponential number of
solutions. However, can we predict the number of solutions in the
whole solution space only based on a finite number of sampled
solutions? This issue has generated broad interests across a variety
of different disciplines such as computer science, probabilistic
reasoning, statistical physics and computational
biology~\cite{Roth-1996,Selman-2007,Kroc-2008,Favier-2009,Montanari-2009,Nature-06,Bialek-09ep,Cocco-2011,Opper-2011}.
An efficient sample-based counting strategy was proposed in
Ref.~\cite{Selman-2007}. This strategy successively sets the most
balanced variable until an exact counter is feasible on the reduced
formula, and provides a lower bound on the true count.
Alternatively, we address the solution counting problem within the
framework of inverse Ising problem in examples of diluted
models\textemdash random $K$-SAT problems and fully-connected
model\textemdash binary perceptron, and show that our method yields
an estimate whose value could be very close to the true count when
the constraint density is small. The constraint density is defined
as the ratio of the number of constraints to that of variables in
the system.
The inverse Ising problem~\cite{Ack-1985} has recently attracted
much attention not only in the development of fast mean field
inverse algorithms~\cite{Mezard-09,Roudi-2009,SM-09,Cocco-2011} but
also in modeling vast amounts of biological
data~\cite{Nature-06,Weigt-2009,Cocco-09,Mora-2010}. The pairwise
Ising model is able to capture most of the correlation structure of
the real neuronal network activity and is much more informative than
the independent model where each neuron is assumed to fire
independently~\cite{Nature-06,Tang-2008}. The observed collective
behavior of a large neuronal network results from interactions of
many individual neurons. The joint activity patterns for a retina
under naturalistic stimuli were reported to convey information about
the visual stimuli~\cite{Tkacik-2010,Bialek-09ep}. Estimating the
information stored by the real neuronal network directly from data
remains an open and important issue. We show in this work the
information can be estimated reliably and sizes of metastable states
for the neuronal network can also be predicted.
The paper is organized as follows. The inverse Ising problem is
introduced in Sec.~\ref{InvIsing}, together with a brief description
of the susceptibility propagation algorithm used to infer the
disordered Ising model. In Sec.~\ref{Method}, we present the belief
propagation to estimate the entropy from the data and apply this
method to predict the entropies of four different examples only from
a limited number of samplings. Finally, the conclusion suggests some
implications of our study as well as potential applications of the
presented methodology.
\section{the inverse Ising problem}
\label{InvIsing}
For a system of $N$ variables, one can collect $P$ configurations or solutions
$\{\sigma_{i}^{\nu}\}(i=1,\ldots,N;\nu=1,\ldots,P)$ either from real
biological experiments (e.g., spike trains in multi-electrode array
recordings~\cite{Nature-06}) or from random walks in the solution
space of a model. We assume $\sigma_{i}$ takes Ising-type value
$\pm1$. The task of the inverse Ising problem is to find couplings
$\{J_{ij}\}$ and fields $\{h_{i}\}$ to construct a minimal model
\begin{equation}\label{Ising}
P_{{\rm
Ising}}(\boldsymbol\sigma)=\frac{1}{Z}\exp\left[\sum_{i<j}J_{ij}\sigma_{i}\sigma_{j}+\sum_{i}h_{i}\sigma_{i}\right]
\end{equation}
such that its magnetizations and pairwise correlations are
compatible with those measured, i.e., $\left<\sigma_{i}\right>_{{\rm
Ising}}=\left<\sigma_{i}\right>_{{\rm data}},
\left<\sigma_{i}\sigma_{j}\right>_{{\rm
Ising}}=\left<\sigma_{i}\sigma_{j}\right>_{{\rm data}}$. $Z$ is the
partition function and the inverse temperature $\beta=1$ as it can
be absorbed in the strength of couplings and fields. Hereafter, we
define the measured magnetization and connected correlation as
$m_{i}\equiv\left<\sigma_{i}\right>_{{\rm data}}$ and
$C_{ij}\equiv\left<\sigma_{i}\sigma_{j}\right>_{{\rm
data}}-m_{i}m_{j}$ respectively where $\left<\cdots\right>_{{\rm
data}}$ denotes the average over the sampled configurations or
solutions.
We use susceptibility propagation (SusProp) to infer the couplings
and fields. SusProp passes messages along the oriented edges of the
network by iterative updating. To run SusProp, two kinds of messages
are needed. One is the cavity magnetization of variable $i$ in the
absence of variable $j$ denoted as $m_{i\rightarrow j}$; the other
is the cavity susceptibility $g_{i\rightarrow j,k}$ that is the
response of cavity field of variable $i$ without variable $j$ to a
local perturbation of external field of variable
$k$~\cite{Mezard-09}. The update rule can be derived using belief
propagation Eq.~(\ref{bpIsing}) and fluctuation-response
relation~\cite{Mezard-09,Marinari-2010,Huang-2010b} and reads as
follows~\cite{Huang-2010b}:
\begin{subequations}\label{SusP}
\begin{align}
m_{i\rightarrow j}&=\frac{m_{i}-m_{j\rightarrow i}\tanh J_{ij}}{1-m_{i}m_{j\rightarrow i}\tanh J_{ij}}\\
g_{i\rightarrow j,k}&=\delta_{ik}+\sum_{l\in \partial i\backslash
j}\frac{1-m_{l\rightarrow i}^{2}}{1-(m_{l\rightarrow i}\tanh
J_{li})^{2}}
\tanh J_{li}g_{l\rightarrow i,k}\\
J_{ij}^{{\rm
new}}&=\frac{\epsilon}{2}\log\left(\frac{(1+\tilde{C_{ij}})(1-m_{i\rightarrow
j}m_{j\rightarrow i})}
{(1-\tilde{C_{ij}})(1+m_{i\rightarrow j}m_{j\rightarrow i})}\right)+(1-\epsilon)J_{ij}^{{\rm old}}\\
\tilde{C_{ij}}&=\frac{C_{ij}-(1-m_{i}^{2})g_{i\rightarrow
j,j}}{g_{j\rightarrow i,j}}+m_{i}m_{j}
\end{align}
\end{subequations}
where $\partial i\backslash j$ denotes neighbors of variable $i$
except $j$, $\delta_{ik}$ is the Kronecker delta function and
$\epsilon\in[0,1]$ is introduced as a damping factor and should be
appropriately chosen to prevent the absolute updated $\tanh(J_{ij})$
from being larger than $1$. In practice, all couplings are initially
set to be zero and for every directed edge of the network, the
message $m_{i\rightarrow j}$ is randomly initialized in the interval
$[-1.0,1.0]$ and $g_{i\rightarrow j,k}=0$ if $i\neq k$ and $1.0$
otherwise. The SusProp rule Eq.~(\ref{SusP}) is then iterated until
either the inferred couplings converge within a predefined precision
$\eta^{(1)}$ or the preset maximal number of iterations $\mathcal
{T}_{{\rm max}}^{(1)}$ is exceeded. After the set of couplings is
obtained, the fields are inferred via $
h_{i}=\tanh^{-1}(m_{i})-\sum_{l\in\partial i}\tanh^{-1}\left[\tanh
J_{li}m_{l\rightarrow i}\right]$.
To ensure a reliable estimate of the parameters, we define a
convergence fraction $R$ as the ratio of the number of converged
couplings to the total number of edges in the network. In the
non-convergent case, we take the inferred parameters corresponding
to $R_{\rm max}=\max\{R_{t},t=1,\ldots,\mathcal {T}_{{\rm
max}}^{(1)}\}$ where $R_{t}$ is the convergence fraction of $t$-th
iteration. $R_{{\rm max}}=1.0$ if the update rule converges. For an
inverse problem, $\{m_{i},C_{ij}\}$ serve as inputs to the update
rule, and they are computed from $P$ sampled solutions or
configurations. We use stochastic local search algorithms to sample
the solution space of random $K$-SAT ($K=2,3$ here) formulas and
that of the binary perceptron. For the retinal network, the
configurations were obtained from the spike trains in the
multi-electrode recording experiments (data courtesy of Gasper
Tkacik, Refs.~\cite{Nature-06,Bialek-09ep}).
\section{Estimating the entropy from the data}
\label{Method}
We derive the entropy of the constructed Ising model
Eq.~(\ref{Ising}) under Bethe approximation (also called cavity
method~\cite{cavity-2001}) assuming sufficiently weak interactions
among variables. We compute the entropy through site contributions
$\Delta S_{i}$ and edge contributions $\Delta S_{ij}$ as $S_{{\rm
Ising}}=Ns_{{\rm Ising}}=\sum_{i}\Delta
S_{i}-\sum_{\left<ij\right>}\Delta S_{\left<ij\right>}$:
\begin{widetext}
\begin{subequations}\label{entropy}
\begin{align}
\begin{split}
\Delta S_{i}&=\log Z_{i}-\frac{1}{Z_{i}}\Biggl[h_{i}e^{h_{i}}\prod_{l\in\partial i}\cosh J_{li}(1+\tanh J_{li}m_{l\rightarrow
i})-h_{i}e^{-h_{i}}\prod_{l\in\partial i}\cosh J_{li}(1-\tanh J_{li}m_{l\rightarrow i})\\
&+e^{h_{i}}\sum_{l\in\partial i}\Bigl[J_{li}\sinh J_{li}(1+\tanh J_{li}m_{l\rightarrow i})+J_{li}\cosh J_{li}(1-\tanh^{2}J_{li})m_{l\rightarrow i}\Bigr]
\cdot\prod_{j\in\partial i\backslash l}\cosh J_{ij}(1+\tanh J_{ij}m_{j\rightarrow i})\\
&+e^{-h_{i}}\sum_{l\in\partial i}\Bigl[J_{li}\sinh J_{li}(1-\tanh J_{li}m_{l\rightarrow i})-J_{li}\cosh J_{li}(1-\tanh^{2}J_{li})m_{l\rightarrow i}\Bigr]
\cdot\prod_{j\in\partial i\backslash l}\cosh J_{ij}(1-\tanh
J_{ij}m_{j\rightarrow i})\Biggr]
\end{split}\\
\Delta S_{\left<ij\right>}&=\log Z_{ij}-J_{ij}\frac{\tanh J_{ij}+m_{i\rightarrow j}m_{j\rightarrow i}}{1+\tanh J_{ij}m_{i\rightarrow j}m_{j\rightarrow i}}
\end{align}
\end{subequations}
\end{widetext}
where $\partial i\backslash l$ denotes neighbors of variable $i$
except $l$. $Z_{i}=e^{h_{i}}\prod_{l\in\partial i}\cosh
J_{li}(1+\hat{m}_{l\rightarrow i})+e^{-h_{i}}\prod_{l\in\partial
i}\cosh J_{li}(1-\hat{m}_{l\rightarrow i})$ and $Z_{ij}=\cosh
J_{ij}\left(1+\tanh J_{ij}m_{i\rightarrow j}m_{j\rightarrow
i}\right)$. The cavity magnetization $m_{i\rightarrow j}$ obeys
simple recursive equations:
\begin{subequations}\label{bpIsing}
\begin{align}
m_{i\rightarrow j}&=\frac{e^{h_{i}}\prod_{l\in\partial i\backslash j}(1+\hat{m}_{l\rightarrow i})-e^{-h_{i}}\prod_{l\in\partial i\backslash j}(1-\hat{m}_{l\rightarrow i})}
{e^{h_{i}}\prod_{l\in\partial i\backslash j}(1+\hat{m}_{l\rightarrow i})+e^{-h_{i}}\prod_{l\in\partial i\backslash j}(1-\hat{m}_{l\rightarrow i})}\\
\hat{m}_{l\rightarrow i}&=\tanh J_{li}m_{l\rightarrow i}
\end{align}
\end{subequations}
We first randomly initialize $m_{i\rightarrow j}\in[-1.0,1.0]$ for
every directed edge of the reconstructed network, then iterate
Eq.~(\ref{bpIsing}) until all messages converge within the precision
$\eta^{(2)}$ or the maximal number of iterations $\mathcal {T}_{{\rm
max}}^{(2)}$ is reached. From the fixed point, the entropy can be
computed via Eq.~(\ref{entropy}). The case where some variable, say
$i$ is positively frozen, i.e., corresponding measured magnetization
$m_{i}=1.0$, can also be handled. In this case, $h_{i}=+\infty$, and
$\Delta S_{i}$ is reduced to be $\log
Z_{i}^{'}-\frac{Z_{i}^{''}}{Z_{i}^{'}}$ where
$Z_{i}^{'}=\prod_{l\in\partial i}\cosh J_{li}\left(1+\tanh
J_{li}m_{l\rightarrow i}\right)$ and $Z_{i}^{''}=\sum_{l\in\partial
i}\Bigl[J_{li}\sinh J_{li}(1+\tanh J_{li}m_{l\rightarrow
i})+J_{li}\cosh J_{li}(1-\tanh^{2}J_{li})m_{l\rightarrow
i}\Bigr]\cdot\prod_{j\in\partial i\backslash l}\cosh J_{ij}(1+\tanh
J_{ij}m_{j\rightarrow i})$. The edge contribution remains unchanged.
The negatively frozen case is similarly treated. In numerical
simulations, we adopt $\mathcal {T}_{{\rm max}}^{(2)}=500$,
$\eta^{(2)}=10^{-4}$, $\mathcal {T}_{{\rm max}}^{(1)}=2000$.
$\eta^{(1)}$ as well as $\epsilon$ depends on the following specific
applications.
We remark here that Eq.~(\ref{entropy}) is used specifically for the
solution counting problem where we now have known the magnetizations
and correlations and additionally some frozen cases (some $m_{i}=+1$
or $-1$) should be treated. On the other hand, the coupling or field
distributions depend on the collected data and Eq.~(\ref{entropy})
is derived only under the weakly-coupled approximation but the fully
connected topology is reserved. The first point is, the entropy we
try to estimate is not only for two-body interaction system (e.g.,
random $2$-SAT) but also for three-body interaction system and
densely-interacted system (e.g., the binary perceptron where each
constraint involves all variables of the system). The second point
is, the sampled solutions come from the zero energy ground state and
the sampling process is always confined in a single cluster
(solutions in it are connected with each other by single variable
flips).
All underlying parameters of pairwise Ising model are predicted
directly from the observed data and the entropy of the original
model is estimated based on the constructed Ising model. We
emphasize here that two layers of approximations are made. The first
one is the disordered Ising model Eq.~(\ref{Ising}) is used to
approximate the original model. When estimating the entropy from the
data, we actually do not know the original model. The second layer
is we use mean-field methods, specifically the message passing
algorithms to infer the underlying parameters of the pairwise Ising
model. Since the computational complexity of SusProp is $\mathcal
{O}(N^{3})$ for the fully-connected network, we focus on small size
networks with $N$ of order $\mathcal {O}(10^{2})$. When the
constraint density is small, the efficiency of our methodology is
supported by two concrete examples: random $K$-SAT problem and the
binary perceptron. For these two examples, we use $s_{{\rm true}}$
to represent the entropy density computed by belief propagation with
the knowledge of the original model (for details, see
Ref.~\cite{Zhou-2009} for random $K$-SAT problem and
Ref.~\cite{Zecchina-2006} for binary perceptron). To show the
efficiency of the pairwise Ising model, we also compute the
independent entropy $S_{{\rm ind}}=Ns_{{\rm
ind}}=-\sum_{i}\left[\frac{1+m_{i}}{2}\log\frac{1+m_{i}}{2}+\frac{1-m_{i}}{2}\log\frac{1-m_{i}}{2}\right]$
assuming $P(\boldsymbol\sigma)\approx\prod_{i}P_{i}(\sigma_{i})$.
For retinal network, we could neither know the true model underlying
the network nor get the true value for the entropy (when the network
is large). Therefore we just compare the result obtained by our
current fast belief propagation with that obtained by time-consuming
Monte Carlo method and show that the belief propagation not only
reproduces the entropy value evaluated by Monte Carlo method but
also yields rich information about the metastable states which are
relevant for neuronal population
coding~\cite{Bialek-09ep,Tkacik-2010}. In this case, we denote
$s_{{\rm BP}}$ as the entropy density estimated by belief
propagation and $s_{{\rm MC}}$ estimated by Monte Carlo method. Note
that both belief propagation and Monte Carlo method under the
reconstructed Ising model yield approximate value for the true
entropy since we consider only up to second-order correlations in
the observed data while the system may develop higher-order
correlations in its solution space or energy landscape. The
different natures of these examples imply wide applications of our
methodology to evaluate the entropy of an unknown model with only a
limited number of samplings.
\begin{figure}
\includegraphics[bb=9 14 311 220,scale=0.85]{figure01a.eps}
\hskip .2cm
\includegraphics[bb=14 15 252 145,scale=1.0]{figure01b.eps}\vskip .2cm
\caption{(Color online)
The quality of the pairwise and independent models versus constraint density $\alpha$. (a) Entropy density difference versus $\alpha$.
The data points connected by dashed line are the differences between $s_{{\rm ind}}$
and $s_{{\rm true}}$, while those connected by solid line are the differences $s_{{\rm Ising}}-s_{{\rm true}}$. The number of variables $N=100$ for random
$K$-SAT (r$2$-SAT or r$3$-SAT)
problem and $101$ for binary perceptron (bperc). $s_{{\rm true}}$ is
computed with the knowledge of the original model using belief
propagation~\cite{Zhou-2009,Zecchina-2006}.
Each point represents the average over eight
random samples. (b) Scatter
plot comparing $s_{{\rm Ising}}$ with $s_{{\rm true}}$. The full line indicates equality.
}\label{compa}
\end{figure}
\subsection{Random $K$-SAT problem}
\label{sat}
The random $K$-SAT problem is finding a solution (an assignment of
$N$ boolean variables) satisfying a random formula composed of
logical AND of $M$ constraints~\cite{Monasson-1996}. Each constraint
is a logical OR function of $K$ randomly chosen distinct variables
(either directed or negated with equal probability). The constraint
density $\alpha=M/N$. For $K=2$, the threshold separating a SAT
phase from an UNSAT phase was confirmed to be $\alpha_{s}=1$ below
which the solution space is ergodic and a simple local search
algorithm can easily identify a solution~\cite{Monasson-1996}. For
$K=3$, the estimated threshold $\alpha_{s}\simeq4.267$ below which
the solution space exhibits richer
structures~\cite{Krzakala-PNAS-2007}. The dynamical transition point
locates at $\alpha_{d}\simeq3.86$ and separates the ergodic phase
from non-ergodic phase. We use {\tt SEQSAT} algorithm of
Ref.~\cite{Zhou-2010epjb} to first find a solution for a given
$\alpha$, then $10^{8}N$ single variable flips are performed in the
current solution space, after that we perform random walks in the
current solution
space to sample one solution every $10^4$ steps. Each step
involves $N$ attempts to move from one solution to its adjacent one
by single variable flip, i.e., a randomly chosen variable is
flipped and if the new configuration is a solution, the flip is
accepted with probability $1/2$; otherwise the movement is
rejected. We sample totally $P=10^{5}$ solutions to estimate the entropy. We choose
$\epsilon=0.128,\eta^{(1)}=10^{-3}$ for random $2$-SAT and $\epsilon=0.002,\eta^{(1)}=10^{-4}$ for random
$3$-SAT. Results are reported in Fig.~\ref{compa}. When $\alpha$ is small, our method can predict the true entropy very well
especially for $K=2$ which can be actually transformed into a pairwise Ising model. As $\alpha$ increases, the difference between $s_{{\rm
Ising}}$ and $s_{{\rm
true}}$~\cite{Zhou-2009} becomes large and this deviation is more obvious for $K=3$,
which manifests the presence of higher-order
correlations in the solution space. At high
$\alpha$ (e.g., $\alpha=3.0$ for $K=3$), the belief propagation Eq.~(\ref{bpIsing}) would yield multiple fixed
points. This signals ergodicity breaking phenomenon in the
energy landscape of the constructed Ising model or indicates that long range correlations develop in the original system~\cite{Montanari-2009}, although our samplings are still confined in a
single cluster of the original model, as a result, the predicted entropy becomes rather inaccurate compared with the true one computed under the original model.
\subsection{Binary perceptron}
\label{bperc}
The binary perceptron with $N$ binary weights
connecting $N$ input nodes to a single output node performs a random
classification of $\alpha N$ random binary patterns
$\{\xi_{i}^{\mu}\}(i=1,\ldots,N;\mu=1,\ldots,\alpha N)$. The
critical constraint density $\alpha_{s}\simeq0.83$ below which the
solution space is non-empty~\cite{Krauth-1989}. Given an input
pattern $\boldsymbol{\xi}^{\mu}$, if the actual output $o^{\mu}={\rm
sgn}\left(\sum_{i=1}^{N}\sigma_{i}\xi_{i}^{\mu}\right)$ is equal to
the desired output $o_{0}^{\mu}$ assigned a value $\pm 1$ with equal
probabilities, the configuration $\boldsymbol{\sigma}$ learns this
pattern. The solution space of the binary perceptron consists of all
configurations learning $\alpha N$ random patterns. Before sampling,
we first learn $\alpha N$ patterns using {\tt DWF}
algorithm of Ref.~\cite{Huang-2010jstat}. The sampling procedure is
the same as that used for random $K$-SAT problems. In numerical
simulations, we choose $\epsilon=0.001,\eta^{(1)}=10^{-4}$. The
deviation of estimated $s_{{\rm Ising}}$ from
$s_{{\rm true}}$~\cite{Zecchina-2006} is plotted against $\alpha$. For
small $\alpha$, our method can predict the true entropy well
without the knowledge of the original model. The large deviation shown in Fig.~\ref{compa}
at high $\alpha$ implies higher-order correlations start to dominate the solution space.
\begin{figure}
\includegraphics[bb=11 13 245 137,scale=1.0]{figure02a.eps}
\hskip .1cm
\includegraphics[bb=11 16 282 216,scale=0.8]{figure02b.eps}\vskip .2cm
\caption{(Color online)
(a) Histograms of inferred couplings and fields (inset) for the
retinal network with the number of neurons $N=40$ (data
courtesy of Gasper Tkacik, Refs.~\cite{Nature-06,Bialek-09ep}).
To infer the network, we choose
$\epsilon=0.128,\eta^{(1)}=10^{-3}$. The network is inferred at
$R_{{\rm max}}=0.959$. (b) Reconstructed $C_{ij}^{{\rm BP}}$ using belief propagation Eq.~(\ref{Corre}) versus the measured one.
We only show the case $i\neq j$ since $C_{ii}^{{\rm BP}}=1-(m_{i}^{{\rm
BP}})^{2}$. The full line indicates equality.
}\label{histo}
\end{figure}
\begin{table}
\caption{Estimated entropy density $s_{{\rm BP}}$ for the inferred
retinal network through belief propagation Eq.~(\ref{bpIsing}).
$q_{0}=\frac{1}{N}\sum_{i}m_{i}^{{\rm BP}}m_{i}$ where $m_{i}^{{\rm
BP}}$ is computed from the fixed point of belief propagation
Eq.~(\ref{bpIsing}). The self-overlap
$q_{1}=\frac{1}{N}\sum_{i}m_{i}^{2}\simeq0.906828$. The last column
gives the probability of appearance for each fixed point during
$1000$ runs of belief propagation with the same inferred
parameters.}\label{table:retina}
\begin{tabular}{c c c c}
\hline
\hline
$s_{{\rm BP}}$ & $q_{0}$ & prob. app \\
\hline
0.09771 & 0.906816 & 0.089\\ 0.1093 & 0.512411 & 0.075\\
0.1224 &0.313566 & 0.066\\ 0.1634 & 0.281124 & 0.400\\ 0.1765 & 0.021960 & 0.354\\
0.1918 & 0.264979 & 0.016\\
\hline
\hline
\end{tabular}
\end{table}
\subsection{Retinal network}
\label{retina} A recording of the activity of $40$ neurons in a
salamander retina under natural movie stimuli could also be analyzed
within the current setting. The total effective number of samplings
$P\simeq7\times10^{4}$~\cite{Bialek-09ep}. Our estimated entropy
density for the retinal network is $s_{{\rm BP}}\simeq0.09771$
consistent with that obtained by Gasper Tkacik
et.al~\cite{Bialek-09ep} using Monte Carlo method which produces an
estimate $s_{{\rm MC}}\simeq0.09479$ but is rather time consuming
for large $N$. Our result implies that the retina under the
naturalistic movie stimuli stores $e^{Ns_{{\rm BP}}}\simeq50$
effective configurations. If we approximate the true entropy using
that calculated by belief propagation (Eq.~(\ref{entropy}) and
Eq.~(\ref{bpIsing})) or Monte Carlo method, then the
multi-information measuring the total amount of correlations in the
network~\cite{Bialek-03} $I_{{\rm BP}}=s_{{\rm ind}}-s_{{\rm BP}}$
or $I_{{\rm MC}}=s_{{\rm ind}}-s_{{\rm MC}}$ where $s_{{\rm ind}}$
is the independent entropy density. The result is that the
difference between these two multi-information values $I_{{\rm
MC}}-I_{{\rm BP}}\simeq0.00292$. The histogram of inferred
parameters is shown in Fig.~\ref{histo} (a). Note that most of
predicted couplings concentrate around zero value with a long tail
of distribution for large negative couplings whose weights are
rather small. Most of the predicted fields are negative since most
of the neurons are silent across the movie presentations. Using the
same inferred parameters, we run belief propagation
Eq.~(\ref{bpIsing}) $1000$ times from different random
initializations. Several fixed points are found and one of them is
consistent with the previous result~\cite{Bialek-09ep} (see
Table~\ref{table:retina}). These fixed points represent different
metastable states and the entropy measures the capacity of neurons
to convey information about the visual stimulus which contains
high-order correlation structure. The visual information could be
encoded by identity of the basin of attraction~\cite{Tkacik-2010}
and these predicted metastable states may code for specific stimulus
features. Therefore, the information of the inputs to the retina can
be stored in the couplings and fields which generate a free energy
landscape with multiple metastable states for redundant error
correction~\cite{Tkacik-2010}. Future research on neuronal
population coding needs to elucidate this point.
In Fig.~\ref{histo} (b), we verify that the inferred pairwise Ising
model reproduces the measured connected correlations with very good
agreement. The reconstructed correlations $\{C_{ij}^{{\rm BP}}\}$
can be computed by the following message passing
algorithm~\cite{Higuchi-2010,Huang-2010b}:
\begin{subequations}\label{bpC}
\begin{align}
m_{i\rightarrow j}&=\frac{e^{h_{i}}\prod_{l\in\partial i\backslash j}(1+\hat{m}_{l\rightarrow i})-e^{-h_{i}}\prod_{l\in\partial i\backslash j}(1-\hat{m}_{l\rightarrow i})}
{e^{h_{i}}\prod_{l\in\partial i\backslash j}(1+\hat{m}_{l\rightarrow i})+e^{-h_{i}}\prod_{l\in\partial i\backslash j}(1-\hat{m}_{l\rightarrow i})}\\
\hat{m}_{l\rightarrow i}&=\tanh J_{li}m_{l\rightarrow i}\\
g_{i\rightarrow j,k}&=\delta_{ik}+\sum_{l\in \partial i\backslash
j}\frac{1-m_{l\rightarrow i}^{2}}{1-(m_{l\rightarrow i}\tanh
J_{li})^{2}} \tanh J_{li}g_{l\rightarrow i,k}
\end{align}
\end{subequations}
where two kinds of messages, $m_{i\rightarrow j}$ and
$g_{i\rightarrow j,k}$ are updated. Once both messages for each
directed link in the network are converged, i.e., iteration of
Eq.~(\ref{bpC}) reaches fixed point, we compute the predicted
connected correlations $\{C_{ij}^{{\rm BP}}\}$ via
\begin{subequations}\label{Corre}
\begin{align}
C_{ij}^{{\rm BP}}&=(\tilde{C_{ij}}-m_{i}m_{j})g_{j\rightarrow i,j}+(1-m_{i}^{2})g_{i\rightarrow j,j}\\
\tilde{C_{ij}}&=\frac{\tanh J_{ij}+m_{i\rightarrow j}m_{j\rightarrow i}}{1+\tanh J_{ij}m_{i\rightarrow j}m_{j\rightarrow i}}
\end{align}
\end{subequations}
where the Ising model is known and all messages needed to compute
$C_{ij}^{{\rm BP}}$ including $m_{i}$ and $m_{j}$ are read from the
fixed point. This message passing strategy to evaluate the
correlations is very fast and takes tens of iterations to converge.
Remarkably, the estimated magnetizations and correlations fit those
measured very well. However, using Monte Carlo samplings to
reconstruct the correlations, we failed to reproduce those measured.
For example, after sufficient thermalization, the configuration is
sampled every $10^{4}$ Monte Carlo sweeps, then the correlations are
computed with $10^{4}$ sampled configurations. The obtained root
mean square error is about $0.23$. Possible reason is, the
reconstructed Ising model by SusProp already develops multiple
states (different fixed points of belief propagation
Eq.~(\ref{bpIsing})), therefore, at temperature $T=1.0$ and $N=40$,
as observed in our simulations, Monte Carlo samplings can have
transitions between different states yielding the average energy
density of sampled configurations $\bar{e}\simeq-3.7119$ with
fluctuation of order $0.0724$ while the state reproducing the
measured correlations in Fig.~\ref{histo} (b) has typical energy
density $-3.62954$ but smallest entropy density $s_{{\rm
BP}}\simeq0.09771$ of all observed states. Actually, in this case,
the correlation computed by Monte Carlo samplings corresponds to the
average over different states while the measured data comes from one
state of the reconstructed Ising model.
\subsection{Convergence patterns and entropy density difference versus $P$}
\label{edp}
There exist three kinds of convergence patterns for different
iterations of SusProp rules. One is the convergence case shown by an
example of random $2$-SAT with $\alpha=0.7$; the second type is the
convergence fraction $R$ first increases then decreases (see in
Fig.~\ref{CovP} an example of binary perceptron with $\alpha=0.5$);
the last type is $R$ reaches a plateau with small fluctuations shown
by an example of retina.
\begin{figure}
\includegraphics[bb=14 15 273 206,scale=0.8]{figure03.eps}
\caption{(Color online)
Convergence patterns possibly appearing in the iteration of
SusProp update rules.
}\label{CovP}
\end{figure}
\begin{figure}
\includegraphics[bb=17 16 283 212,scale=0.85]{figure04a.eps}
\hskip .2cm
\includegraphics[bb=14 15 300 214,scale=0.85]{figure04b.eps}\vskip .2cm
\caption{(Color online)
(a) Entropy density difference versus $P$. The error bar shows the fluctuation across eight random samples. (b) The distribution
of inferred couplings for r$3$-SAT with $\alpha=1.0$ and
$N=100$. Two results for $P=500$ and $P=12500$ are shown.
}\label{Edf}
\end{figure}
In Fig.~\ref{Edf}(a), the influence of the number of samplings $P$ on
the estimation of entropy from the finite samplings is shown. It
seems that the entropy density difference decreases as $P$ becomes
small. Note that we construct the pairwise Ising model based on the
samplings from the ground state (zero energy for these three
constraint satisfaction problems) and the underlying graphical
model (e.g., r$3$-SAT or bperc) may not be a pairwise Ising model.
Our samplings are always confined in a single solution cluster and
the quality may depend on the fine structure of the solution space~\cite{Zhou-2009pre,Hartmann-2010,Huang-2011epl}.
This case is different from studies on the reconstruction of
Sherrington-Kirkpatrick model at high
temperatures~\cite{Marinari-2010}. In our current setting, when the
sampling number $P$ decreases, the collected data seems to have
more correlations (this may be induced by the statistical errors) and the SusProp turns out to be not converged any more (e.g., in the case of r$2$-SAT with $\alpha=0.3$), which
can be justified from the Fig.~\ref{Edf}(b) that the inferred
coupling distribution becomes broader with decreasing $P$ and thus the estimated entropy should take smaller values.
As observed in Fig.~\ref{Edf}(b), once $P$ decreases down to some value, the estimated entropy value would underestimate the true one computed with the knowledge of
the original model. On the other hand, given small $P$, the computed magnetizations and correlations
would have large statistical errors and may not correctly reflect
the correlations in the sampled solution cluster. Safely, we
select $P=10^{5}$ to reduce the statistical error for calculating
the magnetizations and correlations.
\section{Conclusion}
\label{Conc}
In this work, we address the important problem of
counting the total number of solutions (configurations) based on a
limited number of sampled solutions (configurations). We formulate
the solution (configuration) counting problem within the framework
of inverse Ising problem, and this idea is tested on both diluted
models and fully-connected models, as well as on real neural data.
In the first case, we do not know a priori the underlying graphical
models and try to construct a disordered Ising model from the
collected data to evaluate the entropy of those unknown models. Note
that the sampled solutions come from the zero energy ground state
(single solution cluster) and the number is limited to $P$ but the
size of that sampled cluster is evaluated. To this end, the pairwise
model improves substantially the independent model (see
Fig.~\ref{compa}(a)). When the constraint density is small then the
pairwise correlation dominates the solution space, the estimated
entropy gets very close to the true one estimated with the knowledge
of the original model. Another interesting point to be demonstrated
in our further work is, for small $N$, one can compute the
magnetizations and correlations through exact enumeration and
further the real entropy value. The result of Boltzmann learning
algorithm~\cite{Ack-1985} and Monte Carlo
simulation~\cite{Bialek-09ep} (using a large enough amount of Monte
Carlo samplings) should provide an upper bound on the true entropy.
Instead, using the approximate method we proposed, whether the
obtained result provides a bound should be checked for small size
system or proved in the limit of large $N$, provided that the
magnetizations and correlations are less noisy. In the second case,
the susceptibility propagation is applied to infer the retinal
network and belief propagation is used to reproduce the entropy
computed by Monte Carlo method. This message passing scheme is very
fast and efficient especially for large network and the observed
multiple fixed points predict other metastable states in the
inferred retinal network. These metastable states may have intimate
relation with the neuronal population coding~\cite{Tkacik-2010}.
Extensions to the neuronal interaction network organized in a
hierarchical and modular manner would be very
interesting~\cite{Nature-10,Ganmor-11}. Our presented framework
constructs a statistical mechanics description of the system
directly from either artificial data or real data, and has the
potential to describe biological networks more generally and
estimate the size of the solution space in various contexts
especially when pairwise correlation dominates the system.
\section*{Acknowledgments}
We thank Gasper Tkacik for providing us multielectrode recordings of
the salamander retina. The improvement of the manuscript benefited
from comments and suggestions of anonymous referees. The present
work was partially supported by the NSFC Grant 10834014 and the
973-Program Grant 2007CB935903 and HKUST 605010.
|
\section{Introduction}
A lattice $\Lambda=\Lambda^{n} \subseteq \mathbb{R}^n$ is a discrete
set generated by integer combinations of $n$ linearly independents
vectors ${\bm v_1},\ldots,{\bm v_n} \in \mathbb{R}^n$. Its packing
density $\Delta(\Lambda)$ is the proportion of the space
$\matR^{n}$ covered by congruent disjoint spheres of maximum radius
\cite{sloane}. A lattice $\Lambda$ has diversity $m\leq n$ if $m$ is
the maximum number such that for all ${\bm y}=(y_1,\cdots,y_n) \in
\Lambda$, ${\bm y} \neq {\bm 0}$ there are at least $m$
non-vanishing coordinates. Given a full diversity lattice $\Lambda
\subseteq \matR^{n}$ $(m=n)$, the minimum product distance is
defined as $d_{min}(\Lambda) = \min\{\prod_{i=1}^{n}|y_i|\,\,
\mbox{for all}\,\, {\bm y}=(y_1,\cdots,y_n) \in \Lambda,{\bm y}
\neq {\bm 0} \}$ \cite{Oggier}.
Signal constellations having lattice structure have been studied as
meaningful means for \linebreak signal transmission over both
Gaussian and single-antenna Rayleigh fading channel \cite{boutros}.
\linebreak Usually the problem of finding good signal constellations
for a Gaussian channel is associated to the search for lattices
with high packing density \cite{sloane}. On the other hand for a
Rayleigh fading channel the efficiency, measured by lower error
probability in the transmission, is strongly related to the lattice
diversity and minimum product distance \cite{boutros},
\cite{Oggier}. The approach in this work, following \cite{eva2} and
\cite{Oggier} is the use of algebraic number theory to construct
lattices with good performance for both channels.
For general lattices the packing density and the minimum product
distance are usually hard to estimate \cite{mic}. Those parameters
can be obtained in certain cases of lattices associated to number
fields, through algebraic properties.
In \cite{gabriele}, \cite{Oggier} and \cite{upper-bound} some
families of rotated $\matZ^{n}$-lattices with full diversity and
good minimum product distance are studied for transmission over
Rayleigh fading channels. In \cite{suarez} the lattices $A_{p-1}$,
$p$ prime, $E_6$, $E_8$ $K_{12}$ and $\Lambda_{24}$ were realized as
full diversity ideal lattices via some subfields of cyclotomic
fields. In \cite{boutros} rotated $n$-dimensional lattices
(including $D_4$, $K_{12}$ and $\Lambda_{16}$) which are good for
both channels are constructed with diversity $n/2$.
In this work we also attempt to consider lattices which are feasible
for both channels by constructing rotated $D_n$-lattices with full
diversity $n$ and get a closed-form for their minimum product
distance. The results were obtained for $n=2^{r-2},$ $r \geq 5$ and
$n=(p-1)/2$, $p$ prime and $p\geq 7,$ in Propositions \ref{Prop1},
\ref{idealI} and \ref{rotacionado}. As it is known, a $D_n$ lattice
has better packing density $\delta(D_n)$ when compared to
$\matZ^{n}$ ($D_n$ has the best lattice packing density for
$n=3,4,5$ and $\lim_{n\longrightarrow
\infty}\frac{\delta(\mattZ^{n})}{\delta(D_n)} =0$) and also a very
efficient decoding algorithm \cite{sloane}. The relative minimum
product distances $d_{p,rel}(D_n)$ of the rotated $D_n$-lattices
obtained here are smaller than the minimum product distance
$d_{p,rel}(\matZ^{n})$ of rotated $\matZ^{n}$-lattices constructed
for the Rayleigh channels in \cite{and} and \cite{Oggier}, but, as
it is shown in Sections 4 and 5, $\lim_{n\longrightarrow
\infty}\frac{\sqrt[n]{d_{p,rel}(\mattZ^{n})}}{\sqrt[n]{d_{p,rel}(D_n)}}
=\sqrt{2}$, what offers a good trade-off.
In Sections 2 e 3 we summarize some definitions and results on
Algebraic Number Theory. Sections 4 and 5 are devoted to the
construction of full diversity rotated $D_n$-lattices through
cyclotomic fields and the deduction of their minimum product
distance.
\section{Number Fields}
In this section we summarize some concepts and results of algebraic
\linebreak number theory and establish the notation to be used from
now on. The results presented here can be found in \cite{traj},
\cite{pierre}, \cite{stewart} and \cite{was}.
Let $\matK$ be a number field of degree $n$ and $\mathcal
O_{\mattK}$ its ring of integers. It can be shown that every nonzero
fractionary ideal $I$ of $\mathcal O_{\mattK}$ is a free
$\matZ$-module of rank $n$.
There are exactly $n$ distinct $\matQ$-homomorphisms
$\{\sigma_i\}_{i=1}^{n}$ of $\matK$ in $\matC.$ A homomorphism
$\sigma_i$ is said {\it real} if $\sigma_i(\matK)\subset \matR$, and
the field $\matK$ is said {\it totally real} if $\sigma_i$ is real
for all $i=1,\cdots,n.$
Given $x \in \matK,$ the values $N(x) = N_{\mattK|\mattQ}(x) =
\prod_{i=1}^{n} \sigma_i(x), \ \ \ Tr(x)=Tr_{\mattK|\mattQ}(x) =
\sum_{i=1}^{n} \sigma_i(x)$ are called, {\it norm} and {\it trace}
of $x$ in $\matK|\matQ,$ respectively. It can shown that if $x \in
\mathcal{O}_{\mattK}$, then $N(x), Tr(x) \in \matZ$.
Let $\{\omega_1,\ldots,\omega_n\}$ be a $\matZ$-basis of $\mathcal
O_{\mattK}.$ The integer $d_{\mattK} = (det[
\sigma_j(\omega_i)]_{i,j=1}^{n})^2$ is called the {\it discriminant}
of $\matK$.
The {\it norm} of an ideal $I \subseteq \mathcal O_{\mattK}$ is
defined as $N(I) = |\mathcal O_{\mattK}/I|.$
The {\it codifferent} de $\matK|\matQ$ is the fractionary ideal
$\Delta(\matK|\matQ)^{-1}=\{x\in \matK; \,\, \forall \, \alpha \in
{\mathcal{O}}_{\mattK},\,Tr_{\mattK|\mattQ}(x\alpha) \in \matZ\}$ of
${\mathcal{O}}_{\mattK}.$
Let $\zeta=\zeta_m \in \matC$ be a primitive $m$-th root of unity.
We consider here the {\it cyclotomic field} $\matQ(\zeta)$ and its
subfield $\matK=\matQ(\zeta+\zeta^{-1}).$ We have that
$[\matQ(\zeta+\zeta^{-1}):\matQ]= \varphi(m)/2$, where $\varphi$ is
the Euler function; $\mathcal O_{\mattK} = \matZ
[\zeta+\zeta^{-1}];$ $d_{\mattK}=p^{\frac{p-3}{2}}$ if $m=p$, $p$
prime, $p\geq 5$ and $d_{\mattK}=2^{(r-1)2^{r-2}-1}$ if $m=2^r$.
\section{Ideal lattices}
From now on, let $\matK$ be a totally real number
field. Let $\alpha \in \matK$ such that $\alpha_i=\sigma_i(\alpha)
> 0$ for all $i=1,\cdots,n.$ The homomorphism
$$\left.\begin{array}{l} \sigma_{\alpha}:\matK \longrightarrow
\matR^{n} \\
\hspace{0.8cm} x\longmapsto
\left(\sqrt{\alpha_{1}}\sigma_1(x),\ldots,\sqrt{\alpha_{n}}\sigma_{n}(x)\right)\end{array}
\right.$$ is called {\it twisted homomorphism}. When $\alpha=1$ the
twisted homomorphism is the {\it Minkowski homomorphism}.
It can be shown that if $I\subseteq {\matK}$ is a free
$\matZ$-module of rank $n$ with $\matZ$-basis $\{w_1,\ldots,w_n\}$,
then the image $\Lambda = \sigma_{\alpha}(I)$ is a lattice in
$\matR^n$ with basis $\{{ \sigma_{\alpha}(w_1)},\ldots,{
\sigma_{\alpha}(w_n)}\},$ or equivalently with generator matrix
${\bm M}= (\sigma_{\alpha}(w_{ij}))_{i,j=1}^{n}$ where $w_i =
(w_{i1,\cdots,w_{in}})$ for all $i=1,\cdots,n$. Moreover, if
$\alpha I \overline{I} \subseteq \Delta(\matK|\matQ)^{-1}$ where
$\overline{I}$ denote the complex conjugation of $I,$ then
$\sigma_{\alpha}(I)$ is an integer lattice. Since $\matK$ is totally
real, the associated Gram matrix of $\sigma_{\alpha}(I)$ is ${\bm
G}={\bm M}.{\bm M^{t}} = \left( Tr_{\mattK|\mattQ}(\alpha
w_{i}\overline{w_{j}}) \right)_{i,j=1}^{n}$ \cite{Oggier}.
\begin{proposition}\label{detb}\cite{eva2} If $I\subseteq \matK$ is a fractional ideal, then for $\Lambda=\sigma_{\alpha}(I)$ and $det(\Lambda)=det(G)$, we have:
\begin{equation} \label{detb1} det(\Lambda)= det(G) = N(I)^{2}
N_{\mattK|\mattQ}(\alpha)|d_{\mattK}|.\end{equation}
\end{proposition}
\begin{proposition} \label{distanciaminima}\cite{Oggier} Let $\matK$ be a totally real field number with $[\matK:\matQ]=n$ and $I \subseteq
{\matK}$ a fractional ideal. The {\it minimum product distance} of
$\Lambda =\sigma_{\alpha}(I)$ is
\begin{equation} d_{p,min}(\Lambda)= \sqrt{N_{\mattK|\mattQ}(\alpha)}min_{0\neq
y\in I}|N_{\mattK|\mattQ}(y)|.\end{equation} In particular, if $I$
is a principal ideal then $ d_{p,min}(\Lambda)=
\sqrt{\frac{det(\Lambda)}{|d_{\mattK}|}}.$
\end{proposition}
\begin{definition} \label{relative} The {\it relative minimum product distance} of $\Lambda,$ denoted by ${\bm
d_{p,rel}(\Lambda)}$, is the minimum product distance of a scaled
version of $\Lambda$ with unitary minimum norm vector.
\end{definition}
\section{Rotated $D_n$-lattices for $n=2^{r-2}$, $r \geq 5$ via
$\matK=\matQ(\zeta_{2^{r}}+\zeta_{2^{r}}^{-1})$}
In this section we will present some families of rotated
$D_n$-lattices using ideals and modules in the totally real number
field $\matK=\matQ(\zeta_{2^{r}}+\zeta_{2^{r}}^{-1}).$ One of the
strategies to construct these lattices was to start from the
standard characterization of $D_n$ as generated by the basis
\begin{equation}\label{beta} \beta= \{(-1,-1,0,\cdots,0),(1, -1,0,
\cdots,0),\cdots,(0,0,\cdots,1,-1)\}. \end{equation} We derive in
4.2 a rotated $D_n$-lattice as a sublattice of the rotated
$\matZ^{n}$ algebraic constructions presented in \cite{and},
\cite{Oggier} and \cite{gabriele}. Another strategy explored next in
4.1 is to investigate the necessary condition given in Proposition
\ref{detb}, for the existence of rotated $D_n$-lattices.
Let $\zeta=\zeta_{2^r}$ be a primitive $2^r$-th root of unity, $m =
2^{r},$ $\matK=\matQ (\zeta +\zeta^{-1})$ and
$n=[\matK:\matQ]=2^{r-2}$.
\subsection{\bf A first construction:} Let $\alpha \in \mathcal O_{\mattK}$ and $I \subseteq \mathcal O_{\mattK}$ an ideal. If $\sigma_{\alpha}(I)$ is a
rotated $D_n$-lattice scaled by $\sqrt{c}$, then
$det(\sigma_{\alpha}(I))=4\, c^n$. Based on Proposition \ref{detb},
taking $I=\mathcal O_{\mattK}$ and $c=2^{r-1}$, since
$d_{\mattK}=2^{(r-1)2^{r-2}-1}$ and $n=2^{r-2}$ it follows that a
necessary condition to construct a rotated $D_n$-lattice
$\sigma_{\alpha}(I)$ is to find an element $\alpha \in \mathcal
O_{\mattK}$ such that $N(\alpha)=8$. Table $1$ shows some elements
$\alpha \in \mathcal O_{\mattK}$ such that $N(\alpha)=8$ in low
dimensions. From it we got the suggestion for a general expression
for $\alpha$ as \begin{equation} \label{alpha} \alpha=4 +
(\zeta_{2^r}+ {\zeta{_{2^r}^{-1}}})- 2({\zeta{_{2^r}^{2}}}+
{\zeta{_{2^r}^{-2}}})-({\zeta{_{2^r}^{3}}}+ {\zeta{_{2^r}^{-3}}})
\end{equation} and then derive Proposition \ref{Prop1}.
\begin{table}[h]
\begin{center}
\begin{tabular}{c|c|c} \hline \hline
$r $& $\alpha $& ${N}(\alpha) $ \\ \hline $4$ & $4 + (\zeta_{16}+
{\zeta{_{16}^{-1}}})- 2({\zeta{_{16}^{2}}}+
{\zeta{_{16}^{-2}}})-({\zeta{_{16}^{3}}}+ {\zeta{_{16}^{-3}}})$& $8$
\\ $5 $& $4 + (\zeta_{32}+ {\zeta{_{32}^{-1}}})-
2({\zeta{_{32}^{2}}}+ {\zeta{_{32}^{-2}}})-({\zeta{_{32}^{3}}}+
{\zeta{_{32}^{-3}}})$& $8$ \\ $6 $& $4 + (\zeta_{64}+
{\zeta{_{64}^{-1}}})- 2({\zeta{_{64}^{2}}}+
{\zeta{_{64}^{-2}}})-({\zeta{_{64}^{3}}}+ {\zeta{_{64}^{-3}}})$& $8$
\\ \hline \hline \end{tabular} \vspace{0.3cm} \caption{ }
\end{center}
\end{table}
To prove that $\frac{1}{\sqrt{2^{r-1}}}\sigma_{\alpha}(I)$ is a
rotated $D_n$-lattice we need the next preliminary results.
\begin{proposition}\label{traco} \cite{and} If $\zeta=\zeta_{2^r}$ and $\matK=\matQ (\zeta +\zeta^{-1})$, then \newline $
Tr_{\mattK|\mattQ}(\zeta^k + \zeta^{-k})=\left\{\begin{array}{l} 0,
\ \
se \ \ gcd(k,2^r)<2^{r-1}; \\ -2^{r-1}, \ \ se \ \ gcd(k,2^r)=2^{r-1}; \\
2^{r-1}, \ \ se \ \ gcd(k,2^r)=2^{r}. \end{array}\right
.$\end{proposition}
\begin{proposition}\label{quase} If $\matK=\matQ(\zeta+\zeta^{-1})$, $e_0 =1$ and $e_i=\zeta^{i}+\zeta^{-i}$ for $i = 1,\cdots,
{2^{r-2}-1}$, then \\
(a) $Tr_{\mattK|\mattQ}(\alpha e_i e_i) = \left\{\begin{array}{ll} 2^{r}, & \mbox{if }\, i=0,1 \\
2^{r+1}, & \mbox{if }\, 2 \leq i < 2^{r-2}-1 \\ 3.2^{r}, & \mbox{if
}\, i=2^{r-2}-1
\end{array} \right.$ \\
(b) $Tr_{\mattK|\mattQ}(\alpha e_i e_0) = \left\{\begin{array}{ll} 2^{r-1}, & \mbox{if }\, i=1 \\
-2^{r}, & \mbox{if }\, i=2 \\ -2^{r-1}, & \mbox{ if }\,\, i=3
\\ 0, & \mbox{if }\, 3<i \leq 2^{r-2}-1 \end{array}\right.$ \\ (c) If $0 < i < j \leq 2^{r-2}-1$ then $$Tr_{\mattK|\mattQ}(\alpha e_i e_j) =
\left\{\begin{array}{ll} 2^{r-1}, & \mbox{ if }\,\, |i-j|=1\,\, \mbox{ and} \\
& (i,j)\not\in \{(1,2),(2,1),(2^{r-2}-2, 2^{r-2}-1), \\ &
(2^{r-2}-1,2^{r-2}-2)\} \\
-2^{r}, & \mbox{ if }\,\, |i-j|=2 \\ -2^{r-1}, & \mbox{ if }\,\,
|i-j|=3 \\ 2^{r}, & \mbox{ if }\,\, (i,j) \in \{(2^{r-2}-2,
2^{r-2}-1), \\ & (2^{r-2}-1,2^{r-2}-2)\}
\\ 0, & \mbox{ otherwise.}
\end{array}\right.$$
\end{proposition}
{\it Proof:} The proof is straightforward by calculating the
$gcd(k,2^{r})$ for some values of $k$ and applying Proposition
\ref{traco}. For $0 < i < j \leq 2^{r-2}-1$ we have:
\[\begin{split} Tr(\alpha e_i e_i) & = Tr(8)+ 4Tr({\zeta^{2i}}+
{\zeta^{-2i}})+ 2Tr({\zeta}+ {\zeta^{-1}})+
Tr({\zeta^{2i+1}} + {\zeta^{-2i-1}}) \\& + Tr({\zeta^{2i-1}}+
{\zeta^{-(2i-1)}})- 4Tr({\zeta^{2}}+ {\zeta^{-2}})
-2Tr({\zeta^{2i+2}}+ {\zeta^{-(2i+2)}}) \\& -2Tr({\zeta^{2i-2}}+
{\zeta^{-(2i-2)}})-2Tr({\zeta^{3}}+ {\zeta^{-3}}) \\& -
Tr({\zeta^{2i+3}}+ {\zeta^{-(2i+3)}}) - Tr({\zeta^{2i-3}}+
{\zeta^{-(2i-3)}}) \end{split} \] For $2\leq i < 2^{r-2}-1$ since
$gdc(k,2^{r})< 2^{r-1}$ for $k= 2i,2i \pm 1,2i\pm2, 2i \pm 3$ it
follows that $Tr(\alpha e_i e_i) = 2^{r+1}$. For $i=1,2^{r-2}-1$ the
development is analogous. For $i=0$ we have: \[ Tr(\alpha
e_0e_0)=Tr(4)+ Tr(\zeta+ {\zeta^{-1}}) -2Tr({\zeta^{2}}+
{\zeta^{-2}})-Tr({\zeta^{3}}+{\zeta^{-3}})=2^r\] and then it follows
(a).
\[ \begin{split} Tr(\alpha e_ie_0) &
=4Tr({\zeta^{i}}+{\zeta^{-i}})+Tr({\zeta^{i+1}}+{\zeta^{-(i+1)}}) +
Tr({\zeta^{i-1}}+{\zeta^{-(i-1)}}) \\&-2Tr({\zeta^{i+2}}+
{\zeta^{-(i+2)}})-2Tr({\zeta^{i-2}}+{\zeta^{-(i-2)}})\\&-
Tr({\zeta^{i+3}}+{\zeta^{-(i+3)}})-Tr({\zeta^{i-3}}+
{\zeta^{-(i-3)}})\end{split} \] For $i\neq 1,2,3$, since
$gcd(k,2^{r})< 2^{r-1}$ for $k=i, i\pm 1, i\pm 2, i\pm 3$ then
$Tr(\alpha e_ie_0)=0$.\\ For $i=1,2,3$ using $Tr(\zeta^{0}+
\zeta^{0})=2^{r-1}$ it follows (b).
\[ \begin{split} Tr(\alpha e_ie_j)&
=Tr(\zeta^{i-j+1}+\zeta^{-(i-j+1)})+
Tr(\zeta^{i-j-1}+\zeta^{-(i-j-1)})
\\&-2[Tr(\zeta^{i-j+2}+\zeta^{-(i-j+2)})+
Tr(\zeta^{i-j-2}+\zeta^{-(i-j-2)})]\\& -[Tr(\zeta^{i-j+3}+
\zeta^{-(i-j+3)}) + Tr(\zeta^{i-j-3}+\zeta^{-(i-j-3)})]\\& -2
Tr(\zeta^{i+j-2}+\zeta^{-(i+j-2)})-
Tr(\zeta^{i+j+3}+\zeta^{-(i+j+3)})\\&-
Tr(\zeta^{i+j-3}+\zeta^{-(i+j-3)}) \end{split}\]
Since $gcd(k,2^{r})<2^{r-1}$ for $k=i+j+1, i+j+ 2$;
$gcd(i+j+3,2^{r})<2^{r-1}$ for $i+j \neq 2^{r-1}-3$; $gcd(i+j+3,2^r)
=2^{r-1}$, for $i+j = 2^{r-1}-3$ and $gcd(i+j-3,2^r)<2^{r-1}$, for
$i+j \neq 3$, it follows (c).
\bb
\begin{proposition}\label{Prop1} The lattice $\frac{1}{\sqrt{2^{r-1}}}\sigma_{\alpha}(\mathcal
O_{\mattK}) \subseteq \matR^{2^{r-2}}$, $\alpha =4 + (\zeta_{2^r}+
{\zeta{_{2^r}^{-1}}})- 2({\zeta{_{2^r}^{2}}}+
{\zeta{_{2^r}^{-2}}})-({\zeta{_{2^r}^{3}}}+ {\zeta{_{2^r}^{-3}}})$
is a rotated $D_n$-lattice for $n=2^{r-2}$.
\end{proposition}
{\it Proof:} The Gram matrix for
$\frac{1}{\sqrt{2^{r-1}}}\sigma_{\alpha}(\mathcal O_{\mattK})$
related to the $\matZ$-basis $\{e_0,e_1,\cdots,e_{n-1}\}$ is
\[{\bm G}={ \left(
\begin{array}{ccccccccccc}
2 & 1 & -2 & -1 & 0 & \cdots & & & & \cdots & 0 \\
1 & 2 & 0 & -2 & -1 & 0 & & & & & \vdots \\
\vdots & 0 & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & 0 & \vdots \\
& & & & \ddots & 0 & -1 & -2 & 1 & 4 & 2 \\
0 & \vdots & & & & \cdots & 0 & -1 & -2 & 2 & 6
\end{array}
\right)}\] \noindent and it is easy to see that ${\bm G}$ is the
Gram matrix for $D_n$ related to the generator matrix ${\bm T}{\bm
B}$ where
\[{\bm T}={\left(
\begin{array}{cccccccc}
0 & \cdots & & & & \cdots & 0 & -1 \\
0 & \cdots & & & \cdots & 0 & 1 & 0 \\
0 & \cdots & & \cdots & 0 & -1 & 0 & 1 \\
\vdots & \vdots & \vdots & \vdots & \vdots & \vdots & \vdots & \\
0 & 0 & 1 & 0 & -1 & 0 & \cdots & 0 \\
0 & -1 & 0 & 1 & 0 & \cdots & \cdots & 0 \\
1 & -1 & -1 & 0 & \cdots & & \cdots & 0
\end{array}
\right)}\] and ${\bm B}$ is the standard generator matrix for $D_n$
given by basis $\beta$ (\ref{beta}). So, since lattices with the
same Gram matrix must be Euclidean equivalent, then then
$\sigma_{\alpha}(I)$ is a rotated $D_n$-lattice. \bb
We determine next the relative minimum product distance of the
rotated $D_n$-lattice considered in Proposition \ref{Prop1}.
Using Propositions \ref{detb} and \ref{Prop1} we conclude:
\begin{corollary} If $m = 2^{r}$, $r \geq 4$, $\matK=\matQ(\zeta_m+\zeta_m^{-1})$ and $\alpha =4 + (\zeta_{2^r}+ {\zeta{_{2^r}^{-1}}})-
2({\zeta{_{2^r}^{2}}}+ {\zeta{_{2^r}^{-2}}})-({\zeta{_{2^r}^{3}}}+
{\zeta{_{2^r}^{-3}}})$ then $N_{\mattK|\mattQ}(\alpha) = 8.$
\end{corollary}
\begin{proposition} For $n=2^{r-2}$, if $\Lambda = \frac{1}{\sqrt{2^{r-1}}}\sigma_{\alpha}(\mathcal
O_{\mattK})$ and $\alpha$ as in (\ref{alpha}) then the lattice
relative minimum product distance is
$${\bm
d_{p,rel}\left(\frac{1}{\sqrt{2^{r-1}}}\sigma_{\alpha}(\mathcal
O_{\mattK})\right)} =2^{\frac{3-r n}{2}}.$$
\end{proposition} {\it Proof:} The minimum norm of the standard $D_n$ is
$\sqrt{2}.$ $\mathcal O_{\mattK}$ is a principal ideal, therefore
using Proposition \ref{distanciaminima} we have
$d_{p,min}(\sigma_{\alpha}(\mathcal O_{\mattK})) =
\sqrt{N(\alpha)N(\mathcal O_{\mattK})^{2}}.$ Since $N(\alpha)=8$ and
$N(\mathcal O_{\mattK})=1,$ then
$${\bm
d_{p,rel}\left(\frac{1}{\sqrt{2^{r-1}}}\sigma_{\alpha}(\mathcal
O_{\mattK})\right)} =
\frac{1}{\sqrt{2}^{n}}\frac{1}{\sqrt{2^{r-1}}^{n}} \sqrt{8}
=\frac{\sqrt{8}}{2^{r \frac{n}{2}}} = 2^{\frac{3-rn}{2}}.$$ \bb
\vspace{1cm}
\subsection{\bf A second construction:} In \cite{and} and \cite{gabriele}
families of rotated $\matZ^n$-lattices obtained as image of a
twisted homomorphism applied to $\matZ[\zeta+\zeta^{-1}]$ and having
full diversity are constructed. Those constructions consider $\alpha
= 2 + e_1$ and $\alpha=2-e_1,$ respectively, and generate equivalent
lattices in the Euclidean metric by permutations and coordinate
signal changes.
We will use in our construction the rotated $\matZ^{n}$-lattice
$\Lambda=\frac{1}{\sqrt{2^{r-1}}}\sigma_{\alpha}(I)$ with
$\alpha = 2 + e_1$ and $I=\mathcal
O_{\mattK}=\matZ[\zeta+\zeta^{-1}],$ and then consider $D_n$ as a
sublattice of $\Lambda$.
If $e_0 =1$ and $e_i=\zeta^{i}+\zeta^{-i}$ for $i = 1,\cdots,
{2^{r-2}-1}$, by \cite{gabriele} a generator matrix for the rotated
$\matZ^{n}$-lattice $\Lambda=$
$\frac{1}{\sqrt{2^{r-1}}}\sigma_{\alpha}(\mathcal O_{\mattK})$ is
${\bm M_1} = \displaystyle\frac{1}{\sqrt{2^{r-1}}}{\bm N} {\bm A},$
where ${\bm N}=(\sigma_i(e_{j-1})_{i,j=1}^{n}$ and ${\bm A}=
diag(\sqrt{\sigma_{k}(\alpha)}).$ Let ${\bm T}$ the basis change
matrix \[{{\bm T}=\left( \begin{array}{ccccc} 1 & -1 & \cdots & 1 & -1 \\
1 & -1 & \cdots & 1 & 0 \\ \vdots & \vdots & \ddots & \vdots & \vdots \\
1 & 0 & \cdots & 0 & 0
\end{array}\right)}.\]
For ${\bm M} = {\bm T}{\bm M_1}$, ${\bm G}={\bm M} {\bm M^{t}}={\bm
I_n}$ and we will consider the standard lattice $D_n \subseteq
\matZ^{n}$ rotated by ${\bm M}$.
\begin{proposition}\label{idealI} Let $I \subseteq \mathcal O_{\mattK}$ be the $\matZ$-module with
$\matZ$-basis \linebreak $\{-e_1,e_2,\cdots,-e_{n-1}, -2e_0 + 2 e_1
- 2e_2 + \cdots -2e_{n-2} + e_{n-1}\}$ and $\alpha=2+e_1.$ The
lattice $\frac{1}{\sqrt{2^{r-1}}}\sigma_{\alpha}(I) \subseteq
\matR^{2^{r-2}}$ is a rotated $D_n$-lattice.
\end{proposition} \noindent{\it Proof:} Let ${\bm B}$ be the generator matrix of $D_n$ associated to the basis
$\beta$ (\ref{beta}). Using homomorphism properties, a
straightforward computation shows that ${\bm B} {\bm M} =$
$$ {
\frac{1}{\sqrt{2^{r-1}}}\footnotesize\left(\begin{array}{ccc} \sigma_1(-2 e_0 +
\cdots - 2e_{n-2} + e_{n-1} ) & \cdots & \sigma_n(-2 e_0 +
\cdots - 2e_{n-2} + e_{n-1} )
\\ \sigma_1(-e_{n-1}) & \cdots &
\sigma_n( -e_{n-1})
\\ \vdots & \ddots & \vdots \\ \sigma_1(-e_1)
& \cdots & \sigma_n(-e_1) \end{array}\right) {\bm A}} $$ is a
generator matrix for $\frac{1}{\sqrt{2^{r-1}}}\sigma_{\alpha}(I).$
This lattice is a rotated $D_n$-lattice since ${\bm B} {\bm M} ({\bm
B} {\bm M})^{t} = {\bm B} {\bm B^{t}}$ is the standard Gram matrix
of $D_n$ relative to the basis $\beta$. \bb
We show next that the rotated $D_n$-lattice of the last proposition
is associated to a principal ideal of $\mathcal O_{\mattK}$ and then
calculate its relative minimum product distance.
\begin{proposition} Let $I$ be the $\matZ$-module given in Proposition \ref{idealI}.
$I$ is a principal ideal and $I = e_1 \mathcal O_{\mattK}.$
\end{proposition} {\bf Proof:} It is easy to see that $I = 2e_0\matZ + e_1\matZ + \cdots +
e_{n-1}\matZ.$ Let $x \in e_1 \mathcal O_{\mattK}.$ Then $x = e_1
(a_0 e_0 + a_1 e_1 + a_2 e_2 + \cdots + a_{n-1}e_{n-1}) = a_0 (e_1
+ e_{-1}) + a_1 (e_2 + 2e_0) + a_2 (e_3 + e_{-1}) + \cdots +
a_{n-1}(e_n + e_{-n+2}) = a_1 (2 e_0) + (2 a_0 + a_2)(e_1) + (a_1
+a_3)(e_2) + \cdots + (a_{n-2})(e_{n-1}) \in I.$ Now, if $x \in I,$
then $x = a_0 2e_0 + a_1 e_1 + \cdots + a_{n-1} e_{n-1} = (e_1)[a_0
e_1 + a_1 e_2 + (a_2 - a_0) e_3 + ( a_3 - a_1) e_4 + (a_4 - a_2
-a_0) e_5 + (a_5 - a_3 -a_1)e_6 + \cdots + ( a_{n-1})e_{n-2} +
(a_{n-2} - a_{n-4} \cdots - a_0)e_{n-1}] \in e_1 \mathcal
O_{\mattK}.$ So, $I$ is a principal ideal of $\mathcal O_{\mattK}.$
\bb
\begin{remark} It follows from Proposition \ref{distanciaminima} and
Definition \ref{relative} that the \ relative minimum product
distance of $D_n$-lattices constructed from principal ideals in
$\mathcal O_{\mattK}=\matQ(\zeta_m+\zeta_m^{-1})$, $m=2^{r}$, $r
\geq5$, depends only of the determinant of $D_n$ and of the
discriminant of $\matK$. Therefore for any construction of a
rotated $D_n$ lattice from a principal ideal $I$ in $\mathcal
O_{\mattK}$ the relative minimum product distance is
$d_{p,rel}(\sigma_{\alpha}(I)) = 2^{\frac{3-r n}{2}}.$
\end{remark}
It is also interesting to note that besides being Euclidean
equivalent, the lattices obtained through the first and second
constructions are equivalent in the Lee metric since the isometry is
a composition of permutations and coordinate signal changes.
The density $\Delta(\Lambda)$ of a lattice $\Lambda \subseteq
\matR^{n}$ is given by $\Delta(\Lambda) =
\frac{(d/2)^{n}Vol(B(1))}{det(\Lambda)^{1/2}}$ where $Vol(B(1))$ is
the volume of the unitary sphere in $\matR^{n}$ and $d$ is the
minimum norm of $\Lambda$. The parameter $\delta(\Lambda) =
\frac{(d/2)^{n}}{det(\Lambda)^{1/2}}$ is so called center density.
Table $2$ shows a comparison between the normalized $d_{p,rel}$ and
the center density of rotated $\matZ^n$-lattices constructed in
\cite{gabriele} and rotated $D_n$-lattices constructed here via
principal ideals in $\matK=\matQ(\zeta+\zeta^{-1})$, $n=2^{r-2}$.
Asymptotically we have
\begin{equation} \label{limite} \lim_{n\longrightarrow
\infty}\frac{\sqrt[n]{d_{p,rel}(\matZ^{n})}}{\sqrt[n]{d_{p,rel}(D_n)}}
=\sqrt{2}\,\, \, \mbox{and}\,\,\, \lim_{n\longrightarrow
\infty}\frac{\delta(\matZ^{n})}{\delta(D_n)} =0. \end{equation}
\begin{table}[h]
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|}
\hline \hline
$r$ & $n$ & $\sqrt[n]{d_{p,rel}(\matZ^n)}$ & $\sqrt[n]{d_{p,rel}(D_n)}$ & $\delta(\matZ^n)$ & $\delta(D_n)$ \\
\hline
$4$ & $4$ & $ 0.385553 $ & $ 0.324210 $ &$0.062500$ & $0.125000$\\
\hline
$5$ & $8$ & $ 0.261068 $ & $ 0.201311 $ &$0.003906$ & $0.031250$ \\
\hline
$6$ & $16$ & $ 0.180648 $ & $ 0.133393 $ &$0.000015$ & $0.001953$ \\
\hline
$7$ & $32$ & $ 0.126361 $ & $ 0.091307 $ &$2.3 \times 10^{-10}$ & $7.6 \times 10^{-6}$\\
\hline
$8$ & $64$ & $ 0.088868 $ & $ 0.063523 $ &$5.4 \times 10^{-20}$ & $1.1 \times 10^{-10}$\\
\hline
$9$ & $128$ & $ 0.062669 $ & $ 0.044554 $ &$2.9 \times 10^{-39}$ & $2.7 \times 10^{-20}$\\
\hline \hline
\end{tabular}
\vspace{0.3cm} \caption{ }
\end{center}
\end{table}
If the goal is to construct lattices which have good performance on
both Gaussian and Rayleigh channels, we may assert that taking into
account the trade-off density versus product distance, there is some
advantages in considering these rotated $D_n$-lattices instead of
rotated $\matZ^{n}$-lattices, $n=2^{r-2}$, $r \geq 5,$ in high
dimensions.
\section{Rotated $D_n$-lattices for $n=\frac{p-1}{2}$, $p$ prime,
via $\matK=\matQ(\zeta_{p}+\zeta_{p}^{-1})$}
Let $\zeta=\zeta_{p}$ be a primitive $p$-th root of unity, $p$
prime, $\matL=\matQ (\zeta)$ and $\matK=\matQ (\zeta+\zeta^{-1})$.
We will construct a family of rotated $D_n$-lattices, derived from
the construction of a rotated $\matZ^{n}$-lattice in \cite{Oggier},
via a $\matZ$-module that is not an ideal. Let
$e_j=\zeta^{j}+\zeta^{-j}$ for $j = 1,\cdots, {(p-1)/2}.$
By \cite{Oggier} a generator matrix of the rotated
$\matZ^{n}$-lattice
$\Lambda=\frac{1}{\sqrt{p}}\sigma_{\alpha}(\mathcal O_{\mattK})$ is
${\bm M}= \displaystyle\frac{1}{\sqrt{p}}{\bm T} {\bm N} {\bm A} ,
\mbox{ where }$ ${\bm T}=(t_{ij})$ is an upper triangular matrix
with $t_{ij}=1$ if $i\leq j,$ ${\bm N} =
(\sigma_i(e_j))_{i,j=1}^{n}$ and ${\bm A}
=diag(\sqrt{\sigma_{k}(\alpha)}).$ We have ${\bm G}={\bm M} {\bm
M^{t}}={\bm I_n}$ \cite{Oggier}.
\begin{proposition} \label{rotacionado} Let $I \subseteq \mathcal O_{\mattK}$ be a $\matZ$-module with
$\matZ$-basis $\{e_1,e_2,\cdots,e_{n-1},-e_1 - 2e_2 - \cdots
-2e_n\}$ and $\alpha=2-e_1.$ The lattice
$\frac{1}{\sqrt{p}}\sigma_{\alpha}(I) \subseteq
\matR^{\frac{p-1}{2}}$ is a rotated $D_n$-lattice.
\end{proposition} \noindent{\it Proof:} Let ${\bm B}$ be a generator matrix for $D_n$ given by basis $\beta$ \ref{beta}.
Using homomorphism properties, a straightforward computation shows
that ${\bm B}{\bm M}$ is a generator matrix for
$\Lambda=\frac{1}{\sqrt{p}}\sigma_{\alpha}(I).$ This lattice is a
rotated $D_n$ since ${\bm B} {\bm M} ({\bm B} {\bm M})^{t} = {\bm B}
{\bm B^{t}}$ is a Gram matrix of $D_n.$ It has full diversity since
it is contained in $\frac{1}{\sqrt{p}}\sigma_{\alpha}(\mathcal
O_{\mattK})$ \cite{Oggier}. \bb
\begin{proposition} The $\matZ$-module $I \subseteq \mathcal O_{\mattK}$ is not an ideal of $\mathcal O_{\mattK}$. \end{proposition}
\noindent{\it Proof:} The set $\{e_1,e_2,\cdots,e_{n-1},2e_n\}$ is
an another $\matZ$-basis to $I$. We will show that $e_n$ is not in
$I.$ Indeed, if $e_n \in I$, then $I = \mathcal O_{\mattK},$ but
$\left|\frac{\mathcal O_{\mattK}}{I}\right| = 2.$ So, $e_n \not\in
I.$ $e_{n-1} e_{1}$ is not in $I$. In fact, note that $e_{n-1} e_1
= e_{n} + e_{n-2}$ and $e_{n-2} \in I$. If $e_{n-1}e_1 \in I$, then
$e_n = e_{n-1}e_1 - e_{n-2} \in I,$ and this doesn't happen. \bb
\begin{proposition} If $\Lambda =\frac{1}{\sqrt{p}}\sigma_{\alpha}(I) \subseteq
\matR^{\frac{p-1}{2}}$ with $\alpha$ and $I$ as in the Proposition
\ref{rotacionado}, then the relative minimum product distance is
$${\bm d_{p,rel}}(\Lambda)= 2^{\frac{1-p}{4}} p^{\frac{3-p}{4}} .$$
\end{proposition} \noindent{\it Proof:} First note that $|N(e_1)| =1.$
Indeed, $(\zeta+\zeta^{-1})(-\zeta^{p-1} - \zeta^{p-2} - \cdots -
\zeta - 1)=1$ and so
$$N(\zeta+\zeta^{-1})N(-\zeta^{p-1} -
\zeta^{p-2} - \cdots - \zeta - 1)=N(1)=1.$$
Since $e_1 \in \mathcal O_{\mattK}$, then $N(e_1) \in \matZ$, what
implies $|N(e_1)|=1.$ Now, the minimum norm in $D_n$ is $\sqrt{2}.$
By Proposition \ref{distanciaminima}, ${\bm
d_{p}(\sigma_{\alpha}(I))} =\sqrt{N(\alpha)} min_{0\neq y \in
I}|N(y)| = \sqrt{p},$ since $min_{0\neq y \in I}|N(y)| =1.$
Therefore, the relative minimum product distance is
$$d_{p,rel}\left(\frac{1}{\sqrt{p}}\sigma_{\alpha}(I)\right) =
\left(\frac{1}{\sqrt{p}^{\frac{p-1}{2}}}\right)
\left(\frac{1}{\sqrt{2}^{\frac{p-1}{2}}}\right)\sqrt{p} =
2^{\frac{1-p}{4}} p^{\frac{3-p}{4}}.$$ \bb \vspace{0.4cm}
Table $3$ shows a comparison between the normalized $d_{p,rel}$ and
the center density $\delta$ of rotated $\matZ^n$-lattices
constructed in \cite{Oggier} and rotated $D_n$-lattices constructed
here, $n=(p-1)/2$. As in Section \ref{limite} we also have for
$\Lambda = \frac{1}{\sqrt{p}}(\sigma_{\alpha}(I)) \subseteq
\matR^{\frac{p-1}{2}}$ and $p$ prime, the following results:
$$\lim_{n\longrightarrow
\infty}\frac{\sqrt[n]{d_{p,rel}(\matZ^{n})}}{\sqrt[n]{d_{p,rel}(D_n)}}
=\sqrt{2}\,\, \mbox{and} \,\, \lim_{n\longrightarrow
\infty}\frac{\delta(\matZ^{n})}{\delta(D_n)} =0.$$
\begin{table}[h]
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|}
\hline \hline
$p$ & $n$ & $\sqrt[n]{d_{p,rel}(\matZ^n)}$ & $\sqrt[n]{d_{p,rel}(D_n)}$ & $\delta(\matZ^n)$ & $\delta(D_n)$ \\
\hline
$11$ & $5$ & $ 0,38321 $ & $0,27097 $ & $0,03125$ & $0,08838$ \\
\hline
$13$ & $6$ & $ 0,34344 $ & $0,24285 $ & $0,01563$ & $0,06250$ \\
\hline
$17$ & $8$ & $ 0,28952$ & $0,20472 $ & $0,00390$ & $0,03125$ \\
\hline
$19$ & $9$ & $0,27187 $ & $ 0,19105 $ & $0,00195$ & $0,02209$ \\
\hline
$23$ & $11$ & $ 0,24045$ & $ 0,17003 $ & $0,00049 $ & $0,01105$\\
\hline \hline
\end{tabular}
\vspace{0.3cm} \caption{ }
\end{center}
\end{table}
\section{Conclusion}
In this work we construct some families of full diversity rotated
$D_n$-lattices. These lattices present good performance for signal
transmission over both Gaussian and Rayleigh channels. Considering
the trade-off between density and relative product distance we may
assert that the rotated $D_n$-lattices presented here have better
performance than the known rotated $\matZ^{n}$-lattices for
$n=2^{r-2},$ $r \geq 5$ and $n=\frac{p-1}{2}$, $p$ prime, $p \geq
7$.
|
\section{Introduction}
It is well established that massive stars have a significant impact on the dynamics and energetics of the interstellar medium (ISM) surrounding them. In the classical scenario, OB stars are born deeply buried within dense molecular clumps and emit a copious amount of far ultraviolet (FUV) radiation \hbox{($h\nu$ $>$ 13.6 eV)}. FUV photons ionize the neutral hydrogen, creating an H{\sc ii}\, region which expands into the molecular cloud due to the pressure difference between the molecular and ionized gas. At the interface between the ionized and molecular gas, where photons of lower energies \hbox{(6 eV $<$ $h\nu$ $<$ 13.6 eV)} dominate, photodissociation regions (PDR) are known to exist (see \citealt{ht97} for a complete review).
The molecular component of H{\sc ii}\, regions has been studied in detail by many authors \citep{kh89,rc04}. Many efforts have been devoted to observing the interaction between the H{\sc ii}\, regions and their parental molecular clouds and some evidence has been presented showing that molecular gas near H{\sc ii}\, regions may be kinematically disturbed by several \hbox{km s$^{-1}$\,} \citep{ew87}. Understanding the complex interaction between massive young OB stars, H{\sc ii}\, regions, and molecular gas is crucial to the study of massive star formation and the impact of massive stars on their environment.
NGC\,3503 (=Hf 44=BBW 335) \hbox{({\it l,b} = 289\fdg 51, +0\fdg 12)} is a bright and small ($\sim$ 3$'$ in diameter) optical emission nebula \citep{dr88} ionized by early B-type stars belonging to the open cluster Pis 17 \citep{H75, pco10} and it is seen projected onto an extended region of medium brightness in H$\alpha$ known as {\rm RCW}\,54. This last region is centered at \hbox{({\it l, b}) = (289\fdg4, -0\fdg6)}, and has dimensions \hbox{($\Delta l$ $\times$ $\Delta b$) = (3\fdg5 $\times$ 1\fdg0)} \citep{r60}. The point source {\rm IRAS}\,10591-5934, which appears projected onto \rm NGC\,3503\,, is the infrared counterpart of the optical nebula. For the sake of clarity, we show the main components and their relative location in Fig. {\ref{fig:rcw54}} superimposed on a DSSR image of the region of interest. The dashed white square delimits the region studied in this paper.\
\begin{figure*}
\centering
\includegraphics[width=170mm]{rcw54-dssr-lbdc2.eps}
\caption{ DSSR image of the brightest part of RCW 54. The image is $\sim$ 1\fdg5 $\times$ 1\fdg5 in size, centered at \hbox{({\it l,b}) = (289\fdg5, 0\fdg25)}. Our region of interest is delimited by the white dashed square. }
\label{fig:rcw54}
\end{figure*}
In their Southern Hemisphere catalogue of bright-rimmed clouds ({\rm BRC}s) associated with {\rm IRAS} point sources, \citet{SO94} quote {\rm SFO}\,62 as related to {\rm NGC}\,3503.
BRCs are defined as isolated molecular clouds located at the edges of evolved H{\sc ii}\, regions, and are suspected to be potential sites of star formation through the radiation-driven implosion (RDI) process \citep{so91,U09}. Making use of the Australian Telescope Compact Array ({\rm ATCA}), \citet{T04} carried out radio continuum observations towards all {\rm BRC}s catalogued by \citet{SO94}. These authors classified {\rm SFO}\,62 as a broken-rimmed cloud associated with an evolved stellar cluster that is about to disrupt its natal molecular cloud. Latter on, \citet{U09} claimed that SFO62 was incorrectly classified as a BRC, and that the RDI process is not working in this region.
Using the [S{\sc ii}] $\lambda$6716/$\lambda$6731 line ratio of \rm NGC\,3503\,, \citet{c00} detected a significant electron density dependence on position, which according to the authors, could be interpreted as a radial gradient. The authors claimed that \rm NGC\,3503\, may be a candidate of showing a ``champagne flow''. In their H$\alpha$ survey of the Milky Way towards \hbox{{\it l} = 290$^{\circ}$}, \citet{g00} reported a radial velocity\footnote{Radial velocities are referred to the Local Standard of Rest ({\rm LSR})} of \hbox{--21 \hbox{km s$^{-1}$\,}} towards \rm NGC\,3503\,, and proposed that \rm NGC\,3503\, is linked to a complex of H{\sc ii} regions placed at a distance of about 2.7 kpc, with radial velocities of about $-$25 \hbox{km s$^{-1}$\,}. Published distance determinations of {\rm NGC}\,3503 vary between a minimum of 2.6 kpc \citep{H75} and a maximum of 4.2 kpc \citep{mv75}. In this paper we shall adopt a distance of \hbox{2.9 $\pm$ 0.4 kpc} \citep{pco10}.
{\rm NANTEN} $^{13}$CO (J=1$\rightarrow$0) (HPBW = 2.7$'$) line observations were carried out by \citet{Y99} towards 23 H{\sc ii} regions associated with 43 {\rm BRC}s catalogued by \citet{SO94}, with a grid spacing of $\sim$ 4$'$ $\times$ 4$'$. The aim of that work was to investigate statistically the dynamical effects of H{\sc ii} regions on their associated molecular clouds and star formation. A square region of $\sim$ 0\fdg4 in size towards {\rm NGC}\,3503 was observed and a single and slightly extended 6$'$ diameter molecular cloud having a peak radial velocity of --24.9 \hbox{km s$^{-1}$\,} was found. Adopting a distance of 2.9 kpc, this molecular structure has a linear radious of 2 pc and a total mass of $\sim$ 500 M$_\odot$\,. The authors rejected this cloud from the analysis sample, since it was detected at only one position. $^{12}$CO, $^{13}$CO, and $^{18}$CO (J=1$\rightarrow$0) {\rm MOPRA} position-switched observations were carried out by \citet{U09}. The {\rm ON} position was centred on the position of the {\rm IRAS} source associated with {\rm NGC}\,3503. The least abundant of the isotopes was not detected, whilst the emission lines of the other two isotopes show a double-peaked profile. From a Gaussian fitting, their peak radial velocities are $\sim$ +20 \hbox{km s$^{-1}$\,} and --25.6 \hbox{km s$^{-1}$\,}.
Molecular line observations provide an invaluable support in pursuing a better understanding of the interaction of H{\sc ii} regions with their surroundings. Previous observations of \citet{Y99} and \citet{U09}, although providing important information about the molecular gas associated with \rm NGC\,3503\,, do not offer a complete picture of the molecular environment of the nebula. Bearing this in mind, the goals of this paper are twofold, namely: {\it a)} to map the spatial distribution of the molecular gas associated with {\rm NGC}\,3503 and to study its physical characteristics, and {\it b)} to achieve a better understanding of the interaction of {\rm NGC}\,3503 and Pis 17 with their molecular environment. These aims were addressed by analyzing new $^{12}$CO (J=1$\rightarrow$0) data gathered by using the {\rm NANTEN} telescope to observe a square region \hbox{$\sim$ 0\fdg7} in size, centred on \hbox{({\it l,b}) = (289\fdg36, +0\fdg02)}. Molecular observations were combined with unpublished radio continuum data, and optical, mid, and \hbox{far-infrared} archival data.
The structure of the paper is the following. In Sect. 2, the {\rm CO} and radio continuum observations are briefly described along with a short mention to the archival data used in this work. The main observational results of the molecular, ionized and dust emission are detailed in Sect. 3, while both the comparison among these interstellar phases and the action of the stellar ionizing sources of NGC\,3503 over its interstellar environs, as well as the presence of star formation activity are exposed in Sect. 4. Finally, a possible scenario is put forward to explain the stellar and interstellar interactions. A summary is presented in Sect. 5.
\section{Observations and data reductions}
The databases used in this work are:
\begin{enumerate}
\item Intermediate angular resolution, medium sensitivity, and
high-velocity resolution {\rm $^{12}$CO (J=1$\rightarrow$0)} data
obtained with the 4-m {\rm NANTEN} millimeter-wave telescope of Nagoya University. At the time the authors carried out the observations, April 2001, this telescope was
installed at Las Campanas Observatory,
Chile. The half-power beamwidth and the system temperature, including the
atmospheric contribution towards the zenith, were 2\farcm7
($\sim$ 2.3 pc at 2.9 kpc) and $\sim$ 220\,K
(SSB) at 115 GHz, respectively. The data were gathered using the position switching mode. Observations of points devoid of {\rm CO} emission were interspersed among the
program positions. The coordinates of these points were retrieved from a
database that was kindly made available to us by the {\rm NANTEN} staff. The
spectrometer used was an acusto-optical with 2048
channels providing a velocity resolution of $\sim$ 0.055 kms$^{-1}$. For
intensity calibrations, a room-temperature chopper wheel was employed \citep{pb73}.
An absolute intensity calibration
\citep{uh76,ku81} was achieved by observing
Orion {\rm KL} \hbox{(RA(1950.0) = 5$^h$\,32$^m$\,47\fs0},\hbox{Dec (1950.0) = $-$5$^\circ$\,24\arcmin\,21\arcsec )}, and $\rho$ Oph East \hbox{(RA(1950.0) = 16$^h$ \,29$^m$ \,20\fs9}, \hbox{Dec (1950.0) = $-$24$^\circ$\,22\arcmin\,13\arcsec )}. The absolute radiation
temperature, $T_R^\ast$, of
Orion {\rm KL} and $\rho$ Oph East, as observed by the {\rm NANTEN}
radiotelescope were assumed to be 65 K and 15 K, respectively \citep{myom01}.
The {\rm CO} observations covered a region ($\bigtriangleup${\it l}
$\times$ $\bigtriangleup${\it b}) of 35\farcm1 $\times$ 35\farcm1 centred at
({\it l, b}) = {(289\fdg36, +0\fdg02 )} and the observed grid consists of points located every
one beam apart. A total of 169 positions were observed. Typically, the integration time per point was 16s resulting in an rms noise of $\sim$ 0.3 K. A second order degree polynomial was substracted from the observations to account for instrumental baseline effects. The spectra were reduced using CLASS software (GILDAS working group).
\item Unpublished radio continuum data at 4800 and 8640 MHz which were kindly provided by \hbox{J. S. Urquhart} and \hbox{M. A. Thompson}. The images were obtained in March 2005 with the Australia Telescope Compact Array (ATCA) with synthesized beams and rms noises of \hbox{23\farcs 55 $\times$ 18\farcs 62} and \hbox{0.82 mJy beam$^{-1}$} at 4800 MHz and \hbox{14\farcs 73 $\times$ 11\farcs 74} and \hbox{0.56 mJy beam$^{-1}$} at 8640 MHz. Radio continuum archival images were also obtained from the Sydney University Molonglo Sky Survey (SUMSS)\footnote{{\it http://www.astrop.physics.usyd.edu.au/cgi-bin/postage.pl}} \citep{b99} at 843 MHz, with angular resolution of 45$''$$\times$45$''$cosec($\delta$).
\item Infrared data retrieved from the {\it Midcourse Space Experiment (MSX)}\footnote{{\it http://irsa.ipac.caltech.edu/Missions/msx.html}} \citep{p01}, high resolution IRAS images (HIRES)\footnote{http://irsa.ipac.caltech.edu/applications/IRAS/IGA/} at 60 and 100 $\mu$m \citep{fa94}, Spitzer images at 8.0 and 4.5 $\mu$m from the Galactic Legacy Infrared Mid-Plane Survey Extraordinaire (GLIMPSE)\footnote{{\it http://sha.ipac.caltech.edu/applications/Spitzer/SHA//}} \citep{b03}, and Multiband Imaging Photometer for $Spitzer$ (MIPS) images at 24 and 70 $\mu$m from the MIPS Inner Galactic Plane Survey (MIPSGAL)\footnote{{\it http://sha.ipac.caltech.edu/applications/Spitzer/SHA//}} \hbox{\citep{ca05}}.
\item Optical data retrieved from the 2nd Digitized Sky Survey (red plate)\footnote{{\it http://skyview.gsfc.nasa.gov/cgi-bin/query.pl}} \citep{mcl00}.
\end{enumerate}
\section{Results and analysis}
\subsection{Carbon Monoxide}
\subsubsection{Individual velocity components}
In order to illustrate in broad terms the molecular structures detected towards the region under study, a series of {\rm CO} profiles are displayed in Fig.~\ref{fig:perfiles}.
\begin{figure*}
\centering
\includegraphics[width=430pt]{ngc3503espectros3.eps}
\caption{CO emission profiles towards five positions around NGC 3503. The CO profiles are averaged over a square area $\sim$ 3$'$ in size, centered on the black/white dots drawn on the DSSR image (center). The intensities are given as absolute radiation temperature $T_R^\ast$. }
\label{fig:perfiles}
\end{figure*}
Profiles {\it a)} and {\it c)} show the {\rm CO} emission along the line of sight to \rm NGC\,3503\, and the bright edge of \rm SFO\,62\,, respectively. In both spectra the bulk of the molecular emission is detected between \hbox{--30} and \hbox{--20} \hbox{km s$^{-1}$\,}, and +15 to +25 \hbox{km s$^{-1}$\,}, in good agreement with previous results by \citet{U09}.
The medium brightness optical region northeast of \rm NGC\,3503\,\ (from here onwards MBO) though also depicts a double peak structure (see profile {\it b}), displays some small scale structure in the feature peaking at $\sim$ +20 \hbox{km s$^{-1}$\,}, while the most negative CO feature is detected at $\sim$ --18 \hbox{km s$^{-1}$\,}.
The molecular gas along regions of high optical absorption is shown in profiles {\it d)} and {\it e)}. The former resembles the CO spectrum observed towards \rm NGC\,3503\,, and is characteristic of the molecular emission in the region of high optical absorption seen northwest of \rm NGC\,3503\,. The roundish patch of high absorption seen at \hbox{{\it (l,b)} = (289\fdg55,+0\fdg07)} only shows the CO peak at positive velocities.
The spatial distribution of the molecular gas observed in the three velocity intervals mentioned above is shown in the left side panels of Fig. \ref{fig:comp12y3}. In order of increasing radial velocity, the CO components will be referred to as Component 1 (peaking at \hbox{$\sim$ --25 km s$^{-1}$\,}), Component 2 (peaking at \hbox{$\sim$ -16 km s$^{-1}$\,}), and Component 3 (peaking at \hbox{$\sim$ +20 km s$^{-1}$\,}). To facilitate the comparison between the spatial distribution of molecular and ionised gas, the right panels of Fig. \ref{fig:comp12y3} show the same molecular contours as the left panels superimposed to the DSSR image in grayscale. The difference in spatial distribution among the three components is readily appreciated. The molecular gas in the velocity interval --29 to --20 \hbox{km s$^{-1}$\,}\ (Component 1) shows two well developed concentrations, whose emission peaks are located at \hbox{{\it (l,b)} = (289\fdg47,+0\fdg12)} (clump A) and \hbox{{\it (l,b)} = (289\fdg32\arcmin,+0\fdg03)} (clump B). Component 1 has a very good morphological correspondence with the high optical obscuration region seen north and northwest of \rm NGC\,3503\,.
Component 2 peaks at \hbox{{\it (l,b)} = (289\fdg55, +0\fdg17)} near both \rm NGC\,3503\, and MBO, and is projected onto a region without apreciable optical absorption. Finally, Component 3, peaking at \hbox{{\it (l,b)} = (289\fdg45, $-$0\fdg18)}, does not show a clear morphological correspondence either with ionized regions or areas displaying strong absorption.
\begin{figure*}
\centering
\includegraphics[width=378pt]{comp12y3-co-2.eps}
\caption{{\it Upper panels} (Left): Averaged $T_R^\ast$ in the velocity range $\sim$ --29 to -- 20 \hbox{km s$^{-1}$\,} (Component 1). Contour levels start at 0.42 K (23 rms) and the contour spacing is 0.6 K. The beam size of the CO observations is shown by a circle in the lower left corner. (Right): Overlay of the $T_R^\ast$ values in the same velocity interval (contours), superimposed on the DSSR image (greyscale). {\it Middle panels:} Averaged $T_R^\ast$ in the velocity range $\sim$ --17.5 to --15.5 \hbox{km s$^{-1}$\,} (Component 2). Contour levels start at 0.7 K (23 rms) and the contour spacing is 1 K. {\it Lower panels:} Averaged $T_R^\ast$ in the velocity range $\sim$ +14 to +24 \hbox{km s$^{-1}$\,} (Component 3). Contour levels start at 0.35 K (23 rms) and the contour spacing is 0.45 K. In all cases, the molecular emission grayscale goes from 0.35 K to 3.9 K. The orientation of the equatorial system is indicated in the top left panel.}
\label{fig:comp12y3}
\end{figure*}
For the three molecular components, a mean radial velocity ($\bar{V}$) weighted by line temperature, was derived by means of
\begin{eqnarray}
\qquad \qquad \bar{V} \ =\ \frac{\sum_{i}\ T_{Peak_i}\ \times\ V_{Peak_i}}{\sum_{i}\ T_{Peak_i}}
\label{eq:masaH2a}
\end{eqnarray}
where $T_{\rm Peak_i}$ and $V_{\rm Peak_i}$ are the peak $T_R^\ast$ temperature and the peak radial velocity of the $i$-spectrum observed within the 3 rms contour line defining the outer border. Mean radial velocities of \hbox{--24.7 km s$^{-1}$\,}, \hbox{--16.6 km s$^{-1}$\,}, and \hbox{+19.5 km s$^{-1}$\,} were obtained for Component 1, Component 2, and Component 3, respectively.
Based on $\bar{V}$, we tried to estimate the kinematical distance of Component 1. The analytical fit to the rotation curve by \citet{bb93} along {\it l} = 289\fdg5 shows that radial velocities more negative than --15.5 \hbox{km s$^{-1}$\,}\ are forbidden for this galactic longitude. However, the ionised gas along this galactic longitude exhibits velocity departures more negative than $-$7 \hbox{km s$^{-1}$\,} relative to the circular rotation model, as the cases of the H{\sc ii}\,\ regions Gum\,37 \hbox{{\it (l,b)} = (290\fdg65, +0\fdg26)} \hbox{($d$ = 2.7 kpc)} and Gum\,38a \hbox{{\it (l,b)} = (291\fdg28, +0\fdg71)} \hbox{($d$ = 2.8 kpc)}, whose radial velocities are in the range $-$25 to $-$28 \hbox{km s$^{-1}$\,} and $-$25 to $-$ 29 \hbox{km s$^{-1}$\,}, respectively \citep{g00}. Bearing in mind that close to the tangential point along {\it l} = 290\fdg5 the radial velocity gradient is very small and that the H{\sc ii}\,\ regions located at distances between 2.7 and 2.8 kpc exhibit forbidden radial velocities similar to those found for Component 1, we adopt for the latter a distance of 2.9 $\pm$ 0.4 kpc \citep{pco10}, which is also the kinematical distance corresponding to the tangential point.
Component 1 shows an excellent morphological resemblance with a region of high optical absorption next to \rm NGC\,3503\, (see Fig. {\ref{fig:comp12y3}}) and its mean velocity is in good agreement with the velocity of the H$\alpha$ line towards the nebula ($-$21 \hbox{km s$^{-1}$\,}; \citealt{g00}). It is also worth noting that the velocity interval of Component 1 is in excellent agreement with the velocity of SFO\,62 found by \citet{SO94} \hbox{(--28 $\leq\ v_{LSR}\ \leq$ --20 \hbox{km s$^{-1}$\,}}, at $^{12}$CO), by \citet{Y99} (--24.9 km s$^{-1}$\, at $^{13}$CO), and by \citet{U09} ($-$25.6 and $-$25.7 \hbox{km s$^{-1}$\,}, at $^{12}$CO and $^{13}$CO, respectively). Based on above, we conclude that Component 1 is associated with \rm NGC\,3503\, and its environs. The bright rim of \rm SFO\,62\,\ clearly follows the southernmost border of clump A, which very likely indicates that the molecular gas in the southern border of this clump is being ionized, originating the bright optical rim.
On the other hand, considering that the mean radial velocity of Component 2 (--16.5 \hbox{km s$^{-1}$\,}) is close to the radial velocity of the tangential point at {\it l} = 289\fdg5 (--15.5 \hbox{km s$^{-1}$\,}), we suggest that Component 2 may be located in the neighborhood of the tangential point, close to \hbox{Component 1}. From the mean radial velocity of Component 3, a kinematic distance of $\sim$ 8 kpc is determined, indicating that Component 3 is unrelated to \rm NGC\,3503\,. Very likely, Component 3 is associated with the complex of H{\sc ii} regions at a velocity of $\sim$ +20 \hbox{km s$^{-1}$\,}\ reported by \citet{g00}.
A direct comparison of Component 1 with the molecular cloud detected by \hbox{\citet{Y99}} (see Fig. 2u from that work), shows that the angular size of Component 1 is about a factor 3 - 4 greater than the latter. Clearly, only the densest part of the molecular cloud associated with \rm NGC\,3503\, (clump A) was detected in the $^{13}$CO observations of \hbox{\citet{Y99}}. From here onwards, the analysis of the molecular gas associated with \rm NGC\,3503\, will focus on Component 1.
\subsubsection{Kinematics and excitation conditions}
The kinematics of Component 1 was studied by using position-velocity maps across selected strips. The map obtained along \hbox{{\it b} = +0\fdg12} (corresponding to clump A) is shown in the upper panel of Fig.~\ref{fig:pos-vel}. A noticeable velocity gradient is observed at velocities from about $-23$ \hbox{km s$^{-1}$\,} to about $-28$ \hbox{km s$^{-1}$\,}.
\begin{figure}
\centering
\includegraphics[width=240pt]{pos-vel.eps}
\caption{{\it Upper panel}: Velocity-Galactic Longitude map obtained in a strip along {\it b} = +0\fdg11 (corresponding to clump A) showing $T_R^\ast$. Contour levels start at 0.7 K and the contour spacing is 0.7 K. The dotted lines indicate the location of \rm NGC\,3503\,. {\it Lower panel}: Mean velocity map of Component 1. Contours levels go from $-26$ to $-24$ \hbox{\hbox{km s$^{-1}$\,}}, with interval of 0.2 \hbox{\hbox{km s$^{-1}$\,}}. The dotted ellipse depicts approximately the region of clump A. }
\label{fig:pos-vel}
\end{figure}
Additional support to the existence of this gradient can be obtained through a moment analysis. Due to the large angular dimensions of Component 1 and the spatial sampling of NANTEN observations, 28 independient CO profiles were observed towards the region enclosed by the 0.42 K-contour line in Fig.~\ref{fig:comp12y3}. Having these spectra a high signal-to-noise ratio \hbox{(S/N $>$ 10)}, they are well suited to study in some detail possible variations of the CO profiles across Component 1. With this aim, we used the AIPS package to calculate the first three moments (integrated area, temperature weighted mean radial velocity, and velocity dispersion) of the CO profiles. In the lower panel of Fig. ~\ref{fig:pos-vel}, the temperature weighted mean radial velocity distribution of CO is shown. A clear velocity gradient is observed along Component 1, and particulary in the region of clump A (depicted by the dotted ellipse), with mean radial velocities more negative toward the position of \rm NGC\,3503\,. A velocity shift of about $\sim$ 1.4 \hbox{km s$^{-1}$\,} is observed across the region of clump A, which at a distance of 2.9 kpc translates to a gradient $\omega$ $\approx$ 0.15 \hbox{km s$^{-1}$\,} pc$^{-1}$. This panel also shows that the CO emission corresponding to clump B does not show a significant velocity gradient.
In order to offer a complete picture of the kinematical properties of clump A, we show in Fig.~\ref{fig:mosaico1} the spatial distribution of the CO emission within the velocity range from --27.4 to --23.4 \hbox{km s$^{-1}$\,}. Every image depicts mean $T_R^\ast$-values (in contours) over a velocity interval of \hbox{1 \hbox{km s$^{-1}$\,}} superimposed on the 8.13 $\mu$m \rm MSX\,\ emission (in greyscale). In the velocity range from --27.4 to --26.4 \hbox{km s$^{-1}$\,} the molecular emission arising from clump A is slightly displaced from the brightest MSX emission located at {\it (l,b)} $\approx$ (289\fdg5, +0\fdg11). As we move towards more positive velocities, the maximum of the CO emission gradually shifts westwards.
Following \citet{U09}, we analyze the excitation temperature ($T_{\rm exc}$) obtained from the CO data to probe the surface conditions of Component 1.
The excitation temperature of the $^{12}$CO line can be obtained considering that $^{12}$CO is optically thick ($\tau_{\nu}$ $>>$ 1). Then, the peak temperature of this line is given by
\begin{equation}\label{eq:tpeak}
T_{peak}\ (^{12}CO)\ =\ J_{\nu}(T_{exc})\ -\ J_{\nu}(T_{bg})
\end{equation}
\citep{di78} where $J_{\nu}$ is the Planck function at a frequency $\nu$. Assuming gaussian profiles for the $^{12}$CO line, combining the order zero moment map (i.e., integrated area) with the order two moment map (i.e., velocity dispersion), and using Eq.~ \ref{eq:tpeak} we obtain the $T_{\rm exc}$ distribution map \hbox{(Fig. \ref{fig:momento})}. As expected, the $T_{\rm exc}$ distribution is quite similar to the CO emission distribution of Component 1, reaching a maximum of $\sim$ 17.7 K at \hbox{({\it l,b}) $\approx$ (289\fdg45, +0\fdg11)}. It is worth noting that the obtained value of $T_{exc}$ towards the center of clump A is lower than the obtained by \citet{U09} towards \hbox{{\it (l,b)} $\approx$ (289\fdg5, +0\fdg11)} \hbox{($T_{\rm exc}$ = 23.9 K)}. This difference may be explained in terms of a beam smearing of our NANTEN data, which implies that the values of $T_{\rm exc}$ shown in Fig. \ref{fig:momento} must be considered as lower limits.
\begin{figure}
\centering
\includegraphics[width=255pt]{momento1c.eps}
\caption{ Spatial distribution of $T_{\rm exc}$ for Component 1. Contours leves are 5.7, 7.9, 10.2, 12.4, 14.5, and 16.6 K. }
\label{fig:momento}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=175mm]{mosaico-co.eps}
\caption{Overlay of mean $T_R^\ast$-values (contours) in the velocity range from $-$27.4 to $-$23.4 km s$^{-1}$ and the MSX band A emission (greyscale). Every image represents the CO emission distribution averaged in a velocity interval of 1 km s$^{-1}$. The velocity interval is indicated in the bottom right corner of each image. The lowest temperature contour is 0.84 K ($\sim$ 12 rms). The contour spacing temperature is 1.4 K.}
\label{fig:mosaico1}
\end{figure*}
\subsubsection{Masses and densities}
The mass of the molecular gas associated with \rm NGC\,3503\, can be derived making use of the empirical relationship between the molecular hydrogen column density, $N(\rm H_2)$, and the integrated molecular emission, \hbox{$I_{{\rm ^{12}CO}}$ ($\equiv\int\ T^*_{R} \ d{\rm v}$)}. The conversion between $I_{{\rm ^{12}CO}}$ and $N(\rm H_2$) is given by the equation
\begin{equation}\label{eq:cero}
\quad N({\rm H_2})\ =\ (1.9\ \pm\ 0.3)\ \times\ 10^{20}\ I_{{ ^{12}CO}} \ \ \ \quad ({\rm cm}^{-2})
\end{equation}
\citep{ d96,sm96}. The total molecular mass $M_{\rm tot}$, was calculated through
\begin{equation}\label{eq:uno}
\quad M_{\rm tot}\ =\ (m_{sun})^{-1}\ \mu\ m_H\ \sum\ \Omega\ N({\rm H_2})\ d^2 \qquad \quad \ \quad ({\rm M}_{\odot})
\end{equation}
where $m_{sun}$ is the solar mass ($\sim$ 2 $\times$ 10$^{33}$ g), $\mu$ is the mean molecular weight, assumed to be equal to 2.8 after allowance of a relative helium abundance of 25\% by mass \citep{Y99,ya06}, $m_{H}$ is the
hydrogen atom mass ($\sim$ 1.67 $\times$ 10$^{-24}$ g), $\Omega$ is the solid angle subtended by the CO feature in ster, $d$ is the distance, expressed in cm, and $M_{\rm tot}$ is given in solar masses. For Component 1, we obtain a mean column density of $N(\rm H_2)$ = (1.9 $\pm$ 0.3) $\times$ 10$^{21}$ cm$^{-2}$, and a total molecular mass of \hbox{$M_{\rm tot}$ = (7.6 $\pm$ 2.1) $\times$ 10$^3$ M$_\odot$\,}.
The uncertainty in $M_{tot}$ ($\sim$ 28 $\%$) stems from the distance uncertainty and from the error quoted for the coefficient in Eq.~\ref{eq:cero}. Areas having $T^*_{R}$ $\geq$ 0.42 K were taken into account. The difference with the value derived by \citet{Y99}, (500 M$_\odot$\,), is due to both the subsampled observations of these authors and the fact that $^{13}$CO line is a better tracer of high density regions.
The mean volume density ($n_{\rm H_2}$) of Component 1 can be derived from the ratio of its molecular mass and its volume considering that the volume of Component 1 is the result of the addition of an ellipsoid with mayor and minor axis of 7\farcm 5 and 4\farcm 5, respectively, centered at \hbox{{\it (l,b)} $\approx$ (289\fdg42, +0\fdg15)} (which includes clump A), and a sphere of 4\farcm 5 in radius, centered at \hbox{{\it (l,b)} $\approx$ (289\fdg33, +0\fdg04)} (which includes clump B). We derived $n_{\rm H_2}$ $\approx$ 400 cm$^{-3}$. Taking into account the mass and volume uncertainties, we derive a conservative density error of about \hbox{60 $\%$} \hbox{($\sim$ 240 cm$^{-3}$).}
\subsection{Radio continuum emission}
\begin{figure*}
\centering
\includegraphics[width=420pt]{cont-perfil2.eps}
\caption{{\it Upper left panel:} Radio continuum image at 843 MHz (contours) superimposed to the DSSR image (grayscale). Contours levels go from 4 mJy beam$^{-1}$ \hbox{($\sim$ 3 rms)} to 12 mJy beam$^{-1}$ in steps of 4 mJy beam$^{-1}$, and from 20 mJy beam$^{-1}$ in steps of 10 mJy beam$^{-1}$. {\it Upper right panel:} Radio continuum image at 4800 MHz (contours) superimposed to the DSSR image (grayscale). Contours levels go from 2.4 mJy beam$^{-1}$ \hbox{($\sim$ 3 rms)} to 10.4 mJy beam$^{-1}$ in steps of 2 mJy beam$^{-1}$, and from 10.4 mJy beam$^{-1}$ in steps of 4 mJy beam$^{-1}$. The symmetry axis of the nebula is depicted by the dotted line. {\it Lower left panel:} Radio continuum image at 8640 MHz (contours) superimposed to the DSSR image (grayscale). Contours levels go from 15 mJy beam$^{-1}$ \hbox{($\sim$ 3 rms)} in steps of 1 mJy beam$^{-1}$. {\it Lower right panel:} Emission measure profile obtained from the 4800 MHz radio continuum image along the symmetry axis of \rm NGC\,3503\,.}
\label{fig:843}
\end{figure*}
Figure {\ref{fig:843}} shows the radio continuum images at 843, 4800, and 8640 MHz (in contours) superimposed on the DSSR image (in grayscale) in a region of $\sim$ 9$'$ $\times$ 9$'$ centred on \rm NGC\,3503\,.
The continuum images show an extended source coincident with \rm NGC\,3503\,. The maxima at the three frequencies coincide with the brightest optical emission region. The source, which has a good morphological correspondence with the optical emission, exhibits a cometary shape with its symmetry axis indicated by the dotted line in the image at 4800 MHz (Fig. {\ref{fig:843}, top right panel)}. The fainter emission area detected at 843 and 4800 MHz to the northeast of \rm NGC\,3503\, coincides with MBO and appears to be linked to the emission of the nebula at \hbox{({\it l,b}) $\approx$ (289\fdg53, +0\fdg13)}. The emission region detected at \hbox{{\it(l,b)} $\approx$ (289\fdg48, +0\fdg08)} at 843 MHz depicts some morphological correspondence with the bright rim of SFO 62. The point-like source placed at \hbox{({\it l,b}) $\approx$ (289\fdg49,+0\fdg15)} without optical emission counterpart is labelled G289.49+0.15 in Fig. {\ref{fig:843}}. From its flux densities \hbox{$F_{843}$ = 100 mJy}, \hbox{$F_{4800}$ = 17.1 mJy}, and \hbox{$F_{8640}$ = 11.5 mJy}, we derived a spectral index $\alpha$ = --0.92 $\pm$ 0.03 \hbox{(S$_{\nu}$ $\propto$ $\nu^{\alpha}$)}, which suggests that G289.49+0.15 is a non-thermal extragalactic source. We substracted the emission of this source in order to calculate radio continuum flux densities of \rm NGC\,3503\,. The results are given in Table \ref{table:flujos-cont}. The spectral index based on these estimates is \hbox{$\alpha$ = $-$0.12 $\pm$ 0.03}, typical for the thermal free-free radio continuum emission of an H{\sc ii} region.
\begin{table}
\caption{Radio continuum flux densities measurements of NGC\,3503}
\begin{center}
\begin{tabular}{ l c c c}
\hline
Frequency & 843 MHz & 4800 MHz & 8640 MHz \\
\hline\hline
&&&\\
Flux density (mJy) & 320 $\pm$ 30 & 270 $\pm$ 50 & 240 $\pm$ 40 \\
&&&\\
\hline
\label{table:flujos-cont}
\end{tabular}
\end{center}
\end{table}
We estimate the emission measure $EM$ $\left(= \int n_e^2\ dl \right)$ along the symmetry axis of the nebula using the image at 4800 MHz. The brightness temperature ($T_b$) is related to the optical depth ($\tau$) by
\begin{eqnarray}
\quad T_b(\nu) = T_e \times (1-e^{-\tau})
\label{formula:tb}
\end{eqnarray}
where $T_e$ = 8100 $\pm$ 700 K \citep{q06} is the electron temperature and the optical depth $\tau$ is given by
\begin{eqnarray}
\quad \tau = 8.235\ \times\ 10^{-2}\ T_e^{-1.35}\ \nu^{-2.1}\ EM
\label{formula:tau}
\end{eqnarray}
In this last expresion, $\nu$ is given in GHz, and $EM$ in pc cm$^{-6}$. The $EM$ profile is shown in the lower right panel of Fig. {\ref{fig:843}}. As expected, two maxima are observed, one at the brightest section of \rm NGC\,3503\, and the other coincident with MBO, indicating that the electron density is higher towards these regions than in the sorroundings. The $EM$ profile of \rm NGC\,3503\, is consistent with the cometary shape morphology of the nebula seen in the radio continuum and optical images. The $EM$ profile shows a sharp border towards the west of the nebula, while towards the east it decreases smoothly, indicating the existence of an electron density gradient. The observed $EM$ profile is in agreement with the results of \citet{c00}, who reported the existence of an electron density gradient in the east-west direction. Using the peak values \hbox{$EM$ = 18000 pc cm$^{-6}$} (obtained at \hbox{{\it(l,b)} $\approx$ (289\fdg50, +0\fdg11)} for \rm NGC\,3503\,) and \hbox{$EM$ = 7000 cm$^{-6}$} (obtained at \hbox{{\it(l,b)} $\approx$ (289\fdg53, +0\fdg18)} for MBO), and considering a pure hydrogen plasma with a depth along the line of sight equal to the size in the plane of the sky, electron density estimates \hbox{$n_e$ = 75 $\pm$ 14 cm$^{-3}$} and \hbox{$n_e$ = 45 $\pm$ 9 cm$^{-3}$} are obtained for \rm NGC\,3503\, and MBO, respectively.
\begin{figure*}
\centering
\includegraphics[width=500pt]{pdr-total-final2.eps}
\caption{{\it Left panel} Composite image of \rm NGC\,3503\, and its environs. Red and green show emission at 8.13 $\mu$m (MSX) and 24 $\mu$m (MIPSGAL). The color scale goes from 1$\times$10$^{-5}$ to 4$\times$10$^{-5}$ \hbox{W m$^{-2}$ sr$^{-1}$} and from 27 to 100 \hbox{MJy ster$^{-1}$}, respectively. The white rectangle encloses the IR counterpart of \rm NGC\,3503\, (IRK). {\it Right panel}: Composite image of \rm NGC\,3503\,. Red, green and blue show emission at 8 $\mu$m (IRAC-GLIMPSE), 24 and 70 $\mu$m (MIPSGAL), respectively. Colors scales range from 40 to 500 \hbox{MJy ster$^{-1}$} (24 $\mu$m), from 300 to 2000 \hbox{MJy ster$^{-1}$} (70 $\mu$m), and from 30 to 400 \hbox{MJy ster$^{-1}$} (8 $\mu$m). The location of the members of Pis 17 is indicated with black crosses. MSX and 2MASS candidate YSOs are indicated with red and green crosses (see text).}
\label{fig:msx-halfa}
\end{figure*}
As a different approach to \rm NGC\,3503\,, the rms electron density and ionized mass can be obtained using the spherical model of \citet{mh67} and the flux density at 4800 MHz. Assuming a constant electron density, and a radious \hbox{$R_{\rm HII}$ = 1.25 pc}, we obtain \hbox{$n_e$ = 54 $\pm$ 13 cm$^{-3}$} and \hbox{$M_{\rm HII}$ = 11 $\pm$ 3 M$_\odot$\,}.
Thus, the electron density of \rm NGC\,3503\, is about a factor of 5 lower than the density of Component 1, which clearly indicates that the nebula has expanded.
An estimate of the filling factor \hbox{($f$ = $\sqrt{n_e/n_e'}$)} can be obtained by taking into account the maximum electron density derived from optical lines \hbox{($n_e'$ = 154$^{+52}_{-45}$ cm$^{-3}$; \citealt{c00})}. We obtain $f$ = 0.5 - 0.8. Then, the ionized mass for $f$ = 0.5 - 0.8 is in the range 8 - 10 M$_\odot$\,.
Regarding MBO, rms electron densities and masses estimated from the image at 843 MHz are \hbox{$n_e$ $\simeq$ 33 cm$^{-3}$} and \hbox{$M_{\rm HII}$ $\simeq$ 6 M$_\odot$\,}, respectively.
\subsection{Infrared Emission}
The emission distribution at 8.13 $\mu$m (MSX-A band) superposed to the image at 24 $\mu$m (MIPSGAL) is shown in the left panel of Fig.~\ref{fig:msx-halfa}. The \rm MSX\, Band {\rm A} includes the strong emission features at 7.7 and 8.6 $\mu$m attributed to {\rm PAH} molecules, which are considered tracers of {\rm UV}-irradiated {\it Photodissociated Regions} (PDR) \citep{ht97}. This image displays a small and strong feature, from hereonwards dubbed the {\it IR Knot}, or {\rm IRK} for short (indicated with a white rectangle), which is coincident with the location of \rm NGC\,3503\,, and a weaker and more extended emission region detected in the north-western area of the image. The last feature will be referred to as the {\it Extended IR Emission}, or {\rm EIE} for short. IR emission at 8.3 and 24 $\mu$m appears mixed along the whole feature. A region of low IR emission is seen between the IRK and the EIE. The bright rim of SFO 62 is detected in the IR as the bright filament observed from \hbox{($l, b$) $\simeq$ (289\fdg42, +0\fdg03)} to \hbox{($l, b$) $\simeq$ (289\fdg50, +0\fdg08)}, which appears to be the southernmost boundary of EIE. The presence of PAH emission coincident with this bright rim suggests the existence of a PDR at the southern edge of Component 1.
The right panel of Fig.~\ref{fig:msx-halfa} shows a detailed image of the IRK, at 8.0 $\mu$m ({\it Spitzer}-IRAC band 4, in red), 24 $\mu$m (MIPSGAL, in green), and 70 $\mu$m (MIPSGAL, in blue). The emissions at 8 and 70 $\mu$m show a bright half shell-like feature, which encloses the position of the stars of Pis 17, indicated as black crosses. Arc-shaped faint filaments are also detected in the 8 $\mu$m emission in the northeastern section of IRK. The morphology of the IR emission is compatible with the cometary-shape of the H{\sc ii}\, region and very likely indicates the presence of a PDR between \rm NGC\,3503\, and clump A. The MIPSGAL emission at 24 $\mu$m, which is projected at the center of the structure, indicates the existence of warm dust inside the H{\sc ii} region (as the cases of N10 and N21, \citealt{w08}).
\begin{figure}
\centering
\includegraphics[width=250pt]{tdust-dss.eps}
\caption{ {\it Upper panel:} $T_{\rm d}$ distribution estimated from HIRES IRAS images at 60 and 100 $\mu$m. Contour levels go from 25 to 31 K in steps of 1 K, and from 33 K in steps of 2 K. {\it Lower panel:} Overlay of the $T_{\rm d}$ distribution (contours) and the DSSR image.}
\label{fig:tdust}
\end{figure}
IRAS 60 and 100 $\mu$m data show dust with color temperatures between about 20 to 100 K, which corresponds to the ``cool dust component'' (see \citealt{sr11}, and references therein). A dust temperature ($T_{\rm d}$) map was produced using the equation
\begin{equation}\label{td}
\qquad T_{\rm d}\ =\ 95.94\ / \ \textrm{ln}(B_n) \qquad ({\rm K})
\end{equation}
\citep{d90,wh92,ci01}, where $B_n$ = 1.667$^{(3+n)}$$\left(F_{100}/F_{60} \right)$ is the modified Planck function, with $F_{100}$ and $F_{60}$ being the 100 $\mu$m and 60 $\mu$m fluxes, respectively. The parameter $n$ = 1.5 is related to the absorption efficiency of the dust ($k_{\nu}\ \propto\ \nu^n$, normalized to 40 cm$^2$ g$^{-1}$ at 100 $\mu$m). The obtained dust temperature map is shown in the upper panel of Fig. \ref{fig:tdust}. Dust temperature goes from 25 K to 46 K. Fig. \ref{fig:tdust} (lower panel) also shows that the region with the highest dust temperatures ($\sim$ 46 K) coincides with the brightest section of \rm NGC\,3503\,. This temperature is in good agreement with those obtained in RCW 121 and RCW 122 \citep{a08} and in NGC 6357 \citep{cb11} although they are slightly higher than typical values for H{\sc ii} regions \hbox{($\sim$ 30 K)} (see \citealt{cn08,ci09,v10}). The stellar UV radiation field of the stars in NGC 3503 is responsible for the heating of the dust immersed in the H{\sc ii}\, region. A noticeable feature is a fringe of cold ($\sim$ 25 K) dust surrounding the position of \rm NGC\,3503\,, which coincides with the region showing low emission at 8.3 $\mu$m placed between IRK and EIE described above. An extended region of warmer ($\sim$ 36 K) dust is detected to the northeast of \rm NGC\,3503\,, coincident with the position of MBO, indicating that this optical feature has a radiatively heated dust component.
In Table \ref{table:IR} we summarized the IR and dust parameters for IRK and EIE. The averaged dust temperatures ($\bar{T}_{\rm d}$) are calculated using Eq. \ref{td} and flux densities at 60 and 100 $\mu$m obtained integrating several polygons around the sources. Dust masses are derived from
\begin{equation}\label{}
M_d = m_n\ F_{60}\ d^2 \ (B_n^{2.5}\ -\ 1) \qquad ({\rm M}_{\odot})
\end{equation}
where {\it d} the distance in kpc, $F_{60}$ is given Jy, and \hbox{$m_{1.5}$ = 0.3 $\times$ 10$^{-6}$}. Based on the molecular gas mass derived for Component 1, the ionized gas of \rm NGC\,3503\, and MBO, and the dust masses obtained for the IRK and the EIE ($\sim$ 10 M$_\odot$\,), we derive a mean weighted gas-to-dust ratio of about \hbox{$\sim$ 800 $\pm$ 300}. This ratio is a factor of about $\sim$ 5 higher than the average values assumed for the Galaxy (140 $\pm$ 50; \citealp{t00}). Note however that dust with temperatures $T_d\ \lesssim\ 20$ K (``cold dust component''), which is detected at wavelengths not surveyed by IRAS \hbox{($\lambda$ $>$ 100 $\mu$m)}, dominates the dust mass by a factor larger than $\sim$ 70 with respect to the cool dust component \citep{sr11}. This notoriously decreases the gas-to-dust ratio. The dust temperature estimated for EIE may suggest the existence of ionizing sources inside Component 1. However, a search for OB stars embedded in Component 1 performed using the available VIZIER catalogues failed to detect such sources.
To investigate the presence of protostellar candidates possibly related to NGC\,3503, we used data from the MSX, 2\,MASS, and IRAS point source catalogs. We searched for point sources in a region of about 10\arcmin\ in size centered at the position of the H{\sc ii}\,\ region. Taking into account sources with flux quality $q >$ 2, a total of 69 MSX point sources were found projected onto the area. Based on $F_{21}$/$F_{8}$ and $F_{14}$/$F_{12}$ ratios, where $F_8$, $F_{12}$, $F_{14}$, and $F_{21}$ are the fluxes at 8.3, 12, 14, and 21 $\mu$m, respectively, and after applying the criteria summarized by \citet{lu02}, we were left with three sources with $F_{21}$/$F_8 >$ 2 and $F_{14}$/$F_{12} <$ 1, which are classified as compact H{\sc ii}\, regions (CH{\sc ii}\,). This is indicative of stellar formation. The location of these sources is indicated with red crosses in Fig. \ref{fig:msx-halfa}.
\begin{table}
\caption{ Main infrared parameters inferred from IRAS fluxes at 60 $\mu$m and 100 $\mu$m}
\begin{center}
\begin{tabular}{lcc}
\hline
{\it Parameter} & IRK & EIE \\
\hline\hline
&&\\
S$_{60}$ (Jy) & 1470 $\pm$ 20 & 700 $\pm$ 15 \\
S$_{100}$ (Jy) & 2170 $\pm$ 25 & 1970 $\pm$ 25 \\
$\bar{T}_d$ (K) & $\sim$ 37 & $\sim$ 28 \\
$M_d$ (M$_\odot$\,) & $\sim$ 3 & $\sim$ 7 \\
&&\\
\hline
\label{table:IR}
\end{tabular}
\end{center}
\end{table}
\begin{table*}
\begin{center}
\caption{Candidate YSOs obtained from the MSX and 2\,MASS catalogues.}
\label{yso}
\begin{tabular}{ccccccc}
\hline
& ($l,b$) & MSX source & F$_8$ (Jy) & F$_{12}$ (Jy) & F$_{14}$ (Jy) & F$_{21}$ (Jy) \\
\hline
& & & & & & \\
1 & 289.48,+0.14 & G289.4859+00.1420& 1.49250 & 1.8145 & 0.9005 & 0.3180 \\
2 & 289.49,+0.12 & G289.4993+00.1231& 0.82699 & 1.4887 & 1.2237 & 4.3818 \\
3 & 289.50,+0.11 & G289.5051+00.1161& 0.87911 & 2.8632 & 1.6759 & 5.6330 \\
&&&&&&\\
\hline
& ($l,b$) & 2\,MASS source & $J$ (mag) & $H$ (mag) & $K$ (mag) & \\
\hline
& & & & & & \\
4 & 289.49,+0.14 & 11011934-5949422& 13.277& 13.047& 12.806 &\\
5 & 289.50,+0.15 & 11012021-5949123& 15.084& 14.837& 14.506 &\\
6 & 289.50,+0.13 & 11011933-5950037& 14.539& 14.100& 13.696 &\\
7 & 289.50,+0.14 & 11012329-5949269& 14.862& 14.380& 14.027 &\\
8 & 289.54,+0.13 & 11013591-5951037& 15.237& 14.592& 14.106 &\\
9 & 289.53,+0.12 & 11013263-5951140& 15.633& 14.946& 14.404 &\\
10 & 289.52,+0.10 & 11012238-5952449& 12.747& 11.744& 10.836 &\\
11 &289.52,+0.15 & 11013238-5949339& 14.754& 14.243& 13.873 &\\
12 & 289.48,+0.14 & 11011390-5949099& 14.888& 13.661& 12.827 &\\
& & & & & & \\
\hline
\end{tabular}
\end{center}
\end{table*}
Using the 2\,MASS catalog \citep{cu03}, which provides detections in $J$, $H$ and $K_s$ bands, we searched for point sources with infrared excess. Taking into account sources with signal-to-noise ratio (S/N) $>$ 10 (corresponding to quality ``AAA''), we found 5528 sources projected onto a circular region of 5\arcmin\ in radius.
Following \citet{co05}, we determined the parameter \hbox{$q$ = ($J$ - $H$) - 1.83 $\times$ ($H$ - $K_s$)}. Sources with $q <$ --0.15 are classified as objects with infrared excess , i.e. candidate YSOs. By applying the above criteria, we found only 9 sources projected onto \rm NGC\,3503\,. The location of these sources is also indicated in Fig. \ref{fig:msx-halfa} with green crosses. The location of the 2\,MASS source 11011390-5949099 (source $\#$12) is almost coincident with the MSX source G289.4859+00.1420 (source $\#$1). These sources are projected onto an intense mid-IR clump located in the northeastern border of the bright half shell-like feature, at \hbox{($l,b$) = (289\fdg48,+0\fdg14 )} (see Fig. \ref{fig:msx-halfa}, right panel), and onto a small enhancement in the radio continuum emission detected at 4800 MHz (see Fig. \ref{fig:champ-ir-halfa} below). These characteristics make this object an excellent candidate for investigating star formation with high angular resolution observations. The names of the YSO candidates, their position, fluxes and magnitudes at different IR wavelengths are listed in Table \ref{yso}.
\section{Discussion}
\subsection{The ionized gas}
The number of ionizing Lyman continuum photons ($N_{\rm Lyc}$) needed to sustain the ionization in \rm NGC\,3503\, can be calculated using the equation given by \citet{sr90}
\begin{eqnarray}
\quad N_{\rm Lyc}\ =\ 4.45 \times 10^{48}\ T_e^{-0.45}\ S_{(4800\ {\rm MHz})}\ d^2
\label{formula:nlym}
\end{eqnarray}
where $d$ is the distance in kpc, and $S_{(4800\rm \ MHz)}$ is the flux density at 4800 MHz in Jy. We obtain \hbox{$N_{\rm Lyc}$ = (1.8 $\pm$ 0.4) $\times$ 10$^{47}$ s$^{-1}$}. This number is a lower limit to the total number of Lyman continuum photons required to maintain the gas ionized, since about 25 - 50 $\%$ of the UV photons are absorbed by interstellar dust in the H{\sc ii} region \citep{i01}. Consequently, we need \hbox{$N_{\rm Lyc}$ $\approx$ 3.6 $\times$ 10 $^{47}$ s$^{-1}$}. The number of Lyman continuum photons emitted by a B0 V star is $N_{\rm Lyc}^*$ = 1.07 $\times$ 10$^{48}$ s$^{-1}$ \citep{st03}, which is capable of ionizing \rm NGC\,3503\,, in agreement with previous results by \citet{T04} and \citet{pco10}. The last authors asserted one B0 V star and three B2 V stars, belonging to the open cluster Pis17, inside the nebula as ionizing sources. Since the ionizing photons emitted by a B2 V star are significantly fewer than those of a B0 V star, their contribution to the energetics of the nebula can be neglected.
In Sec. 3.2 we pointed out on the discrepancy between the electron and molecular densities. The density of Component 1 is about a factor of 5 higher than the electron density of \rm NGC\,3503\,, which clearly indicates that \rm NGC\,3503\, has been expanding as a result of unbalanced pressure between ionised and molecular gas.
In order to estimate the dynamical age ($t_{\rm dyn}$) of the H{\sc ii}\, region we used the model of \citet{dw97}. The radious of an H{\sc ii}\, region ($R_{\rm HII}$) in a uniform medium is given by
\begin{equation}
\quad \frac{R_{\rm HII}}{R_{\rm S}}\ =\ \left(1\ + \frac{7\ {\rm v}_{\rm s}\ {t}_{\rm dyn}}{4\ R_{\rm S}}\right)^{4/7}
\end{equation}
were $R_S$ is the radious of the Str\"omgren sphere \citep{s39} before expanding, given by \hbox{$R_{\rm S} = \left(3 N^*_{Lyc} / 4 \pi\ (2n_{\rm H_2})^2 \alpha_{\beta} \right)^{1/3}$}, and v$_{\rm s}$ is the sound speed in the ionized gas \hbox{($\sim$ 10 \hbox{km s$^{-1}$\,})}. For \hbox{$R_{\rm HII}$ $\simeq$ 1.25 pc} (1\farcm5 at a distance of 2.9 kpc), an ambient density \hbox{$n_{\rm H_2}$ = 400 $\pm$ 240 cm$^{-3}$} (see Section 3.1.3), and $N^*_{Lyc}$ = 1.07 $\times$ 10$^{48}$ s$^{-1}$, we infer \hbox{$t_{\rm dyn}$ $\approx$ 2 $\times$ 10$^5$} yr. The expansion velocity of the H{\sc ii}\, region can be estimated by means of $\dot{R}_{\rm HII}$ = v$_s$ $\left( \frac{R_{\rm HII}}{R_s} \right)^{-3/4}$, yielding an expansion velocity of about $\sim$ 5 km s$^{-1}$. The obtained dynamical age and expansion velocity are in agreement with those obtained in typical H{\sc ii}\, regions (Gum 31, \citealt{cn08}; Sh2-173, \citealt{ci09}).
The total number of ionizing Lyman continuum photons needed to sustain the ionization in MBO is \hbox{$N_{\rm Lyc}$ $\approx$ 2.5 $\times$ 10$^{47}$ seg$^{-1}$}. To search for stars that can provide the necessary UV photons, we used the available VIZIER catalogues. CP-59 2951 is a B1 V star placed at \hbox{{\it(l,b)} = (289\fdg511, +0\fdg195)} \citep{bu99}. Using the catalogued spectral type, the visual magnitude, and the calibration of \citet{sk82}, we estimated a distance \hbox{$d$ = 2.7 $\pm$ 0.8 kpc}. The continuum ionizing photons emitted by the star are \hbox{$N_{\rm Lyc}^*$ = 3.2 $\times$ 10$^{46}$ s$^{-1}$} \citep{sm02}. Although the distance of this star is in agreement (within errors) with the distance of \rm NGC\,3503\,, the value of $N^*_{\rm Lyc}$ is almost one order of magnitude lower than the required to ionize MBO. The presence of Pis17 at a projected distance of 3.4 pc ($\sim$ 4$'$ at a distance of 2.9 kpc) indicates that the contribution of the star cluster to the ionization of MBO can not be ruled out.
\subsection{Star formation process}
The presence of 12 candidate YSOs projected onto the \rm NGC\,3503\,-region suggests that they may have been triggered by the expansion of the H{\sc ii}\, region through the ``collect and collapse'' model, which indicates that expanding nebulae compress gas between the ionization and the shock fronts, leading to the formation of molecular cores where new stars can be embedded. Using the analytical model of \citet{wi94} for the case of expanding H{\sc ii}\, regions, we derived the time when the fragmentation may have occurred ($t_{frag}$), and the size of the H{\sc ii}\, region at $t_{frag}$ ($R_{frag}$), which are given by
\begin{eqnarray}
t_{frag} [10^6 yr]\ =\ 1.56 a_2^{4/11}\ n_3^{-6/11}\ N_{49}^{-1/11}
\end{eqnarray}
\begin{eqnarray}
R_{frag} [pc]\ =\ 5.8 a_2^{4/11}\ n_3^{-6/11}\ N_{49}^{1/11}
\end{eqnarray}
where $a_2$ is the sound velocity in units of 0.2 km s$^{-1}$\,, $n_3 \equiv n_{H_2}/1000$, and $N_{49}\equiv N_{Lyc}^*/10^{49}$. Adopting for this region 0.3 km s$^{-1}$\, for the sound velocity, which corresponds to temperatures of 10-15 K in the surrounding molecular clouds (see Section 3.1.2), we obtained \hbox{$t_{frag} \sim$ 3.5 $\times$ 10$^6$ yr}, and \hbox{$R_{frag} \sim$ 7.5 pc}. Considering that the values of $t_{frag}$ and $R_{frag}$ are larger than $t_{dyn}$ and $R_{HII}$ (see Section 4.1), we can conclude that the fragmentation in the edge of NGC\,3503 is doubtful.
\subsection{Kinematics of clump A}
In section 3.2 we reported the existence of a velocity gradient across clump A. Velocity gradients in molecular cores/clumps were usually interpreted as gravitationally bound rotation motions (particulary in dense and small molecular cores). Several theoretical models predict cloud flattening perpendicular to the rotation axis in response to centrifugal stress, which at first glance seems to be suitable for clump A.
In order to investigate the dynamical stability of clump A, we use the parameter $\beta$ defined by \citet{go93} to quantify the dynamical role of rotation by comparing the rotational kinetic energy to the gravitational energy. Thus, $\beta$ can be written as
\begin{equation}\label{eq:Tr}
\qquad \beta = \frac{(1/2)\ I\ \omega'^2}{q\ G\ M^2/R}\ =\ (1/2)\ (p/q)\ \frac{\omega'^2\ R^3}{G\ M}
\end{equation}
where $I$ is the moment of inertia ($I=pMR^2$), $q G M^2/R$ is the gravitational potential energy, and $\omega'=\omega/sin(i)$, where $i$ is the inclination of the cloud along the line of sight. Considering \hbox{$p/q$=0.22} (see \citealt{go93}), \hbox{sin($i$) = 1}, \hbox{$\omega$=0.15 km s$^{-1}$\, pc$^{-1}$} (see Sect 3.1.2), \hbox{$R$ $\sim$ 4.5 pc}, and a lower limit mass to clump A \hbox{($M$ $\sim$ 3 $\times$ 10$^{3}$ M$_\odot$\,)}, we obtain \hbox{$\beta\ \approx$ 0.02}. This extremely low value of $\beta$ indicates that the effect of rotation, if exists, is not significant in mantaining the dynamical stability of \hbox{clump A}. This might weaken the rotating cloud interpretation. Furthermore, Figs. \ref{fig:pos-vel} and \ref{fig:mosaico1} show that only molecular gas at more negative velocities ($\sim$ $-$27 \hbox{km s$^{-1}$\,} / $-$26 \hbox{km s$^{-1}$\,}) coincide with \rm NGC\,3503\,, which suggests that the velocity gradient might be a direct consequence of an interaction between clump A and \rm NGC\,3503\,.
Therefore, although rotation can not be entirely ruled out with the present data, we consider different origins for the velocity gradient observed across clump A, namely, 1) the expansion of the nebula at $\sim$ 5 \hbox{km s$^{-1}$\,} (see Sect 4.1) has been accumulating molecular gas behind the shock front which originated the expansion of clump A at approximately the same velocity of the ionized gas, as expected according to the models of \citet{hi96}, 2) the velocity gradient of clump A is the consequence of a collision between Component 1 and another molecular cloud, which in turn, might have induced the formation of Pis 17 and the candidate YSOs reported in Sect. 3.3. Although they are rare, cloud-cloud collisions can lead to gravitational instabilities in the dense, shocked gas, resulting in triggered star formation (see \citealt{elm98}, and references therein), 3) clump A is actually composed by different subclumps at different velocities, which are not resolved by the NANTEN observations. In this case, \rm NGC\,3503\, might be related with a subclump at more negative velocities.
Further high-resolution studies with instruments like APEX may help to clarify this question.
\subsection{The PDR at the interface between NGC 3503 and clump A}
As mentioned in Sect 3.1.2, clump A exhibits an excitation temperature \hbox{$T_{exc}$ $\geq$ 17.7 K}, which is also the highest excitation temperature along Component 1. This temperature is higher than expected inside molecular cores if only cosmic ray ionization is considered as the main heating source \hbox{($T$ $\sim$ 8 - 10 K, \citealt{vdt00})}, which implies that additional heating processes are present close to clump A. Very likely, clump A is being externally heated through the photoionisation of its surface layers \citep{U09} as a consequence of its proximity to \rm NGC\,3503\,. This scenario is in line with the presence of the PDR at the interface between \rm NGC\,3503\, and clump A (see \hbox{Sect. 3.3}). In this context, it would be instructive to make a simple comparison of the PDR dust surface temperature, with the dust temperature obtained before towards the edge of \rm NGC\,3503\, (see Fig.\ref{fig:tdust}).
The structure of a PDR is governed by the intensity of the UV radiation field impinging onto the cloud surface ($G_0$) in units of the \citet{ha68} FUV flux (1.6 $\times$ 10$^{-3}$ ergs cm$^{-2}$ s$^{-1}$), with $G_0\propto N_{\rm Lyc}^{-2/3}\ \ n_e^{4/3}\ \ L_{\star}\ \chi$ \citep{ti05}, where $\chi$ is the fraction luminosity over 6 eV, and $L_{\star}$ is the luminosity of the star. Adopting $N_{\rm Ly}$ = 1.07$\times$10$^{48}$ s$^{-1}$ and $L$ = 7.6 $\times$ 10$^4$ $L_{\odot}$ \citep{st03}, $\chi$ = 1, and considering $n_e\simeq$ 75 -- 154 cm$^{-3}$ (see Sect. 3.2), a range of $G_0\simeq$ (0.5 - 1.5) $\times$ 10$^3$ is obtained. The distribution of dust temperature in the PDR ($T_d^{pdr}$) is governed by the absorption of the stellar photons which are reemited by dust grains as IR photons. Following \citet{ti05}, we can relate $T_d^{pdr}$ with the visual absorption along the PDR ($A_{\rm v}$) as
\begin{equation}\label{g0}
(T_d^{pdr})^5 \simeq \nu_0\ G_0\ e^{(-1.8A_{\rm v})} +\ \ln(3.5\times10^{-2}\tau_{100}\ T_0)\ \tau_{100}\ {T_0}^6
\end{equation}
\noindent{where $\nu_0$ is the frequency at 0.1 $\mu$m, $\tau_{100}=10^{-3}$ is the effective optical depth at 100 $\mu$m, and $T_0$ is the temperature of the slab estimated as \hbox{$T_0$ = 12.2 {$G_0$}$^{1/5}$}}. Considering both, the position of NGC\,3503 at the edge of clump A, and that the low value of visual absorption derived for the members of Pis\,17 ($\bar{A}_{\rm v}$ = 1.6 mag, \citealt{pco10}) is probably due to the interstellar absorption in the line of sight of the H{\sc ii}\, region, we can assume $A_{\rm v}\sim$ 0 for the surface of the PDR. Taking into account the range of $G_0$ and Eq.~\ref{g0}, we obtain \hbox{$T_d^{pdr}$ $\simeq$ 40 - 55 K}, in accordance with the observed dust color temperature estimated for NGC\,3503 \hbox{($\sim$46 K)}. This gives additional support to the existence of a PDR between \rm NGC\,3503\, and clump A.
\begin{figure*}
\centering
\includegraphics[width=170mm]{modelo-finalb2.eps}
\caption{ Composite image of \rm NGC\,3503\, and its environs. Red and green show emission at 8 and 4.5 $\mu$m (IRAC-GLIMPSE), respectively. White and light blue contours show the radiocontinuum 4800 MHz and CO line emission, respectively. }
\label{fig:champ-ir-halfa}
\end{figure*}
\subsection{Possible scenario}
Figure \ref{fig:champ-ir-halfa} displays a composite image of NGC3503 and its environs. White and light blue contours show the radio continuum emission at 4800 MHz corresponding to NGC3503 and MBO, and the CO emission, respectively. The color scale shows the emission at 8 $\mu$m (red) and 4.5 $\mu$m (green).
As described in Sect. 3.3 and 4.3, emission attributed to PAHs encircles the southern and western borders of the H{\sc ii}\, region, indicating the location of the PDR. The position of NGC 3503 near the border of clump A and the existence of an electron density gradient along the symmetry axis of the H{\sc ii}\, region (with the higher electron densities close to the strongest CO emission) suggest that the ionised gas is being streamed away from the densest part of the molecular cloud. This is indicative that \rm NGC\,3503\, is a blister-type H{\sc ii}\, region \citep{i78} that probably has undergone a champagne phase.
The so-called {\it Champagne flow model} \citep{tt79,btt81,ttb82} proposes that an expanding H{\sc ii}\, region placed at the edge of a molecular cloud eventually reaches the border of the cloud and expands freely in the lower density surrounding gas, originating an extended H{\sc ii}\, region with a characteristic density distribution. The H{\sc ii}\, region becomes density bounded towards the lower density region, while it is ionization bounded towards the molecular cloud. This scenario was formerly proposed by \citet{c00} to explain the electron density gradient observed across NGC 3503. The location of Pis 17 close to the brightest radio continuum region is consistent with a projection effect \citep{ytb83}. In this scenario, Pis 17 originated NGC 3503, which reached the northeastern border of \hbox{clump A} after 2$\times$ 10$^5$ yr. This time represents a lower limit to the age of the H{\sc ii}\, region, since the inferred main-sequence life time of the main ionizing star is about \hbox{(3 -5) $\times$ 10$^6$ yr} \citep{m98}. The leakage of ionized gas and UV photons might have contributed to the formation of MBO, since its location (along the symmetry axis of \rm NGC\,3503\,) and its low density (about half of \rm NGC\,3503\,) suggest that this feature may consist of ionized gas which has scaped from \rm NGC\,3503\, after a time of 2$\times$ 10$^5$ yr.
The velocities of both the molecular and ionized gas are compatible with the champagne scenario. The mean velocity of Component 1 (--24.7 \hbox{km s$^{-1}$\,}, see Sect. 3.1.1) corresponds to the natal cloud where NGC 3503 originated. Molecular gas having more negative velocities is mainly linked to clump A and may represent material moving away from the ionized gas placed in front of the H{\sc ii}\, region (as a result of expanding motions or an external impact over Component 1). The ionized gas, having a velocity of --21 \hbox{\hbox{km s$^{-1}$\,}} \citep{g00}, might be receding the observer. We note, however, that the Fabry-Perot observations by \citet{g00} have a spectral resolution of 5 \hbox{km s$^{-1}$\,}.
\rm NGC\,3503\, resembles the two well studied cases of S305 and S307. These H{\sc ii}\, regions show a non spherical morphology in the 1465 MHz map of \citet{fi93}, and they are placed close to $^{12}$CO emission peaks, which suggests that they are located close to the hottest part of their parental molecular clouds (see \citealt{ru95}, and references therein). Both objects show significant electron density dependence on position, which can also be interpreted as radial gradients \citep{c00}. Furthermore, both H{\sc ii}\, regions show discrepancies of about \hbox{$\sim$ 12 \hbox{km s$^{-1}$\,}} between the velocities of the ionized and molecular gas, probably due to a champagne effect producing a flow toward the observer \citep{ru95}. In the case of S307, the champagne scenario is reinforced by its half-shell shape, and the high density in the brightest part of the nebula \citep{fh81,as86}
According to the champagne scenario, a velocity gradient along the symmetry axis of the nebula can be expected \citep{glg94}. In the case of \rm NGC\,3503\, increasing velocities with the distance to the molecular cloud are expected. Thus, high spectral resolution radio recombination are required to analyze the ionized gas velocities and may give additional support to our interpretation.
\section{Summary}
NGC\,3503 is a bright H{\sc ii}\,\ region of $\sim$ 3\arcmin\ in size centered at \hbox{{\it(l,b)} = (289\fdg51, +0\fdg12)} located at a distance of 2.9 $\pm$ 0.4 kpc. The ionizing sources are B-stars belonging to the open cluster Pis\,17.
With the aim of investigating the molecular gas and dust distribution in the environs of the H{\sc ii}\,\ region and analyzing the interaction of the nebula and Pis\,17 with their molecular environment, we analyzed $^{12}$CO(1-0) data of a region of 0\fdg 7 in size obtained with the NANTEN telescope with an angular resolution of 2\farcm 7, radio continuum data of NGC\,3503 at 4800 and 8640 MHz obtained with the ATCA telescope (with synthesized beams of 21\arcsec\ and 13\arcsec, respectively), and data at 843 MHz retrieved from SUMSS, and available IRAS, MSX, IRAC-GLIMPSE, and MIPSGAL images.
The analysis of the CO data allowed the molecular gas linked to the nebula to be mapped. This molecular gas (Component 1) has a mean velocity of --24.7 \hbox{km s$^{-1}$\,}, a total mass and density of (7.6 $\pm$ 2.1) $\times$ 10$^3$ M$_\odot$\, and 400 $\pm$ 240 cm$^{-3}$, respectively, and displays two clumps centered at \hbox{{\it(l,b)} = (289\fdg47, +0\fdg12)} (clump A) and \hbox{{\it(l,b)} = (289\fdg32, +0\fdg03)} (clump B). The morphological correspondence of the molecular emission with a large patch of high optical absorption adjacent to NGC\,3503 is excellent. NGC\,3503 is projected near the border of clump A, with the strongest molecular emission adjacent to the highest electron density regions. The agreement of the molecular velocities with the velocity of the ionizad gas (--21 \hbox{km s$^{-1}$\,}) as well as the morphological correspondence with the nebula indicate that clump A is associated with NGC\,3503. The more negative velocities of the gas in clump A, coincident with the H{\sc ii}\,\ region, are probably due to the expansion of the H{\sc ii}\,\ region or an external impact over Component 1.
The analysis of the radio continuum images confirms the electron density gradient previously found by \citet{c00}. The images show that NGC\,3503 exhibits a cometary morphology, with the higher density region near the maximum of clump A. The rms electron density and the ionized mass amount to 54 $\pm$ 13 cm$^{-3}$ and 9 $\pm$ 3 M$_\odot$\,, respectively. A low density ionized region (MBO) located close to the lower electron density area of NGC\,3503 is also identified in these images.
Strong emission at 8 $\mu$m surrounds the bright radio continuum region of the nebula, indicating the presence of a photodissociated region at the interface between the ionized region and clump A. MIPSGAL emission at 24 $\mu$m shows the existence of warm dust inside the H{\sc ii}\,\ region. Based on high resolution IRAS images at 60 and 100 $\mu$m, a mean dust color temperature and dust mass of 37 K and 3 M$_\odot$\, were estimated for \rm NGC\,3503\,. The presence of candidate YSOs projected onto the HII region, detected using MSX and 2MASS point source catalogues, suggests the existence of protostellar objects in the neighbourhood of NGC 3503, although there is no clear evidence of triggered star formation.
The location of NGC\,3503 at the edge of a molecular clump and the electron density gradient in the H{\sc ii}\,\ region suggests that NGC\,3503 is a blister-type H{\sc ii}\,\ region that has undergone a champagne phase. In this scenario, the massive stars of Pis\,17 created NGC\,3503, which reached the border of the molecular cloud after 2 $\times$ 10$^5$ yr. The leakage of ionized gas and UV photons have probably contributed in the formation of MBO. Thus, the spatial distribution of the molecular gas and PAHs give additional support to a scenario first proposed by \citet{c00}. The proposed scenario for \rm NGC\,3503\, may explain the slight difference between the main velocity of its parental molecular cloud \hbox{(--24.7 \hbox{km s$^{-1}$\,})} and the velocity of its ionized gas \hbox{(--21 \hbox{km s$^{-1}$\,})}. High resolution radio recombination line observations may help to confirm the proposed scenario.
.
\begin{acknowledgements}
We especially thank Dr. James S. Urquhart and Dr. Mark A. Thompson for making their unpublished radio continuum images at 4800 MHz and 8640 MHz available to us. We acknowledge the anonymous referees for their helpful comments that improved the presentation of this paper. This project was partially financed by the Consejo Nacional de Investigaciones Cient\'ificas y T\'ecnicas (CONICET) of Argentina
under projects \hbox{PIP 112-200801-02488} and \hbox{PIP 112-200801-01299}, Universidad
Nacional de La Plata (UNLP) under project \hbox{11G/093}, and
Agencia Nacional de Promoci\'on Cient\'ica y Tecnol\'ogica
(ANPCYT) under projects \hbox{PICT 2007-00902} and \hbox{PICT 14018/03}.
This research has made use of the VIZIER database, operated at CDS, Strasbourg, France.
We greatly appreciate the hospitality of all staff members of Las Campanas Observatory of the Carnegie Institute of Washington. We thank all members of the NANTEN staff, in particular Prof. Yasuo Fukui, Dr. Toshikazu Onishi, Dr. Akira Mizuno, and students Y. Moriguchi, H. Saito, and S Sakamoto. We also would like to thank Dr. D. Miniti (Pont\'{\i}fica Universidad Cat\'olica, Chile) and Mr. F Bareilles (IAR) for their involment in early stages of this project
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
\label{sec.intro}
The possibility to manipulate the electron spin by electric currents was first proposed by Dyakonov and Perel, forty years ago.\cite{Dyakonov-1971fr} They noted that, in analogy with the anomalous Hall effect, an electric current must create a spin flow perpendicular to the charge current. This ``spin Hall'' effect is a consequence of the asymmetry, driven by the disorder potential through spin-orbit interaction, in the spin and momentum of the electrons scattered off impurities. Renewed interest in this effect arose more recently, when an intrinsic counterpart was predicted in strongly spin-orbit coupled semiconductors.\cite{Murakami-2003pb,Sinova-2004fb} The intrinsic effect results from the spin-orbit coupling in the underlying topology of the Bloch states (in momentum space), that gives rise to an anomalous component in the group velocity, perpendicular to the applied electric field.\cite{Murakami-2006df,Inoue-2009fk} It is based entirely on the one-particle band structure, and therefore subject to the influence of disorder. In the case of an n-doped semiconductor quantum well, the intrinsic contribution to the spin Hall conductivity turns out to be exactly cancelled by the effect of scattering.\cite{Schwab-2002vn,Mishchenko-2004ys,Inoue-2004fk,Dimitrova-2005uq,Chalaev-2005fk} The exact cancellation of the intrinsic spin Hall effect is particular to the Rashba spin-orbit coupling, linear in momentum, and is not present for instance in the Luttinger model of a p-type semiconductor.\cite{Murakami-2004kx}
Experimental observation of the spin Hall effect was achieved first by optical techniques to detect the spin accumulation,\cite{Kato-2004fu,Wunderlich-2005dq} and more recently by direct all-electrical measurements.\cite{Garlid-2010fj} The inverse intrinsic effect, that is the generation of a longitudinal charge current by the injection of polarized electrons has also been reported.\cite{Werake-2011fk}
The vanishing of the spin conductivity in the Rasbha model of a two-dimensional electron gas, for arbitrarily small disorder, can be traced back to the fact that the spin current operator \(\bm J^z=(\hbar/4)\{\sigma^z,\bm v\}\), is proportional to the spin \(\bm s\) derivative
\[
J^z_y\sim\dot{s}_y=\frac{\mathrm{i}}{\hbar}[H,s_y]\,,
\]
where \(\bm\sigma=(\sigma_x,\sigma_y,\sigma_z)\) is the vector of the Pauli matrices, \(\bm v\) is the electron velocity, \(H\) is the Hamiltonian, and \((x,y,z)\) refer to the directions of the applied electric field, the generated transverse spin current, and the out of plane spin polarization (Fig.~\ref{fig.hall-V}). The commutator \([H,s_y]\) is not affected by the presence of randomly distributed impurities. In a stationary state, the time derivative of the averaged spin density vanishes, and hence the spin current must also vanish.\cite{Dimitrova-2005uq,Chalaev-2005fk,Rashba-2004zr} However, if the impurities were magnetic this argument is no longer valid, the spin current would depend on additional contributions, and could therefore remain finite. As a consequence, magnetic impurities should in principle be able to restore the intrinsic Hall effect in a two-dimensional n-doped semiconductor.
\begin{figure}[!b]
\centering%
\includegraphics[width=0.48\textwidth]{fig1}
\caption{(Color online) Schematic geometry of the spin Hall effect; a longitudinal current \(j_x\) driven by an external electric field induces a transversal spin current that results in a polarization of the carrier spins following the sign of their momentum (spheres with arrows).}
\label{fig.hall-V}
\end{figure}
The influence of magnetic impurities on the intrinsic Hall effect was investigated using numerical computations of the Landauer formula,\cite{Liu-2006fk} linear response theory and Kubo formula,\cite{Inoue-2006fk,Wang-2007uq,Moca-2007tg} and quasiclassical Green functions.\cite{Gorini-2008kx} It is well established that, contrary to the non-magnetic case, the spin Hall effect does not vanish in this case, although different values of the spin conductivity in the clean limit were found.\cite{Inoue-2009fk} Interest in materials combining the electronic properties of semiconductors with ferromagnetism was enhanced by their potential applications in spintronics.\cite{Fabian-2007kx,dietl2008spintronics,Sato-2010fk} The main focus in the study of the so-called ``diluted magnetic semiconductors'' is on the mechanisms governing the interactions between impurities, that determine their thermodynamical and magnetic properties.\cite{Jungwirth-2006ca} However, it would also be interesting to investigate the spin dependence of the carriers response, to an applied electric field. Indeed, while the topological component of the anomalous Hall effect is widely documented in magnetic semiconductors, its spin current counterpart, the intrinsic spin Hall effect, is still a developing topic.\cite{Nagaosa-2010vn} This is the question we address in the present paper. Our main goal is to compute, using linear response theory, the intrinsic spin conductivity in a confined electron gas taking into account Rashba spin-orbit coupling and magnetic disorder. In particular, we want to obtain an analytical expression of the dynamical spin conductivity, to calculate first the clean limit for fixed frequency, and then the static limit. The order of limits is important here because the large scattering time limit does not commute with the small frequency limit.\cite{Murakami-2004kx}
In Sec.~\ref{sec.system}, we consider a simple model suitable to describe the linear response of a confined electron gas to an external electric field, as determined by the spin-orbit coupling (inversion asymmetry) and scattering on magnetic impurities (assumed to be in a paramagnetic state). We neglect extrinsic effects and non-magnetic impurities, and take into account an effective mass conduction band. In Sec.~\ref{sec.numerics} we generalize the model to a finite band tight-binding Hamiltonian in a square lattice. As in the effective mass model, we consider spatially distributed magnetic impurities with random orientations. We use a Chebyshev pseudo-spectral method to compute the density of states (total and local) and the spin Hall conductivity.\cite{Weisse-2006fk} Finally, we discuss and summarize our results (Sec.~\ref{sec.conclusion}).
\section{Spin Hall dynamical conductivity}
\label{sec.system}
In this section we compute the spin conductivity \(\sigma_{sH}(\omega)\) as a function of the frequency \(\omega\), starting with a simple model Hamiltonian \(H\). The Kubo formula is analytically calculated in the weak disorder limit, taking into account the ladder diagrams.
\subsection{Model Hamiltonian}
\label{subsec.model}
We consider a two-dimensional system of noninteracting electrons of effective mass \(m\), charge \(-e\), and Rashba spin-orbit coupling constant \(\lambda\), in the presence of randomly distributed magnetic impurities,
\begin{equation}
\label{Hcont}
H = \frac{{\bm p}^2}{2m} -
\frac{\lambda}{\hbar}{\bm \sigma}\cdot(\hat{z}\times {\bm p}) +
V(\bm x)\,.
\end{equation}
where \(\bm p\) and \(\bm x\) are the momentum and position operators of the electron in the plane \((x,y)\). The impurity potential energy is assumed to be short ranged and characterized by an exchange interaction constant \(J_s>0\) and a microscopic length \(a\),
\begin{equation}
\label{Vm}
V(\bm x)=J_s a^2\sum_{i\in I} \hat{\bm n}_i\cdot \bm\sigma\,
\delta(\bm x - \bm x_i)
\end{equation}
where \(I\) is the set of impurity sites \(\bm x_i\), and \(\hat{\bm n}_i\) the unit vector in the direction of the impurity magnetic moment; we assume that the impurities are in a paramagnetic state, so \(\hat{\bm n}_i=(\sin \theta_i \cos\phi_i,\sin \theta_i \sin\phi_i, \cos\theta_i) \) is uniformly distributed over the sphere, with \(\theta_i\in(0,\pi)\) and \( \phi_i \in (0,2\pi) \) independent random variables. Typical orders of magnitude, for a III-V semiconductor heterostructure, are: effective mass \(m = 0.067\, m_e\) (\(m_e\) is the electron mass), lattice length \(a\sim 0.56\,\mathrm{nm}\), Rashba constant \(\lambda\sim 10^{-11}\,\mathrm{eV\,m}\), and s-d exchange constant (for Mn impurities, for example) \(J_s\sim 0.1\,\mathrm{eV}\).\cite{Fabian-2007kx,Wu-2010fk} It is convenient to choose units such that \(\hbar=m=a=e=1\).
The clean Hamiltonian becomes diagonal in the chiral base,\cite{Arii-2007kx}
\begin{equation}
\label{eqn.eigenS}
|\alpha_\pm \rangle = \frac{1}{\sqrt{2}}
\begin{pmatrix}
\pm \mathrm{i} \mathrm{e}^{- \mathrm{i} \varphi}\\
1
\end{pmatrix}\,,
\end{equation}
where \( (k,\varphi) \) are polar coordinates in momentum space \(k_x+\mathrm{i} k_y=k\mathrm{e}^{\mathrm{i} \varphi}\). The corresponding eigenvalues are,
\begin{equation}
\label{eqn.eigenE}
\epsilon_{\pm}(k) = \frac{k^2}{2}\pm \lambda k\,.
\end{equation}
The transformation matrix between the spin and the chiral basis is
\begin{equation}
\label{eqn.u}
U (\varphi) = \frac{1}{\sqrt{2}}
\begin{pmatrix}
- \mathrm{i} \mathrm{e}^{- \mathrm{i} \varphi} & \mathrm{i} \mathrm{e}^{- \mathrm{i} \varphi} \\
1 & 1
\end{pmatrix} \,.
\end{equation}
For a given Fermi energy \( \epsilon_F\), the spin splitting is given by \(\Delta=2\lambda k_F\), where \(k_F=\sqrt{2\epsilon_F}\) is the Fermi wavenumber (energy is measured in units of \( \hbar^2/ma^2 \)). In the chiral base the velocity operator \(v_x=\mathrm{i}\,[H,x]\) is given by,
\begin{equation}
v_{x}=\left(
\begin{array}{cc}
(k-\lambda) \cos \varphi & -\mathrm{i}\lambda \sin \varphi \\
\mathrm{i} \lambda \sin \varphi & (k+\lambda) \cos \varphi
\end{array}
\right)\,,
\end{equation}
\(j_x=-v_x\) is the current in the \(x\) direction, and the spin current operator \(j^z_y=(1/4)\{\sigma_z,v_y\}\) is,
\begin{equation}
j^{z}_{y}=\left(
\begin{array}{cc}
0 & -\frac{k}{2} \sin \varphi \\
-\frac{k}{2} \sin \varphi & 0
\end{array}
\right)\,.
\end{equation}
Note that the velocity operator, at variance with the spin current operator, explicitly depends on the spin-orbit coupling constant. We use the conventional definition of the spin current operator \(j^z_y\), but other choices are possible (in the presence of spin-orbit coupling it is not a conserved quantity).\cite{Shi-2006ys}
In the presence of isotropic magnetic impurities, the scattering strength is characterized by a spin-flip relaxation time
\begin{equation}
\label{eqn.tau}
\frac{1}{\tau} = 2 \pi N_F c J_s^2\approx c J_s^2
\end{equation}
where \(c\) is the concentration of impurities (per unit area), and \(N_F\) is the density of states at the Fermi surface; it approaches the value \(N_F\approx 1/2\pi\) when the Fermi energy is much larger than any other energy in the system (\(N_F\approx m/2\pi\hbar^2\) in dimensional units).\cite{Moca-2007tg,Gorini-2008kx}
\subsection{Kubo conductivity}
\label{subsec.kubo}
\begin{figure}
\centering%
\includegraphics[width=0.48\textwidth]{fig2}
\caption{(Color online) The conductivity is calculated in the non-crossing approximation, as a zero order loop with disorder averaged Green's functions, plus the ladder diagrams. The ladder diagrams are calculated by renormalization of the spin current vertex.}
\label{fig.loops}
\end{figure}
\begin{figure}
\centering%
\includegraphics[width=0.48\textwidth]{fig3a}\\
\includegraphics[width=0.48\textwidth]{fig3b}
\caption{(Color online) Real and imaginary parts of the intrinsic spin Hall conductivity for a pure system (\protect\ref{eqn.anaRe})-(\protect\ref{eqn.anaIm}). The four branches \(\pm\omega_\pm\) are indicated by vertical dashed lines, their separation is \( |\omega_+ + \omega_-| = 4 \lambda^2 \) (arrows). }
\label{fig.exact}
\end{figure}
The spin Hall conductivity can be calculated in the framework of linear response theory,\cite{rammer1998quantum} using the following Kubo formula,\cite{Schliemann-2004fk}
\begin{equation}
\label{eqn.kubo}
\sigma_{sH}=\sigma^z_{xy}(\omega) = \frac{-\mathrm{i}}{L^2} \sum_{\alpha, \alpha '}
\frac{f_{\alpha}- f_{\alpha '}}
{\epsilon_{\alpha} - \epsilon_{\alpha '}} \,
\frac{\langle \alpha | j^{z}_y | \alpha ' \rangle
\langle \alpha ' | j_x |\alpha \rangle}
{\epsilon_{\alpha} - \epsilon_{\alpha '} + \omega + \mathrm{i} o}
\,,
\end{equation}
where \(\omega\) is the frequency of the applied external electric field, \(\mathrm{i} o\) is a small imaginary energy added to ensure causality, \(L^2\) is the system's area, \(\epsilon_{\alpha}\) and \(|\alpha\rangle\) are eigenvalues and eigenstates of \(H\), respectively, and \(f_\alpha=f(\epsilon_{\alpha})\) is the Fermi energy function. Formula (\ref{eqn.kubo}) is well suited for numerical computations of the tight-binding model\cite{Nomura-2005fk,Sheng-2005kl} (see Sec.~\ref{sec.numerics} below); for the conduction band model (\ref{Hcont}) it is convenient to express it in terms of retarded \(G^R\), and advanced \(G^A\), Green functions,\cite{Dimitrova-2005uq,Chalaev-2005fk}
\begin{align}
\label{eqn.sigma(G)}
\sigma^z_{xy} (\omega) = & \frac{1}{\omega}
\left\langle\mathrm{Tr} \left[
j_y^{z} (\bm{k}) G^R (\epsilon+\omega,\bm{k})
j_x (\bm{k}) G^< (\epsilon, \bm{k}) \right. \right.
\nonumber \\
& + \,\left. \left.
j_y^{z} (\bm{k}) G^< (\epsilon+\omega,\bm{k})
j_x (\bm{k}) G^A(\epsilon, \bm{k})
\right] \right\rangle,
\end{align}
where
\[
\mathrm{Tr}[\dots]=\int \frac{d^2 \bm{k}}{(2 \pi)^2}
\int \frac{d\epsilon}{2 \pi} \mathrm{Tr}_s \,(\dots)
\]
\(\mathrm{Tr}_s\) is the trace over the spin index, \(\langle\dots\rangle\) denotes the average over the spatial disorder and magnetic moment orientations, and \( G^< (\epsilon) = f(\epsilon) (G^R(\epsilon)-G^A(\epsilon)) \). Calculation of ({\ref{eqn.sigma(G)}) can be done using the diagrams of Fig.~\ref{fig.loops}, where solid lines represent the disorder averaged Green functions, \(\langle G^{R,A}(\epsilon,\bm{k})\rangle\):\cite{Wang-2007uq}
\begin{equation}
\label{eqn.GR}
G^R_k (\epsilon)=\begin{pmatrix}
\displaystyle\frac{1}{\epsilon-\epsilon_{-}(k) + \frac{\mathrm{i} }{2\tau}} & 0 \\
0 & \displaystyle\frac{1}{\epsilon-\epsilon_{+}(k) + \frac{\mathrm{i} }{2\tau}}
\end{pmatrix}
\end{equation}
and
\begin{equation}
\label{eqn.GA}
G^A_k (\epsilon)=\begin{pmatrix}
\displaystyle\frac{1}{\epsilon-\epsilon_{-}(k) - \frac{\mathrm{i} }{2\tau}} & 0 \\
0 & \displaystyle\frac{1}{\epsilon-\epsilon_{+}(k) - \frac{\mathrm{i} }{2\tau}}
\end{pmatrix}
\end{equation}
both written in the chiral basis. The perturbation expansion of Fig.~\ref{fig.loops} contains the lowest order one-loop contribution (the first diagram) and higher order one-loop contributions, containing disorder averaged terms, from which we only retain the non-crossing diagrams (ladder approximation).
Before proceeding with the computation of these diagrams, we
present the spin Hall conductivity in the clean limit, for which the Kubo formula can be calculated exactly. The real part of \(\sigma_{sH}\) is,
\begin{equation}
\label{eqn.anaRe}
\mathrm{Re}\,\sigma_{sH}(\omega)=\frac{-1}{8\pi}\left[1-\frac{\omega}{8\lambda^2}
\log \left|\frac{
\omega(\omega+4\lambda^2)-\Delta^2}{
\omega(\omega-4\lambda^2)-\Delta^2}
\right|\right],
\end{equation}
which has branch cuts at the roots \(\pm\omega_\pm\) with \(\omega_\pm=2\lambda^2 \pm \sqrt{4\lambda^4+\Delta^2}\) (Fig.~\ref{fig.exact}). This formula coincides with the result of Ref.~\onlinecite{Erlingsson-2005fk} in the small frequency limit. The imaginary part is given by
\begin{equation}
\label{eqn.anaIm}
\mathrm{Im}\,\sigma_{sH}(\omega)=\frac{-\omega}{64\lambda^2}
\Theta\big[1-\frac{
(\omega^2-\Delta^2)^2}
{(4\lambda^2\omega)^2}\big]
\end{equation}
where \(\Theta\) is the Heaviside function (see Fig.~\ref{fig.exact}, bottom).
For the evaluation of \(\sigma=\sigma^{z}_{xy}\), using the diagram expansion,
\[
\sigma_{sH}(\omega)=\sigma_{sH}^0(\omega)+\sigma_{sH}^L(\omega)\,,
\]
of Fig.~\ref{fig.loops}, we consider the weak scattering limit \(1/\tau,\,\Delta\ll \epsilon_F\), in which the products \(G^RG^R\) or \(G^AG^A\), of like Green functions are small compared to the product \(G^RG^A\):
\begin{equation}
\label{eqn.GrGa}
\sigma^z_{xy} (\omega) = \big\langle \mathrm{Tr}
\big[\frac{f(\epsilon + \omega) - f(\epsilon)}{\omega}
j_y^{z}
G^R (\epsilon+\omega)j_x
G^A(\epsilon) \big] \big\rangle ,
\end{equation}
where we omitted the momentum variable \(\bm{k}\). Moreover, we consider the low frequency limit \(\omega\ll \epsilon_F\), so the difference of Fermi distributions is different from zero only in a small neighborhood of \(\epsilon_F\). The calculation of the first digram is straightforward,
\begin{align}
\label{eqn.cleansW}
\sigma_{sH}^{0}(\omega) &= \frac{-1}{8\pi}
\left[ 1 - \frac{(1-
\mathrm{i} \tau \omega)^2}{(1-\mathrm{i} \tau \omega)^2 + \Delta^2 \tau^2 } \right]
\nonumber\\
&= \frac{-1}{8\pi}\frac{\Delta^2 \tau^2}{\Delta^2 \tau^2+w^2}\,,\quad
w=1-\mathrm{i} \tau \omega
\end{align}
where the first term gives the intrinsic contribution (\(\tau\gg1\)) to the spin Hall conductivity (\(-e/8\pi\) in dimensional units). To this order, for which electron-impurity correlations are neglected, we obtain the same result as in the non-magnetic case.\cite{Chalaev-2005fk}
\begin{figure}
\begin{centering}
\includegraphics[width=\linewidth]{fig4}
\end{centering}
\caption{(Color online) Dynamical conductivity \(\sigma_{sH}\) as a function of \(\omega\), for \( \tau \) varying between \(40\) and \( 190\) (light to dark lines). (Inset) As \( \omega \rightarrow 0 \) the real part tends to \(1/8\pi\) for \(\omega>1/\tau\), but for \(\omega<1/\tau\) it increases up to \( 8/7 \approx 1.14\). }
\label{fig.s0sL}
\end{figure}
\begin{figure}
\centering%
\includegraphics[width=\linewidth]{fig5}
\caption{(Color online) Static spin Hall conductivity as a function of the disorder strength from Eq.~(\protect{\ref{eqn.sigma}}), for two values of \(\Delta\). The dot at \(\tau^{-1}=0\) corresponds to the clean static limit.}
\label{fig.s0tau}
\end{figure}
Next, we compute the vertex correction (ladder diagrams, and third line in Fig.~\ref{fig.loops}) in the same approximation, where the Fermi energy \(\epsilon_F\), is much larger than any other energy scale:
\begin{equation}
\label{eqn.sLexpression}
\sigma_{sH}^{L}(\omega) = \langle \mathrm{Tr} [
J_y^{z}(\omega) G^R (\epsilon_F+\omega) j_x G^A(\epsilon_F)
] \rangle
\end{equation}
where the trace now runs over momentum and spin, and \(J_y^{z} =J_y^{z} (\omega,\bm{k}) \) is the corrected spin current vertex, which is a function of the energy \(\omega\)). The algebraic equation that determines the corrected vertex, the third line of Fig.~\ref{fig.loops}, is conveniently expressed in the spin basis,
\begin{align}
\label{eqn.selfconsistentvertex}
J_y^{z} = &\,c J_s^2 \int \frac{kdk d\varphi}{(2\pi)^2}
\int\frac{\sin \theta d\theta d\phi}{4\pi}
\begin{pmatrix}
\cos \theta & \mathrm{e}^{- \mathrm{i} \phi}\sin \theta \\
\mathrm{e}^{\mathrm{i} \phi }\sin \theta & -\cos \theta
\end{pmatrix}
\nonumber \\
& \times \, U(\varphi)G^A (0, \bm{k}) U^{\dagger}(\varphi)
\left[ j^{\sigma_z}_y + J_y^{z} \right]
\\
& \times \, U(\varphi)G^R(\omega,\bm{k})U^{\dagger}(\varphi)
\begin{pmatrix}
\cos \theta & \mathrm{e}^{- \mathrm{i} \phi} \sin \theta \\
\mathrm{e}^{\mathrm{i} \phi }\sin \theta & -\cos \theta
\end{pmatrix},
\nonumber
\end{align}
where we used the unitary matrix (\ref{eqn.u}); we note that \(J_y^{z}\), in the spin basis, has the structure,
\begin{equation}
J_y^{z} =
\begin{pmatrix}
0 & J_{\uparrow \downarrow} \\
J_{\downarrow \uparrow} & 0
\end{pmatrix}\,,
\end{equation}
with \(J_{\uparrow \downarrow}=J_{\downarrow \uparrow}\) given by,
\begin{equation}
\label{eqn.vertexJ}
J_{\uparrow \downarrow} = \frac{\mathrm{i} k_F \Delta \tau w}
{2 w^2 + 2(7 w + \mathrm{i} \tau \omega)(\Delta^2 \tau^2 + w^2) }\,.
\end{equation}
Introducing this expression into (\ref{eqn.sLexpression}) and computing the trace, results in the ladder contribution,
\begin{equation}
\label{eqn.sLadder}
\sigma^{L}_{sH} (\omega)= \sigma^{0}_{sH}(\omega)
\frac{\Delta^2 \tau^2 + 4 \mathrm{i} \tau \omega }
{w^2 + ( 7w+ \mathrm{i} \tau\omega)(\Delta^2 \tau^2 + w^2)}\,.
\end{equation}
Collecting the two contributions (\ref{eqn.cleansW}) and (\ref{eqn.sLadder}) we finally obtain the dynamical spin Hall conductivity in the presence of magnetic impurities, in the weak scattering strength approximation,
\begin{align}
\label{eqn.sigma}
\sigma_{sH} (\omega)= &\,
\frac{-1}{8\pi} \frac{\Delta^2 \tau^2}{\Delta^2 \tau^2+w^2}
\bigg[1 + \,
\nonumber \\
& +\, \frac{\Delta^2 \tau^2 + 4 \mathrm{i} \tau \omega }
{2w^2(4-3\mathrm{i} \tau \omega) + \Delta^2 \tau^2( 7- 6\mathrm{i} \tau\omega)}
\bigg]\,.
\end{align}
This result is similar to the one obtained in Ref.~\onlinecite{Gorini-2008kx}, although it has an extra term \(4 \mathrm{i} \tau \omega\) in the numerator of the last term; this difference, that does not change the limiting behavior for small frequency or large scattering times, can be attributed to the different approximation methods used. We show in Fig.~\ref{fig.s0sL} the real and imaginary parts of \(\sigma_{sH}\)
for different values of the disorder scattering time, as a function of the frequency. The real part of \(\sigma_{sH}\), an even function of the frequency, has a sharp change of sign at \(\omega=\Delta\) in the clean limit, that smoothes out with finite \(\tau\). The imaginary part shows, at the same frequency, a peak that broadens with increasing disorder. Comparing with Fig.~\ref{fig.exact}, we note that the in the weak scattering limit the splitting \(\omega_-+\omega_+\) is neglected.
The static spin current, as deduced from the real part of (\ref{eqn.sigma}) at \(\omega\rightarrow0\),
\begin{equation}
\sigma_{sH}(0)=\frac{-1}{8\pi}\frac{8\Delta^2 \tau^2}{8+7\Delta^2 \tau^2}\,,
\label{s0s}
\end{equation}
is shown in Fig.~\ref{fig.s0tau}. The continuous band model predicts an enhancement of the spin conductivity near the clean limit \(\tau\gg \Delta^{-1}\), and a slow reduction for smaller scattering times.
\subsection{The clean limit of \(\sigma_{sH}\)}
\label{subsec.limits}
\begin{figure}
\centering%
\includegraphics[width=\linewidth]{fig6}
\caption{(Color online) Convergence of the conductivity for values of \( \omega \) near zero, ranging from \( 10^{-3} \) to \( 10^{-9} \). The expression of \( 8 \pi Re \left[ \sigma (\omega)\right] \) for \( \omega = 10^{-n} \) converges to zero only for \( \tau > 10^{n+1}\). The horizontal line in the graph corresponds to the numerical factor \( 8/7 \). }
\label{fig.limits}
\end{figure}
The imaginary part of \(\sigma_{sH}\) is an odd function of the frequency, and thus it must vanish at zero. Moreover, the real part, contrary to the non-magnetic case, does not vanish in the clean limit. For \(\tau \rightarrow \infty\) we find,
\begin{equation}
\lim_{\tau \rightarrow \infty} \sigma_{xy}^z (\omega) = \frac{-1}{8 \pi} \frac{\Delta^2}{\Delta^2 - \omega^2}\,;
\label{eqn.cleansW0}
\end{equation}
if now we put \(\omega \rightarrow 0\) we recover the intrinsic value of the static spin Hall effect, as might be expected.\cite{Gorini-2008kx} However, the dynamical spin conductivity near the origin is singular,\cite{Arii-2007kx} as can be inferred from the behavior of \(\mathrm{Re}\,\sigma_{sH}\) near \( \omega=0\) (inset in Fig.~\ref{fig.s0sL}). Indeed, we observe that convergence to the static limit is not uniform: the larger the value of \(\tau\), the steeper the slope at the origin. This is related to the fact that the limits, \(\tau \rightarrow \infty\) and \( \omega \rightarrow 0\) do not commute:
\begin{equation}
\lim_{\tau \rightarrow \infty} \sigma_{xy}^z (0) = \frac{-1}{8 \pi} \frac{8}{7}\,,
\label{eqn.cleansW-tw}
\end{equation}
as seen from formula (\ref{s0s}). Various different results for the clean limit of \(\sigma_{sH}(0)\) were reported.\cite{Inoue-2006fk,Wang-2007uq,Gorini-2008kx} The reason can be attributed to the fact that the static value \(\sigma_{sH}(0)\) cannot be computed at \(\omega=0\), for \(\tau\) tending to infinity. The order of limits starting with \(\omega \rightarrow 0\), would violate the causality condition contained in the linear response formula.
We present in Fig.~\ref{fig.limits} a sequence showing the spin conductivity as a function of the scattering time for the series \( \omega_n = 10^{-n} \), with \(n=2,\dots,9\). We observe that, for a fixed value of the frequency \(\omega_n\), convergence to \(-1/8\pi\) is reached for \(\tau>\tau_{n+1}\) (where \(\tau_n=10^n\)). Therefore, the limit \( \omega \rightarrow 0 \) does not converge for any finite value of \( \tau \). In fact, one can erroneously interpret the transitory value of \(8/7\) as being the static spin conductivity.\cite{Wang-2007uq}
We conclude that in the presence of magnetic impurities, and in the weak disorder approximation, the static spin conductivity jumps from its intrinsic value to a value about \(8/7\) for finite disorder, and then it smoothly decreases when the disorder increases.
\section{The tight binding model and numerical results}
\label{sec.numerics}
We will now go beyond the effective mass model and the perturbation approximation in the disorder strength, to compute the spin conductivity. We introduce a discrete tight-binding model to investigate, in particular, the influence of a finite band width on the spin transport. The Hamiltonian is defined on a finite square lattice of size \(L \times L\) and lattice constant \(a\), with Rashba spin-orbit coupling \(\lambda\), and exchange interaction on impurities \(J_s\),
\begin{align}
\label{eqn.Hdiscrete}
H = - & t \sum_{\langle i,j \rangle} c_i^{\dagger} c_j\, +
\frac{\lambda}{a} \sum_i (
\mathrm{i} c_i^{\dagger} \sigma_x c_{i+\hat{y}} -
\mathrm{i} c_i^{\dagger} \sigma_y c_{i+\hat{x}} )
\nonumber \\
&+ \,J_s \sum_{i \in I} c^{\dagger}_i \hat{\bm n}_i \cdot \bm {\sigma} c_i,
\end{align}
where the hermitian conjugate terms of the first two sums are implicit; \(c_i^\dagger = (c^\dagger_{i\uparrow}\; c^\dagger_{i\downarrow})\) is the creation operator at site \(i\), of spin up \(\uparrow\) or down \(\downarrow\), and \(c_i\) is the corresponding annihilation operator. We denoted \(\hat x\) and \(\hat y\) the two orthogonal directions in the lattice. The kinetic term is limited to nearest neighbours and its bandwidth is \(4t\). This model is a straightforward generalization of (\ref{Hcont}), with discrete kinetic, spin-orbit and exchange terms. In the following we take \(t =1\) as the energy unit, and \(\hbar=a=e=1\) as before.
The computation of the Kubo formula (\ref{eqn.kubo}) needs in principle the knowledge of the whole spectrum and the full set of eigenvectors of \(H\). One can avoid this numerically expensive task, by reducing the formula to a trace, which can be evaluated with a stochastic method, in any basis. Specifically, we adopted a pseudospectral implementation of the kernel polynomial method.\cite{Weisse-2008uq}
Because the real and imaginary parts of the spin conductivity (\ref{eqn.kubo}) are related by the Kramers-Kronig relations, it is enough to compute the optical spin conductivity \(\sigma_{o}=\mathrm{Im}\,\sigma_{sH}\), defined by,
\begin{equation}
\label{eqn.sigma(J)}
\sigma_{o}(\omega)=\frac{1}{L^2} \sum_{\alpha,\alpha'}J^z_{xy}(\alpha,\alpha')\frac{f_{\alpha} - f_{\alpha'}}{\omega}\delta(\omega-\epsilon_{\alpha} + \epsilon_{\alpha'})\,,
\end{equation}
where
\[
J^z_{xy} (\alpha, \alpha ')= \mathrm{Im} \left[ \langle \alpha | j^{z}_y | \alpha ' \rangle \langle \alpha ' | j_x |\alpha \rangle \right],
\]
can be written in the form,
\begin{equation}
\label{eqn.jzxydelta}
\sigma_{o} (\omega)= \frac{1}{ \omega} \int d\epsilon \ j(\epsilon + \omega, \epsilon) \left[ f(\epsilon) - f(\epsilon + \omega) \right]\,,
\end{equation}
using the definition,
\begin{equation}
\label{eqn.matrixj}
j(\epsilon, \epsilon ') = \frac{1}{L^2} \sum_{\alpha, \alpha '} J^z_{xy} (\alpha, \alpha ') \delta(\epsilon - \epsilon_{\alpha}) \delta(\epsilon ' - \epsilon_{\alpha '}).
\end{equation}
The advantage of this representation is that the matrix \(j(\epsilon,\epsilon')\) is a function of energies only which can be expanded in Chebyshev polynomials products \(T_n(\epsilon)T_m(\epsilon')\) with expansion moments \(\mu_{nm}\), which are easily computed using a stochastic trace evaluation method. We now briefly detail the algorithm.
The first step is to scale and translate the energy, in such a way that the Hamiltonian spectrum is mapped into the interval \(\epsilon \in [-1,1]\). The series expansion of (\ref{eqn.matrixj}) can then be written,
\begin{align}
j(\epsilon,\epsilon') & = \sum_{n,m=0}^{\infty}
\frac{\mu_{nm} h_{nm} T_n(\epsilon) T_m(\epsilon') }{
\pi^2 \sqrt{1-\epsilon^2} \sqrt{1-\epsilon'^2} } \nonumber \\
& \approx \sum_{n,m=0}^{M-1}
\frac{\mu_{nm} h_{nm} g_n g_m T_n(\epsilon) T_m(\epsilon') }{
\pi^2 \sqrt{1-\epsilon^2} \sqrt{1-\epsilon'^2} }\,,
\label{eqn.chebyshevexp}
\end{align}
where \(M^2\) is the number of moments we use, \(T_m\) is the Chebyshev polynomial of order \(m\),
\[
h_{nm} = \frac{2}{1+\delta_{n,0}} \frac{2}{1+ \delta_{m,0}}\,,
\]
and \(g_m\) is a filter that minimizes the Gibbs oscillations arising in truncating the series to a finite order (we use the Jackson kernel, see Ref.~\onlinecite{Weisse-2008uq}).
The second step is to compute \(\mu_{nm}\). The presence of the radicals in the denominator allows to express the moments as usual scalar products,
\begin{equation}
\label{eqn.moments1}
\mu_{nm} = \int^{1}_{-1} \int^{1}_{-1} d\epsilon d\epsilon'
j(\epsilon,\epsilon')
T_n(\epsilon) T_m(\epsilon')
\end{equation}
which after rearranging,
\begin{eqnarray*}
\mu_{nm} &=& \frac{1}{L^2} \mathrm{Im} \sum_{\alpha,\alpha'}
\langle \alpha | j^{z}_y | \alpha ' \rangle
\langle \alpha ' | j_x |\alpha \rangle
T_n(\epsilon_{\alpha})
T_m(\epsilon'_{\alpha'})
\\
&=& \frac{1}{L^2} \mathrm{Im} \sum_{\alpha,\alpha'}
\langle \alpha | T_n (H) j^{z}_y | \alpha' \rangle
\langle \alpha ' | T_m(H) j_x |\alpha \rangle \,,
\end{eqnarray*}
can be written as a trace formula,
\begin{equation}
\label{eqn.moments}
\mu_{nm} = \frac{1}{L^2} \mathrm{Im} \mathrm{Tr} \left[
T_n(H) j^{z}_y T_m (H) j_x \right]
\end{equation}
In the third step, the trace is evaluated using an average over random states \(|r\rangle\),
\begin{equation*}
\label{eqn.moments1}
\mu_{nm} \approx \frac{1}{N_r} \sum_r \langle r|
T_n(H) j^{z}_y T_m (H) j_x |r\rangle\,,
\end{equation*}
where \(N_r\) is the number of random vectors needed for averaging. Next, one uses the recurrence relations between the Chebyshev polynomials to obtain the operator products to order \((n,m)\). In our case, even if the spin current operator is traceless, fast decrease of statistical errors and a rapid convergence of the algorithm is obtained. In order to optimize the Chebyshev recursion algorithm we adapted a pseudospectral procedure:
\begin{equation*}
H|r\rangle=\mathrm{IFFT}\big[H_0(\bm k)\mathrm{FFT}[|r\rangle]\big]+V|r\rangle\,,
\end{equation*}
where \(H_0\) contains the kinetic and spin-orbit terms, diagonal in the \(k\)-representation, \(V\) is the impurity potential, diagonal in the \(x\)-representation, and FFT, IFFT denote direct and inverse fast Fourier transforms, respectively. Once \(\mu_{m n}\) are known, we can efficiently obtain \(j(\epsilon,\epsilon')\) from formula (\ref{eqn.matrixj}) using \(M_\epsilon\) collocation points \(\epsilon_l = \cos[\pi(l+1/2)/M_\epsilon]\). One drawback of this method is that the \(\omega\)-dependent integral~(\ref{eqn.jzxydelta}), cannot be directly computed, since \(\epsilon_l + \omega\) is not necessarily a collocation point. One therefore has to perform an interpolation of some sort.
The last step is thus to compute the convolution integral (\ref{eqn.jzxydelta}). We insert into this equation a \(\delta\)-function in its integral form, in order to separate the energy integrals. The expression for \(\sigma_{o}\) becomes,
\begin{equation*}
\sigma_{o} (\omega) = \int \frac{dt}{2\pi}
\int d\epsilon d\epsilon' \frac{f_{\epsilon'} - f_\epsilon }{\omega}\,
\mathrm{e}^{\mathrm{i} t (\epsilon - \epsilon' + \omega)}
j (\epsilon, \epsilon ')
\end{equation*}
We discretize the frequencies as well as the energies and the auxiliary time variable \(t\). The sampled \(\sigma_{o}\), finally takes the form of a Fourier transform
\begin{equation*}
\sigma_{o} (\omega_n)=\frac{1}{\omega_n} \sum_m
\mathrm{e}^{\mathrm{i} t_m \omega_n} F (t_m)\,,
\end{equation*}
where
\begin{equation*}
F (t_m)=\frac{1}{\omega_n} \sum_{i,j}
\mathrm{e}^{\mathrm{i} t_m (\epsilon_i - \epsilon_j)}
j(\epsilon_i, \epsilon_j)(f_i - f_j)\,,
\end{equation*}
and so \(\sigma_{o}(\omega_n)\) is straightforwardly evaluated by a fast Fourier transform of \(F\). From the expression of the optical spin conductivity we calculate its static value using the Kramers-Kronig relation,
\begin{equation}
\label{s0}
\sigma_{sH}(0)=\Xint-_{-\infty}^{\infty} \frac{d\omega}{\pi} \frac{\sigma_{o} (\omega)}{\omega}\,.
\end{equation}
The density of states,
\begin{equation}
\rho(\epsilon)=\sum_\alpha \delta(\epsilon-\epsilon_\alpha)\,,
\end{equation}
may be computed using the same method, and the expansion
\[
\rho(\epsilon) = \sum_{m=0}^{M-1} \frac{ \mu_m h_m g_m T_n(\epsilon)}{\pi \sqrt{1-\epsilon^2}} \,,
\]
where
\[
\mu_m = \frac{1}{L} \mathrm{Tr}[T_n(H)]\, .
\]
In order to avoid finite size artifacts and to approach the thermodynamic limit faster, we use twisted boundary conditions: \(\Psi (x+L,y) = \mathrm{e}^{i \varphi_x} \Psi(x,y)\) and \(\Psi (x,y+L) = \mathrm{e}^{i \varphi_y} \Psi(x,y)\), for any state \(\Psi\). Physical quantities are averaged over many phase configurations \((\varphi_x , \varphi_y)\in [0,2\pi]^2\). This method has already been applied to determine the finite-size spin Hall conductivity, which in the case of non magnetic impurities, should vanish in the thermodynamic limit.\cite{Sheng-2005kl, Nomura-2005fk} In our case, the spin Hall conductivity being finite even at the thermodynamic limit, twisted boundary conditions constitute a useful tool to improve numerical convergence. For instance, the translational invariance of the disorder in the limit of infinite number of configurations, is recovered by averaging over twisted boundary conditions.
\begin{figure}
\centering%
\includegraphics[width=\linewidth]{fig7}
\caption{(Color online) Density of states for \(c J_s^2= 0\) (clean system), and \(c J_s^2=0.01,\,0.06\); \(\lambda=0.4\). The spin-orbit interaction is responsible for the splitting of the chiral states (central peaks). }
\label{fig.rho}
\end{figure}
\begin{figure}
\centering%
\includegraphics[width=\linewidth]{fig8}
\caption{(Color online) Optical spin conductivity as a function of the frequency, for different disorder strengths. For each \(\tau^{-1}=c J_s^2= 0.002,0.01,0.06\), three values of the concentration \(c=0.04,0.06,0.08\), were used; \(\epsilon_F=2\), \(\lambda=0.2\).}
\label{fig.sW_taudep}
\end{figure}
Let us make some remarks about the choice of the number of moments \(M\). Filtering the truncated Chebyshev expansion, introduces a broadening of sharp features that scales as \(1/M\), causing a reduction in the energy resolution. The optimal number of moments can be calibrated from the clean (\(J_s=0\)) case, and it turns out to be a few times the linear system size \( M = 2L,\, 4L\), therefore \(4L^2,\, 16L^2\) coefficients are involved in the \(j(\epsilon,\epsilon')\) expansion. We obtained a a well detailed density of states, without finite-size artifacts. This choice of numerical parameters, has been verified by comparing small samples averaged over twisted boundary conditions, with equivalent samples of greater size.
\subsection{Numerical results}
\label{subsec.sW}
In the following, we describe the properties of spin transport through the numerical computation of the density of states, and the static and optical spin conductivities. Simulations were performed on lattices of size \(64^2,\,128^2\), but higher sizes were used to test convergence. Typical numerical parameters are: \(\lambda = 0.2, \, 0.4\), in units of \(at\), \(c\) from \(2\%\) to \(10\%\), and \(\tau\) from \(5\) to \(500\), in units of \(\hbar/t\).
We will limit the range of parameters in order to preserve the qualitative shape of the clean density of states, avoiding for instance an impurity band regime (strong disorder regime). This choice should allow us to compare the tight-binding model to the analytical results of the preceding section, and to determine its validity. One observes in Fig.~\ref{fig.rho} that, within the range \(\tau^{-1}=0,\dots,0.06\), the density of states \(\rho(\epsilon)\) remains close to its clean limit, and the spin-orbit splitting, proportional to \(\lambda\), fades out with stronger scattering (to enhance the effect of the spin-orbit coupling, we put \(\lambda=0.4\)).
\begin{figure}
\centering%
\includegraphics[width=\linewidth]{fig9}
\caption{(Color online) Optical spin conductivity as a function of frequency and \(\epsilon_F\), for two different disorder strengths \(\tau^{-1}=0.002,\,0.2\). As the disorder increases, the optical conductivity decreases and spreads out over \(\omega\). }
\label{fig.sW_Efwdep}
\end{figure}
The spin conductivity depends on the Fermi energy \(\epsilon_F\), the spin-orbit coupling \(\lambda\), and in principle, the concentration \(c\) and the exchange constant \(J_s\). In fact, the scattering off impurities is characterized by the sole parameter \(\tau^{-1}=cJ_s^2\), at least in the weak disorder limit \(\tau^{-1}\ll 1\), hence we would expect that \(\sigma_o=\sigma_o(\omega;\tau)\). In Fig.~\ref{fig.sW_taudep} we represented the optical spin conductivity as a function of the frequency, for different sets of pairs \(c\) and \(J_s\), giving the same values of \(\tau\). We chose the Fermi energy at the middle of the band, \(\epsilon_F=2\). We observe that the curves coincide according to their \(\tau\) values, confirming that in the regime studied, the relevant parameter is the product \(cJ_s^2\). The maximum of \(\sigma_o\) corresponds to \(\omega\approx \Delta\) as in the continuous case, and is almost independent of \(\tau\) (compare with Fig.~\ref{fig.exact}). In Fig.~\ref{fig.sW_Efwdep}, we present \(\sigma_o\) as a function of the frequency and the Fermi energy. Its maximum appears at Fermi energies in the middle of the band, and follows a law in \(\omega\sim\sqrt{\epsilon_F}\), compatible with the estimation \(\omega\approx \Delta\) given above. The intensity of the peak decreased by a factor 2, for a factor of 100 in \(\tau\) (it is 500 for the top figure and 5 for the bottom one). The optical spin conductivity vanishes for small \(\omega\ll\Delta\), and large frequencies \(\omega\gg\Delta\). Therefore, the effect of increasing the disorder is, as expected, to broaden and damp the peak on the spectrum. However, at variance with the conduction band model for which the peak width is proportional to \(\tau^{-1} \), in the tight-binding model, the effect of a finite band width quantitatively change the scaling with the scattering time (the large Fermi energy limit is non longer applicable).
The computation of the static spin conductivity (\ref{s0}), appears to be particularly sensitive to statistical errors, as already noted.\cite{Sheng-2005kl} A large set of boundary conditions and random impurities distributions, is necessary. Results are shown in Fig.~\ref{fig.s0_v0dep}, where \(\sigma_{sH}(0)\) is represented as a function of the Fermi energy and the disorder strength. The spin Hall conductivity is an odd function, it vanishes at \(\epsilon_F=0\) and outside the energy band, but is nearly constant over a large range of Fermi energies. Its slow fall off with disorder underlines the weakness of vertex corrections in the magnetic case, as shown by the analytical calculation of the previous section.\cite{Moca-2007tg} More importantly, we observe large fluctuations of the spin conductivity at low frequencies, depending on the sample distribution. Strong spin conductance fluctuations are known to arise at mesoscopic scales,\cite{Ren-2006fk} and, provided that scattering ensures ergodicity, their universal statistics can be related to that of random matrices.\cite{Bardarson-2007fk} We found a noticeable amplification of the fluctuations in the region \(\omega\lesssim\tau^{-1}\) that explains the large dispersion of \(\sigma_{sH}(0)\) values around its mean. In the inset of Fig.~\ref{fig.sW_taudep}, the integrand \(\sigma_o/\omega\) in (\ref{s0}), is represented together with error bars proportional to the sample to sample dispersion,
\begin{equation}
\label{deltas}
\Delta\sigma_{sH}(\omega)=\langle \sigma_{sH}(\omega)^2-
\langle \sigma_{sH}(\omega)\rangle^2\rangle^{1/2}\,,
\end{equation}
where the angle brackets stand for the random potential averaging. The averaged static spin Hall conductivity, is then not enough to characterize the spin transport. In fact, even in the present weak disorder limit, where localization of the carriers can be neglected, we find a Gaussian distribution of the spin conductivity fluctuations, as illustrated in Fig.~\ref{fig.hist}. The standard deviation of the static spin conductivity \(\Delta \sigma_sH(0)/\langle \sigma_{sH}(0)\rangle \approx 0.1\), giving an incertitude of 10\% about the mean value.
\begin{figure}
\centering%
\includegraphics[width=\linewidth]{fig10}
\caption{(Color online) Static spin Hall conductivity as a function of \(\epsilon_F\), for disorder strengths \(\tau^{-1}=0,\,0.2\). (Inset) Integrand of (\protect\ref{s0}), used to compute \(\sigma_{sH}(0)\); the error bars represent the amplitude of \(\sigma_o(\omega)/\omega\) fluctuations near the zero frequency.}
\label{fig.s0_v0dep}
\end{figure}
\begin{figure}
\centering%
\includegraphics[width=\linewidth]{fig11}
\caption{(Color online) Normalized histogram of the static spin conductivity for \(\tau^{-1}=0.2\), showing a Gaussian distribution with mean value \(\langle \sigma_{sH}(\omega)\rangle=0.98\, [-1/8\pi]\) and standard deviation \(\Delta \sigma_{sH}(0) =0.1\) (solid line).}
\label{fig.hist}
\end{figure}
\section{Conclusion}
\label{sec.conclusion}
In this paper we investigated the influence of magnetic impurities in the intrinsic spin Hall effect. A first model, amenable to analytical calculations, was introduced. We considered a simple one conduction band Hamiltonian including the effects of Rashba spin-orbit coupling and exchange interaction with randomly distributed magnetic moments. Using standard linear response methods we obtained, in the weak scattering and low frequency approximation, the expression of the dynamical spin conductivity as a function of the system's parameters (spin-orbit coupling, scattering time, Fermi energy). We emphasized the relevance of the order of the large scattering time and small frequency limits, and found that when frequency is going to zero, one must recover the spin conductivity of an ideal two dimensional electron gaz, \(\sigma=-e/8\pi\). However, an arbitrarily small amount of magnetic impurities will increase the static spin conductivity by a factor of about \(8/7\), showing the singular nature of the clean limit.
We also investigated, using numerical methods, the tight-binding generalization of the conduction band Hamiltonian. We adapted the kernel polynomial method using a pseudospectral evaluation of the Chebyshev recursion to compute the density of states and the optical spin conductivity. As a function of frequency, the spin Hall conductivity shows a peak at \(\omega\sim \lambda\sqrt{\epsilon_F}\), that broadens and weakens with the increase of the disorder strength. These changes are determined by the combination \(\tau^{-1}\sim cJ_s^2\), that characterize the disorder strength.
Within the regime of weak disorder, that is in the absence of separate impurity bands, the deviation with respect to the clean system is small. The main effect of magnetic impurities in contrast with non-magnetic ones, is to preserve and reinforce the intrinsic mechanism that would otherwise be destroyed. In the finite band model, we observe an amplification of spin conductivity fluctuations at low frequencies, which are at the origin of a significant incertitude of the static \(\sigma_{sH}\) mean value. A detailed theory of the spin Hall conductivity in the presence of magnetic impurities should consider these sample to sample fluctuations, that could become the main physical effect in the description of spin transport in the strong disorder regime.
In spite of the absence of magnetic order (we treated the paramagnetic state), the sole presence of spin-flip processes, guarantees the emergence of a spin current proportional to the applied electric field. For weak disorder, and strong spin-orbit coupling, the intrinsic spin Hall effect is reinforced by multiple scattering interactions, described by the vertex correction to the spin current. Manipulation of the carrier's spin in a semiconductor impurities, is then possible by doping it with magnetic impurities and applying an electric field. This result naturally raises the question of the interplay between intrinsic spin Hall currents and ferromagnetic impurities as present in a diluted magnetic semiconductor.
\acknowledgments
We would like to thank A. Wei\ss e for his kind advice, and TM, AC for useful discussions.
|
\section{Introduction}
In the last two decades, discoveries of the Kuiper belt \citep{1993Natur.362..730J}, as well as planets orbiting stars other than the Sun \citep{1995Natur.378..355M}, have supplied the centuries-old quest to understand the formation of the solar system with fresh constraints and insights into physical processes at play. Among a multitude of newly proposed formation scenarios, the ``Nice" model \citep{2005Natur.435..459T, 2005Natur.435..466G, 2005Natur.435..462M} is particularly notable, as it has attained a considerable amount of success in reproducing the various observed features of the solar system. Within the context of the scenario envisioned by the Nice model, giant planets start their post-nebular evolution in a compact, multi-resonant configuration, and following a brief period of dynamical instability, scatter onto their current orbits \citep{2007AJ....134.1790M, 2010ApJ...716.1323B, 2011Natur.475..206W}.
The first success of the Nice model lies in its ability to quantitatively reproduce the observed orbits of the giant planets, as well as their dynamical architecture (i.e. secular eigenmodes of the system) \citep{2005Natur.435..459T, 2009A&A...507.1041M}. Simultaneously, the brief instability, inherent to the model, provides a natural trigger to the Late Heavy Bombardment \citep{2005Natur.435..466G}, as well as a transport mechanism for emplacement of dynamically ``hot" Kuiper belt objects (KBOs) from inside of $\sim 35$AU \citep{2008Icar..196..258L}. Meanwhile, it has been recently demonstrated that survival of a dynamically ``cold" primordial population between Neptune's current 3:2 and 2:1 exterior mean-motion resonances (MMRs) is fully consistent with a Nice model-like evolution of the planets, implying an \textit{in-situ} formation of the cold classical population of the Kuiper belt \citep{2011arXiv1106.0937B}. Finally, the presence of Jupiter's and Neptune's Trojan asteroids has been attributed to chaotic capture of planetesimals during the instability \citep{2005Natur.435..462M, 2007AJ....133.1962N}.
\begin{figure*}[t]
\includegraphics[width=1\textwidth]{f1.pdf}
\caption{Orbital evolution of planets. Each planet's semi-major axis, as well as perihelion and apohelion distances are shown as functions of time. The actual perihelion and apohelion distances of the planets are also shown for comparison as black error bars. In both cases, the innermost ice-giant ejects during the transient phase of instability, leaving behind 4 planets, whose orbits resemble that of the solar system. See main text for a description of the initial conditions.}
\end{figure*}
All successful realizations of the Nice model to date have been comprised exclusively of the four currently present giant planets. However, there exists no strong evidence that suggests that additional planets were not present in the solar system, at the epoch of the dispersion of the nebula. In fact, theoretical arguments, presented by \citet{2004ApJ...614..497G} point to a possibility of initially forming as many as five ice giants, three of which get subsequently removed via ejections (see however \citet{2007Icar..189..196L}). The dynamical sensibility of such a scenario is further strengthened by the fact that a considerable fraction of standard Nice model simulations result in an ejection of an ice-giant after an encounter with at least one of the gas giants.
In this paper, we explore an instability-driven dynamical evolution of a 5-planet system (2 gas giants + 3 ice giants) with an eye towards identifying a pathway towards reproduction of solar system-like dynamical architecture. In principle, the realm of possibility available to this study is enormous. Consequently, rather than performing a comprehensive parameter-search, here we limit ourselves to systems that contain an additional Uranus-like planet, with the aim of presenting a few proof-of-concept numerical experiments. The plan of the paper is as follows. In section 2, we describe our numerical setup. In section 3, we show that in our model, the planetary orbits, their secular eigenmodes, as well as various populations of the Kuiper belt are approximately reproduced. We conclude and discuss our results in section 4.
\section{Numerical Experiments}
The numerical setup of the simulations performed here was qualitatively similar to those presented by \citet{2010ApJ...716.1323B} and \citet{2011arXiv1106.0937B}. Particularly, the five giant planets were initialized in a compact, multi-resonant initial condition, surrounded by a massive planetesimal disk that extended between its immediate stability boundary and $30$AU.
Two of the three ice-giants were taken to have the same mass as Uranus and were initially placed next to Saturn and as the outermost planet respectively. The middle ice-giant was taken to have Neptune's mass. In all simulations, Jupiter and Saturn started out in a 3:2 MMR, in accord with the results of hydrodynamical simulations of convergent migration of the planets in the solar nebula \citep{2001MNRAS.320L..55M, 2007Icar..191..158M, 2008A&A...482..333P}. The ice giants were also sequentially assembled into first order MMRs by applying dissipative forces, designed to mimic the presence of the nebula \citep{2002ApJ...567..596L}. Following resonant locking, each assembled multi-resonant initial condition was evolved in isolation for 10Myr, as an immediate test of orbital stability. This procedure yielded a total of 81 stable multi-resonant initial conditions.
The search for adequate dynamical evolutions was performed in two steps. First, we evolved 10 permutations of each initial condition, with planetesimal disks composed of $N=1000$ planetesimals. Disk masses were chosen randomly between $M_{\rm{disk}}^{\rm{min}} = 25M_{\oplus}$ and $M_{\rm{disk}}^{\rm{max}} = 100M_{\oplus}$. The density profiles followed a power-law distribution, $\Sigma \propto r^k$ where the power-law index, $k$, was chosen randomly between $k_{\rm{min}}= 1$ and $k_{\rm{max}} = 2$. The planetesimals were initialized on near-circular orbits ($e \sim \sin i \sim 10^{-3}$). To reduce the already substantial computational cost, self-gravity of the planetesimal swarm was neglected.
Subsequently, we eliminated all initial conditions that did not yield any final systems that were comprised of 4 planets, reducing the number of viable initial conditions to 25. Then, an additional 30 permutations of these initial conditions were integrated with disks composed of $N=3000$ planetesimals (but otherwise identical to those described above). Each integration was performed using the \textit{mercury6} integration software package \citep{1999MNRAS.304..793C} and spanned 50Myr\footnote{Note that here, we make not attempt to time the onset of instability with the late heavy bombardment.}. The calculations were performed on Caltech's \textit{PANGU} super-computer.
After their completion, simulations that were deemed successful were reintegrated with the use of tracer simulations (see \citet{2008Icar..196..258L, 2011arXiv1106.0937B}), to address the dynamical evolution of a locally formed population of KBOs. In particular, each run was supplemented with an additional disk of mass-less particles that resided in the cold classical region of the Kuiper belt (i.e. between the final exterior 3:2 and 2:1 MMRs of Neptune).
\section{Results}
\begin{figure*}[t]
\includegraphics[width=1\textwidth]{f2.pdf}
\caption{Eccentricity distribution of the remnant planetesimal disk. Red dots represent objects that have been dynamically emplaced, while the blue dots depict the locally formed cold classical belt at $t=50$Myr. The red and blue triangles show objects whose orbits are stable on a $500$Myr timescale. The scattered disk is shown with two solid curves and Neptune's exterior 3:2 and 2:1 MMR's are labeled with dashed lines. In the simulation presented in panel A of Figure (1), the cold belt suffers numerous encounters with the ejecting ice-giant, yielding considerable orbital excitation. Moreover, the inner cold belt is further dynamically depleted over $500$Myr of evolution. In the simulation presented in panel B of Figure (1), there is only a single close encounter between the cold belt and the ejecting ice-giant, yielding a dynamically cold orbital structure.}
\end{figure*}
Out of the 810 integrations that were initially performed, 214 ($\sim 25\%$) cases featured an ejection of a single ice-giant, yielding a system composed of 4 planets. Perhaps unsurprisingly, in most cases the ejected planet is the ice-giant that neighbors Saturn. Of the 750 simulations that were performed following the elimination of initial conditions, 33 evolutions resulted in orbits reminiscent of the solar system. Specifically, we searched for solutions where the Saturn-Jupiter period ratio exceed $2$, while the final semi-major axes of the ice-giants were within $3$AU of their observed counterparts. No strong requirements were placed on the planetary eccentricities and inclinations.
It is noteworthy that evolving the giant planets onto solar system-like orbits is insufficient for a simulation to be deemed successful for indeed, there are additional constraints that must be satisfied. The first orbital constraint is the reproduction of the secular architecture of the system. The secular orbital angular momentum exchange (i.e. eccentricity evolution) of a planetary system containing $N$ secondaries can be approximately represented as a superposition of $N$ eigenmodes, each corresponding to a fundamental frequency of the system (see \citet{1999ssd..book.....M, 2009A&A...507.1041M}). Physically, the maximum eccentricity that a given planet attains in its secular cycle is equal to the sum of all of its corresponding eigenmode amplitudes. In the context of the traditional Nice model, some difficulty has been noted in correctly reproducing the dynamical character of Jupiter's and Saturn's eccentricity evolution. Particularly, it has been shown that smooth passage of Jupiter and Saturn through the 2:1 MMR has a tendency to under-excite the $g_5$ eccentricity eigenmode as well as their mutual inclinations \citep{2009A&A...507.1041M}. This difficulty can be overcome sometimes by invoking a close-encounter between an ice-giant and the gas giants. Consequently, we have checked, using Fourier analysis (see for example \citet{2009A&A...507.1041M}), the relative strength of the $g_5$ and $g_6$ eigenmodes in all simulations whose orbital end-state resembled the solar system. We did not restrict the success criteria of our simulations to include the correct reproduction of the mean eccentricities of the planets and in some cases, the mean final eccentricities (namely those of the ice-giants) exceeded their observed counter-parts by as much as a factor of $\sim 2$. This is, however, likely an artifact of the coarse representation of the planetesimal disk and the resulting dynamical friction, that we employed in our calculations and should not be viewed as a major drawback.
In total, we found ten cases (corresponding to eight different initial conditions) where their amplitudes are satisfactorily reproduced. Specifically, in these ten cases, the amplitude of Jupiter's $g_5$ eigenmode exceeds that of the $g_6$ eigenmode, while the amplitudes are roughly the same for Saturn. We made no attempt at quantitatively matching the pair's inclination eigenmodes, however their reproduction does not appear to be problematic \citep{2009A&A...507.1041M}. We also examined the amplitudes of the secular eigenmodes of Uranus and Neptune. Generally it appears that the dynamical architecture of the ice-giants is set in an essentially random manner, depending on the particular encounter history. Consequently, we decided to not use ice-giant secular architecture as a distinctive property in our analysis.
Successful formation of the Kuiper belt is another important constraint of the Nice model. \citet{2008Icar..196..258L} have shown that the excited populations of the Kuiper belt are naturally emplaced from inside of $\sim 35$AU during the instability. Given that this aspect of the planetary evolution is not particularly different between the 4-planet and the 5-planet scenarios, there is little reason to speculate that the dynamical pathway for formation of the resonant, scattered, and hot classical populations of the Kuiper belt will be inhibited. The same is likely to be true for chaotic capture of Trojan asteroids. The cold classical population of the Kuiper Belt, however, is a different story.
A series of observational dissimilarities between the cold classical population and the rest of the Kuiper belt (e.g. uniquely red colors \citep{2002ApJ...566L.125T, 2005EM&P...97..107L}; strongly enhanced wide binary fraction \citep{2006AJ....131.1142S, 2010ApJ...722L.204P}) suggest that the cold classicals formed in situ, and maintained dynamical coherence despite Neptune's temporary acquisition of high eccentricity and inclination (a characteristic orbital feature of the cold classical population is inclination that does not exceed $\sim 5^{\circ}$ \citep{2001Icar..151..190B}). In a recent study, \citet{2011arXiv1106.0937B} showed that local formation of the cold classicals is fully consistent with an instability-driven evolution of the planets, given favorable conditions during the instability. Particularly, \citet{2011arXiv1106.0937B} required the apsidal precession and nodal recession rates of Neptune to be comparatively fast to prevent secular excitation of the cold classical orbits, in addition to a sufficiently small apohelion distance of Neptune, to avoid orbital excitation due to close encounters.
\begin{figure*}[t]
\includegraphics[width=1\textwidth]{f3.pdf}
\caption{Inclination distribution of the remnant planetesimal disk. Red dots represent objects that have been dynamically emplaced, while the blue dots depict the locally formed cold classical belt at $t=50$Myr. The red and blue triangles show objects whose orbits are stable on a $500$Myr timescale. The tentative $i=5^{\circ}$ boundary between the cold and hot classical belts is shown with a solid line, while Neptune's exterior 3:2 and 2:1 MMR's are labeled with dashed lines. In the simulation presented in panel A of Figure (1), the cold belt suffers numerous encounters with the ejecting ice-giant, yielding considerable orbital excitation. Moreover, the inner cold belt is further dynamically depleted over $500$Myr of evolution. In the simulation presented in panel B of Figure (1), there is only a single close encounter between the cold belt and the ejecting ice-giant, yielding a dynamically cold orbital structure.}
\end{figure*}
In a 5-planet scenario, the retention of unexcited orbits of the cold population can be jeopardized by the ejecting planet. This possibility served as a premise for recalculation of the dynamical evolution of test particles in the cold classical region, with the aid of tracer simulations. We found that in only three of the ten simulations where the secular eigenmodes were successfully reproduced, a primordial cold classical population of the Kuiper belt also retained unexcited orbits. Consequently, it appears that the retention of an unexcited cold belt is not the norm of the 5-planet scenario. This is not surprising, since dynamical excitation can only be avoided if ejection and close encounter timescales are sufficiently short. From a quantitative point of view, our simulations imply that dynamically cold orbits can only be sustained if the ejecting ice-giant spends $\sim 10^4$ years or less crossing the classical Kuiper belt region. As a result, we can expect that the incorporation of yet additional ice-giants into the primordial solar system (as suggested by \citet{2004ApJ...614..497G}) will further diminish the chances of reproducing the cold classical Kuiper belt, since all ejecting planets would have to do so very rapidly.
Two of the three successful simulations discussed above are presented in Figures (1-3). Specifically, Figure (1) shows the orbital evolutions of the runs, while Figures (2) and (3) show the eccentricity and inclination distributions of planetesimals in the Kuiper belt region respectively. The red and blue dots, shown in Figures (2) and (3) depict the orbital distribution of emplaced and local planetesimal populations respectively at $t=50$Myr. The red and blue triangles depict the planetesimals whose orbits remain stable at $t=500$Myr.
The starting multi-resonant initial condition of the simulation presented in panels (A) of the figures is one where Saturn and the first ice-giant are locked in a 2:1 MMR, while both pairs of ice-giants are locked in 5:4 MMRs. The planetesimal disk in this simulation was comprised of $N=3000$ particles and contained a total of 26$M_{\oplus}$. In this evolution, the local population of test particles suffers numerous short close encounters with the ejecting ice giant, yielding a more excited and depleted cold classical population, compared to that of run (B). Note also that at $t = 50$Myr, the ice-giant eccentricities are considerably greater than that of Uranus and Neptune. These high eccentricities do not get damped away by dynamical friction in the following 500Myr of dormant evolution. Consequently, in this simulation, the inner edge of the cold belt gets dynamically depleted over the following 500Myr. It is furthermore noteworthy that another simulation that originated from the same initial condition reproduced the eigenmodes of the system correctly, although the primordial cold Kuiper belt in this integration was entirely destroyed by close encounters.
The starting multi-resonant initial condition of the simulation presented in panels (B) of the figures is one where Saturn and the first ice-giant are locked in a 3:2 MMR, the inner pair of ice-giants is locked in a 2:1 MMR, and the outer pair of the ice-giants is locked in a 4:3 MMR. In this simulation, the disk consisted of $N=1000$ particles and had a cumulative mass of $42M_{\oplus}$. Incidentally, the frequency spectrum of the eccentricity vectors of Jupiter and Saturn produced in this simulation, matches that of the real Jupiter and Saturn exceptionally well, signaling a nearly ideal reproduction of the secular eigenmodes. Particularly, the simulation yields (in the notation of \citep{1999ssd..book.....M}) $e_{55}^{(sim)}/e_{56}^{(sim)} = 2.28$ and $e_{65}^{(sim)}/e_{66}^{(sim)} = 0.51$ where as the solar system is characterized by $e_{55}/e_{56} = 2.81$ and $e_{65}/e_{66} = 0.68$. The scaling of the eigen-vectors are also well reproduced: $e_{55}^{sim} = 0.0465, e_{66}^{sim} = 0.067$ while the solar system has $e_{55} = 0.0442, e_{66} = 0.0482$ \citep{2009A&A...507.1041M}. In this simulation, the cold Kuiper belt suffers only a single short encounter with the escaping ice-giant, allowing for the orbits (inclinations in particular) to remain dynamically cold.
\section{Discussion}
In this paper, we have presented a successful realization of the Nice model which starts out with 5 planets. The numerical experiments presented here explicitly show that such an evolution is plausible since the resulting 4-planet systems can closely resemble the solar system. Particularly, in both simulations presented here, the secular architecture of the outer solar system is well reproduced. Furthermore the demonstrated survival of a local, primordial cold classical Kuiper belt suggests that all constraints that can be matched with a 4-planet model can also be matched with a 5-planet model to an equal degree of satisfaction.
It is noteworthy that ejection was not always necessary in our simulations to generate a 4 planet system. In a handful of runs (one of which successfully reproduced the secular eigenmodes, but not the cold Kuiper belt), one of the ice-giants ended up merging with Saturn. In principle, such a scenario may help explain Saturn's enhanced metallicity in comparison with Jupiter. Although, here again the explanation is not unique (see \citet{1982P&SS...30..755S} and the references therein).
In a traditional realization of the Nice model, the rate of successful reproduction of the secular eigenmodes is rather low i.e. $\sim 10\%$ of the integrations for a favorable initial condition \citep{2010ApJ...716.1323B}. This is in part because an ice-giant/gas giant encounter often leads to an ejection of the ice-giant, leaving behind only three planets. Thus, the need for an ice-giant/gas-giant encounter in the orbital history of the solar system is in itself motivation for a 5-planet model.
The statistics of simulations presented in this work suggest that a 5-planet model is neither more nor less advantageous. Recall that the probability of ending up with only 4 planets is $214/810 \sim 25\%$. The probability of reproducing the secular eigenmodes of Jupiter and Saturn is $10/750 \sim 1.5 \% $. Naively, this yields an overall probability of success of only $\sim 0.4 \%$. However, it is important to keep in mind that the characteristic outcomes are generally dependent on initial conditions\footnote{Interestingly, we do not observe any correlation between the degree to which a given simulation is successful and disk mass.} and runs that originated from the initial condition presented in panels (A) correctly reproduced the secular eigenmodes in 2 out of 30 simulations. This statistic is similar to the 4-planet model. Finally, the $1/30$ probability of also retaining an unexcited cold classical Kuiper belt puts the 5-planet model and the 4-planet model on equal footing in terms of success rate \citep{2011arXiv1106.0937B}. That said, it is important to note that this success rate is only characteristic of the particular 5-planet model that we have constructed. In other words, it is likely that if one allows the mass of the ejected planet to also be a variable parameter, tuning of the initial state may in principle lead to a more frequent reproduction of the solar system.
The results presented in this work imply that the solar system is one of many possible outcomes of dynamical evolution, and can originate from many possible initial conditions. As a result, the possibility of having an extra planet initially present in the system, yet its ejection leaving no observable signature erases any hope for construction of a deterministic model for solar system evolution. The forward process-like nature of the Nice model is not surprising, given that the solar system exhibits large-scale chaos, characterized by Lypunov times that are comparable to orbital timescales, during the instability. Moreover, the similarity between the orbital architectures of simulations whose outcomes were deemed unsuccessful in this work and those of extra-solar planetary systems further confirms that planet-planet scattering is likely to be the physical process responsible for shaping the orbital distribution of planets \citep{2008ApJ...686..603J}. Consequently, we conclude that an instability-driven dynamical history remains a sensible choice as a baseline scenario for solar system's early dynamical evolution. \\
\textbf{Acknowledgments} We thank Alessandro Morbidelli, Hal Levison, David Nesvorny and Peter Goldreich for useful conversations. We thank Naveed Near-Ansari for operational help with the \textit{PANGU} supercomputer. K. Batygin acknowledges supported from NASA's NESSF graduate fellowship.
|
\section{An autoequivalence criterion}
\label{section.autoequivalence_criterion}
We give a condition on a functor $F$ which implies that the corresponding twist $T_F$ is an autoequivalence.
Although we will only apply this to our specific Grassmannian case, we present it in the general triangulated category setting to underscore the formal nature of the proof, and to make the key points more transparent.
\begin{remark} To avoid overwhelming the reader with unnecessary notation, in the following section we write exact triangles of integral functors where we mean triangles of the corresponding Fourier-Mukai kernels \cite{Huybrechts:2007tf}. \end{remark}
\subsection{Calabi-Yau spherical functors}
\label{section.CY_spherical_functors}
\begin{definition}\label{definition.Calabi-Yau_spherical} We say that an integral functor $F : \mcD' \To \mcD$ from an indecomposable, non-trivial triangulated category $\mcD'$ is \defined{Calabi-Yau spherical} if it satisfies:
\begin{enumerate}
\item Adjoint and twist existence conditions
\begin{itemize}
\item $F$ has integral adjoints $L \dashv F \dashv R$;
\item The adjoint $L$ is full;
\item There exist a twist $T_F$ and a cotwist $C_F$ with adjoints on both sides, such that there exist distinguished triangles $$\ExactTriangleWithMaps{FR}{\epsilon}{\id,}{}{T_F}{} \qquad \ExactTriangleWithMaps{\id}{\eta}{RF;}{}{C_F}{}$$
\end{itemize}
\item Serre duality conditions
\begin{itemize}
\item $\mcD$ and $\mcD'$ have Serre functors $S$ and $S'$ respectively;
\end{itemize}
\item Compatibility conditions
\begin{itemize}
\item $F$ intertwines $S'$ with an autoequivalence $S^*$ of $\mcD$ (so that $S^* F \iso F S'$);
\item $C_F$ commutes with $S'$;
\end{itemize}
\item Local Calabi-Yau condition
\begin{itemize}
\item $\mcD$ is locally $n$-Calabi-Yau with respect to $F$ (in the sense that $SF \iso F[n]$) for some $n$;
\end{itemize}
\item Sphericity condition
\begin{itemize}
\item There is an isomorphism of functors $R \overset{\sim}{\To} C_F L $ induced by the natural morphism $R \overset{R \eta}{\To} RFL$.
\end{itemize}
\end{enumerate}
\end{definition}
\begin{remark}We will see that a Calabi-Yau spherical functor $F$ is in fact a \remarkEmphasis{spherical functor} in the sense of Anno \cite{Anno:2007wo}. However the above conditions turn out to be easier to check in our situation. This simplifies our work considerably in Section \ref{section.autoequivalence_property}.\end{remark}
\begin{remark}Although the definition is somewhat unwieldy, most of the conditions are immediately satisfied in our case, and should follow very naturally in cases of interest. The final condition is the one we spend almost all of our time proving.
\end{remark}
\begin{remark}The sphericity condition in Definition \ref{definition.Calabi-Yau_spherical} is a direct generalization \cite[Section 8.4]{Huybrechts:2007tf} of Horja's spherical condition: in that case $L$ is always surjective, so the condition translates to $C_F \iso S'[-n]$ under our Calabi-Yau assumptions. The intertwinement condition $S^* F \iso FS'$ corresponds to the requirement that Horja's line bundle $\mathcal{L}$ is a restriction from $X$.
\end{remark}
\subsection{Twists of Calabi-Yau spherical functors}
To show that Calabi-Yau spherical functors $F$ give autoequivalences $T_F$, we first show that the cotwist $C_F$ is an autoequivalence.
\begin{remark}\label{remark.spanning_set} Our first step will be to construct a \remarkEmphasis{spanning set} for the triangulated category $\mcD'$. Assuming the existence of a Serre functor, a set $\Omega' \subset \mcD'$ is said to \remarkEmphasis{span} $\mcD'$ if any non-trivial object of $\mcD'$ has a non-trivial $\Hom$ from some element of $\Omega'$. See \cite[Definition 1.47]{Huybrechts:2007tf} for precise definitions.\end{remark}
\begin{lemma}\label{proposition.Calabi-Yau_spherical_functor_cotwist}For $F$ Calabi-Yau spherical, $C_F$ is an autoequivalence.
\Proof{
\stepDescribed{1}{Spanning set for $\mcD'$} We take $$\Omega' := \Im L \cup \ker F.$$ For each $\mcA \in \mcD'$ we seek $\omega \in \Omega'$ such that $\Hom_\mcD(\omega, \mcA) \neq 0$. If $\mcA \in \ker F$ then we simply take $\omega := \mcA$. Otherwise we have $$\Hom_{\mcD'}(LF\mcA, \mcA) \iso \Hom_\mcD(F\mcA, F\mcA) \not\iso 0$$ and we may take $\omega := LF\mcA$. This completes the verification of the spanning set $\Omega'$.
\stepDescribed{2}{Action of $C_F$ on $\Omega'$} We claim that \beqa C_F |_{\ker F} & \iso & [1], \\ C_F |_{\Im L} & \iso & S'[-n]. \eeqa The first follows directly from the definition of $C_F$. For the second we use Serre duality in the form $R \iso S'LS^{-1}$ to note that \beqa C_F L & \iso & R \\ & \iso & S'LS^{-1} \\ & \iso & S'L[-n] \\ & \iso & S'[-n]L \eeqa where we use the left adjoint of the local Calabi-Yau condition, namely $$L S^{-1} \iso L[-n].$$ The claim now follows from the fullness of $L$.
\stepDescribed{3}{Preservation of $\Omega'$ by $C_F$} Using the previous step, it is immediate that $C_F$ takes $\ker F$ to itself, as it simply acts by a shift. Also $C_F$ takes $\Im L$ to itself: this follows by using the left adjoint of the intertwinement assumption, namely $$LS^{*-1} \iso S'^{-1}L,$$ which gives $$C_FL\mcA \iso S'L\mcA[-n] \iso LS^*\mcA [-n].$$
\stepDescribed{4}{Vanishing of $\Hom$s between parts of $\Omega'$} We note that for $\mcB \in \ker F$ we have $$\Hom_{\mcD'}(L\mcA, \mcB) \iso \Hom_\mcD(\mcA, F\mcB) \iso 0.$$ For $\Hom$s in the other direction we use Serre duality to evaluate \beqa \Hom_{\mcD'}(\mcB, L\mcA) & \iso & \Hom_{\mcD'}(S' \mcB, S'LS^{-1} S\mcA) \\ & \iso & \Hom_{\mcD'}(S' \mcB, R S \mcA) \\ & \iso & \Hom_\mcD(FS' \mcB, S \mcA) \\ & \iso & \Hom_\mcD(S^* F \mcB, S\mcA) \\ & \iso & 0, \eeqa where we use our intertwinement assumption $FS' \iso S^*F$.
\stepDescribed{5}{Autoequivalence property} We first note that $C_F$ is integral and therefore exact. Then \cite[Corollary 1.56]{Huybrechts:2007tf} gives the result if
\begin{itemize}
\item $\mcD'$ is indecomposable and non-trivial,
\item $C_F$ has adjoints on both sides and commutes with the Serre functor $S'$,
\item and for all $\omega_i \in \Omega'$ the induced morphism $$\Hom_{\mcD'}(\omega_1, \omega_2) \To \Hom_{\mcD'}(C_F (\omega_1), C_F (\omega_2))$$ is a bijection.
\end{itemize}
The first two conditions follow by assumption, so it remains to check the criterion on $\Hom$s between elements of the spanning set $\Omega' = \Im L \cup \ker F$. The condition holds for $\omega_i \in \Im L$ or $\omega_i \in \ker F$ by Step $2$. The other cases follow from the following $2$ steps, as all $\Hom$s involved vanish. This completes the proof.
}
\end{lemma}
We then deduce:
\begin{proposition}\label{proposition.autoequivance_criterion} For $F$ Calabi-Yau spherical, $T_F$ is an autoequivalence of $\mcD$.
\Proof{We simply note that $F$ is a spherical functor in the sense of Definition \ref{theoremDefinition.spherical_functor_and_twist} by combining the assumptions and the lemma above, and so $T_F$ is an autoequivalence by Theorem \ref{theoremDefinition.spherical_functor_and_twist}.}
\end{proposition}
\subsection{Action on a spanning set}
Here we describe the action of the twist $T_F$ on a spanning set $\Omega$ for the triangulated category $\mcD$. We will use this to understand the action of the Grassmannian twist in Section \ref{section.twist-properties}.
\begin{remark}In the definition of a spanning set in Remark \ref{remark.spanning_set} we required every non-trivial object of our category to have a non-trivial $\Hom$ \remarkEmphasis{from} some element of the set. In the presence of a Serre functor we can equivalently require a non-trivial $\Hom$ \remarkEmphasis{to} some element of the set \cite[Exercise 1.48]{Huybrechts:2007tf}. This will be used in the following proposition.\end{remark}
\begin{proposition}\label{proposition.spanning_set} Assume $\mcD$ has a Serre functor $S$, and take a functor $F : \mcD' \To \mcD$ of triangulated categories, with a left adjoint $L$ as follows: $$\xymatrix{\mcD' \ar@/_/[r]_F^{\perp} & \mcD \ar@/_/[l]_L }$$ Then $$\Omega := \Im(FL) \cup \ker(L)$$ is a spanning set for $\mcD$. If furthermore $\mcD$ is locally $n$-Calabi-Yau with respect to $F$ (in the sense that $SF \iso F[n]$) then there are no $\Hom$s between the two parts of $\Omega$.
\Proof{To show that $\Omega$ spans, take a non-zero object $\mcA \in \mcD$. We give a suitable $\omega$ with $\Hom_{\mcD}( \mcA, \omega) \not\simeq 0$ in the following cases:
\begin{itemize}
\item {\bf Case $\mcA \in \ker(L)$:} Take $\omega := \mcA$ and use $\Hom_{\mcD}(\mcA, \mcA) \not\simeq 0$.
\item{\bf Case $\mcA \notin \ker(L)$:} We then have $L \mcA \not\simeq 0$ so $$0 \not\simeq \Hom_{\mcD'}(L \mcA, L \mcA) \iso \Hom_{\mcD}(\mcA, FL \mcA),$$ and so we may take $\omega := FL \mcA$.
\end{itemize}
This proves that $\Omega$ spans. We now show the vanishing of $\Hom$s between the two parts of $\Omega$.
\stepDescribed{1}{backward $\Hom$s} Taking $\mcA \in \ker(L)$ and any $\mcB \in \mcD$ we have $$\Hom_{\mcD}(\mcA, FL\mcB) \iso \Hom_{\mcD'}(L\mcA, L\mcB) \iso 0,$$ by adjunction.
\stepDescribed{2}{forward $\Hom$s} We similarly observe \beqa \Hom_{\mcD}(FL\mcB, \mcA) & \iso & \Hom_{\mcD}(\mcA, SFL\mcB)^\vee \\ & \iso & \Hom_{\mcD}(\mcA, FL\mcB [n])^\vee \\ & \iso & 0, \eeqa where we use the local Calabi-Yau condition and the previous step.}
\end{proposition}
Note in particular that the proposition applies to a Calabi-Yau spherical functor $F$. We now describe the action of the associated twist $T_F$ on $\Omega$.
\begin{remark} Here and elsewhere we use the \remarkEmphasis{triangular identities} for the units $\eta$ and counits $\epsilon$ of our adjunctions. For example for the adjunction $F \dashv R$ we have \beq{equation.triangular}\epsilon F \circ F \eta = \id_F.\eeq We briefly explain how this arises. The crucial observation is that the functorial adjunction isomorphism $$\psi: \Hom(F- , F-) \overset{\sim}{\To} \Hom(-,RF-)$$ can be explicitly inverted in terms of the counit $\epsilon$ by $\psi^{-1} := (\epsilon F) \circ (F -).$ Now \eqref{equation.triangular} follows from the definition of the unit $\eta := \psi(\id_F)$ \cite[Section IV.1]{:SaundersMacLane}.
\end{remark}
\begin{proposition}\label{proposition.Calabi-Yau_spherical_functor_action}For $F$ Calabi-Yau spherical we have \begin{enumerate} \item $\mcA\in \ker L \so T_F\mcA \iso \mcA,$ and \item $\mcA \in \Im FL \so T_F\mcA \iso S^* \mcA[-n+1]$. \end{enumerate}
\Proof{First note that $$L S^{-1} \iso L [-n],$$ from the local Calabi-Yau condition by uniqueness of left adjoints.
For the first part, if $\mcA \in \ker L$ then \beqa R\mcA & \iso & S' L S^{-1} \mcA \\ & \iso & S' L \mcA [-n] \\ & \iso & 0. \eeqa The result follows by definition of $T_F$.
For the second part, we emulate \cite[Section 8.4]{Huybrechts:2007tf} and observe that by the definitions of $T_F$ and $C_F$ we have a diagram of distinguished triangles:
$$
\xymatrix{
F \ar[r]_{F \eta } \ar@{=}[d] & FRF \ar[r] \ar[d]^{\epsilon F} & F C_F \ar@{-->}[r] & \\
F \ar@{=}[r] & F \ar[d] & & \\
& T_F F \ar@{-->}[d] & & \\
& & & \\
}
$$
The commutativity of the top left-hand square follows from a triangular identity for the adjunction $F \dashv R$. Applying the octahedral axiom \cite[Chapter 10]{Kashiwara:2005vw} gives a diagram as follows:
$$
\xymatrix{
F \ar[r] \ar@{=}[d] & FRF \ar[r] \ar[d] & F C_F \ar@{-->}[r] \ar[d] & \\
F \ar@{=}[r] & F \ar[r] \ar[d] & 0 \ar@{-->}[r] \ar[d] & \\
& T_F F \ar@{-->}[d] \ar^{\sim\quad}[r] & F C_F [1] \ar@{-->}[d] & \\
& & & \\
}
$$
We then have that \beqa T_F FL & \iso & F C_F L[1] \\ & \iso & F R[1] \qquad\qquad \text{(sphericity condition)} \\ & \iso & F S' L S^{-1} [1]\\ & \iso & S^* F L S^{-1} [1] \qquad \text{(intertwinement of $S'$ and $S^*$)} \\ & \iso & S^* F L [-n+1], \eeqa which yields the result.}
\end{proposition}
\section{Autoequivalence property for the twist}
\label{section.autoequivalence_property}
\subsection{Orientation}
\begin{remark} From now on we restrict to the case $r=2$, so that the tautological hyperplane bundle $H$ on $\hat{B}$ is just a line bundle, which we denote by $l$. \end{remark}
We find in this case that:
\begin{observation}The twist base $X_0$ given in Definition \ref{definition.twist_base_resolution} is simply the total space of the bundle $$\xymatrix{\Hom(V,l) \ar[d]^p \\ \PP V.}$$ \end{observation}
$C_F$ acts on $D^b(X_0)$ by a non-trivial autoequivalence. To understand the action of $C_F$ we use a tilting generator $\mathcal{T}_0$ for $D^b(X_0)$ given by $$\mathcal{T}_0 := p^*\left(\O \oplus l^\vee \oplus \ldots \oplus l^{\vee (d-1)}\right).$$ We explain why this is a tilting generator in Section \ref{section.base_tilting_object}.
Our next step is to understand $F$ applied to the summands $p^* l^{\vee k}$. In Section \ref{section.applying_functor_F} we show how to calculate them all at once, using a geometrical method. We use this to deduce the required properties of the cotwist in Section \ref{section.cotwist_on_image}. The proof concludes in Section \ref{section.autoequivalence_proof}.
\subsection{A tilting generator for $X_0$}
\label{section.base_tilting_object}
We give the straightforward proof of the above tilting claim, deferring some other tilting results which we will need until Appendix \ref{section.tilting_object}.
\begin{proposition}\label{proposition.base_tilting_object}$D^b(X_0)$ has a tilting generator (Definition \ref{definition.tilting_generator}) given by $\mathcal{T}_0 := p^* \mcT_\PP$ where $$\mcT_\PP := \O \oplus l^\vee \oplus \ldots \oplus l^{\vee (d-1)} \in \Coh(\PP V).$$
\Proof{Note first that $X_0$ is projective over $\End^{<r}(V)$ by Proposition \ref{proposition.resolutions_and_flattening}, and that $\End^{<r}(V)$ is a Noetherian affine of finite type, as required in Definition \ref{definition.tilting_generator}.
We now show that $\mathcal{T}_0$ is tilting. We have \beqa \RHom_{X_0} (p^* \mcT_\PP, p^* \mcT_\PP) & \iso & \RHom_{\PP V}(\mcT_\PP, p_* p^* \mcT_\PP) \\ & \iso & \RHom_{\PP V}(\mcT_\PP, \mcT_\PP \otimes p_* \O_{X_0} ) \\ & \iso & \RHom_{\PP V}(\mcT_\PP, \mcT_\PP \otimes \Sym^\bullet \Hom(V,l)^\vee) \\ & \iso & \RHom_{\PP V}\left(\mcT_\PP, \mcT_\PP \otimes \bigoplus_k \left(\Sym^k V \otimes l^{\vee k}\right)\right), \eeqa which splits into terms of the form $$\RHom_{\PP V}(l^{\vee a}, l^{\vee b} \otimes l^{\vee k}) \iso \RDerived\Gamma_{\PP V}(l^{\vee b+k-a}),$$ where $0 \leq a,b \leq d-1$. We note that $b+k-a > -d,$ so Kodaira vanishing gives the result.
We then show that $\mathcal{T}_0$ spans $D^b(X)$. By adjunction, we have $$\Hom_X(\mathcal{T}_0, -) = \Hom_X(p^* \mcT_\PP, -) \iso \Hom_{\PP V}(\mcT_\PP, p_* -).$$ Now $p$ is affine hence $p_*$ is injective, and $\mcT_\PP$ is the Beilinson tilting generator for $\PP V$ \cite[Example 7]{Toda:2008wl}. We deduce that $\mathcal{T}_0^{\perp} \iso 0$ and this completes the proof.
}
\end{proposition}
\subsection{Preliminary: pushdowns from resolution $\hat{B}$}
We consider the bundle on $\hat{B}$ with fibre $l \take \{0\}$, with its natural $\C^*$ action. We have:
$$\xymatrix{\C^* \ar@/^1pc/[r] & l \take \{ 0 \} \ar[d]_\hat{q} & \\
& \hat{B} \ar[r]_j & X}$$
Now we observe that, by definition of $F$, \begin{eqnarray}\label{equation.F_as_pushforward}
F (l^{\vee k}) & = & \RDerived j_* (\pi^* l^{\vee k}) \nonumber \\
& = & \RDerived j_* (l^{\vee k}) \nonumber \\
& = & \RDerived j_* (\hat{q}_* \O_{l \take \{ 0 \}})_k \nonumber \\
& \iso & \left( \RDerived (j\hat{q})_* \O_{l \take \{ 0 \}} \right)_k.
\end{eqnarray} (The subscript $k$ denotes taking equivariants of weight $k$ for $k \in \Z$, and we note that the bundle $l \take \{0\}$ is a family of affine schemes, so $\hat{q}$ has no higher pushdowns.)
We now define a morphism of schemes $i$ fitting into the following diagram:
\beq{equation.factoring_resolution_pushdown}\xymatrix{l \take \{ 0 \} \ar[d]_\hat{q} \xyhook[r]_i & S \take \{ 0 \} \ar[d]^q \\
\hat{B} \ar[r]_j & X}\eeq (We write $S \take \{0\}$ for the total space of the tautological bundle $S$, with the zero section removed.)
\begin{definition} The morphism $i$ is defined affine locally (we omit an explicit presentation) so that it maps a closed point $(x,v)$ of the bundle $l \take \{ 0 \}$ which we write as $$ (x,v) = \left( \xymatrix{ 0 \ar[r]^1 & l \ar[r]^1 & S \ar[r] & V \ar@/^/[ll]^A }, \qquad 0 \neq v \in l \right),$$ to a closed point $i(x,v)$ of $S \take \{ 0 \}$ given by $$i(x,v) := \left( \xymatrix{ 0 \ar[rr]^2 & & S \ar[r] & V \ar@/^/[l]^{\iota_l \circA }}, \qquad 0 \neq \iota_l (v) \in S \right),$$ where $\iota_l$ denotes the inclusion $l \into S$.
\end{definition}
\begin{lemma}\label{lemma.cut_out_by_section} The map $i$ is a closed embedding, with $\Im i$ cut out scheme\discretionary{-}{-}{-}theoretically by a section $\alpha \in \Gamma(\mcN)$ of the bundle $$\mcN := \Hom(V, \wedge^2 S \{1\}),$$ where $\{1\}$ denotes a shift of weight under the $\C^*$-action.
\Proof{
We first show that $i$ is injective on closed points (the scheme-theoretic result that $i$ is a closed embedding follows by a local calculation, which we omit). If $i$ were not injective, so that $i(x,v) = i(x',v')$ say, then by definition of the map we would have $\iota_l (v) = \iota_{l'} (v')$ for $l \neq l'$, which would imply $v=v'=0$, a contradiction.
We now define $\alpha$ as the section induced by the following composition of tautological morphisms $$\alpha : V \overset{A}{\To} S \overset{\wedge v}{\To} \wedge^2 S \{1\},$$ between tautological bundles on the bundle $S \take \{ 0 \}$. The map $\alpha$ is zero at a closed point $(x,v)$ precisely when \beq{equation.section_vanishing_criterion} w || v, \quad \forall w \in \Im A. \eeq At such a point $A$ factors through $\langle v \rangle$, hence the point lies in $\Im i$. Conversely, if a point $(x, v)$ is in $\Im i$ then $A$ factors through $l \ni v$ and \eqref{equation.section_vanishing_criterion} holds. Working on the pull-up $(p q)^{-1} (U)$ of an open affine $U \subset \mathbbm{Gr}$ we see that indeed $\Im i$ is the subscheme of zeroes of $\alpha$.
}
\end{lemma}
Now we observe:
\begin{lemma} The square given in \eqref{equation.factoring_resolution_pushdown} commutes, that is: $$\xymatrix{\ar@{}[dr]|{\Box} l \take \{ 0 \} \ar[d]_\hat{q} \xyhook[r]_i & S \take \{ 0 \} \ar[d]^q \\
\hat{B} \ar[r]_j & X}$$ In particular, the composite map $j \hat{q}$ factors as a closed embedding $i$ followed by a flat projection $q$.
\Proof{The commutativity is clear from the definitions: the horizontal maps forget $l$, and the vertical maps forget $v$.
}
\end{lemma}
\begin{remark} The method used here is similar to that in \cite[Proposition 11.12]{Huybrechts:2007tf}, where the derived pullback via a blow-up map is computed by factoring it into a closed embedding and a flat projection.
\end{remark}
\begin{remark} The embedding $i$ is $\C^*$-equivariant for the natural $\C^*$-action on the bundle $S \take \{0\}$.
\end{remark}
Our calculation now reduces to evaluating \begin{eqnarray}\label{equation.pushforward_is_pushdown_of_structure_sheaf} \RDerived (j \hat{q})_* \O_{l \take \{0\}} & \iso & \RDerived q_* (i_* \O_{l \take \{0\}}) \nonumber \\
& = & \RDerived q_* \O_{\Im i}, \end{eqnarray}
and so to calculate the derived pushdown $\RDerived q_*$ we Koszul resolve $\O_{\Im i}$. To this end, Proposition \ref{proposition.transverse_section} checks that $\alpha$ cuts out the subscheme $\Im i$ \emphasis{transversally}. We begin by observing:
\begin{lemma}$\codim_X B = d-1$.
\Proof{For $S$ fixed we consider the space $\Hom^{\leq \rho}(V,S)$ of homomorphisms with $\rk \leq \rho$. By \cite[Prop 1.1(b)]{Bruns:198wwa} we have that \beqa \dim \Hom^{\leq \rho}(V,S) & = & (\dim V + \dim S)\rho - \rho^2 \\ & = & d+1, \eeqa having set $\rho=1$, and hence the codimension of $\Hom^{\leq \rho}(V,S)$ in $\Hom(V,S)$ is $2d - (d+1) = d-1$. Applying this in a family over the base $\mathbbm{Gr}$ gives the result.}
\end{lemma}
\begin{remark}We can see this result explicitly: locally on an open affine $p^{-1} (U)$, the subscheme $B$ of $X$ is cut out by $d-1$ independent minors of the $2 \times d$ matrix representing $A \in \Hom(V,S)$.
\end{remark}
We now give a more complete description of $\Im i$:
\begin{lemma}The restriction of $\Im i$ to the fibre over a closed point $x = (S, A) \in X$ is $$\Im i|_{q^{-1} \{ x \}} = \left\{ \begin{array}{cc} S \take \{ 0 \} & x \in \mathbbm{Gr} \\ \Im A \take \{ 0 \} & x \in B \take \mathbbm{Gr} \\ \emptyset & x \in X \take B \end{array} \right\}$$ and furthermore $$\dim \Im i = \dim \mathbbm{Gr} + d + 2.$$
\Proof{For the first part we use that a closed point $(x, v)$ in $S \take \{0\}$ lies in $\Im i$ precisely when \begin{itemize} \item there exists a line $l \subset S$ such that $v \in l$, and \item $A: V \to S$ factors through the inclusion $\iota_l : l \into S$. \end{itemize}
When $x \in X \take B$ this is impossible, as $A$ is surjective. When $x \in B \take \mathbbm{Gr}$ we have $\rk A = 1$ and so we are forced to have non-zero $v \in l = \Im A$. Finally when $x \in \mathbbm{Gr}$ we can take any non-zero $v \in l \subset S$, hence the result.
For the second part we decompose $\Im i$ with respect to the natural stratification of $X$ so that $$\Im i = \Im i|_{q^{-1} (B\take\mathbbm{Gr})} \cup \Im i|_{q^{-1} (\mathbbm{Gr})}.$$ Note that $B \take \mathbbm{Gr}$ is a large open subset of $B$ and so, using that the fibre of $\Im i$ over a point $x \in B \take \mathbbm{Gr}$ has dimension 1, we have \beqa \dim \Im i|_{q^{-1} (B\take\mathbbm{Gr})} & = & \dim B + 1 \\ & = & \dim X - \codim B + 1 \\ & = & (\dim \mathbbm{Gr} + 2d) - (d - 1) + 1 \\ & = & \dim \mathbbm{Gr} + d + 2, \eeqa which is the dimension claimed for $\dim \Im i$. To conclude we note that for the other stratum we have $$\dim \Im i|_{q^{-1} (\mathbbm{Gr})} = \dim \mathbbm{Gr} + 2 < \dim \mathbbm{Gr} + d + 2.$$}\end{lemma}
\begin{proposition}\label{proposition.transverse_section} The subscheme $\Im i$ is cut out \theoremEmphasis{transversally} by the section $\alpha \in \Gamma(\mcN)$ of the bundle $\mcN := \Hom(V, \wedge^2 S \{1\}).$
\Proof{This follows from Lemma \ref{lemma.cut_out_by_section}. We use the previous proposition to check that \beqa \codim_{S \take \{0\}} \Im i & = & \dim \Tot({S \take \{0\}}) - \dim \Im i \\ & = & (\dim \mathbbm{Gr} + 2d) + 2 - \dim \Im i \\ & = & d, \eeqa as expected. This suffices by smoothness of $S \take \{0\}$.
}
\end{proposition}
\subsection{Technical digression: convolutions}
To present some of our intermediate results more compactly, we choose to use the language of \emphasis{convolutions} for bounded complexes $(\mcA_\bullet, \partial)$ of objects in the derived category. For the reader who does not want to delve into the details, we note the following facts:
\begin{remarks}
\begin{enumerate}
\item Convolutions generalize the mapping cone: for a two-term complex $(\mcA_1 \overset{\partial}{\To} \mcA_0)$ we just have $$\Cone(\mcA_\bullet, \partial) \iso \{\mcA_1 \overset{\partial}{\To} \mcA_0\}.$$
\item Convolutions are defined using \remarkEmphasis{Postnikov systems} \cite[Section IV.2, Exercise 1]{Gelfand} of exact triangles involving the objects and morphisms of the complex. Details are given below.
\item Given a general complex $(\mcA_\bullet, \partial)$ it is not a priori possible to say that the convolution exists or is unique. However if \beq{equation.convolution_condition} \Hom(\mcA_{k+l+1}, \mcA_k[-l]) \iso 0, \qquad k\geq0, \: l>0, \eeq then a unique convolution exists \cite[Section 3.4]{Cautis:2009vz}. In particular if the $\mcA_k$ are sheaves in $\Coh(X)$, this follows immediately from the vanishing of negative $\Ext$s.
\end{enumerate}
\end{remarks}
We now give a formal definition for the case we will require, following \cite[Section 3.4]{Cautis:2009vz}.
\begin{definition}For a complex $(\mcA_\bullet, \partial)$ of objects in $D^b(X)$ whose non-zero terms are given by $$\left(\mcA_d \To \mcA_{d-1} \To \ldots \To \mcA_1 \To \mcA_0\right)$$ we say that $\mcC_d$ is a \defined{convolution} of the complex if there exists a \defined{Postnikov system} as follows:
$$\verticalPostnikov{\mcA_d}{\mcA_{d-1}}{\mcA_1}{\mcA_0}{\mcC_{d}[-d]}{\mcC_{d-1}[-d+1]}{\mcC_{d-2}[-d+2]}{\mcC_1[-1]}{\mcC_0}$$ The triangles on the left are commutative, and those on the right are exact.
\end{definition}
\begin{definition}If the convolution for the complex $(\mcA_\bullet, \partial)$ is unique up to isomorphism, we write it as $\Cone(\mcA_\bullet, \partial)$. \end{definition}
\begin{remark}As for the mapping cone itself, in general uniqueness of convolutions is up to non-unique isomorphism.\end{remark}
\begin{remark}Notice that $\mcA_0 \overset{\sim}{\From} \mcC_0$. It immediately follows that for the two-term complex $(\mcA_1 \overset{\partial}{\To} \mcA_0)$ we have $$\Cone(\mcA_\bullet, \partial) \iso \{ \mcA_1 \overset{\partial}{\To} \mcA_0 \},$$ as expected.\end{remark}
\subsection{Applying the functor $F$}
\label{section.applying_functor_F}
The lemma in this section is analogous to Remark \ref{remark.spherical_object_proof_plan}, step 1, as we describe resolutions for certain sheaves in $\Im F$.
\begin{remark} We identify $Fl^{\vee k}$ as a \remarkEmphasis{non-unique} convolution of a certain complex. However, after applying $R$ to the corresponding Postnikov system in Lemma \ref{proposition.image_within_window} we obtain a complex whose convolution \remarkEmphasis{is} unique. The advantage of this (perhaps unusual) approach is that we avoid the need to keep track of all the data of the Postnikov systems in the intermediate stages: most of the objects will be killed by $R$.\end{remark}
\begin{lemma} \label{proposition.generalized_Koszul_complexes} $ F l^{\vee k}$ is a convolution of a complex of objects $(\mcE_{k,\bullet}, \partial)$ where $$\mcE_{k,j} = \left\{ \begin{array}{cc} \Sym^{k-j} S^\vee & 0 \leq j \leq k,d \\ \Sym^{j-k-2} S(-1)[-1] & 0, k+2 \leq j \leq d \\ 0 & \text{otherwise} \end{array} \right\} \otimes \wedge^j V(j).$$ Here we define $\O(-1) := \wedge^2 S$.
\Proof{The Koszul resolution for $\O_{\Im i}$ on the total space of the bundle $S\take\{0\}$, justified in Proposition \ref{proposition.transverse_section}, gives an isomorphism $$\left\{ \wedge^d \mcN^\vee \overset{ \alpha}{\To} \wedge^{d-1} \mcN^\vee \overset{ \alpha}{\To} \ldots \overset{ \alpha}{\To} \mcN^\vee \overset{ \alpha}{\To} \underline{\O} \right\} \overset{\sim}{\To} \O_{\Im i},$$ where the differentials in the complex are given by wedging with the section $\alpha$, and the underline denotes the degree $0$ term. We consider now the objects corresponding to successive truncations of this complex, as follows: \beqa
\left\{ \underline{\O} \right\} & =: & \mcC_0 \\
\left\{ \mcN^\vee \To \underline{\O} \right\} & =: & \mcC_1 \\
& \vdots & \\
\left\{ \wedge^{d-1} \mcN^\vee \To \ldots \To \mcN^\vee \To \underline{\O} \right\} & =: & \mcC_{d-1} \\
\left\{ \wedge^d \mcN^\vee \To \wedge^{d-1} \mcN^\vee \To \ldots \To \mcN^\vee \To \underline{\O} \right\} & =: & \mcC_d \\
\eeqa
These form a Postnikov system
\beq{equation.koszul_Postnikov} \verticalPostnikov{\wedge^d \mcN^\vee}{\wedge^{d-1} \mcN^\vee}{\mcN^\vee}{\O}{\mcC_{d}[-d]}{\mcC_{d-1}[-d+1]}{\mcC_{d-2}[-d+2]}{\mcC_1[-1]}{\mcC_0}\eeq
with $\mcC_d \iso \O_{\Im i}$. Now we have that $Fl^{\vee k} \iso \RDerived q_* ( \O_{\Im i} )_k$ by the isomorphisms \eqref{equation.F_as_pushforward} and \eqref{equation.pushforward_is_pushdown_of_structure_sheaf}. Therefore applying the functor $\RDerived q_* ( - )_k$ to the system above and writing
$$\mcE_{k,j} := \RDerived q_* \left(\bigwedge{}^j \mcN^\vee\right)_k$$
we find that $Fl^{\vee k}$ is a convolution of the complex $\left(\mcE_{k,\bullet}, \partial \right)$. It only remains to show that $\mcE_{k,j}$ takes the form given above. We have \beqa \mcE_{k,j} & = & \RDerived q_* \left(\bigwedge{}^j \Hom(V, \wedge^2 S\{1\})^\vee \right)_k \\ & \iso & \RDerived q_* \left(\O\{-j\} \otimes \wedge^j V \otimes (\wedge^2 S^\vee)^j\right)_k \\ & \iso & \RDerived q_* \left(\O\{k-j\}\right)_0 \otimes \wedge^j V \otimes (\wedge^2 S^\vee)^j, \eeqa for $0 \leq j \leq d$, and knowing the cohomology of $\PP S$ then gives $$\mcE_{k,j} \iso \left\{ \begin{array}{cc} \Sym^{k-j} S^\vee & k \geq j \\ 0 & k = j-1 \\ \Sym^{-2-(k-j)} S \otimes \wedge^2 S[-1] & k \leq j-2 \end{array} \right\} \otimes \wedge^j V \otimes (\wedge^2 S^\vee)^j.$$
This rearranges to give the required result.
}
\end{lemma}
\begin{remark} The differentials $\partial$ are naturally determined by following the Koszul differentials $\ip \alpha$ through the functor $\RDerived q_* ( - )_k$ and the functorial isomorphisms used in the proof. We will not need to do this in our argument, so we omit an explicit description.
\end{remark}
\begin{remark}At least for $k \geq 0$, the convolutions obtained here are in fact examples of \remarkEmphasis{generalized Koszul complexes} \cite{Bruns:1998ty} associated to the degeneracy locus $B$ of the tautological map of bundles $V(1) \to S^\vee$ given by the following composition: $$\xymatrix{V(1) \ar[dr] \ar[r]^A & S(1) \ar[d] \ar[r]^\sim & S^\vee \ar[dl] \\ & X & }$$
\end{remark}
\subsection{Applying the adjoint $R$}
\label{section.applying_adjoint}
As in Remark \ref{remark.spherical_object_proof_plan}, steps 2 and 3, we hope to apply $R$ and find that only the first and last terms of the complexes $(\mcE_{k,\bullet}, \partial)$ survive. This does indeed carry through in Lemma \ref{proposition.image_within_window}, with an important caveat: the vanishing only works within a certain range $0 \leq k \leq d-2$.
First we offer some explanation for this phenomenon. The corresponding sheaves $l^{\vee k}$ do not generate $D^b(X_0)$, however they do generate the proper subcategory $\Im L$, as recorded in the proposition below.
\begin{proposition}\label{lemma.generators_for_subcategory} $\Im R = \Im L = \langle l^{\vee k} \rangle_{0 \leq k \leq d-2}$
\Proof{We use some technical results from the Appendices. Proposition \ref{proposition.tilting_object} gives us a tilting generator $\mathcal{T}$ for $D^b(X)$, derived from Kapranov's exceptional collection for $\mathbbm{Gr}$ \cite{Kapranov:2009wf}. Its summands are explicitly described in Section \ref{section.explicit_tilting_object} and are given by:
\begin{center} \begin{tikzpicture}
\bigWindowNodes{1.3}
\end{tikzpicture} \end{center}
Appendix \ref{section.concepts_of_generation} then gives that $\mathcal{T}$ split-generates $D^b(X)$. Applying $L$ to the summands of $\mathcal{T}$ using Proposition \ref{proposition.action_of_right_adjoint} we obtain
\begin{center} \begin{tikzpicture}
\bigWindowNodesResults{1.3}
\end{tikzpicture} \end{center}
so that all the terms vanish except for $L \Sym^k S^\vee \iso l^{\vee k}$. We then deduce the result for $\Im L$, and the result for $\Im R$ follows similarly (with a shift).
}
\end{proposition}
We also observe:
\begin{proposition}\label{proposition.adjoints_full} The functors $L$ and $R$ are full.
\Proof{We show in Lemma \ref{lemma.preimages_for_Homs_on_X} that the natural map $$\RHom_X(\Sym^a S^\vee, \Sym^b S^\vee) \To \RHom_{X_0}(l^{\vee a} , l^{\vee b})$$ induced by the functor $L$ for $a,b \in \Z$ is \emphasis{surjective}. We deduce that $L$ is full. The case of the functor $R$ is similar, with a shift.
}
\end{proposition}
From our new autoequivalence criterion described in Section \ref{section.autoequivalence_criterion}, we see that it suffices to understand the composition $C_FL$ to conclude that $T_F$ is an autoequivalence. Restricting therefore to the generators of $\Im L$ given in the proposition above we have:
\begin{lemma}\label{proposition.image_within_window} For $0 \leq k \leq d-2$, $$RF l^{\vee k} \iso l^{\vee k} \oplus l^{\vee k}[-s],$$ where as before $s := \dim \pi - \dim j$.
\Proof{Applying $R$ to the result of Lemma \ref{proposition.generalized_Koszul_complexes} we have that $RF l^{\vee k}$ is a convolution of a complex $(\mcF_{k,\bullet}, \partial)$ where $\mcF_{k,j} := R \mcE_{k,j}$. We claim that only the $\mcF_{k,d}$ and $\mcF_{k,0}$, corresponding to the left-most and right-most terms of the complexes, are non-zero. Now the convolution is defined using a Postnikov system as follows:
$$\verticalPostnikov{\mcF_{k,d}}{\mcF_{k,d-1}}{\mcF_{k,1}}{\mcF_{k,0}}{\mcC_{k,d}[-d]}{\mcC_{k,d-1}[-d+1]}{\mcC_{k,d-2}[-d+2]}{\mcC_{k,1}[-1]}{\mcC_{k,0}}$$ The $\mcC_{k,j}$ are \emphasis{partial convolutions} and $\mcC_{k,d} \iso RFl^{\vee k}$. Our vanishing assumption gives that most of the vertical right-hand maps are isomorphisms and so we see that $$\xymatrix@!C@C=1.6em{\mcC_{k,0} \ar[r]^\sim & \mcC_{k,1} \ar[r]^{\sim\quad} & \quad \ldots \quad \ar[r]^\sim & \mcC_{k,d-2} \ar[r]^\sim & \mcC_{k,d-1}}.$$ The uppermost distinguished triangle then reads $$\FlatExactTriangle{\mcC_{k,d}[-d]}{\mcF_{k,d}}{\mcF_{k,0}[-d+1]},$$ which gives $$\FlatExactTriangle{\mcF_{k,0}}{RFl^{\vee k}}{\mcF_{k,d}[d]}.$$
Now we claim specifically that $$\mcF_{k,j} \iso \left\{ \begin{array}{cc} R(\Sym^k S^\vee) & j=0 \\ R(\Sym^{d-k-2}S^\vee(k+1))[-1] & j=d \\ 0 & \text{otherwise} \end{array} \right\} \otimes \wedge^j V.$$ From Proposition \ref{proposition.action_of_right_adjoint} we have that \beq{equation.right_adjoint_vanishing} R(\Sym^{k-j} S^\vee(j)) \iso 0, \quad 0 < j \leq k \leq d-2.\eeq This suffices to show that $R(\mcE_{k,j}) \iso 0$ for $0 < j \leq k$, so it remains to consider $R(\mcE_{k,j})$ with $k+2 \leq j < d$: as we might expect (see Remark \ref{remark.spherical_object_proof_plan}), the vanishing here is dual to the vanishing \eqref{equation.right_adjoint_vanishing}. In this remaining case we have $$\mcE_{k,j} \iso\Sym^{j-k-2} S(-1)[-1] \otimes \wedge^j V(j).$$ Using Lemma \ref{lemma.exceptional_isomorphism}, which follows this one, we have $$\Sym^{j-k-2} S(j-1) \iso \Sym^{j-k-2} S^\vee (k+1).$$ We then see that $R(\mcE_{k,j}) \iso 0$ by applying \eqref{equation.right_adjoint_vanishing} with $$j'=k+1, \: k'=j-1,$$ and verifying that indeed $0 < j' \leq k' \leq d-2$ (under the assumption $k+2 \leq j < d$). Combining all this vanishing with Lemma \ref{proposition.generalized_Koszul_complexes} gives the claim.
Now from the results in Appendix \ref{section.calculations_on_generator}, we see that \beqa \mcF_{k,0} & \iso & R( \Sym^k S^\vee ) \\ & \iso & l^{\vee k} [\dim i - \dim \pi] \\ & = & l^{\vee k} [-s], \\ \mcF_{k,d} & \iso & R( \Sym^{d-k-2} S^\vee (k+1) )[-1] \\ & \iso & l^{\vee k} [\dim i] [-1] \\ & \iso & l^{\vee k} [-d].\eeqa Consequently $RFl^{\vee k}$ is an extension of the following two objects: \beqa \mcF_{k,0} & \iso & l^{\vee k}[-s], \\ \mcF_{k,d}[d] & \iso & l^{\vee k}. \eeqa There is no non-trivial extension of these sheaves, as each $l^{\vee k}$ is a summand of the tilting bundle $\mcT$. This gives the isomorphism.
}
\end{lemma}
\begin{lemma}\label{lemma.exceptional_isomorphism} $S \iso S^\vee(-1)$.
\Proof{As $\rk S = 2$ the natural map $S^\vee \otimes \wedge^2 S \To S$ is an isomorphism, and then $\O(-1) :=\wedge^2 S$ gives the result.}
\end{lemma}
\subsection{The cotwist on the image of $L$}
\label{section.cotwist_on_image}
Finally we can characterize the composition $C_FL$, as follows:
\begin{proposition}\label{proposition.RF_within_window} We have $C_FL \iso L[-s]$, indeed there exists a natural isomorphism of functors \beq{equation.natural_between_adjoints}\phi: R \overset{\sim}{\To} C_F L,\eeq induced by the natural transformation \beq{equation.inducing_natural_between_adjoints}R \overset{R\eta}{\To} RFL.\eeq
\Proof{For $\mcA \in D^b(X)$ the component of the claimed natural isomorphism is given by the morphism which makes the following diagram commute: $$\xymatrix{L\mcA \ar[r]^{\eta L_\mcA \quad} & RFL\mcA \ar[r] & C_FL\mcA \ar@{-->}[r] & \\ & R\mcA \ar[u]^{R \eta_\mcA} \ar@{..>}[ur]_{\phi_\mcA} &}$$
It will suffice to check that this is an isomorphism on the summands of our split-generator $\mathcal{T}$ for $D^b(X)$. As in Proposition \ref{lemma.generators_for_subcategory} we use the summands of $\mcT$ given in Proposition \ref{proposition.tilting_object} and described explictly in Section \ref{section.explicit_tilting_object}. As before the only summands which give non-zero objects after applying $L$ or $R$ are the $\Sym^k S^\vee$ for $0 \leq k \leq d-2$. By Lemma \ref{proposition.image_within_window} for these the left-hand part of the diagram then reads as follows: $$\xymatrix{l^{\vee k} \ar[r]^{\rho_l\qquad} & l^{\vee k} \oplus l^{\vee k}[-s] \\ & l^{\vee k}[-s] \ar[u]^{\rho_r}}$$ We determine the $\rho$'s. First observe that $$\Hom(l^{\vee k}, l^{\vee k}[-s]) \iso \Hom(l^{\vee k}[-s], l^{\vee k}) \iso 0,$$ and so $$\rho_l = \twobyone{z_l}{0}, \qquad \rho_r = \twobyone{0}{z_r},$$ for $z_l, z_r \in \End(l^{\vee k})$. Now note that our entire setup is invariant, and in particular the morphisms in question, under the $\C^*$-action given by scaling $A$ (this just scales the fibres on all our bundles). Now with this action $l^{\vee k}$ is exceptional in the sense that $\Hom^{\C^*}(l^{\vee k}, l^{\vee k}) \iso \C$.
We next prove that the morphisms are non-trivial. Firstly $\rho_l$ is a component of $\eta$, so it is necessarily non-trivial (otherwise $\epsilon F \circ F\eta \neq 1$). Secondly $L\eta$ is non-trivial (otherwise $\epsilon L \circ L\eta \neq 1$) so $R\eta$ is non-trivial by Proposition \ref{proposition.right_adjoint_description}, which gives the result for $\rho_r$. Consequently, using the scaling automorphisms of the $l^{\vee k}$, we have:
$$\xymatrix{l^{\vee k} \ar[r]^{\smalltwobyone{1}{0}\qquad} & l^{\vee k} \oplus l^{\vee k}[-s] \ar[r]^{\quad\smallonebytwo{0}{1}} & l^{\vee k}[-s] \ar@{-->}[r] & \\ & l^{\vee k}[-s] \ar[u]^{\smalltwobyone{0}{1}} \ar@{..>}[ur]_{\phi_{\Sym^k S^\vee}} &}$$ It then follows immediately that $\phi_{\Sym^k S^\vee}$ is an isomorphism.
}
\end{proposition}
\subsection{Autoequivalence proof}
\label{section.autoequivalence_proof}
\begin{theorem} $T_F$ is an autoequivalence.
\label{theorem.autoequivalence}
\Proof{We claim that $F : D^b(X_0) \to D^b(X)$ is \emphasis{Calabi-Yau spherical} as in Definition \ref{definition.Calabi-Yau_spherical} and apply our autoequivalence criterion, Proposition \ref{proposition.autoequivance_criterion}. The previous proposition gives the sphericity condition $R \overset{\sim}{\To} C_F L$. We explain why the other technical conditions hold:
\begin{itemize}
\item The category $D^b(X_0)$ is irreducible by \cite[Proposition 3.10]{Huybrechts:2007tf} because $X_0$ is smooth, and in particular normal.
\item The existence of $T_F$ and $C_F$ is covered in Section \ref{section.twist_construction}. They possess adjoints because they are of Fourier-Mukai type, and the left adjoint $L$ is full by Proposition \ref{proposition.adjoints_full}.
\item $D^b(X)$ and $D^b(X_0)$ are Calabi-Yau categories by Section \ref{section.Calabi-Yau} and hence have Serre functors \beqa S & = & [\dim X], \\ S' & = & [\dim X_0], \eeqa respectively. We therefore take $n := \dim X$ and $S^* := [\dim X_0]$. The local Calabi-Yau and compatibility conditions are then immediately satisfied, and $S^*$ is clearly an autoequivalence as required.
\end{itemize}
This shows that $F$ is Calabi-Yau spherical, and completes the proof.
}
\end{theorem}
\begin{remark}It follows immediately from the proof of Proposition \ref{proposition.autoequivance_criterion} that $F$ is a \emphasis{spherical functor} as in \cite{Anno:2007wo} and that $C_F$ is an autoequivalence: we give a full description of $C_F$ in \cite{Donovan:2011ufa}.\end{remark}
\section{Concepts of generation}
\label{section.concepts_of_generation}
We clarify two related concepts of generation for the derived category:
\begin{definition} We say that an object $\mcE$ \defined{split-generates} (or simply \defined{generates}) a triangulated category $\mcD$ if the smallest full triangulated subcategory closed under taking direct summands and containing $\mcE$ is $\mcD$ itself.
\end{definition}
Our goal is to show that an object $\mcE$ split-generates $\mcD = D^b(X)$ for a scheme $X$ if $\mcE$ is a \emphasis{tilting generator} in the sense explained below. To do this, we place appropriate smoothness and finite-dimensionality assumptions on $X$.
It will turn out that the tilting generator condition is easy for us to check in our examples. We explain this in the case of $X_0$ in the following Section \ref{section.base_tilting_object}. The case of $X$ is more elaborate, and is deferred to Appendix \ref{section.tilting}.
\begin{definition}\label{definition.tilting_generator} (cf. \cite[Definition 6]{Toda:2008wl})
We say that a locally free sheaf $\mcE$ on a scheme $X$, where $X$ is projective over a Noetherian affine of finite type, is a \defined{tilting generator} for $D^b(X)$ if
\begin{enumerate}
\item $\mcE$ is \defined{tilting} in that it satisfies $\RHom^{>0}_X(\mcE,\mcE) \iso 0$;
\item $\mcE$ is \defined{spanning} in the sense that $0 \iso \mcE^{\perp} \subset D^-(X)$.
\end{enumerate}
\end{definition}
It is standard that:
\begin{proposition}For $\mcE$ a tilting generator as above there exist quasi-inverse equivalences $$ \xymatrix@C=3em{D^b(A) \xybend[rr]^{\quad \Psi} & & D^b(X) \xybend[ll]^{\quad \Phi}} $$ Here $A := \End_X(\mcE)$, we write $D^b(A):=D^b(A\operatorname{-mod})$, and we define \beqa \Psi(-) & := & - \underset{A}{\Ltimes} \mcE, \\ \Phi(-) & := & \RHom_X(\mcE,-).\eeqa
\Proof{This is \cite[Lemma 8]{Toda:2008wl}.}
\end{proposition}
\subsection{Boundedness for tilting functors}
We now record some boundedness properties of the functors $\Phi$ and $\Psi$:
\begin{proposition}\label{proposition.boundedness_for_Phi} For $\mcF^\bullet \in D^b(X)$ we have: $$\HH^{\geq m}\mcF^\bullet \iso 0 \so \HH^{\geq m+\dim X}\Phi(\mcF^\bullet) \iso 0.$$
\Proof{To evaluate $\Phi(\mcF^\bullet)$ we use $$\Phi(-) := \RHom_X(\mcE, -) \iso \R\Gamma_X \hom_X(\mcE, -).$$ The $\hom$ need not be derived because $\mcE$ is a locally free sheaf. Our assumption $\HH^{\geq m}\mcF^\bullet \iso 0$ implies that $\HH^{\geq m}\hom_X(\mcE, \mcF^\bullet) \iso 0$. Now for any sheaf $\mcG$ we have $\HH^{>\dim X} \R\Gamma_X (\mcG) \iso 0$ by Grothendieck vanishing \cite[Theorem III.2.7]{Hartshorne:1977}, and so the result follows from vanishing in the following spectral sequence for $\R\Gamma_X$ \cite[Equation 2.6]{Huybrechts:2007tf}: $$E^2_{p,q} = \HH^p \R\Gamma_X(\HH^q (-)) \so \HH^{p+q} \R\Gamma_X(-).$$}
\end{proposition}
\begin{proposition}\label{proposition.boundedness_for_Psi} For $M \in A\operatorname{-mod}$ we have $$\HH^{>0}\Psi(M) \iso 0,$$ $$\HH^{<-\dim X}\Psi(M) \iso 0.$$
\Proof{The first vanishing follows directly from the definition of $\Psi$. We show how to deduce the second from Proposition \ref{proposition.boundedness_for_Phi}. Following \cite[Lemma 8]{Toda:2008wl}, we consider the canonical map $$\rho : \tau_{< m} \Psi(M) \To \Psi(M)$$ where $\tau_{< m}$ is a truncation functor. This is defined by $\tau_{<m} := \tau_{\leq m-1}$ where $$(\tau_{\leq n} \mcF^\bullet)^i := \left\{ \begin{array}{ll}\mcF^i & i < n \\ \ker \partial & i = n \\ 0 & i > n \end{array} \right\}$$ as in \cite[Section 1]{Toda:2008wl}. The crucial property of this functor for us is that $$\HH^{\geq m} \tau_{< m} \mcF^\bullet \iso 0,$$ whereas $\HH^i \rho$ is an isomorphism for $i < m$. Now applying $\Phi$ we obtain $$\Phi(\rho) : \Phi(\tau_{< m} \Psi(M)) \To \Phi\Psi(M) \iso M.$$ If we put $m:=-\dim X$ then Proposition \ref{proposition.boundedness_for_Phi} gives that $$\HH^{\geq 0} \Phi(\tau_{< m} \Psi(M)) \iso 0,$$ and then we see that $\Phi(\rho)$ must be zero, as its codomain $M \in D^b(A)$ is a complex concentrated in degree $0$. It follows that $\rho$ itself is zero, as $\Phi$ is an equivalence. Applying $\HH^i$ to $\rho$ for $i < -\dim X$ then allows us to deduce that $\HH^{<-\dim X} \Psi(M)\iso 0$ as required.}
\end{proposition}
\subsection{Consequences of smoothness}
Now assuming furthermore that $X$ is smooth we obtain the following lemma:
\begin{lemma}\label{lemma.higher_ext_vanishing} Given $X$ as above and additionally \theoremEmphasis{smooth}, for $M\in A\operatorname{-mod}$ we have $$\Ext^i_A(M,-) \iso 0$$ for $i \gg 0$, where the placeholder stands for an element of $A\operatorname{-mod}$.
\Proof{We note that \beq{equation.exts_by_smoothness}\Ext^i_A(M,-) \iso \HH^i \RHom_A(M,-) \iso \HH^i \RHom_X(\Psi(M),\Psi(-))\eeq using the equivalence $\Psi$. We want to show that this functor vanishes for $i \gg 0$:
\step{1} We show that for sufficiently large $i$ we have that $$\HH^i \RHom_X(\Psi(M),-)\iso 0,$$ where the placeholder now stands for a coherent sheaf on $X$. Proposition \ref{proposition.boundedness_for_Psi} gives that $\HH^{<-\dim X} \Psi(M) \iso 0$. We now use the spectral sequence \cite[Equation 2.8]{Huybrechts:2007tf} $$E^{p,q}_2 = \HH^p \RHom_X(\HH^{-q} \Psi(M), -) \so \HH^{p+q} \RHom_X(\Psi(M), -).$$ Any coherent sheaf on $X$ has a locally free resolution of length at most $\dim X+1$ by smoothness \cite[Proposition 3.26, and remarks following]{Huybrechts:2007tf}, and it follows that there exists $N$ such that $$\HH^{>N}\RHom_X(-,-)\iso 0,$$ where once again the placeholders stand for coherent sheaves on $X$. (We see this by using the locally free resolutions to evaluate the $\RHom$.) The resulting vanishing in the spectral sequence above suffices to deduce that $$\HH^{>\dim X+N} \RHom_X(\Psi(M),-)\iso 0,$$ as required.
\step{2} Now we consider the spectral sequence \cite[Equation 2.7]{Huybrechts:2007tf} $$E^{p,q}_2 = \HH^p \RHom_X(\Psi(M), \HH^q\Psi(-)) \so \HH^{p+q} \RHom_X(\Psi(M), \Psi(-)).$$ We have that $\HH^{>0} \Psi(-) \iso 0$ and so the previous step gives that \eqref{equation.exts_by_smoothness} vanishes for $i>\dim X+N$, and we are done. }
\end{lemma}
\begin{remark}We briefly indicate how locally free resolutions of length $\dim X+1$ for coherent sheaves $\mcG$ on $X$ are obtained. By \cite[Exercise III.6.8]{Hartshorne:1977} we can construct a locally free resolution $\mcF^\bullet \to \mcG$. We can then truncate this to give an acyclic complex $$0 \To \Im \partial \To \mcF^{-\dim X+1} \To \ldots \To \mcF^0 \To \mcG \To 0.$$ It follows by smoothness of $X$ that $\Im \partial$ is in fact locally free \cite[Proof of Lemma 2.5]{Burban:kWm0LpP3}, so this yields the required resolution.
\end{remark}
We then have:
\begin{proposition}\label{proposition.tilting_generator_criterion} If $\mcE$ is a tilting generator for $D^b(X)$ for $X$ smooth, then $\mcE$ also split-generates $D^b(X)$.
\Proof{Considering $A \in A\text{-mod}$ we have that $$\Phi(A) = A \underset{A}{\Ltimes} \mcE \iso \mcE$$ and we deduce that $\mcE$ split-generates $D^b(X)$ precisely when $A$ split-generates $D^b(A)$. We prove the latter claim as follows:
\step{1} We use the lemma to deduce that every $A$-module $M$ has a \emphasis{finite} projective resolution. For this we first note that the category of $A$-modules has enough projectives, so every $A$-module $M$ has a resolution by projective $A$-modules. Following \cite[Section III.5.9]{Gelfand} we write $\operatorname{pdim} M$ for the largest integer $i$ such that $\Ext^i_A(M,-) \not\iso 0$: this exists because of the smoothness of $X$ by Lemma \ref{lemma.higher_ext_vanishing}. Using \cite[Corollary III.5.12(a)]{Gelfand}, we find that $M$ then has a projective resolution of length $\operatorname{pdim} M + 1$.
\step{2} The previous step can be used to yield finite projective resolutions of more general objects $M^\bullet$ in $D^b(A)$. These are given by bounded complexes of $A$-modules $M^i$. We may resolve each $M^i$ separately to produce a bounded double complex \cite[Section 11.5]{Kashiwara:2005vw} of projective $A$-modules: the total complex of this is then a finite projective resolution of $M^\bullet$.
\step{3} Finally we show that $A$ split-generates $D^b(A)$. Consider then the smallest full triangulated subcategory $\mcC$ of $D^b(A\text{-mod})$ closed under taking direct summands and containing $A$. This contains the free $A$-modules $A^{\oplus i}$ (as these are iterated extensions of $A$), and the projective $A$-modules (as these are direct summands of the frees). It then follows from the previous step that $\mcC = D^b(A\text{-mod})$. This suffices to conclude.
}
\end{proposition}
\section{Calculations on the tilting generator}
\label{section.calculations_on_generator}
\subsection{Outline}
We investigate now what the adjoint functors $L$ and $R$ do to the summands of our tilting generator $\mathcal{T}$ from Appendix \ref{section.tilting_object} for the case $r=2$.
These results allow us to characterise the subcategory $\Im(L) \subset D^b(X_0)$ in Proposition \ref{lemma.generators_for_subcategory}. We also use them while understanding the action of the cotwist $C_F$ on $\Im L$ in Lemma \ref{proposition.image_within_window}, to apply $R$ to the convolution expressions for the $Fl^{\vee k}$ arising in Lemma \ref{proposition.generalized_Koszul_complexes}. All terms of each convolution, except the left-most term, are isomorphic to direct sums of the summands of $\mathcal{T}$. The left-most terms are isomorphic to directs sums of summands of $\mathcal{T} \otimes \O(1)$, and so we calculate the action of the adjoints on these too.
\begin{figure}[H]
\label{figure.generator_summands}
\caption{Summands of tilting generators $\mathcal{T}$ and $\mathcal{T} \otimes \O(1)$}
\begin{center} \begin{tikzpicture}
\twoBigWindowNodes{1.3}
\twoBigWindowLabelling
\end{tikzpicture} \end{center}
\end{figure}
\subsection{Calculation of images under $L$ and $R$}
Our calculation is routine, but quite elaborate. We first calculate $L (\O(k))$ for sheaves $\O(k)$ in the bottom row of the diagram:
\begin{lemma}\label{lemma.bottom_row_vanishing} We have $$L (\O(k)) \iso \left\{ \begin{array}{cc} \O & k=0 \\ 0 & 0<k<d-1 \\ l^{\vee d-2} \otimes \det V^\vee [\dim \pi] & k=d-1 \end{array} \right\}.$$
\Proof{Using our setup $$\xymatrix{\hat{B} \ar[r]^f \ar@/_{1em}/[rr]_j \ar[d]_\pi & B \xyhook[r]^i & X \\ X_0 & }$$ the first equality follows directly from the description of $L$ in Proposition \ref{proposition.right_adjoint_description}: \beqa L (\O) & = & \RDerived\pi_* (\omega_\pi \otimes \LDerivedj^* \O) [\dim \pi] \\ & \iso & \RDerived\pi_* \omega_\pi [\dim \pi] \\ & \iso & \O. \eeqa Now we saw in Proposition \ref{proposition.resolutions_and_flattening} that $\pi$ is a projective bundle $\PP(V/H)$. In the $r=2$ case, $H$ is a line bundle so we write $l:=H$ as before and obtain \beqa \omega_\pi & \iso & \O_\pi(-d+1) \otimes \det \left( \frac{V}{l} \right)^\vee \\ & \iso & \O_\pi(-d+1) \otimes l \otimes \det V^\vee. \eeqa We now pull back our sheaves $\O(k)$ from $X$ to $\hat{B}$ and write them in terms of the tautological bundle $\O_\pi(1)$, \beqa \LDerivedj^* \O(k) = \O(k) & = & (\wedge^2 S^\vee)^k \\ & \iso & (S/l)^{\vee k} \otimes l^{\vee k} \\ & = & \O_\pi(k) \otimes l^{\vee k}, \eeqa using $\wedge^2 S^\vee \iso l^\vee \otimes (S/l)^\vee$ which follows from the short exact sequence: $$0 \To (S/l)^\vee \To S^\vee \To l^\vee \To 0.$$ We then obtain \beqa L (\O(k)) & \iso & \RDerived\pi_*(\omega_\pi \otimes \O_\pi(k) \otimes l^{\vee k} )[\dim \pi] \\ & \iso & \RDerived\pi_*\left(\O_\pi(k-d+1) \otimes l^{\vee k-1} \otimes \det V^\vee\right) [\dim \pi] \\ & \iso & \RDerived\pi_*\left(\O_\pi(k-d+1)\right) \\ & & \qquad \otimes l^{\vee k-1} \otimes \det V^\vee [\dim \pi], \eeqa and apply standard vanishing to give the result.
}
\end{lemma}
We have now applied $L$ to the bottom row of our diagram. We apply $L$ to the rest by a simple induction argument (Proposition \ref{proposition.action_of_right_adjoint}) which relies on the following lemma:
\begin{lemma}\label{lemma.induction_SES} We have $$0 \To \Sym^b S^\vee (a+1) \otimes l \To \Sym^{b+1} S^\vee (a) \To l^{\vee a+b+1} \otimes \O_\pi(a) \To 0$$ on $\hat{B}$ where $a,b \geq 0$
\Proof{We have a short exact sequence $$0 \To (S/l)^\vee \To S^\vee \To l^\vee \To 0$$ which yields $$\xymatrix{0 \ar[r] & \Sym^b S^\vee \otimes (S/l)^\vee \ar[r] & \Sym^{b+1} S^\vee \ar[r] & l^{\vee (b +1)} \ar[r] & 0. \\ & \Sym^{b} S^\vee (1) \otimes l \xyequals[u] }$$ Multiplying this by $\O(a) \iso l^{\vee a} \otimes \O_\pi(a)$ we get the result.
}
\end{lemma}
We now perform our induction, according to the strategy shown below:
\begin{center} \begin{tikzpicture}
\twoBigWindowNodesArrows{1.3}
\inductionStrategyLabelling
\end{tikzpicture} \end{center}
\begin{proposition}\label{proposition.action_of_left_adjoint} For $0 \leq b \leq d-2$ we have $$L (\Sym^b S^\vee (a)) \iso \left\{ \begin{array}{cc} 0 & a>0, \: \: a+b \leq d-2 \\ l^{\vee b} & a=0 \\ l^{\vee d-2-b} \otimes \det V^\vee [\dim \pi] & a+b=d-1 \end{array} \right\}.$$
\Proof{
\stepDescribed{1}{vanishing} Fix $a+b=k \leq d-2$. We prove the vanishing results first by increasing induction on $b$. Vanishing is known for $b=0$ by Lemma \ref{lemma.bottom_row_vanishing}. Now Lemma \ref{lemma.induction_SES} gives $$0 \To \LDerivedj^* (\Sym^b S^\vee(a+1)) \otimes l \To \LDerivedj^* (\Sym^{b+1} S^\vee(a)) \To l^{\vee a+b+1} \otimes \O_\pi(a) \To 0.$$ If $b < k$ then $0 < a \leq d-2$ and hence $\RDerived\pi_* (\O_\pi(a) \otimes \omega_\pi) \iso 0$ and we deduce $$L (\Sym^b S^\vee(a+1)) \otimes l \iso L (\Sym^{b+1} S^\vee(a))$$ which yields the vanishing by induction.
\stepDescribed{2}{right-most sheaves} Now for $b>0$, Lemma \ref{lemma.induction_SES} gives $$0 \To \LDerivedj^* (\Sym^{b-1} S^\vee (1)) \otimes l \To \LDerivedj^* (\Sym^b S^\vee) \To l^{\vee b} \To 0,$$ and now the vanishing just proved gives $$L(\Sym^b S^\vee) \iso \pi_* (l^{\vee b} \otimes \omega_\pi) [\dim \pi] \iso l^{\vee b}$$ by the projection formula, as required.
\stepDescribed{3}{left-most sheaves} Here we prove the result by increasing induction on $b$. It holds for $b=0$. As before the proposition gives $$0 \To \LDerivedj^* (\Sym^b S^\vee (d-1-b)) \otimes l \To \LDerivedj^*(\Sym^{b+1} S^\vee (d-2-b))\To$$$$\qquad \qquad \To l^{\vee d-1} \otimes \O_\pi(d-2-b) \To 0.$$ This gives $$L(\Sym^b S^\vee (d-1-b)) \otimes l \iso L(\Sym^{b+1} S^\vee (d-2-b)),$$ which allows us to complete the induction.
}
\end{proposition}
\begin{corollary}\label{proposition.action_of_right_adjoint} For $0 \leq b \leq d-2$ we have $$R (\Sym^b S^\vee (a)) = \left\{ \begin{array}{cc} 0 & a>0, \: \: a+b \leq d-2 \\ l^{\vee b}[\dim j-\dim \pi] & a=0 \\ l^{\vee d-2-b} \otimes \det V^\vee [\dim j] & a+b=d-1 \end{array} \right\}.$$
\Proof{Use $R \iso L[\dim j - \dim \pi].$}
\end{corollary}
\begin{samepage}
Summarizing our results we have:
\begin{figure}[H]
\caption{Images under adjoints $L$ and $R$ of sheaves in Figure \ref{figure.generator_summands} (omitting the shifts and twists by $\det V$).}
\begin{center} \begin{tikzpicture}
\twoBigWindowNodesResults{1.3}
\end{tikzpicture} \end{center}
\end{figure}
\end{samepage}
\subsection{Fullness of functor $L$}
\label{section.fullness}
\begin{lemma}\label{lemma.preimages_for_Homs_on_X}We have $$L(\Sym^b S^\vee) \iso l^{\vee b}$$ for $0 \leq b \leq d-2$ and the natural map $$\RHom_X(\Sym^a S^\vee, \Sym^b S^\vee) \overset{\phi_{a,b}}\To \RHom_{X_0}(l^{\vee a} , l^{\vee b})$$ induced by the functoriality of $L$ is surjective for $0 \leq a,b \leq d-2$.
\Proof{The first part comes from Proposition \ref{proposition.action_of_left_adjoint}. We now analyse the map $\phi_{a,b}$ to show surjectivity, as follows:
\step{1} Working as in Proposition \ref{proposition.base_tilting_object} we find that \beqa
\RHom_{X_0}(l^{\vee a} , l^{\vee b}) & \iso & \bigoplus_k \R\Gamma_{\PP V}(l^a \otimes l^{\vee b} \otimes \Sym^k(V \otimes l^\vee)) \\ & \iso & \bigoplus_k \R\Gamma_{\PP V}(l^{\vee b+k-a} \otimes \Sym^k V)
\eeqa
Similarly working as in Proposition \ref{proposition.tilting_object}, Step 1, we see that
\beqa \RHom_X(\Sym^a S^\vee, \Sym^b S^\vee) & \iso & \bigoplus_k \R\Gamma_{\mathbbm{Gr}}(\underbrace{\Sym^a S \otimes \Sym^b S^\vee \otimes \Sym^k (V \otimes S^\vee)}_{P_{a,b,k}}) \eeqa where $$P_{a,b,k} := \Sym^a S \otimes \Sym^b S^\vee \otimes \Sym^k (V \otimes S^\vee).$$ The map $\phi_{a,b}$ respects the summation and so it suffices to show that its $k^{\text{th}}$ summand, say $\phi_{a,b,k}$, is surjective.
\step{2} We identify a particular direct summand in $P_{a,b,k}$: the map $\phi_{a,b,k}$ will factor through this summand (after taking sections). First note that the $\Sym^k(\ldots)$ factor of $P_{a,b,k}$ can be decomposed into a direct sum of irreducibles by the Cauchy formula \cite[Theorem 2.3.2]{Weyman:wu}: we will only need that \beq{equation.Sym^k_decomposition} \Sym^k V \otimes \Sym^k S^\vee \into \Sym^k(V \otimes S^\vee).\eeq Similarly the Littlewood-Richardson rule \cite[Formula A.8]{Fulton:1996tk} can be used to decompose the complementary $\Sym^a S \otimes \Sym^b S^\vee$ factor of $P_{a,b,k}$: we just note that \beq{equation.Sym^a,b_decomposition} \Sym^{b-a} S^\vee \into \Sym^a S \otimes \Sym^b S^\vee.\eeq Observe that \eqref{equation.Sym^a,b_decomposition} does indeed make sense for $b-a < 0$, providing that we take $$\Sym^{-k} S^\vee := \Sym^k S.$$ Now putting \eqref{equation.Sym^k_decomposition} and \eqref{equation.Sym^a,b_decomposition} together, and using the Littlewood-Richardson rule once again, we have an inclusion $i_{a,b,k}$ as follows $$\xymatrix@C=4em{ \Sym^{b+k-a} S^\vee \otimes \Sym^k V \: \xyhook@<0.5ex>[r]^{\qquad \qquad i_{a,b,k}} & P_{a,b,k} \ar@<0.5ex>[l]^{\qquad \qquad \pi_{a,b,k}}},$$ as well as a corresponding projection, which we denote $\pi_{a,b,k}$.
\step{3} We can now describe the map $\phi_{a,b,k}$: it corresponds to $\R\Gamma_{\mathbbm{Gr}} (\pi_{a,b,k})$ under the following chain of isomorphisms: $$\R\Gamma_{\PP V}(l^{\vee b+k-a} \otimes \Sym^k V) \iso \Sym^{b+k-a} V^\vee \otimes \Sym^k V \iso \R\Gamma_{\mathbbm{Gr}} ( \Sym^{b+k-a} S^\vee \otimes \Sym^k V ) $$
Now $\R\Gamma_{\mathbbm{Gr}} (\pi_{a,b,k})$ is clearly a surjection, split by $\R\Gamma_{\mathbbm{Gr}} (i_{a,b,k})$, and hence the result follows.}
\end{lemma}
\section{Introduction}
\subsection{Background and construction}
Beginning with the spherical twist of Seidel-Thomas \cite{Seidel:2000} a number of `twist' autoequivalences have now been studied, including \cite{Horja:2001,Huybrechts:2005wn, Anonymous:2008ti, Cautis:2010vi, Addington:2011tla}. These are endofunctors of the bounded derived categories $D^b(X)$ of certain varieties $X$, thought of as mirror to symplectic monodromies \cite{Thomas:2010tg} with Seidel's symplectic twist being the prototypical example \cite{Seidel:1999}. In the simplest case, this symplectic twist is just the topological Dehn twist for an embedded curve in a Riemann surface, hence the name.
In this paper we construct new examples of twist autoequivalences, using the technology of \emphasis{spherical functors} \cite{Anno:2007wo}. We work with a Grassmannian $\mathbbm{Gr} = \mathbbm{Gr}(r,V)$ of $r$-dimensional subspaces of a $d$-dimensional vector space $V$. Let $S$ denote the tautological subspace bundle and consider the bundle $\Hom(V,S)$, where $V$ denotes a constant bundle:
\begin{keyDefinition} {\bf [Definition \ref{definition.maximal_stratum}]} The total space $X := \Tot(\Hom(V,S))$ is stratified by the rank of the tautological map $V \To p^* S$, where $p$ denotes the projection $p : X \to \mathbbm{Gr}$. The \defined{{big stratum}}, denoted $B$, is the locus where the rank of the map is not full.
\end{keyDefinition}
We will exhibit an autoequivalence of the derived category $D^b(X)$ which we think of as a twist around the big stratum $B$.
\begin{remark} $X$ is Calabi-Yau (Section \ref{section.Calabi-Yau}).\end{remark}
\begin{remark}\label{remark.spherical_case} In the case $r=1$, we have $X = \Tot ( V^\vee \otimes \O_{\PP V}(-1) )$ and $B = \PP V$, the zero section. We have an autoequivalence of $D^b(X)$ given by the spherical twist around a spherical object, namely the sheaf $\O_B$ (see Section \ref{section.spherical_twist} for more details). Our work here generalises this. \end{remark}
We write a point of $X$ as a pair $(S,A)$ with $S \in \mathbbm{Gr}$ and $A \in \Hom(V,S)$. Composing $A$ with the inclusion of $S$ we obtain a map $$X \To \End^{\leq r}(V),$$ where $\End^{\leq r}(V)$ denotes the space of endomorphisms of $V$ with rank at most $r$. This map is a resolution of singularities. Restricting it to $B$ gives a map $$\pi_{\operatorname{sing}}: B \To \End^{<r}(V)$$ to the space of matrices $\End^{<r}(V)$ whose rank is less than $r$. This space is singular, and $\pi_{\operatorname{sing}}$ is not flat. We show how we may resolve the spaces $B$ and $\End^{<r}(V)$ and `flatten' the map $\pi_{\operatorname{sing}}$ by a commutative diagram as follows: \beq{diagram.flattening}\xymatrix{\hat{B} \ar[r]^f \ar[d]_\pi & B \xyhook[r]^i \ar[d]^{\pi_{\operatorname{sing}}} & X \\ X_0 \ar[r] & \End^{<r}(V)}\eeq
Here $\hat{B}$ and $X_0$ are smooth, and $\pi$ is a $\PP^{d - r}$-bundle. In particular $\pi$ is flat, which will be crucial in our proof. We then have:
\begin{keyProposition}\label{proposition.Grassmannian_twist} {\bf [Propositions \ref{proposition.F_well-defined}, \ref{proposition.twist_well-defined}]} The functor $$ F := i_* \RDerived f_* \pi^* : D^b(X_0) \To D^b(X)$$ is well-defined with a right adjoint $R$, and there exists a twist functor $T_F$ such that $$T_F\mcA \iso \{ FR\mcA \overset{\epsilon_\mcA}{\To} \mcA \}$$ where $\epsilon$ is the counit of the adjunction, and the braces denote taking the mapping cone.
\end{keyProposition}
We then show:
\begin{keyTheorem}\label{keyTheorem.autoequivalence} {\bf [Theorem \ref{theorem.autoequivalence}]} For $r=2$, the twist functor $T_F$ is an autoequivalence of $D^b(X)$.
\end{keyTheorem}
\begin{remark} In the course of the proof we find that $F$ is a \emphasis{spherical functor}. \end{remark}
\begin{remark} Similar results follow for $r>2$ by an extension of our construction, given in \cite{Donovan:2011ufa}. \end{remark}
\subsection{Relation with other results}
The bundle $\Hom(V,S)$ naturally contains $$\cotangentBundle{\mathbbm{Gr}(r,V)} \iso \Hom(V/S,S)$$ as a subbundle, with the inclusion given by composing with the projection $V \to V/S$. Derived autoequivalences of the total space of $\cotangentBundle{\mathbbm{Gr}(r,V)}$ have been found by \cite{Cautis:2009vz} in the case that $2r = \dim V$, using methods of $\mathfrak{sl}_2$-categorification: it would be very interesting to establish some connection with their results.
In other directions, we would like to see how our autoequivalences act on the non-commutative desingularisation of $\End^{\leq r}(V)$, recently demonstrated by \cite{Buchweitz:2011ug}. It would also be interesting to understand the mirror to $T_F$: we make some preliminary comments on this in \cite[Section 1]{Donovan:2011vc}.
\subsection{Outline}
We outline the contents of the paper:
\begin{itemize}
\item Section \ref{section.construction} describes the construction of the diagram \eqref{diagram.flattening}, the functor $F$ and the twist $T_F$.
\item Section \ref{section.autoequivalence_criterion} presents a condition, which we refer to as $F$ being \emphasis{Calabi-Yau spherical}, which implies that $T_F$ is an autoequivalence, and allows us to describe its action on the derived category in terms of a \emphasis{spanning set}.
\item Section \ref{section.autoequivalence_property} gives a proof that $F$ is Calabi-Yau spherical in our case.
\item Section \ref{section.twist-properties} describes the action of $T_F$ on $D^b(X)$ and the associated $K$-theory.
\end{itemize}
The Appendices gather technical results needed in our discussion:
\begin{itemize}
\item Appendix \ref{section.existence_of_twist_kernel} explains the Fourier-Mukai techniques needed to define $T_F$.
\item Appendix \ref{section.concepts_of_generation} explains the tilting generator technology which we use in our proof, and shows in detail that such generators are also \emphasis{split-generators} under suitable assumptions.
\item Appendix \ref{section.tilting} constructs a tilting generator $\mathcal{T}$ on $X$, having recalled the necessary Schur functor formalism.
\item Appendix \ref{section.calculations_on_generator} contains crucial but routine calculations of the action of the relevant functors on $\mathcal{T}$.
\end{itemize}
\subsection{Acknowledgements}
I am grateful to EPSRC for their financial support during my doctoral work, of which this forms a part. I am deeply indebted to many colleagues for useful conversations, suggestions and inspiration, in particular to Nicolas Addington, and to my examiners, Tom Bridgeland and Alessio Corti. Finally I am immensely grateful to my co-supervisor Ed Segal, whose ideas have made much of this work possible, and to my supervisor Richard Thomas for introducing me to this problem and for all his hard work and unflagging support over the years.
\section{Review: Spherical twists}
\label{section.spherical_twist}
In this section we review the spherical twist and spherical functors.
\subsection{Spherical twists (case $r=1$)}
\label{section.spherical_twist_example}
In this simple case we describe the construction of the twist and a proof that it gives an autoequivalence, using a method analogous to our later argument for $r=2$: the main ideas are similar, so we hope it will serve as a guide for the reader.
We have $X = \Tot ( V^\vee \otimes \O_{\PP V}(-1) )$ with maps as follows $$\xymatrix{\PP V \iso B\: \xyhook[r]^{\quad i} & X \ar[d]^p \\ & \PP V}$$ where $i$ is the inclusion of the zero section. We identify a \emphasis{spherical object} $i_* \O_{\PP V} \in D^b(X)$, and apply the following theorem. For simplicity, we specialise to the case of a Calabi-Yau variety.
\begin{theoremDefinition}\label{theoremDefinition.spherical_object_and_twist} \cite{Seidel:2000} \, Given a variety $X$ of dimension $n$ with $\omega_X \iso \O$, we say that an object $\mcE \in D^b(X)$ is \defined{spherical} if $$\RHom^\bullet_X (\mcE, \mcE) \iso H^\bullet(S^n, \C) \iso \C \oplus \C[-n],$$ where $S^n$ is the topological $n$-sphere. In this case there is an induced autoequivalence $T_\mcE$, the \defined{spherical twist}, given by $$T_\mcE \mcA \iso \{ \mcE \Ltimes \RHom_X(\mcE, \mcA) \To \mcA \}.$$
\end{theoremDefinition}
It follows that to establish an autoequivalence $T_\mcE$ we just have to show:
\begin{lemma}\label{lemma.zero_section_spherical} $\mcE := i_* \O_{\PP V} \in D^b(X)$ is spherical.
\Proof{We need to calculate $\RHom_X(\mcE, \mcE)$. By the adjunction $i_* \dashv i^!$ \cite[Section 3.4]{Huybrechts:2007tf} we have \beqa \RHom_X(i_* \O_{\PP V}, i_* \O_{\PP V}) & \iso & \RHom_{\PP V}(\O_{\PP V}, i^! i_* \O_{\PP V}) \\ & \iso & \RDerived\Gamma_{\PP V} (i^! i_* \O_{\PP V}), \eeqa and therefore we proceed as follows:
\step{1} The zero section $\PP V$ has normal bundle $\mathcal{N} \iso p^* (V^\vee \otimes \O_{\PP V}(-1))$. We write this as $V^\vee(-1)$ for brevity, and take the corresponding Koszul resolution: $$\{\wedge^d V(d) \To \wedge^{d-1} V(d-1) \To \ldots \To V(1) \To \underline{\O}\} \overset{\sim}{\To} i_* \O_{\PP V}$$
(The underline indicates the term in the complex which lies in degree $0$.)
\step{2} We then (twisted) restrict the resolution by applying \beqa i^! & =& \det \mathcal{N}[-\codim i] \otimes i^* \\ & \iso & \wedge^d V^\vee(-d)[-\codim i] \otimes i^*. \eeqa This gives that \beq{equation.restricted_zero_section_resolution} i^! i_* \O_{\PP V} \iso \{ \underline{\O} \To V^\vee(-1) \To \ldots \To \wedge^{d-1} V^\vee(-d+1) \To \wedge^d V^\vee(-d) \}.\eeq
\step{3} We evaluate $\RDerived\Gamma_{\PP V} (i^! i_* \O_{\PP V})$ by taking derived sections of \eqref{equation.restricted_zero_section_resolution}. The middle terms have no cohomology, the left-most term gives just $\C$, and the right-most term gives $\C[-\dim \PP V]$ by duality. We hence obtain
$$\RDerived\Gamma_{\PP V} (i^! i_* \O_{\PP V}) = \C \oplus \C[-\dim \PP V-\codim i] \iso H^\bullet(S^{\dim X}, \C),$$ as required.
}
\end{lemma}
\begin{remark}\label{remark.spherical_object_proof_plan}
Observe that we:
\begin{enumerate}
\item resolve the spherical object;
\item (twisted) restrict the resolution to the twisting locus $ B \iso \PP V$;
\item take derived sections and find that
\subitem the middle terms vanish,
\subitem one required piece comes from sections,
\subitem and the other from higher cohomology by duality.
\end{enumerate}
We will follow a similar plan in our argument: all these steps are reflected, albeit in more complicated ways. (Specifically, we work relative to the base $X_0$, and the vanishing becomes more subtle, see Section \ref{section.applying_adjoint}.)
\end{remark}
\subsection{Spherical functors}
\label{section.spherical_functors}
We give a categorical reformulation of the above twist which encompasses our Grassmannian twist $T_F$.
\begin{theoremDefinition}\label{theoremDefinition.spherical_functor_and_twist} \cite{Anno:2007wo}
An exact functor $F:\mcD_0 \To \mcD$ between triangulated categories is \defined{spherical} if \begin{enumerate}\item the \defined{cotwist} $$C_F \iso \{ \id \overset{\eta}{\To} RF \}$$ is an autoequivalence of $\mcD_0$, and \item the natural transformation $R \To C_F L$ induced by $R \overset{R\eta}{\To} RFL$ is an isomorphism of functors. \end{enumerate}
In this case there is an induced autoequivalence $T_F$ of $\mcD$, also known as a \defined{spherical twist}, given by $$T_F \iso \{ FR \overset{\epsilon}{\To} \id \}.$$\end{theoremDefinition}
\begin{remark}For brevity, we leave implicit the requirements that adjoints exist, and that twist and cotwist are well-defined. See \cite{Anno:2007wo} for a full formulation.\end{remark}
\begin{remark}This reduces to the previous theorem by taking $\mcD_0=D^b(\pt)$ and taking $F$ such that $F:\O_{\pt} \longmapsto \mcE$ and $R=\RHom_X(\mcE, -)$ \cite[Section 3, Example 1]{Anno:2007wo}.\end{remark}
\begin{remark}Observe that if our twisting locus $B$ were smooth, and the resolution map $f$ could therefore be taken as the identity, then our twist functor $T_F$ would reduce to a family spherical twist \cite{Horja:2001} \cite[Section 8.4]{Huybrechts:2007tf}. Unfortunately the presence of the map $f$ means that the proof method for the family spherical twist does not transfer to our case. See \cite[Section 2.4]{Donovan:2011vc} for discussion.
\end{remark}
\section{Tilting generator construction}
\label{section.tilting}
\subsection{Schur functors}
We briefly introduce \emphasis{Schur functors}, as they occur in the description of our tilting generator:
\begin{definition} \defined{(Schur functor for vector spaces)} Take a weight $\lambda$ for $GL(W)$. We write $\Sigma^\lambda W$ for the \defined{Schur power} which is the $GL(W)$-representation with highest weight $\lambda$, or $0$ if such a representation does not exist.
\end{definition}
\begin{remark} Weights for $GL(W)$ correspond to sequences $( \lambda_1, \ldots, \lambda_{\dim(W)} )$ of integers, and are ordered lexicographically. The weights occurring as \remarkEmphasis{highest weights} in $GL(W)$-representations are given by \remarkEmphasis{non-increasing} sequences of integers.
\end{remark}
\begin{example} Take $W := V$, $\dim V = 4$. Then we have for example:
\beqa
\Sigma^{1,0,0,0} V & = & V \\
\Sigma^{1,1,0,0} V & = & \wedge^2 V \\
\Sigma^{1,1,1,0} V & = & \wedge^3 V \\
\Sigma^{1,1,1,1} V & = & \wedge^4 V \\
\Sigma^{k,0,0,0} V & = & \Sym^k V
\eeqa
It follows from the definitions that the representations given have the required highest weight: for their irreducibility we refer to \cite{Fulton:1996tk}.
For a general dominant weight, the description of the Schur functor will be more complex. For an example of a non-dominant weight we have:
\beqa
\Sigma^{0,0,0,1} V & = & 0
\eeqa
\end{example}
In general the rule for multiplying Schur powers is quite elaborate, however we quote:
\begin{fact} \defined{(Pieri formula)} Given a weight of the form $\mu = (1, \ldots, 1, 0, \ldots, 0)$ we have $$\Sigma^\lambda W \otimes \Sigma^\mu W = \bigoplus_{\mu' \in S_d(\mu)} \Sigma^{\lambda + \mu'} W,$$ where $\mu'$ ranges over the orbit of $\mu$ under the natural permutation action of $S_d$, where $d=\dim W$. Weights are added component-wise. \cite[Appendix A, equation A.7]{Fulton:1996tk}
\end{fact}
\begin{definition} {\bf (Schur functor for vector bundles)} Take a vector bundle $E$ with structure group $GL(W)$. We can view $E$ as a principal $GL(W)$-bundle via the frame bundle construction. Given a weight $\lambda$ of $GL(W)$, we defined the \defined{Schur power} $\Sigma^\lambda E$ by $$\Sigma^\lambda E := E \underset{GL(W)}{\otimes} \Sigma^\lambda W.$$
\end{definition}
\begin{example}\label{example.Schur_expansion_for_Grassmannian} Take $E := S^\vee$, the dual of the tautological subspace bundle on the Grassmannian $\mathbbm{Gr}(2,V)$. The structure group here is $GL(2)$, and the highest weights are given by pairs $(\lambda_1, \lambda_2)$ with $\lambda_1 \geq \lambda_2$. In this simple case, the Pieri rule shows that the Schur powers decompose into products of $\Sym$s and $\wedge$s and we have: \beqa
\Sigma^{1,0} S^\vee & = & S^\vee \\
\Sigma^{1,1} S^\vee & = & \wedge^2 S^\vee = \O(1) \\
\Sigma^{2,1} S^\vee & = & \Sigma^{1,0} S^\vee \otimes \Sigma^{1,1} S^\vee = S^\vee(1) \\
\Sigma^{2,2} S^\vee & = & \Sigma^{1,1} S^\vee \otimes \Sigma^{1,1} S^\vee = \O(2) \\
\Sigma^{2,0} S^\vee & = & \Sym^2 S^\vee \\
\vdots & & \vdots
\eeqa (We work with the dual $S^\vee$ so that signs match between the left- and right-hand sides under our chosen polarization.)
\end{example}
\subsection{Construction}
\label{section.tilting_object}
We quote:
\begin{proposition}\label{proposition.kapranov_exceptional_collection}\cite[Section 3]{Kapranov:2009wf} There exists a full strong exceptional collection for $D^b(\mathbbm{Gr}(r,V))$ given by suitable Schur powers $$\{ \Sigma^\alpha S^\vee \}_{0 \leq \alpha \leq \alpha_{top} },$$ for $GL(r)$-weights $\alpha$ where $\alpha_{top} := (d-r, \ldots, d-r)$, for $d=\dim V$.
\end{proposition}
By standard arguments this yields a tilting generator for the Grassmannian $\mathbbm{Gr}$: $$\mathcal{T}_\Grassmannian := \bigoplus_{0 \leq \alpha \leq \alpha_{top} } \Sigma^\alpha S^\vee \in D^b(\mathbbm{Gr}).$$ We then obtain:
\begin{proposition}\label{proposition.tilting_object} There exists a tilting generator for $D^b(X)$ (Definition \ref{definition.tilting_generator}), constructed by pullback from the base $\mathbbm{Gr}$ as follows: $$\mathcal{T} := p^* \mathcal{T}_\Grassmannian = p^* \left( \bigoplus_{0 \leq \alpha \leq \alpha_{top} } \Sigma^\alpha S^\vee \right).$$
\Proof{Using a similar approach to \cite[Proposition 4.1]{Bridgeland:2005wi}, we first show that $\mathcal{T}$ is tilting, then demonstrate that it spans the derived category. This is an elaboration of Proposition \ref{proposition.base_tilting_object}. As indicated in the introduction, $X$ is a resolution of the affine singularity $\End^{\leq r}(V)$: in particular it is projective over a Noetherian affine of finite type as required in Definition \ref{definition.tilting_generator}.
\\
\step{1} To show that $\mathcal{T}$ is tilting we require $\RHom^{>0}_X (\mathcal{T}, \mathcal{T}) \iso 0$. First observe that \beqa \RHom_X(\mathcal{T},\mathcal{T}) & \iso & \RHom_X\left(p^* \mathcal{T}_\Grassmannian, p^*\mathcal{T}_\Grassmannian\right) \\ & \iso & \RHom_X\left(\mathcal{T}_\Grassmannian, p_* p^*\mathcal{T}_\Grassmannian\right) \\ & \iso & \RDerived\Gamma_X\left(\mathcal{T}_\Grassmannian^\vee \Ltimes p_* p^*\mathcal{T}_\Grassmannian\right) \\ & \iso & \RDerived\Gamma_X\left(\mathcal{T}_\Grassmannian^\vee \Ltimes\mathcal{T}_\Grassmannian\Ltimes p_* \O_X \right) \\ & \iso & \RDerived\Gamma_X\left(\mathcal{T}_\Grassmannian^\vee \Ltimes \mathcal{T}_\Grassmannian \Ltimes \Sym^\bullet \Hom(V,S)^\vee \right). \eeqa Note that in fact the tensor products of these locally free sheaves do not need to be derived. It suffices to show then that the following bundle has no higher cohomology: $$B_{\alpha,\alpha'} := \Sigma^{-\alpha} S^\vee \otimes \underbrace{\Sigma^{\alpha'} S^\vee \otimes \Sym^\bullet \Hom(V,S)^\vee}_{P_{\alpha'}}.$$ We consider in particular the highlighted bundle $P_{\alpha'}$ with $$P_{\alpha'} := \Sigma^{\alpha'} S^\vee \otimes \Sym^\bullet \Hom(V,S)^\vee.$$ This bundle $P_{\alpha'}$ may be decomposed into terms $\Sigma^\mu S^\vee$ with positive weight $\mu \geq 0$: this follows immediately from the \emphasis{Littlewood-Richardson rule} for calculating tensor products of Schur powers \cite[Formula A.8]{Fulton:1996tk} using that $\alpha' \geq 0$. Similarly the whole bundle $B_{\alpha,\alpha'}$ may be decomposed into Schur powers $\Sigma^\mu S^\vee$ with $\mu \geq -\alpha \geq -\alpha_{top}$. The higher cohomology of these bundles vanishes by the proof of Proposition \ref{proposition.kapranov_exceptional_collection}, see \cite[Lemma 3.2(a)]{Kapranov:2009wf} for details.
\\
\step{2} We now show that $\mathcal{T}$ spans $D^b(X)$. By adjunction we have $$\Hom_X(\mathcal{T}, -) \iso \Hom_X(p^* \mathcal{T}_\Grassmannian, -) \iso \Hom_\mathbbm{Gr}(\mathcal{T}_\Grassmannian, p_* -).$$ Now $p$ is affine hence $p_*$ is injective. $\mathcal{T}_\Grassmannian$ is a tilting generator for $\mathbbm{Gr}$ hence by Proposition \ref{proposition.tilting_generator_criterion} we have $\mathcal{T}_\Grassmannian^{\perp} \iso 0$. We deduce that $\mathcal{T}^{\perp} \iso 0$ as required.}
\end{proposition}
\begin{remark} Note that this construction does \remarkEmphasis{not} work with $X$ replaced by for instance the cotangent bundle $\cotangentBundle{\mathbbm{Gr}}$. In particular, for the simplest non-degenerate Grassmannian $\mathbbm{Gr}(2,4)$, it is noted in \cite[Remark 3.6(1)]{Kawamata:2005} that $$\Ext^2_{\cotangentBundle{\mathbbm{Gr}}}(p^* \mathcal{T}_\Grassmannian, p^* \mathcal{T}_\Grassmannian) \neq 0.$$ (We reuse the notation $p$ for the projection $\cotangentBundle{\mathbbm{Gr}} \to \mathbbm{Gr}$.)
For a construction of a tilting generator on $\cotangentBundle{\mathbbm{Gr}(2,4)}$ by another method see \cite{Toda:2008wl}.
\end{remark}
\subsection{Explicit descriptions}
\label{section.explicit_tilting_object}
\begin{samepage}
We restrict now to the $r=2$ case. The summands of $\mcT$ are as follows:
\begin{center} \begin{tikzpicture}
\bigWindowNodesSchurPowers{1.3}
\end{tikzpicture} \end{center}
\end{samepage}
\begin{samepage}
Expanding the Schur powers as in Example \ref{example.Schur_expansion_for_Grassmannian} we obtain:
\begin{center} \begin{tikzpicture}
\bigWindowNodes{1.3}
\end{tikzpicture} \end{center}
\end{samepage}
\section{Grassmannian twist construction}
\label{section.construction}
\subsection{Resolving singular strata}
\label{section.resolutions}
In this section we show how to `flatten' the map $\pi_{\operatorname{sing}}$ given in the introduction by constructing the commutative diagram \eqref{diagram.flattening}. First we describe the geometry involved and the resolutions required in our construction. We recall:
\begin{definition} \label{definition.maximal_stratum} The total space $X := \Tot(\Hom(V,S))$ is stratified by the rank of the tautological map $V \To p^* S$, where $p$ denotes the projection $p : X \to \mathbbm{Gr}$. The \defined{{big stratum}}, denoted $B$, is the locus where the rank of the map is not full.
\end{definition}
For $r>1$, $B$ is singular, with fibre over a point $S$ of $\mathbbm{Gr}$ given by the singular affine cone of homomorphisms \beqa B_S & = & \Hom^{< r}(V,S). \eeqa A natural way to resolve this space was suggested to us by \cite{Cautis:2009vz}. Following their notation we write \beqa B_S & = & \left\{ \varconds{ \xymatrix{ 0 \ar[r]^r & S \ar[r] & V \ar@/^/[l]^A } }{ \rk A \leq r-1 } \right\}.\eeqa In \cite{Cautis:2009vz}, inclusions are marked by their codimension: we will omit these when they are clear from context. Now to resolve this space we simply add, for each point, the data of a hyperplane $H \subset S$ containing $\Im(A)$. This is always possible because $\rk A \leq r-1$. We denote the resulting resolution by $$\hat{B}_S = \left\{ \xymatrix{ 0 \ar[r]^{r-1} & H \ar[r]^1 & S \ar[r] & V \ar@/^/[ll]^A } \right\},$$ with the obvious projection map $f_S : \hat{B}_S \To B_S$. Now we observe:
\begin{lemma} $\hat{B}_S$ is smooth.
\Proof{The space of hyperplanes of S, written as $$\{ \xymatrix {0 \ar[r]^{r-1} & H \ar[r]^1 & S } \},$$ is just the projective space $\PP^\vee S$, and we may reuse notation and denote its tautological hyperplane bundle by $H$. Then $\hat{B}_S$ is the total space of the bundle $$\xymatrix{ \Hom(V, H) \ar[d] \\ \PP^\vee S. }$$ Everything here is smooth, so we are done.}
\end{lemma}
Observe now that we can perform this construction in a family, by letting $S$ vary as a subspace of a fixed $V$. We then obtain:
\begin{definition}We have a \defined{resolution} $f: \hat{B} \To B,$ where we define $$\hat{B} := \left\{ \xymatrix{ 0 \ar[r]^{r-1} & H \ar[r]^1 & S \ar[r] & V \ar@/^/[ll]^A } \right\}.$$ The morphism $f$ is the natural one which forgets $H$.
\end{definition}
We can perform a similar construction on $\End^{<r}(V)$, the space of endomorphisms of rank less than $r$. This gives:
\begin{definition}\label{definition.twist_base_resolution} We have a \defined{resolution} $X_0 \To \End^{<r}(V),$ where we define $$X_0 := \left\{ \xymatrix{0 \ar[r]^{r-1} & H \ar[r] & V \ar@/^/[l]^A } \right\}.$$ As before the morphism is the one which forgets $H$.
\end{definition}
Putting this all together yields:
\begin{proposition}\label{proposition.resolutions_and_flattening} The resolutions defined above fit into a commutative square: $$\xymatrix{ \hat{B} \ar[r]^f \ar[d]_\pi & B \ar[d]^{\pi_{\operatorname{sing}}} \\ X_0 \ar[r] & \End^{<r}(V)}$$ The map $\pi$ is flat, being the projection map for the bundle $\PP(V/H)$.
\Proof{The maps are the natural forgetful ones, forgetting $H$ in the horizontal direction and $S$ in the vertical. The square commutes because forgetting $H$ and $S$ in either order gives the same result.
For the last part, we once again reuse the notation $H$ to denote the tautological bundle on $X_0$, and we then observe that $\hat{B}$ is isomorphic to the total space of the projective bundle $$\xymatrix{ \PP (V/H) \ar[d]^\pi \\ X_0. } $$ In particular the projection $\pi$ is flat as claimed.
}
\end{proposition}
\subsection{Calabi-Yau property}
\label{section.Calabi-Yau}
We show:
\begin{lemma}\label{lemma.Calabi-Yau} The total space $X$ of our bundle $$\xymatrix{\Hom(V,S) \ar[d]^p \\ \mathbbm{Gr}(r,V)}$$ is Calabi-Yau.
\Proof{The tangent bundle $T_X$ fits in an exact sequence $$0 \To p^* \Hom(V,S) \To T_X \To p^* T_\mathbbm{Gr} \To 0.$$ We then find \beqa \det T_X & \iso & \det p^* T_\mathbbm{Gr} \otimes \det p^* \Hom(V,S) \\ & \iso & p^* ( \det \Hom(S,V/S) \otimes \det \Hom(V,S) ), \eeqa where we use the fact that $T_\mathbbm{Gr} \iso \Hom(S,V/S)$. We also have $$0 \To \Hom(V/S,S) \To \Hom(V,S) \To \End(S) \To 0,$$ and so we deduce that $$\det T_X \iso p^* \det \End(S).$$ This is trivial because it is self-dual, and so we are done.
}
\end{lemma}
\begin{remark}\label{remark.Sigma-hat-Calabi-Yau} Note that $X_0$ is also Calabi-Yau. One way to see this is by setting $r=1$ in the above lemma.
\end{remark}
\subsection{Twist functor definition}
\label{section.twist_construction}
Using Proposition \ref{proposition.resolutions_and_flattening} we now have a diagram of schemes: $$\xymatrix{\hat{B} \ar[r]^f \ar[d]_\pi & B \xyhook[r]^i \ar[d] & X \\ X_0 \ar[r] & \End^{<r}(V)}$$
\begin{definition} We define a functor $F$ as the following composition: $$ F : \qquad \xymatrix{ D^b(X_0) \ar[r]^{\pi^*} & D^b(\hat{B}) \ar[r]^{\RDerived f_*} & D^b(B) \ar[r]^{i_*} & D^b(X). } $$
\end{definition}
For brevity we write $j := i f$.
\begin{proposition} $F$ is well-defined, and has a right adjoint given by $$F \dashv R := \RDerived\pi_* j^!.$$
\label{proposition.F_well-defined}
\Proof{As $\pi$ is flat, and $i$ the inclusion of a closed (albeit singular) subscheme $B$, we have that $\pi^*$ and $i_*$ are exact functors, and do not have to be derived. Finally the derived functor $\RDerived f_*$ preserves the bounded derived category because $f$ is a proper morphism of noetherian schemes, see \cite[Theorem 3.23 and discussion following]{Huybrechts:2007tf}.
For the existence of the adjoint we observe that $F \iso \RDerivedj_* \pi^*$ and use the adjunctions \beqa \pi^* & \dashv & \RDerived\pi_*, \\ \RDerivedj_* & \dashv & j^! .\eeqa The second of these is Grothendieck duality \cite[Corollary 3.35]{Huybrechts:2007tf}, which applies as $j = i f$ is a composition of proper morphisms, and hence proper \cite[Corollary II.4.8b]{Hartshorne:1977}. Composing the adjoints we obtain an adjoint for $F$.
}
\end{proposition}
We can now properly define our twist:
\begin{propositionDefinition} \label{proposition.twist_well-defined} The \defined{twist} $T_F : D^b(X) \To D^b(X)$ and the \defined{cotwist} $C_F : D^b(X_0) \To D^b(X_0)$ can be defined as functors of Fourier-Mukai type such that \beqa T_F\mcA & \iso & \{ FR\mcA \To \mcA \}, \\ C_F \mcB & \iso & \{ \mcB \To R F \mcB \}. \eeqa The morphisms are induced by the (co)unit of the adjunction $F \dashv R$.
\Proof{See Appendix \ref{section.existence_of_twist_kernel} for technical details of why suitable kernels exist.}
\end{propositionDefinition}
To end this section we prove a more concrete description of the right adjoint functor $R$, and the left adjoint $L$, for use later:
\begin{proposition} \label{proposition.right_adjoint_description} We have $$R \simeq \RDerived\pi_* (\omega_\pi \otimes -) \LDerivedj^* [\dim j],$$ where we use the relative canonical bundle $\omega_\pi$ given by $\omega_\pi := \omega_{\hat{B}} \otimes \pi^* \omega^{-1}_{X_0}.$ Furthermore we have a left adjoint $L$ with \beqa L & \simeq & R [s] \\ & \simeq & \RDerived\pi_* (\omega_\pi \otimes -) \LDerivedj^* [\dim \pi],\eeqa where $s := \dim \pi - \dim j$.
\Proof{We have relative canonical bundles for the morphisms $\pi$ and $j$ because the spaces involved are smooth, so we can write \beqa \omega_{\hat{B}} & \iso & \omega_\pi \otimes \pi^* \omega_{X_0} \\ & \iso & \omega_{j} \otimes j^*\omega_X. \eeqa Both $X$ and $X_0$ are Calabi-Yau (Lemma \ref{lemma.Calabi-Yau}), so we deduce that $\omega_\pi \iso \omega_{j}$. Using \cite[Corollary 3.35]{Huybrechts:2007tf} $$j^! - \iso \omega_{j} [\dim j] \otimes \LDerivedj^* -,$$ the expression for $R$ is immediate. We now express $L$ in terms of $R$ and the Serre functors, denoted $S_{X_0}$ and $S_X$, for the categories in question \cite[Remark 1.31]{Huybrechts:2007tf}. By the Calabi-Yau property, the Serre functors are simply shifts: \beqa L & \iso & S_{X_0}^{-1} R S_X \\ & \iso & S_{X_0}^{-1} S_X R \\ & \iso & R [\dim X - \dim X_0] \\ & = & R [\dim \pi - \dim j]. \eeqa The result follows.}
\end{proposition}
\section{Existence of the twist kernel}
\label{section.existence_of_twist_kernel}
We prove the following technical lemma regarding our functor $F$:
\begin{lemma} The functor $F$ is of Fourier-Mukai type with kernel $$K := \O_{(\pi \times j) \hat{B}} \in D^b(X_0 \times X).$$ The kernel is perfect, and its support is proper over both factors $X_0$ and $X$.
\Proof{First note that $\pi^*$ is Fourier-Mukai type with kernel given by the graph of the morphism $\pi$ \cite[Exercise 5.4(ii)]{Huybrechts:2007tf}, that is $$\RDerived(\pi \times \id)_* \O_{\hat{B}} \in D^b(X_0 \times \hat{B}).$$ Now by \cite[Exercise 5.12(i)]{Huybrechts:2007tf} the composition $\RDerived j_* \pi^*$ is Fourier-Mukai type with kernel \beqa \RDerived (\id \times j)_* \RDerived (\pi \times \id)_* \O_{\hat{B}} & \iso & \RDerived (\pi \times j)_* \O_{\hat{B}} \\ & \iso & \O_{(\pi \times j)\hat{B}}. \eeqa
The last step follows because $\pi \times j$ is a closed embedding. We check this on closed points: by definition $\pi \times j$ takes points to points as follows:
\beqa \hat{B} & \To & X_0 \times X \\ \left( \xymatrix{ 0 \ar[r]^{r-1} & H \ar[r]^1 & S \ar[r] & V \ar@/^/[ll]^A } \right) & \longmapsto & \left( \xymatrix{ 0 \ar[r]^{r-1} & H \ar[r] & V \ar@/^/[l]^A }, \xymatrix{ 0 \ar[r]^{r} & S \ar[r] & V \ar@/^/[l]^A }\right) \eeqa
For the second part note the kernel is perfect because it is a sheaf on a smooth space, and consequently has a finite resolution by locally frees. The projection maps $p$ and $q$ from $\Supp(K) = (\pi \times j)\hat{B}$ are shown below:
$$\xymatrix{ & \hat{B} \ar[d]_{\wr} \ar[ld]_\pi \ar[rd]^j \\ X_0 & (\pi \times j)\hat{B} \ar[l]^{p\quad} \ar[r]_{\quad q} \xyhookdown[d] & X \\ & X_0 \times X \ar[ul]^p \ar[ur]_q & } $$
They are proper because $\pi$ and $j$ are proper. This is clear because $\pi$ is projective, and hence proper, and $j = i \circ f$ is a composition of proper morphisms, hence proper.
}
\end{lemma}
\begin{remark}The assumptions on the kernel suffice to guarantee that the Fourier-Mukai transform and its adjoints preserve boundedness and coherence. We will refer to \cite{Anno:2010ws} for the proof that the twist $T_F$ exists: these assumptions are the ones used there.
\end{remark}
We can now define our twist:
\begin{proposition} \label{proposition.twist_well-defined_appendix} The twist $T_F : D^b(X) \To D^b(X)$ and the cotwist $C_F : D^b(X_0) \To D^b(X_0)$ can be defined as functors of Fourier-Mukai type such that \beqa T_F\mcA & \iso & \{ FR\mcA \overset{\epsilon_\mcA}{\To} \mcA \}, \\ C_F \mcB & \iso & \{ \mcB \overset{\eta_\mcB}{\To} R F \mcB \},\eeqa with the morphisms given by the (co)unit of the adjunction $F \dashv R$.
\begin{remark} We note that the cone construction is non-functorial, so we cannot simply define $T_F$ as the cone on the counit morphism. Instead we follow the standard procedure of constructing a Fourier-Mukai kernel which yields a functor $T_F$ with the required property, and similarly for $C_F$.
\end{remark}
\Proof{
\stepDescribed{1}{twist} From the lemma we have that $F$ is of Fourier-Mukai type with kernel $K$. To obtain a functor $T_F$ as required, we use \cite[Corollary 3.5]{Anno:2010ws} under the assumptions that
\begin{samepage}
\begin{itemize}
\item $K$ is perfect, and
\item $\Supp(K)$ is proper over $X_0$ and $X$.
\end{itemize}
\end{samepage}
This gives us a morphism of kernels $Q \overset{\overline{\epsilon}}{\To} \O_\triangle$ where $\Phi_Q \mcA \iso FR\mcA$ and the following diagram commutes: $$\xymatrix{ \Phi_Q \mcA \ar[r]^{{\Phi_{\overline{\epsilon}}}_\mcA} \ar[d]_\wr & \Phi_{\O_\triangle}\mcA \ar[d]^\wr \\ FR\mcA \ar[r]_{\epsilon_\mcA} & \mcA }$$ Here $\epsilon$ is the counit. The required conditions hold by the lemma above, so we may define $$T_F := \Phi_{\{Q \overset{\overline{\epsilon}}{\To} \O_\triangle\}}.$$
\stepDescribed{2}{cotwist} The result for the cotwist follows from the dual result by taking adjoints. Specifically \cite[Theorem 3.1]{Anno:2010ws} similarly gives us a morphism of kernels $Q' \overset{\overline{\epsilon}}{\To} \O_\triangle$ where $\Phi_{Q'} \mcA \iso LF\mcA$ and the following diagram commutes: $$\xymatrix{ \Phi_{Q'} \mcA \ar[r]^{{\Phi_{\overline{\epsilon}}}_\mcA} \ar[d]_\wr & \Phi_{\O_\triangle}\mcA \ar[d]^\wr \\ LF\mcA \ar[r]_{\epsilon_\mcA} & \mcA }$$ We reuse the notation $\epsilon$ for the counit morphism. Now we use \cite[Proposition 5.9]{Huybrechts:2007tf} to produce kernels which induce right adjoints of the functors $LF$ and $\id$, which are given by: \beqa LF & \dashv & RF, \\ \id & \dashv & \id. \eeqa Noting the Calabi-Yau condition on $X$, the proposition tells us that these are given by applying the functor $\mathbb{D} := (-)^\vee [\dim X]$ to the kernels. We then have: $$\xymatrix{ \Phi_{\mathbb{D} \O_\triangle} \mcA \ar[r]^{{\Phi_{\mathbb{D} \overline{\epsilon}}}_\mcA} \ar[d]_\wr & \Phi_{\mathbb{D} Q'} \mcA \ar[d]^\wr \\ \mcA \ar[r]_{\eta_\mcA} & RF \mcA }$$
This commutes because the counit $\epsilon$ and unit $\eta$ are taken to each other by the adjunction isomorphism for the adjunction $LF \dashv RF$. Finally, observing that $\mathbb{D} \O_\triangle \iso \O_\triangle$ we may define $$C_F := \Phi_{\{\O_\triangle \overset{\mathbb{D} \overline{\epsilon}}{\To} \mathbb{D} Q'\}}.$$}
\end{proposition}
\section{Properties of the twist}
\label{section.twist-properties}
\subsection{Action on the spanning set $\Omega$}
We observe that Proposition \ref{proposition.spanning_set} applies to $D^b(X)$ to yield a spanning set $\Omega$. The action of the twist $T_F$ on this set is as follows:
\begin{proposition} We have
\begin{enumerate} \item $\mcA \in \ker L \so T_F\mcA \iso \mcA,$ and \item $\mcA \in \Im FL \so T_F\mcA \iso \mcA[-s+1]$. \end{enumerate}
\Proof{This follows immediately from Proposition \ref{proposition.Calabi-Yau_spherical_functor_action}. For the second part we use that by definition $S^*[-n+1] \iso [\dim X_0 - \dim X + 1] = [-s+1]$.
}
\end{proposition}
\subsection{Action on $K$-theory}
\label{section.action_on_K-theory}
The autoequivalence $T_F$ induces an endomorphism of the algebraic $K$-theory $K(X)$, which we write $T^K_F$. See \cite[Section 5.2]{Huybrechts:2007tf} for details.
We show now that the spanning set $\Omega$ induces a decomposition of $K(X)$, and exhibit the action of $T^K_F$ on this decomposition.
\begin{definition}\cite[Section 5.2]{Huybrechts:2007tf} We write $K(X)$ for the \defined{algebraic $K$-theory} of $X$. This is the free abelian group generated by the locally free sheaves $\mcE$ on $X$, modulo the equivalence relation that $\mcE \sim \mcF_1 + \mcF_2$ if $\mcE$ is an extension of the $\mcF_i$.
\end{definition}
\begin{definition}For $\mcE^\bullet \in \Perf(X)$ we write $[\mcE^\bullet]$ for its \defined{$K$-theory class} which is given by $$[\mcE^\bullet] := \sum_i (-1)^i \mcE^i.$$
\end{definition}
\begin{remark} Here $\Perf(X) \subseteq D^b(X)$ denotes the subcategory of perfect complexes, that is those complexes isomorphic to bounded complexes of locally frees. For $X$ smooth these categories coincide.
\end{remark}
\begin{remark} A spanning set need not generate the $K$-theory in general. For instance the sheaves $\{ \O_p \}_{p \in \PP^1}$ span $D^b(\PP^1)$ but all have the same $K$-theory class, and yet $\rk (K(\PP^1)) = 2$.
\end{remark}
\begin{remark} Note that $\Omega$ spans $D^b(X)$ but we do \remarkEmphasis{not} make the stronger claim that $\Omega$ \remarkEmphasis{generates} $D^b(X)$. In the latter case it would follow immedately that $[\Omega]$ spans the $K$-theory.
\end{remark}
\begin{lemma}$\rk [\Im FL] = d-1$.
\Proof{From Proposition \ref{lemma.generators_for_subcategory} we have that $\Im L = \langle l^{\vee k} \rangle_k$ where $k$ varies within $0 \leq k \leq d-2$. This gives $\Im FL = \langle Fl^{\vee k} \rangle_k.$ It remains to check that the $[Fl^{\vee k}]$ are $\Z$-linearly independent in $K(X)$: using Lemma \ref{proposition.image_within_window} we have that \beqa \Im RFL & = & \langle RFl^{\vee k} \rangle_k \\ & = & \langle l^{\vee k} \oplus l^{\vee k} \rangle_k \\ & = & \langle l^{\vee k} \rangle_k \: \subset \: D^b(X_0) \eeqa and so immediately we see that $\rk [\Im RFL] = d-1$, and we deduce the result.
}
\end{lemma}
Finally we can show:
\begin{proposition}We have $K(X) \iso [\Im FL] \oplus [\ker L]$ and according to this decomposition $$T^K_F = \begin{pmatrix} \mathbbm{1} & 0 \\ 0 & -\mathbbm{1} \end{pmatrix}.$$
\Proof{$[\ker L]$ is generated by the summands of $\mathcal{T}$ which vanish under $L$: there are $$\rk K(X) - (d-1)$$ of these. The subcategories $\Im FL$ and $\ker L$ are orthogonal under the Euler pairing on $K(X)$ by Proposition \ref{proposition.spanning_set}, and so the decomposition follows. The action of $T^K_F$ then follows from the proposition above.}
\end{proposition}
|
\section{Introduction}
Importance of solvable models in physics is enormous. We can acquire
qualitative understanding of the complicated realistic systems
by analyzing simplified models that grab the essence of a physical reality.
These models can serve as a test field for approximative
methods, or can be used as initial solvable systems in a
perturbative treatment. In this paper, we will focus on the construction
and analysis of exactly solvable models described by the $(1+1)$-dimensional Dirac equation.
Such systems lie in the overlap of the quantum field theory with the
condensed matter physics. The one-dimensional Dirac Hamiltonian appears
in the study of the gap equation of the $1+1$ dimensional version of the
Nambu-Jona-Lasinio (chiral Gross-Neveu) model \cite{Dunne}, \cite{Feinberg},
\cite{Thies}, or in the study of the fractionally charged solitons
\cite{Jackiw}, \cite{Jackiw2}. It is used in the effective description of
the non-relativistic fermions: in \cite{Takayama}, the Hamiltonian describes
fermions coupled to solitons in the continuum model of a linear molecule of
polyacetylene. It is employed in the analysis of the quasi-particle bound
states associated with the planar solitons in superfluid $^3\mbox{He}$
\cite{Ho}. It appears in the description of inhomogeneous superconductors
\cite{inhomogeneous} and in the analysis of the vortex in the
extreme type-II superconductors in the mean field approximation \cite{Waxman}.
Last but not least, it is used in the description of carbon nanotubes. In the
low energy regime, the band structure obtained by tight-binding approach can
be approximated very well with the use of the one-dimensional Dirac operator
\cite{Wallace}, \cite{Semenoff}. The stationary equation \cite{Paulimatrix}
\begin{equation}\label{eq1}
(i\sigma_2\partial_x+\Delta_1(x)\sigma_1)\phi=\lambda\phi
\end{equation}
describes dynamics of the low-energy charge-carriers in single wall carbon
nanotubes in presence of magnetic field \cite{Roche}, \cite{KaneMele}.
The Green's function (or its spatial trace called diagonal resolvent or
Gorkov Green's function) plays an important role in the above mentioned systems.
It is used in solution of the gap equation \cite{Dunne} or in the extremal
analysis of the effective action \cite{Feinberg} in quantum field systems.
It is employed in computation of the free energy of the inhomogeneous
superconductors \cite{Kos}. It serves in derivation of the local density of states
(LDOS), the quantity that can be measured in carbon nanostructures
by the spectral tunneling microscopy \cite{STM}, \cite{STM2}. The results
obtained in this paper will be primarily discussed
in the latter context.
The carbon nanotubes are cylinders of small radius rolled up from graphene.
They can be classified as either metallic or semiconducting, in dependence on
their electronic properties. When no external potential is present, the semi-conducting
nanotube has a spectral gap which is related to a constant value of the potential,
$\Delta_1=p_y\neq 0$, where $p_y$ is the value of the canonical momentum in the
compactified direction. For $\Delta_1=p_y=0$, the nanotube is metallic as it has no
gap in the spectrum. In this case, an infinitesimally small excitation is sufficient
to move the electrons from valence to conduction band. The actual value of $p_y$ is
related to the orientation of the crystal lattice in the nanotube, see e.g.,
\cite{Roche}, \cite{KaneMele}, \cite{nasKlein}.
We suppose that the potential $\Delta_1(x)$ is smooth on the scale of the interatomic
distance. Otherwise, it would be necessary to work with an extended, $4\times 4$,
Hamiltonian that would describe mixing of the states between the valleys associated
with two inequivalent Dirac points \cite{graphene}, \cite{Ando1}. The matrix degree
of freedom of $\phi$ in (\ref{eq1}) is the so-called pseudo-spin and is associated
with the two triangular sublattices that build up the hexagonal structure of the
graphene crystal; the wave function with either spin-up or -down is identically
zero on one of the sublattices.
The inhomogeneous magnetic field can appear due to an external source.
Alternatively, it can emerge as a consequence of mechanical deformations of the
lattice. Let us make this point clear. Deformation of the lattice is described by
the vector $\mathbf{d}=(d_x(x,y),d_y(x,y))$ which represents displacement of the
atoms in the crystal. The associated strain tensor $s_{ij}$ is defined as
\begin{equation}\label{strain}
s_{xx}=\partial_xd_x,\quad s_{yy}=\partial_yd_y,\quad s_{xy}=s_{yx}=\frac{\partial_xd_y+\partial_yd_x}{2}.
\end{equation}
The effective Dirac Hamiltonian which describes dynamics of quasi-particles in the
low-energy regime gets the form
$\sigma_2(i\partial_x+\Delta_2(x))+\sigma_1(p_y+\Delta_1(x))+\mathbf{1}\Delta_0$,
where we fixed the Fermi velocity $v_F=1$.
The gauge fields are related to the strain tensor (\ref{strain}) in this way:
$\Delta_2(x)= (s_{xx}-s_{yy})$, $\Delta_1(x)= 2s_{xy}$ and $\Delta_0(x)= s_{xx}+s_{yy}$
up to multiplicative constants,
see \cite{KaneMele}, \cite{ando}, \cite{Vozmediano}.
In this context, the potential $\Delta_1(x)$ in (\ref{eq1}) can be interpreted as
the gauge field generated by the twist perpendicular to the axis of the \textit{metallic}
nanotube. The angle of the twist $\vartheta(x)$ is related to the displacement
$\mathbf{d}=(0,\int \Delta_1(x)dx)$ by $d_y(x)=r\vartheta(x)$ where $r$ is a radius
of the nanotube. In this way, the constant potential $\Delta_1(x)=\beta>0$ can be
associated with a linear displacement ${\mathbf d}=(0,\beta x)$ that would be generated
by the constant twist illustrated in Figure 1. It opens a gap in the spectrum of the
metallic nanotube, however, it does not confine charge carriers. Indeed, constant potential
can be understood as a mass term in the Hamiltonian describing the free particle.
\newsavebox{\figlinear}
\savebox{\figlinear}{
\scalebox{1}{
\includegraphics[scale=.7]{figure1.eps}
}
}
\begin{figure}\begin{center}\label{figlinear}
\usebox{\figlinear}\caption{The nanotube with the twist corresponding to $ d_y\sim x$.
In the untwisted nanotube, the black line would be straight (horizontal).}
\end{center}\end{figure}
In general, the electromagnetic field causes nontrivial scattering of the quasi-particles
and can even cause the appearance of bound states in the system \cite{Hartmann}. It is
well known that the quasi-particles in metallic nanotubes are not backscattered by
electrostatic potential. This is understood as a manifestation of the Klein tunneling
\cite{Klein} and it has been discussed extensively in the literature
\cite{Ando1}, \cite{KatsnelsonKlein}. It was found recently that the phenomenon can be attributed
to the peculiar supersymmetric structure that relates the
Hamiltonian of the system to that of the free Dirac particle \cite{nasKlein}.
Here, we will construct exactly solvable
models described by (\ref{eq1}) where, despite the presence of the effective magnetic
field, the scattering will be reflectionless and the bound states will be confined
in the regions where the twist gets altered. In the construction, the techniques known
in the supersymmetric quantum mechanics will be employed. We will focus to the spectral
properties and Green's function of the new systems. The latter one will be used for
computation of the LDOS. We will provide an analytical formula for bound state energies
in dependence on the twist of the nanotubes.
The work is organized as follows. In the next section, we briefly review the construction
of solvable models based on Darboux transformation with focus on the application in the
context of carbon nanotubes. Then the formulas for Green's function and LDOS of these
models are provided. We discuss reflectionless systems and present two models of twisted
carbon nanotubes. The last section is left for the discussion.
\section{Spectral design via Darboux transformations}
We summarize here the main points of the construction of new solvable models which is
based on the intertwining relations.
This scheme is well known in the context of supersymmetric (SUSY) quantum mechanics
\cite{JunkerKhare}. There, the intertwined second order Schr\"odinger operators give
rise to the supersymmetric Hamiltonian while the intertwining operator, identified
as the Crum-Darboux transformation, is associated with the supercharges of the system.
In the current case, we will discuss briefly the technique in the context of the
first order, one-dimensional Dirac equation. We refer to \cite{DiracDarboux} for
more details.
Let us have a physical system described by a solvable hermitian Hamiltonian
\begin{equation}\label{seed}
h=i\sigma_2\partial_x+\Delta
\end{equation}
with real and symmetric matrix potential $\Delta=\Delta(x)$ and $x$ extending to the
whole real axis. The physical eigenstates (solutions complying with prescribed boundary
conditions) form a basis of the Hilbert space. Besides, the (formal) solutions of the
stationary equation $hu=\lambda u$ are supposed to be known for any complex $\lambda$.
We define the operator $L$ by
\begin{equation}\label{L}
L=U\frac{\partial}{\partial x} U^{-1}=\mathbf{1}\partial_x-U' U^{-1},
\end{equation}
where $U'=\partial U/\partial x$. The matrix $U=(u_1,u_2)$ is a chosen solution of the
equation $h\, U=U\,\Lambda$ where the matrix $\Lambda=diag(\lambda_1,\lambda_2)$ has
fixed real elements. The vectors $u_{1(2)}$ satisfy $hu_{1(2)}=\lambda_{1(2)} u_{1(2)}$
and are chosen to be real. They form the kernel of $L$, $LU=0$ and do not need to be physical.
Next, we define the hermitian operator $\tilde{h}$ with the potential term explicitly
dependent on $u_1$ and $u_2$ and corresponding eigenvalues $\lambda_1$ and $\lambda_2$,
\begin{eqnarray}\label{htilde}
\tilde{h}&=&h+i[\sigma_2,U'U^{-1}]=\sigma_2h\sigma_2+\sigma_2[\sigma_2,U\Lambda U^{-1}]\nonumber\\
&=&\sigma_2 h\sigma_2+\left(\frac{u_1^T\sigma_1u_2}{\det U}\sigma_3-\frac{u_1^T\sigma_3u_2}{\det U}\sigma_1\right)(\lambda_1-\lambda_2).\nonumber\\
\end{eqnarray}
We used here the identity
\begin{equation}\label{derid}
\mathbf{1}\partial_{x}=-i\sigma_2(h-\Delta),
\end{equation}
which will be employed extensively in the following text. Notice that as long as $u_1$
and $u_2$ correspond to the same eigenvalue $\lambda_1=\lambda_2$, $\tilde{h}$ reduces
to a unitary transformed seed Hamiltonian, $\tilde{h}=\sigma_2h\sigma_2$, for any $\Delta$.
The Hamiltonians (\ref{seed}) and (\ref{htilde}) satisfy the following intertwining
relations mediated by $L$ and $L^{\dagger}$,
\begin{equation}\label{intertwining}
Lh=\tilde{h}L,\quad L^{\dagger}\tilde{h}=hL^{\dagger}.
\end{equation}
The conjugate operator $L^{\dagger}$ can be written as $L^{\dagger}=-\partial_x+V'V^{-1}$
where $V=(U^{\dagger})^{-1}=(v_1,v_2)$. The columns $v_1$ and $v_2$ satisfy
$L^{\dagger}v_{1(2)}=0$ and are solutions of $\tilde{h}v_{1(2)}=\lambda_{1(2)} v_{1(2)}$.
There holds
\begin{equation}\nonumber
\tilde{h}\, V=V\,\Lambda.
\end{equation}
Each of the equations $(h-\lambda)\varphi=0$ and $(\tilde{h}-\lambda)\tilde{\varphi}=0$
has two independent formal solutions, let us denote them $\psi_{\lambda}$, $\xi_{\lambda}$
and $\tilde{\psi}_{\lambda}$, $\tilde{\xi}_{\lambda}$, respectively.
For $\lambda\neq \lambda_{1(2)}$, the operators $L$ and $L^{\dagger}$ work as one-to-one
mappings between the two subspaces spanned by $\psi_{\lambda}$, $\xi_{\lambda}$ and
$\tilde{\psi}_{\lambda}$, $\tilde{\xi}_{\lambda}$. They transform the (formal) eigenvectors
of $h$ into the formal eigenvectors of $\tilde{h}$ and vice versa.
Let us consider now the four-dimensional subspace spanned by the solutions of
$(h-\lambda_{1(2)})\varphi=0$. Two of the solutions, the vectors $u_1$ and $u_2$, compose
the matrix $U$. We can use the other two vectors to define the matrix
$\underline{U}=(\underline{u}_1,\underline{u}_2)$, which satisfies
$h\underline{U}=\underline{U}\Lambda$ but is not annihilated by $L$. Similarly, we can define
the matrix $\underline{V}=(\underline{v}_1,\underline{v}_2)$ from the solutions of
$(\tilde{h}-\lambda_{1(2)})\tilde{\varphi}=0$ which satisfies
$\tilde{h}\underline{V}=\underline{V}\Lambda$, but no linear combination of $\underline{v}_1$
and $\underline{v}_2$ is annihilated by $L^{\dagger}$. The intertwining operators then
transform the matrices as $L\underline{U}\sim V$, $L^{\dagger}\underline{V}\sim U$.
Hence, we get $L^{\dagger}LU=L^{\dagger}L\underline{U}=LL^{\dagger}V=LL^{\dagger}\underline{V}=0$.
The latter equalities can be understood as the implication of the
alternative presentation for the products of the intertwining operators,
\begin{equation}\label{factorization}
LL^{\dagger}=(\tilde{h}-\lambda_1)(\tilde{h}-\lambda_2),\quad L^{\dagger}L=(h-\lambda_1)(h-\lambda_2).
\end{equation}
The spectrum of $\tilde{h}$ is identical with the spectrum of $h$ up to a possible difference
in the energy levels $\lambda_1$ and/or $\lambda_2$. These energies are in the spectrum of
either $h$ or $\tilde{h}$ if and only if the associated eigenvectors comply with the boundary
conditions of the corresponding stationary equation. We will discuss specific examples where
the spectrum of the new Hamiltonian $\tilde{h}$ contains additional discrete energies that are
absent in the spectrum of $h$.
The eigenvector $\tilde{\phi}_{k}$ of $\tilde{h}$ corresponding to the energy level
$\lambda_k\neq\lambda_1,\ \lambda_2$ can be expressed in terms of $L$ and the eigenvectors
$\phi_k$ of $h$ ($h\phi_{k}=\lambda_k\phi_k$),
\begin{equation}\label{normalization}
\tilde{\phi}_{k}=\frac{L\phi_{k}}{\sqrt{(\lambda_k-\lambda_1)(\lambda_k-\lambda_2)}}
,\quad \tilde{h}\tilde{\phi}_{k}=\lambda_k\tilde{\phi}_{k}.\nonumber
\end{equation}
When defined in this way, the probability densities of $\tilde{\phi}_k$ and $\phi_k$ coincide.
It is worth noticing that the system $\tilde{h}$ inherits integrals of motion of $h$. Indeed, if $S$
commutes with $h$, then the operator $\tilde{S}=LSL^{\dagger}$ generates a symmetry of $\tilde{h}$,
$[\tilde{h},\tilde{S}]=0$. We will discuss this point in more detail in the context of the
reflectionless models.
The potential term of $\tilde{h}$ in (\ref{htilde}) ceases to have a direct interpretation in the
context of carbon nanotubes with the radial twist. As we are interested in the analysis of namely
such systems, we require $\tilde{h}$ to be equivalent to the Hamiltonian in (\ref{eq1}); the
term proportional to either $\sigma_1$ or $\sigma_3$ in (\ref{htilde}) should vanish. As these
coefficients depend both on the potential of the seed Hamiltonian $h$ and on its eigenvectors,
it is rather difficult to meet this requirement in general. Instead, let us consider two special cases.
First, let us fix the initial Hamiltonian as
\begin{equation}\label{hI}h_{I}=i\sigma_2\partial_x+m\sigma_3+\Delta_1\sigma_1,\end{equation}
where $m>0$. We take $\lambda_1=m$ and $\lambda_2=0$ and denote $U_I\equiv U= (u_1,u_2)$ and
$V_I\equiv V=(v_1,v_2)$, where explicitly $u_1=(u_{11},0)^T$, $u_2=(u_{12},u_{22})^T$ and
\begin{equation} \label{UVI}
U_I=\left(\begin{array}{cc}u_{11}&{u_{12}}\\0&u_{22}\end{array}\right),\quad V_I=\left(\begin{array}{cc}\frac{1}{u_{11}}&0\\\frac{-u_{12}}{u_{11}u_{22}}& \frac{1}{u_{22}}\end{array}\right). \end{equation}
Comparison of the two matrices tells that if $u_1$ (or $u_2$) is a bound state of $h$,
then $v_1$ (or $v_2$) cannot be bound state of $\tilde{h}$. Vice versa, if $v_1$ (or $v_2$)
is a bound state of $\tilde{h}$, then $u_1$ (or $u_2$) is not normalizable. Using (\ref{htilde})
and $u^T_1\sigma_1u_2/\det U_I=1$, we get the Hamiltonian $\tilde{h}_I$,
\begin{equation}\label{HI}
\tilde{h}_I=i\sigma_2\partial_x-\left(\Delta_1+m\frac{u_{12}}{u_{22}}\right)\sigma_1,
\end{equation}
with the required form of the potential.
In the second case, we take the seed Hamiltonian as
\begin{equation}\label{hII}h_{II}=i\sigma_2\partial_x+(\Delta_1+m)\sigma_1\end{equation}
and fix $\lambda_1=-\lambda_2>0$. The vectors $u_{1(2)}$ are chosen as $u_1=(u_{11},u_{21})^T$
and $u_2=\sigma_3u_1$. They satisfy $u_1^T\sigma_1u_2=0$. The matrices $U_{II}\equiv U$ and
$V_{II}=V$ are in this case
\begin{equation}\label{UVII}
U_{II}=\left(\begin{array}{cc}u_{11}&{u_{11}}\\u_{21}&-u_{21}\end{array}\right),\quad V_{II}=\frac{1}{2}\left(\begin{array}{cc}{u_{11}^{-1}}&{u_{11}^{-1}}\\{u_{21}^{-1}}& -{u_{21}^{-1}}\end{array}\right) \end{equation}
and the Hamiltonian (\ref{htilde}) acquires the form
\begin{equation}\label{HII}
\tilde{h}_{II}=i\sigma_2\partial_x-\left(\Delta_1+m-\lambda_1\frac{u_{11}^2+u_{21}^2}{u_{11}u_{21}}\right)\sigma_1.
\end{equation}
We can deduce that if $u_1$ is a bound state of $h$, so is the vector $u_2$ and neither $v_1$
or $v_2$ can be normalized. Vice versa, if $v_1$ and $v_2$ are bound states of $\tilde{h}$,
the vectors $u_1$ and $u_2$ are not normalizable.
In the end of the section, let us notice that there is an alternative interpretation in
dealing with the intertwining relations and the involved operators. Inspired by the
SUSY quantum mechanics, we can define the extended, first order matrix operators
\begin{equation}
\mathcal{H}=\left(\begin{array}{cc}\tilde{h}&0\\0&h\end{array}\right),\quad \mathcal{Q}_1=\left(\begin{array}{cc}0&L\\L^{\dagger}&0\end{array}\right),\quad \mathcal{Q}_2=i\left(\begin{array}{cc}0&L\\-L^{\dagger}&0\end{array}\right),
\end{equation}
which establish the N = 2 (nonlinear) supersymmetry. The grading operator
$\Gamma=\mbox{diag}({\bf 1},-{\bf 1})$ classifies the Hamiltonian $\mathcal{H}$
as bosonic ($[\mathcal{H},\Gamma]=0$), while both $\mathcal{Q}_1$ and
$\mathcal{Q}_2=i\Gamma\mathcal{Q}_1$ are fermionic, $\{\mathcal{Q}_a,\Gamma\}=0$ for $a=1,2.$
Contrary to the SUSY quantum mechanics based on the second order matrix Hamiltonian,
here both the Hamiltonian $\mathcal{H}$ and the supercharges
$\mathcal{Q}_{1,2}$ are the first order differential operators. The associated
(nonlinear) superalgebra
\begin{equation}\label{susyextend}
[\mathcal{H},\mathcal{Q}_a]=0,\quad \{\mathcal{Q}_a,\mathcal{Q}_b\}=2\delta_{ab}(\mathcal{H}-\lambda_1)(\mathcal{H}-\lambda_2)
\end{equation}
encodes the intertwining relations (\ref{intertwining}) together with the factorization
(\ref{factorization}).
\section{Green's function and LDOS for the twisted nanotubes}
We shall derive formula for the Green's function of $\tilde{h}$ in terms of the intertwining
operator $L$ and the Green's function of the initial Hamiltonian $h$. In the end of the section,
we will discuss the explicit form of the LDOS for the systems described by $\tilde{h}_I$ and
$\tilde{h}_{II}$ of the form (\ref{HI}) and (\ref{HII}) corresponding to the twisted carbon nanotubes.
Let us start with the hermitian Hamiltonian $h=i\sigma_2\partial_x+\Delta(x)$. The potential term is
required to be real and symmetric. The (generalized) eigenstates $\phi_{\lambda}$ of $h$ have to
satisfy the following boundary conditions
\begin{equation}\label{heigen}
h\phi_{\lambda}=\lambda\phi_{\lambda},\quad \phi_{\lambda}(x)|_{x\rightarrow\pm\infty}\sim f_{\pm}(\lambda,x),\quad \lambda\in\mathbb{R}.
\end{equation}
The symbol $\sim$ means here that the elements of the eigenvector $\phi_{\lambda}$ are proportional
asymptotically to the function $f_{\pm}(x,\lambda)$. We prefer to leave the boundary conditions
unspecified explicitly at the moment.
They will be discussed for the reflectionless models later in the text.
The Green's function associated with the Hamiltonian $h$ is defined as a solution of the equation
\begin{equation}\label{Gdefiningeq}
(h-\lambda)G(x,y;\lambda)=\delta(x-y),\quad \lambda\in \mathbb{C}.
\end{equation}
It has to satisfy the same boundary conditions as the eigenstates of $h$, i.e. the matrix
elements of the Green's function are proportional to $f_{\pm}(\lambda,x)$ in the limit
${x\rightarrow \pm\infty}$. Being effectively the inverse of $(h-\lambda)$, the Green's
function is not well defined for $\lambda$ from the spectrum $\sigma(h)$ of $h$. It has
simple poles for $\lambda$ corresponding to discrete energies.
If $\lambda$ is in the continuous spectrum, then we can find the limit
$G^{\pm}(x,y;\lambda)=\lim_{\eta\rightarrow0}G(x,y;\lambda\pm i\eta)$, see e.g. \cite{economou}.
The differential equation in (\ref{heigen}) has two formal independent solutions
$\psi_{\lambda}(x)$ and $\xi_{\lambda}(x)$ for any $\lambda\in\mathbb{C}$.
For $\lambda\notin \sigma(h)$, we can fix $\psi_{\lambda}$ and $\xi_{\lambda}$ such that
each of the functions complies with the boundary condition in one of the boundaries; i.e.
we fix $\psi_{\lambda}(x)|_{x\rightarrow +\infty}\sim f_{+}(x,\lambda)$ and
$\xi_{\lambda}(x)|_{x\rightarrow -\infty}\sim f_{-}(x,\lambda)$.
These functions can be employed in the construction of the Green's function in the following way
\begin{equation}\label{GreenG}G(x,y;\lambda)=\frac{\psi_{\lambda}(x)\xi_{\lambda}(y)^T\theta(x-y)+\xi_{\lambda}(x)\psi_{\lambda}(y)^T\theta(y-x)}{W(\psi_{\lambda},\xi_{\lambda})},\end{equation}
where $\theta$ is the step function. The quantity
\begin{equation}
W(\psi,\xi)=i\psi(x)^T\sigma_2\xi(x)
\end{equation}
is the analog of Wronskian for Dirac equation. It is constant for two independent
solutions $\psi_{\lambda}$ and $\xi_{\lambda}$ corresponding to the eigenvalue
$\lambda$ of $h$. Indeed, direct calculation with the use of (\ref{derid}) shows that
$\partial_xW(\psi,\xi)=0$. The Green's function defined in (\ref{GreenG}) then solves
(\ref{Gdefiningeq}) and manifestly satisfies the prescribed boundary conditions for $x\rightarrow\pm\infty$.
Let us pass to the system described by $\tilde{h}$ and construct its Green's function with
the use of (\ref{GreenG}). We suppose that $L$ transforms appropriately the boundary
conditions associated with $h$ to the boundary conditions prescribed for the eigenstates
of $\tilde{h}$. We can define the functions
$\tilde{\psi}_{\lambda}=\frac{L\psi_{\lambda}}{\sqrt{(\lambda-\lambda_1)(\lambda-\lambda_2)}}$ and
$\tilde{\xi}_{\lambda}=\frac{L\xi_{\lambda}}{\sqrt{(\lambda-\lambda_1)(\lambda-\lambda_2)}}$.
They solve $\tilde{h}\tilde{\psi}_{\lambda}=\lambda\tilde{\psi}_{\lambda}$,
$\tilde{h}\tilde{\xi}_{\lambda}=\lambda\tilde{\xi}_{\lambda}$
and satisfy the prescribed boundary condition in $+\infty$ or $-\infty$, respectively.
The Green's function associated with $\tilde{h}$ can be written then as
\begin{eqnarray}\label{GreentildeG}\tilde{G}(x,y;\lambda)&=&\frac{\tilde{\psi}_{\lambda}(x)\tilde{\xi}_{\lambda}(y)^T\theta(x-y)+\tilde{\xi}_{\lambda}(x)\tilde{\psi}_{\lambda}(y)^T\theta(y-x)}{W(\tilde{\psi}_{\lambda},\tilde{\xi}_{\lambda})}\nonumber\\
&=&\frac{1}{W(\psi_{\lambda},\xi_{\lambda})}\left[\frac{(L\psi_{\lambda})(x)(L\xi_{\lambda})^T(y)\theta(x-y)}{(\lambda-\lambda_1)(\lambda-\lambda_2)}\right.\nonumber\\
&&\left.+\frac{(L\xi_{\lambda})(x)(L\psi_{\lambda})^T(y)\theta(y-x)}{(\lambda-\lambda_1)(\lambda-\lambda_2)}\right].
\end{eqnarray}
We used the fact that the Wronskian is invariant with respect to the Darboux transformation
(\ref{L}), $W(\tilde{\psi}_{\lambda},\tilde{\xi}_{\lambda})=W(\psi_{\lambda},\xi_{\lambda})$.
We refer to \cite{Dunne} or \cite{halberg} where the proof of this relation can be found.
Let us mention that the a different supersymmetric approach to Green's
functions of Dirac operators was examined in \cite{Feinbergsusy} where a modification of the
standard supersymmetry (based on second-order Hamiltonians) was discussed.
The eigenvectors of $\tilde{h}$ can be written as
\begin{equation}\label{algebraicL}
\tilde{\psi}_{\lambda}={\cal L}(\lambda,x)\psi_{\lambda},\quad {\cal L}(\lambda,x)=-i\sigma_2\frac{\lambda-U(x)\Lambda U^{-1}(x)}{\sqrt{(\lambda-\lambda_1)(\lambda-\lambda_2)}},
\end{equation}
where we used (\ref{derid}) again.
This allows us to rewrite the Green's
function (\ref{GreentildeG}) in particularly simple form
\begin{equation}\label{LGL}
\tilde{G}(x,y;\lambda)={\mathcal L}(\lambda,x)G(x,y;\lambda){\mathcal L}^T(\lambda,y).
\end{equation}
Hence, $\tilde{G}(x,y;\lambda)$ can be obtained by purely algebraic means without the use of
any differential operator; it can be obtained just by multiplication of $G(x,y;\lambda)$ with
simple matrix operators (\ref{algebraicL}).
The local density of states $\rho(x,\lambda)$ associated with $h$ is computed in the following manner
\begin{eqnarray}\label{LDOS}
\rho(x,\lambda)&=&-\frac{1}{\pi}\lim_{\mbox{Im}\lambda\rightarrow 0_+}\mbox{Im}\,Tr \,G(x,x;\lambda),
\end{eqnarray}
where the trace is taken over the matrix degrees of freedom. Using (\ref{LGL}), we can write
LDOS $\tilde{\rho}$ for $\tilde{h}$ as
\begin{equation}\label{tildeLDOS}
\tilde{\rho}(x,\lambda)=-\frac{1}{\pi}\lim_{\mbox{Im}\lambda\rightarrow 0_+}\mbox{Im}\, Tr\left(\mathcal{L}(\lambda,x)^T\mathcal{L}(\lambda,x)\,G(x,x;\lambda)\right).
\end{equation}
Notice that the formulas (\ref{LGL}) and (\ref{tildeLDOS}) are valid for a general class of
the seed Hamiltonians with real and symmetric potential.
In the literature (see, e.g. \cite{Dunne}, \cite{Feinberg}, \cite{Kos}), the operator
$G(x,x;\lambda)$ is called Gorkov Green's function
or diagonal resolvent of $h$. The
Green's function of the Schr\"odinger operators and generalized Sturm-Liouville equation was
studied in \cite{greenSamsonov} and \cite{halberg2} in the context of intertwining relations.
We turn our attention to the systems represented by $\tilde{h}_I$ and $\tilde{h}_{II}$ which
describe the carbon nanotubes with the radial twist. It is supposed that the Green's functions
of both $h_I$ and $h_{II}$ are known. We denote them $G_I(x,y;\lambda)$ and $G_{II}(x,y;\lambda)$.
The operators $\mathcal{L}_I(\lambda,x)$ and $\mathcal{L}_{II}(\lambda,x)$ based on $U_{I}$ and
$U_{II}$ respectively acquire particularly simple form
\begin{equation}\nonumber
\mathcal{L}_I(\lambda,x)= \frac{1}{\sqrt{\lambda(\lambda-m)}}\left(\begin{array}{cc}0&-\lambda\\-m+\lambda&m\frac{u_{12}}{u_{22}}\end{array}\right)
\end{equation}
and
\begin{equation}\nonumber
\mathcal{L}_{II}(\lambda,x)= \frac{1}{\sqrt{(\lambda^2-\lambda_1^2)}}\left(\begin{array}{cc}\lambda_1\frac{u_{21}}{u_{11}}&-\lambda\\\lambda&-\lambda_1\frac{u_{11}}{u_{21}}\end{array}\right).
\end{equation}
The trace of the $\tilde{G}_{I}$ can be computed directly in terms of the vectors $u_1$ and $u_2$.
A straightforward computation gives
\begin{eqnarray}
\mbox{Tr}(\tilde{G}_{I}(x,x;\lambda))&=&g_0-\frac{m}{\lambda-m}g_3+\frac{m^2\,g_0(u_1^{\dagger}u_1)(u_2^{\dagger}u_2)}{2\lambda(\lambda-m)(\det U_I)^2}\nonumber
\end{eqnarray}
\begin{equation}\label{GGI}
+\frac{m^2u_1^{\dagger}u_1}{2\lambda(\lambda-m)(\det U_I)^2}\left(-g_3u_2^{\dagger}\sigma_3u_2+g_1\frac{\lambda-m}{m}u_2^{\dagger}\sigma_1u_2\right).
\end{equation}
Here we used the abbreviated notation $g_0=\mbox{Tr}\,G_I(x,x;\lambda)$ and
$g_j=\mbox{Tr}(\sigma_jG_I(x,x;\lambda))$ for $j=1,3$.
We can obtain similar expression for the trace of the $\tilde{G}_{II}(x,x;\lambda)$:
\begin{eqnarray}
\mbox{Tr}(\tilde{G}_{II}(x,x;\lambda))&=&g_0+\frac{2\lambda_1^2g_0(u_1^{\dagger}u_1)^2}{(\lambda^2-\lambda_1^2)(\det U_{II})^2}\nonumber
\end{eqnarray}
\begin{equation}\label{GGII}
+\frac{2\lambda_1^2u_1^{\dagger}u_1}{(\lambda^2-\lambda_1^2)(\det U_{II})^2}\left(-g_3u_1^{\dagger}\sigma_3u_1-g_1\frac{\lambda}{\lambda_1}u_1^{\dagger}\sigma_1u_1\right).
\end{equation}
The notation used here is like in (\ref{GGI}) with the replacement of
${G}_{I}(x,x;\lambda)$ by ${G}_{II}(x,x;\lambda)$.
\section{\label{reflectionless}Perfect tunneling in the twisted carbon nanotubes}
There exists an exceptional class of exactly solvable systems whose Hamiltonian
$\tilde{h}$ is intertwined with the Hamiltonian of the free particle. The peculiar
and simple properties of the latter model are manifested in these systems as well.
In particular, they share the trivial scattering characteristics of the interaction-free
model, i.e. they are \textit{reflectionless}. The eigenstates of both the free-particle
system and the reflectionless models are
subject to the same boundary conditions; the scattering states have to be oscillating
in the infinity while the bound states should decay exponentially for $|x|\rightarrow\infty$.
Additionally, the reflectionless systems inherit the integral
of motion that in the free particle system plays
the role of generator of translations.
The stationary equation $h\phi=\lambda\phi$, where
$h=i\sigma_2\partial_x+m\sigma_3$,
is translationally invariant, i.e. the Hamiltonian commutes with $p=-i\partial_x$.
We can find the common eigenstates of $h$ and $p$.
The latter operator distinguishes the two scattering states corresponding to each
doubly degenerate energy level. It annihilates the singlet states $u_{+}=(1,0)^T$
and $u_-=(0,1)^T$ that correspond to the edges $\lambda=\pm m$ of the positive and
negative part of the continuous spectrum (which are called the conduction and the
valence band respectively in the context of nanotubes). The involved operators close
the nonlinear superalgebra
\begin{equation}\label{freehiddensusy}
[p,h]=0,\quad \{p,p\}=2(h-m)(h+m),
\end{equation}
which is graded by the parity operator $\Gamma=R\sigma_3$ ($RxR=-x$, $\Gamma^2=1$).
Let us stress that this supersymmetric structure is completely different from (\ref{susyextend}).
In this case, the supersymmetry is rather hidden; the two fold degeneracy of energy levels,
distinguished by the integral of motion $p$, emerges within the spectrum of the unextended
Hamiltonian $h$.
The Hamiltonian $\tilde{h}$ inherits a modified version of the nontrivial integral of
motion $p$. It can be found by dressing of the initial symmetry operator,
\begin{equation}\label{A}
\tilde{p}=L\, p\,L^{\dagger},\quad [\tilde{p},\tilde{h}]=0.
\end{equation}
It annihilates the states $v_1$ and $v_2$ together with the vectors $\tilde{v}_{\pm}$
which are defined as $\tilde{v}_{\pm}=Lu_{\pm}$. The operator $\tilde{p}$, like
$p$ in the free particle model, reflects the degeneracy
of the spectrum; it can distinguish the scattering states corresponding to the same energy
level.
The superalgebra (\ref{freehiddensusy}) can be recovered in the modified form
\begin{equation}\label{dressedhiddensusy}
[\tilde{p},\tilde{h}]=0,\quad \{\tilde{p},\tilde{p}\}=2(\tilde{h}^2-m^2)(\tilde{h}-\lambda_1)^2(\tilde{h}-\lambda_2)^2.
\end{equation}
Hence, the square of $\tilde{p}$ is the spectral polynomial of $\tilde{h}$.
It is worth noticing that the same algebraic structure, the hidden supersymmetry,
was discussed in detail for both relativistic and nonrelativistic finite-gap systems
in \cite{BdG}, \cite{hiddensusy}, \cite{mirror}, \cite{AdS2}. In this context, the
integral $\tilde{p}$ can be identified as the Lax operator of the system represented by $\tilde{h}$.
\subsection{Single-kink system }
The first model will be derived with the use of the seed Hamiltonian $h_I$ in (\ref{hI})
with $\Delta_1(x)=0$,
\begin{equation}\nonumber
h_I=i\sigma_2\partial_x+m\sigma_3.
\end{equation}
We will compute its LDOS and discuss the realization of the parity operator of the hidden supersymmetry.
We require that the new Hamiltonian $\tilde{h}$ has a single bound state with zero energy.
To meet this requirement, we fix the matrix $U_I$ as
\begin{equation}\nonumber
U_I=\sqrt\frac{2}{m}\left(\begin{array}{cc}1&-\sinh mx\\0&\cosh mx\end{array}\right),
\end{equation}
and the intertwining operator $L$ as
\begin{equation}\label{LI}
L_I=\mathbf{1}\,\partial_x+m\left(\begin{array}{cc}0&1\\0&-\tanh mx\end{array}\right).
\end{equation}
The explicit form of the matrix $V_I$ is then
\begin{equation}V_I=\sqrt{\frac{m}{2}}\left(\begin{array}{cc}1&0\\\tanh\, mx&\mbox{sech}\, mx\end{array}\right). \nonumber \end{equation}
The associated Hamiltonian $\tilde{h}_I$ then reads
\begin{equation}\label{hIex}
\tilde{h}_I=i\sigma_2\partial_x+m\,\sigma_1\,\tanh mx.
\end{equation}
The operator $\tilde{h}_I$ has the normalized bound state $v_2$
\begin{equation}\nonumber
v_2=\left(0,\sqrt{\frac{m}{2}}\,\mbox{sech}\, mx\right)^T.
\end{equation}
Let us notice that the operator (\ref{hIex}) appears in description of many physical systems,
e.g. in the continuum model for solitons in polyacetylene \cite{Takayama} or in the analysis
of the static fermionic bags of the Gross-Neveu model \cite{Feinberg}.
We can use (\ref{LDOS}) together with (\ref{GGI}) to compute the LDOS of the system.
It acquires the following simple form
\begin{equation}\label{exILDOS}
\tilde{\rho}_I(x,\lambda)=\frac{2|\lambda|^2-m^2\mbox{sech}^2 mx}{2\pi|\lambda|\sqrt{|\lambda^2-m^2|}}\theta(\lambda^2-m^2).
\end{equation}
The presence of the step function $\theta$ reflects that fact that imaginary part of
(\ref{GGI}) for $|\lambda|<m$ is zero and, hence, $\rho(x,\lambda)$ vanishes identically.
The formula (\ref{exILDOS}) can be rewritten with the use of the LDOS of the free
particle \begin{equation}\nonumber\rho_I(x,\lambda)=\frac{|\lambda|}{\pi\sqrt{|\lambda^2-m^2|}}\,\theta(\lambda^2-m^2)\end{equation}
and the density of probability of the bound state $v_2$,
\begin{equation}\label{LDOSI}
\tilde{\rho}_{I}(x,\lambda)=\rho_I(x,\lambda)\left(1-\frac{m}{|\lambda|^2}v_2^{\dagger}v_2\right).
\end{equation}
The coefficient of the second term is just the difference of the densities of states of $h$ and $\tilde{h}$,
\begin{equation}\nonumber
\int_{\mathbb{R}}(\rho_I-\tilde{\rho}_I)dx=\frac{m\,\theta(\lambda^2-m^2)}{\pi|\lambda|\sqrt{|\lambda^2-m^2|}}.
\end{equation}
Let us notice that the difference of densities of states for Dirac particle on the \textit{finite} interval
with Dirichlet boundary conditions was discussed in \cite{halberg}.
The hidden superalgebra (\ref{dressedhiddensusy}), closed by $\tilde{h}_I$ and $\tilde{p}_I=L_IpL^{\dagger}_I$,
reads explicitly
\begin{equation}\nonumber
[\tilde{h}_I,\tilde{p}_I]=0,\quad \{\tilde{p}_I,\tilde{p}_I\}=2(\tilde{h}_I-m)^3\tilde{h}_I^2(\tilde{h}_I+m).
\end{equation}
The parity operator $\tilde{\Gamma}=R\, \sigma_3$, $\tilde{\Gamma}^2=1$,
classifies $\tilde{h}_I$ and $\tilde{p}_I$ as, respectively, bosonic and fermionic operators,
$[\tilde{h}_I,\tilde{\Gamma}]=\{\tilde{p}_I,\tilde{\Gamma}\}=0$.
The potential in (\ref{hIex}) can be associated with the displacement vector
$\mathbf{d}=(0,\ln\cosh mx).$
The corresponding twist of the metallic nanotube is illustrated in Figure \ref{bubu}.
Hence, the nanotube is twisted in one direction up to the center (origin)
where the orientation of the twist gets changed.
\newsavebox{\figPT}
\savebox{\figPT}{
\scalebox{1}{
\includegraphics[scale=.7]{figure2.eps}
}
}
\begin{figure}\begin{center}
\usebox{\figPT}\caption{\label{bubu}The metallic nanotube with the twist associated with $ d_y\sim\ln\cosh mx$
and the Hamiltonian (\ref{hIex}). In the untwisted nanotube, the black line would be straight.}
\end{center}\end{figure}
\subsection{Double-kink model}
Here we construct the system with two bound states. We shall employ the scheme
discussed in (\ref{hII})-(\ref{HII}). Fixing $\Delta_1(x)=0$ in (\ref{hII}), we get
the free particle Hamiltonian
\begin{equation}\nonumber
h_{II}=i\sigma_2\partial_x+m\sigma_1.
\end{equation}
We choose the components of $U_{II}$ as
\begin{equation}\nonumber
u_{11}=\frac{1}{\sqrt{k}}\cosh k x,\quad u_{21}=\frac{1}{\sqrt{k}}\cosh (k x+a)
\end{equation}
where
\begin{equation}\nonumber
a=\frac{1}{2}\log\frac{m-k}{m+k}, \quad k=\sqrt{m^2-\lambda_1^2},\quad 0<\lambda_1<m.
\end{equation}
The intertwining operator acquires a diagonal form
\begin{equation}\label{Lreflectionless}
L_{II}=\mathbf{1}\,\partial_x-k\left(\begin{array}{cc}\tanh(kx)&0\\0&\tanh (kx+a)\end{array}\right).\end{equation}
The formula (\ref{HII}) then provides the explicit form of the Hamiltonian $\tilde{h}_{II}$
\begin{eqnarray}\label{hIIex}
\tilde{h}_{II}&=&i\sigma_2\partial_x+\left(-m+\lambda_1\frac{\cosh^2 kx+\cosh^2(kx+a)}{\cosh kx\cosh(kx+a)}\right)\sigma_1\nonumber\\
&=&i\sigma_2\partial_x+(m-k\tanh k x+k\tanh(k x +a))\sigma_1.
\end{eqnarray}
The potential term is asymptotically equal to $m\sigma_1$. The system has two bound states
represented by the normalized vectors $v_1$ and $v_2=\sigma_3\,v_1$ where
\begin{equation}\nonumber
v_1=\frac{\sqrt{k}}{2}(\mbox{sech} kx,\mbox{sech}(kx+a))^T.
\end{equation}
Notice that the equation (\ref{hIIex}) appeared in the analysis of the Dashen, Hasslacher, and Neveu
kink-antikink baryons in Gross-Neveu model \cite{kink-antikink}.
The local density of states in the current system can be computed directly with the
use of (\ref{GGII}). We get
\begin{eqnarray}\nonumber
\tilde{\rho}_{II}(x,\lambda)&=&\frac{|\lambda|\left(1-\frac{k^2}{2(\lambda^2-\lambda_1^2)}(\mbox{sech}^2kx+\mbox{sech}^2(kx+a))\right)}{\pi\sqrt{|m^2-\lambda^2|}}\nonumber\\&&\times\theta(\lambda^2-m^2).
\end{eqnarray}
Likewise in the preceding example, it can be written as the LDOS of the free particle corrected
by the term proportional to the probability density of the bound states,
\begin{equation}\label{LDOSII}
\tilde{\rho}_{II}(x,\lambda)=\rho_{II}(x,\lambda)\left(1-\frac{2\,k\,v_1^{\dagger}v_1}{(\lambda^2-\lambda_1^2)}\right),
\end{equation}
where $\rho_{II}(x,\lambda)=\rho_{I}(x,\lambda)$.
This time, the difference of the densities of states is
\begin{equation}\nonumber
\Delta DOS=\int_{\mathbb{R}}(\rho_0-\rho_1)dx=\frac{2k\,|\lambda|\theta(\lambda^2-m^2)}{\pi\sqrt{|m^2-\lambda^2|}(\lambda^2-\lambda_1^2)}.
\end{equation}
The hidden superalgebra associated with the system,
\begin{equation}
[\tilde{h}_{II},\tilde{p}_{II}]=0,\quad
\{\tilde{p}_{II},\tilde{p}_{II}\}=2(\tilde{h}_{II}^2-m^2)(\tilde{h}_{II}^2-\lambda_1^2)^2,\nonumber\end{equation}
is graded by the operator
$\tilde{\Gamma}=R\, R_{\alpha} \sigma_1$ where $ R_{\alpha}f(x)=f(x+\alpha)R_{\alpha}
$, $R_{\alpha}R=RR_{-\alpha}=R(R_{\alpha})^{-1}$ and $\alpha=-\frac{a}{k}$.
This grading operator (represented in another form) was also discussed
in \cite{mirror}.
The vector potential in (\ref{hIIex}) corresponds to the displacement
$
d_y=mx-\ln\cosh kx+\ln\cosh(kx+a).
$
The corresponding twist of the metallic nanotube does not change its orientation asymptotically,
see Figure \ref{fig3} for illustration.
\newsavebox{\figdva}
\savebox{\figdva}{
\scalebox{1}{
\includegraphics[scale=.65]{figure3.eps}
}
}
\begin{figure}\begin{center}
\usebox{\figdva}\caption{\label{fig3}The metallic nanotube with the twist associated with
$ d_y\sim mx+\ln\frac{\cosh(kx+a)}{\cosh kx}$
and the Hamiltonian (\ref{hIIex}).}
\end{center}\end{figure}
We can find another physically interesting setting described by $\tilde{h}_{II}$.
We can divide the vector potential in (\ref{hIIex}) into two parts. The
first part is associated with the asymptotically vanishing twist of the nanotube,
$\tilde{\Delta}_T=-k\tanh k x+k\tanh(k x +a)$. The second part is constant,
$\tilde{\Delta}_{MG}=m$, and corresponds to the homogeneous external
magnetic field which is parallel with the axis of the nanotube. Hence, $\tilde{h}_{II}$
describes the metallic nanotube which is asymptotically free of twists, however, the external
constant magnetic field is present. See Figure \ref{fig4} for illustration.
\newsavebox{\figctyri}
\savebox{\figctyri}{
\scalebox{1}{
\includegraphics[scale=.65]{figure4.eps}
}
}
\begin{figure}\begin{center}
\usebox{\figctyri}\caption{\label{fig4}The nanotube associated with the Hamiltonian
(\ref{hIIex}) and the twist corresponding to (\ref{twistII}).
The constant part of the magnetic field in (\ref{hIIex}), $\tilde{\Delta}_{MG}=m$, can be attributed to
the external magnetic field or to the semi-conducting character of the nanotube.}
\end{center}\end{figure}
The uniform external field opens a gap of the width $2m$ in the spectrum while the
asymptotically vanishing twist induces two bound states in the gap. The model allows
to compute the bound state energies as a function of an asymptotic (global) twist.
Indeed, the twist associated with $\tilde{\Delta}_T $ is
\begin{equation}\label{twistII}
d_y= \ln\frac{\cosh(kx+a)}{\cosh kx}.
\end{equation}
The asymptotic twist corresponds to
\begin{equation}\label{asymptotictwist}
\delta d=|\lim_{x\rightarrow\infty}d_y-\lim_{x\rightarrow-\infty}d_y\,|=2|a|=-\ln\frac{m-\sqrt{m^2-\lambda_1^2}}{m+\sqrt{m^2-\lambda_1^2}}.
\end{equation}
The dependence of the bound state energies on $\delta d$ then acquires the following simple form
\begin{equation}\label{induced}
\lambda_1=\pm 2m \frac{e^{\frac{\delta d}{2}}}{1+e^{\delta d}}
\end{equation}
and is plotted in Figure \ref{fig5}.
\newsavebox{\figpet}
\savebox{\figpet}{
\scalebox{1}{
\includegraphics[scale=1.15]{figure5.eps}
}
}
\begin{figure}\begin{center}
\usebox{\figpet}\caption{\label{fig5}The spectrum of the Hamiltonian (\ref{hIIex}).
The asymptotic twist of the nanotube (\ref{asymptotictwist}) induces bound states of
energies (\ref{induced}). The parameter $m$ is proportional to the inverse of the radius
of the nanotube, see \cite{scale}.}
\end{center}\end{figure}
Up to now, the twisted nanotubes were considered to be metallic. The analysis can be
extended to semi-conducting nanotubes without any difficulties; a constant, nonzero,
part of the vector potential $\tilde{\Delta}_1$ has to be associated with the internal
characteristics (the orientation of the hexagonal lattice) of the nanotube.
Let us notice in this context that the metallic nanotube can be converted into the
semi-conducting one just by switching on the constant magnetic flux parallel to the
axis of the nanotube. This fact was experimentally confirmed in \cite{ABoscillations}
and coined as Aharonov-Bohm oscillations of the carbon nanotubes. Turning back to (\ref{hIIex}),
we can interpret the Hamiltonian as the energy operator of the semiconducting nanotube
with a radial twist associated to $\tilde{\Delta}_{T}$. The constant part
$p_y\equiv\tilde{\Delta}_{MG}$ of the potential appears due to the semiconducting nature of the nanotube.
\section{Discussion}
The expressions (\ref{LDOSI}) and (\ref{LDOSII}) can be written in the unified form
\begin{equation}\label{obecny}\nonumber
\tilde{\rho}_{I(II)}=\rho_{I(II)}\left(1-\sum_j\frac{\sqrt{m^2-\lambda_j^2}}{(\lambda^2-\lambda_j^2)}v_j^{\dagger}v_j\right),
\end{equation}
where the sum is taken over the \textit{normalized} bound states of $\tilde{h}_{I(II)}$
annihilated by $L_{I(II)}$. It manifests a decrease of the LDOS in the regions where the
bound states are localized. However, it is rather just a peculiar property of the
discussed reflectionless models \cite{feinbergformule}. In general case, the LDOS (\ref{GGII}) of $\tilde{h}_{II}$
acquires the following form in terms of the vectors $v_1$ and $v_2$,
\begin{eqnarray}
\mbox{Tr}(\tilde{G}_{II}(x,x;\lambda))&=&\mbox{Tr}({G}_{II}(x,x;\lambda))\nonumber
\end{eqnarray}
\begin{equation}
+\frac{2\lambda_1^2v_1^{\dagger}v_1\left(g_0v_1^{\dagger}v_1+g_3v_1^{\dagger}\sigma_3v_1-g_1\frac{\lambda}{\lambda_1}v_1^{\dagger}\sigma_1v_1\right)}{(\lambda^2-\lambda_1^2)(\det V)^2}.\nonumber
\end{equation}
When $h_{II}$ is equal to the free particle Hamiltonian, the coefficient of $v_1^{\dagger}v_1$
reduces to a constant. Nevertheless, this apparently does not hold in the general case.
We restricted our consideration just to the systems described by the
Hamiltonian $\tilde{h}=i\sigma_2\partial_x+\Delta_1(x)\sigma_1$. However, the potential
term of the seed Hamiltonian $h$ in (\ref{seed}) can acquire quite generic form, yet
keeping valid the formulas (\ref{LGL}) and (\ref{tildeLDOS}) for the Green's functions and
for the LDOS. Other results are more sensitive to the explicit form of the potential.
The term $\Delta_2(x)\sigma_2$ cannot cause any substantial modifications; it
would play just the role of non-physical gauge field. In contrary, impact of the diagonal term
$\mathbf{1}\Delta_0+\Delta_3(x)\sigma_3$ in $\tilde{h}$ would be much deeper: in general, it would break the
symmetry $\sigma(\tilde{h})=-\sigma(\tilde{h})$ of the spectrum $\sigma(\tilde{h})$
of $\tilde{h}$. In the context of Dirac particles in the carbon nanotubes, the potential $\Delta_3(x)\sigma_3$ would
have different sign for the spin -up and -down components of wave function, i.e. this potential would change
the sign on the two sublattices that form the crystal. Physical realization of such a scenario
in the considered condensed matter system is not clear. It is remarkable that the
Darboux transformation (\ref{L}) does not alter the
form of $\Delta_0$; the new Hamiltonian $\tilde{h}$ shares the same electrostatic potential
as the seed Hamiltonian. We notice that the electrostatic potential can be also
altered via the so-called 0-th order supersymmetry, as it was discussed in \cite{nasKlein}.
In the discussed systems represented by the stationary equation (\ref{eq1}),
the analysis of the bound states can be facilitated by the fact that the square of the Dirac
Hamiltonian takes the form $-\partial_x^2+\Delta_1^2+\sigma_3\Delta_1'$. The existing tools
(see, e.g. \cite{Schubin}) for the analysis of the Schr\"odinger operators can be exploited to
reveal spectral properties of the Dirac Hamiltonian. In this context, let us mention that
interesting results were obtained by the spectral analysis of general class of deformed quantum
waveguides described by Schr\"odinger equation \cite{quantumwaveguides}. We believe that
similar analysis for the carbon nanostructures described by the one- or two-dimensional Dirac
Hamiltonian would be fruitful.
The presented analysis is qualitative. The equation (\ref{eq1}) is a good approximation for
the quasiparticles in carbon nanotubes only for small region of the momentum space where the linear
dispersion relation is valid. When the gap opened by the pseudo-magnetic field in the spectrum
is too big, nonlinear (the so-called trigonal warping) terms \cite{trigonal} have to be included
into the Hamiltonian.
In the article, we neglected surface curvature of the nanotubes. The tubular surface prevents
the $\pi$-orbitals of the carbon atoms to be parallel to each other. This implies presence of
additional pseudo-magnetic fields in the Hamiltonian. However, in case of armchair nanotubes,
these additional gauge fields can be transformed out \cite{KaneMele}.
The examples presented in the text suggest that the non-uniform radial twist can induce bound
states in the nanotube.
In particular, the second model with double-kink potential provides an interesting qualitative
insight into realistic experimental setting: the nanotube with asymptotically vanishing twist
is immersed into the homogeneous magnetic field. Besides the explicit formula (\ref{LDOSII})
for LDOS, the model predicts the appearance of bound states and the formula (\ref{induced})
fixes their energies in dependence on the asymptotic twist. The model can be simply tuned
with the use of perturbation techniques.
The supersymmetry can be very useful for construction of the models with more complicated
(yet asymptotically constant) twist inducing richer spectral properties.
The formalism presented in the second section can be repeated to produce a chain of solvable
Hamiltonians, $h$, $\tilde{h}$, $\tilde{\tilde{h}},...$, by taking the last constructed operator
as the seed Hamiltonian for the new system. These new solvable systems shall provide insight
into the \textit{deformation-induced spectral engineering} of carbon nanotubes.
The reflectionless models are particularly important in this context; they are analytically
feasible and possess nontrivial (super)symmetry, analog of (\ref{A}) and (\ref{dressedhiddensusy}).
The considered double-kink example suggests that the number of bound states could be in a simple
relation to the vector potential of the Hamiltonian; the number of bound states might be
proportional to the number of minima of the potential. Verification of this hypothesis goes
beyond the scope of the current paper and should be discussed elsewhere.
{\bf Acknowledgements:}
The work has been partially supported
by FONDECYT Grant No. 1095027, Chile,
and by the GA\v CR Grant P203/11/P038, Czech Republic.
VJ thanks the Department of Physics of the Universidad de
Santiago de Chile for hospitality.
|
\section{Introduction}
Light elements are important tracers of stellar internal mixing and rotation.
Since they are burned at relatively low temperatures they constrain
how the material inside stars is mixed with the hotter
interior. Rotation and angular momentum loss are among the leading processes to
explain the mixing that leads to depletion of light elements in
solar-type stars \citep[e.g.][]{stephens,Bouvier} although gravitational waves
may also affect the abundances of those elements \citep{Montalban}. However,
available models for evolution of Be (considering rotation) do not predict a
significant depletion of Be during the main sequence for stars with 6000 K $>$
T$_{\rm eff}$ $>$ 4000 K \citep{Pinsonneault}. On the other hand, models which take
into account gravitational waves predict significant Be depletion only for stars
cooler than 4500 K \citep{Montalban}.\\
In a recent work, \citet{israelian09} confirmed that Li was severely depleted in
solar-type stars (with T$_{\rm eff}$ between 5600 K and 5850 K) with planets
when compared with similar stars without detected planets although this result
is controversial \citep{Baumann,sousa10}; for a complete discussion see
\citet{Delgado11}. This difference in Li abundance seems to be related to the
different rotational history of both groups of stars due to the presence of
protoplanetary disks \citep[e.g.][]{Bouvier}. However, beryllium needs
a greater temperature to be burned so we would expect to see the onset
of this effect in cooler stars where convective envelopes are deep enough to
reach those higher temperatures.\\
In a previous paper, \citet{Delgado11} found two cool planet host stars
with an extra depletion of Be when compared with analog stars without detected
planets. This encourages us to try to investigate this process in cool stars. In
this work we continue that analysis by extending the sample with 15 new cool
stars. We refer the reader to that paper for further information and a more
extensive introduction.\\
\begin{figure*}[ht!]
\centering
\includegraphics[width=16cm]{fig1.eps}
\caption{Synthetic spectra (red lines) and observed spectra (dots) for the
planet host star HD 93083 and the stars without detected planets HD 213042, HD
35854, HD 15337, HD 8389A and HD 21019. The position of Be lines are indicated
by the arrows.}
\label{ajustes_be}
\end{figure*}
\begin{figure*}[ht!]
\centering
\includegraphics[width=8cm]{fig2.eps}
\includegraphics[width=8cm]{fig3.eps}
\includegraphics[width=8cm]{fig4.eps}
\includegraphics[width=8cm]{fig5.eps}
\caption{Panel a: Spectral synthesis of HD 213042 using different line lists. Panel b: Spectral synthesis of HD 16270 showing
the contributions of atomic, molecular and Be lines. Panel c: Spectral synthesis of HD 63454 with solar line list and different
values of Be abundance. Panel d: Spectral synthesis of the star R1 in the young cluster IC 2602 using different line lists.}
\label{be_test}
\end{figure*}
\section{Observations and spectral synthesis}
In this study we obtained high resolution spectra for 15 new stars with
magnitudes V between 6 and 10 using the UVES spectrograph at the 8.2-m Kueyen
VLT (UT2) telescope (run ID 86.D-0082A) between October 2010 and March 2011. The
dichroic mirror was used to obtain also red spectra and the slit width was 0.5
arcsec. These new spectra have a spectral resolution \textit{R} $\sim$ 70000 and
\textit{S/N} ratios between 100 and 200. All the data were reduced with the
pipeline of \textit{UVES/VLT}. Standard background correction, flat-field and
extraction procedures were used. The wavelength calibration was made using a
ThAr lamp spectrum taken during the same night. Finally we manually normalized
the continuum by dividing the spectra by a spline function with three
pieces (and the parameter $low-reject$ set to 1) in the whole blue region (3040\AA{}-3800\AA{}).
This normalization does not present any bias for stars with and without planets. When plotting
together observed spectra of stars with and without planets we only had to multiply the flux of
comparison stars by 0.9-1.1 in order to make them match up. We note that in our previous works we have analyzed high metallicity (up to 0.4) solar type stars with Teff down to 5300 K
for which we could make a good normalization. The main problem is the normalization of spectra of stars with Teff less
than $\sim$ 5300 K. These stars could have spots and inhomogeneous atmospheres that make the spectral
synthesis more difficult.\\
The uniform stellar atmospheric parameters were taken from \citet{sousa08} with
typical errors of 25 K for $T_{\rm eff}$, 0.04 dex for $\log g$, 0.03 km s$^{\rm
-1}$ for $\xi_{t}$ and 0.02 dex for metallicity. We refer to that work for
further details in these parameters and their uncertainties.\\
Be abundances were derived by fitting the spectral region around the
\ion{Be}{2} line at 3131.06 \AA{} and then using the \ion{Be}{2} line 3130.42
\AA{} to check the consistency of the fit. We used an empirical line list from
\citet{Garcia-Lopez95} tuned to reproduce the solar spectrum (see Table \ref{lista}).
These synthetic spectra were convolved with a rotational profile. We made a standard LTE
analysis with the revised version of the spectral synthesis code MOOG2002
\citep{sneden} and a grid of Kurucz ATLAS9 atmospheres with overshooting
\citep{kur93}. Examples of synthetic spectra and the parameters used
in the synthesis are shown in Figure \ref{ajustes_be}.\\
The final sample is composed of 70 and 30 stars with and without planets,
respectively, from \citet{santos_be1,santos_be3,santos_be2,galvez}, 14 stars
with planets from \citet{Delgado11} and 5 and 10 stars with and without detected
planets, respectively, from this work. This gives a total sample of 89 stars
with planets and 40 comparison sample stars. All Be abundances for these
89+40 stars were analyzed by our team using the same methodology,
making this a very uniform sample.\\
\begin{deluxetable}{rrrrrrrr}
\tablecaption{Kurucz line list tuned to reproduce solar spectrum.\label{lista}}
\tablewidth{0pt}
\tablehead{
\colhead{$\lambda (\AA)$} & \colhead{Atomic number} & \colhead{$\chi$ (eV)} &
\colhead{$gf$} & \colhead{$\lambda (\AA)$} & \colhead{Atomic number} & \colhead{$\chi$ (eV)} &
\colhead{$gf$}}
\startdata
3127.968 & 108.0 & 1.670 & 0.461E-03 & \vline \hspace{0.35cm} 3130.420 & 4.1 & 0.000 & 0.670E+00 \\
3128.060 & 108.0 & 0.541 & 0.376E-02 & \vline \hspace{0.35cm} 3130.433 & 108.0 & 1.756 & 0.428E-02 \\
3128.101 & 108.0 & 0.210 & 0.131E-02 & \vline \hspace{0.35cm} 3130.439 & 25.0 & 3.772 & 0.303E-02 \\
3128.154 & 108.0 & 1.599 & 0.101E-03 & \vline \hspace{0.35cm} 3130.473 & 108.0 & 1.609 & 0.603E-03 \\
3128.166 & 25.1 & 6.914 & 0.190E-02 & \vline \hspace{0.35cm} 3130.476 & 26.0 & 3.573 & 0.092E-02 \\
3128.172 & 25.0 & 7.822 & 0.228E-03 & \vline \hspace{0.35cm} 3130.549 & 25.1 & 6.494 & 0.714E-01 \\
3128.237 & 108.0 & 0.442 & 0.475E-03 & \vline \hspace{0.35cm} 3130.562 & 26.1 & 3.768 & 0.612E-05 \\
3128.269 & 21.1 & 7.424 & 0.675E+00 & \vline \hspace{0.35cm} 3130.569 & 24.1 & 5.330 & 0.349E-02 \\
3128.286 & 108.0 & 0.210 & 0.104E-01 & \vline \hspace{0.35cm} 3130.570 & 108.0 & 0.683 & 0.298E-01 \\
3128.289 & 108.0 & 0.442 & 0.728E-03 & \vline \hspace{0.35cm} 3130.575 & 23.0 & 1.218 & 0.543E-03 \\
3128.304 & 23.1 & 2.376 & 0.134E+00 & \vline \hspace{0.35cm} 3130.577 & 73.0 & 1.394 & 0.117E+01 \\
3128.307 & 607.0 & 0.513 & 0.288E-04 & \vline \hspace{0.35cm} 3130.585 & 24.0 & 3.556 & 0.108E-01 \\
3128.308 & 607.0 & 0.513 & 0.249E-04 & \vline \hspace{0.35cm} 3130.637 & 25.0 & 4.268 & 0.982E-01 \\
3128.356 & 108.0 & 1.714 & 0.245E-02 & \vline \hspace{0.35cm} 3130.648 & 106.0 & 0.034 & 0.02009 \\
3128.377 & 108.0 & 1.714 & 0.333E-03 & \vline \hspace{0.35cm} 3130.780 & 41.1 & 0.439 & 0.257E+01 \\
3128.393 & 77.0 & 1.728 & 0.110E+00 & \vline \hspace{0.35cm} 3130.791 & 45.0 & 0.431 & 0.776E-02 \\
3128.394 & 106.0 & 0.558 & 0.731E-01 & \vline \hspace{0.35cm} 3130.803 & 22.1 & 0.012 & 0.589E-01 \\
3128.394 & 106.0 & 0.558 & 0.800E-01 & \vline \hspace{0.35cm} 3130.813 & 64.1 & 1.157 & 0.826E+00 \\
3128.406 & 66.1 & 1.314 & 0.226E+01 & \vline \hspace{0.35cm} 3130.842 & 23.0 & 1.955 & 0.119E-02 \\
3128.488 & 22.1 & 7.867 & 0.151E+01 & \vline \hspace{0.35cm} 3130.851 & 26.2 & 11.595 & 0.527E-03 \\
3128.495 & 22.1 & 2.590 & 0.908E-05 & \vline \hspace{0.35cm} 3130.871 & 58.1 & 1.090 & 0.957E-01 \\
3128.518 & 108.0 & 0.786 & 0.817E-02 & \vline \hspace{0.35cm} 3130.905 & 26.1 & 7.487 & 0.100E-02 \\
3128.524 & 108.0 & 0.102 & 0.500E-03 & \vline \hspace{0.35cm} 3130.928 & 106.0 & 0.002 & 0.231E-03 \\
3128.546 & 24.1 & 4.757 & 0.859E-02 & \vline \hspace{0.35cm} 3130.928 & 108.0 & 1.907 & 0.946E-03 \\
3128.568 & 64.1 & 1.134 & 0.787E+00 & \vline \hspace{0.35cm} 3130.933 & 108.0 & 0.683 & 0.294E-03 \\
3128.617 & 22.0 & 5.959 & 0.658E-03 & \vline \hspace{0.35cm} 3130.997 & 108.0 & 1.569 & 0.137E-03 \\
3128.618 & 22.0 & 1.067 & 0.883E-03 & \vline \hspace{0.35cm} 3131.015 & 25.1 & 6.112 & 0.607E-01 \\
3128.626 & 22.0 & 6.065 & 0.643E+00 & \vline \hspace{0.35cm} 3131.037 & 25.0 & 3.773 & 0.596E+00 \\
3128.641 & 25.1 & 6.672 & 0.714E-01 & \vline \hspace{0.35cm} 3131.059 & 25.1 & 6.672 & 0.191E-02 \\
3128.648 & 26.1 & 12.966 & 0.855E-02 & \vline \hspace{0.35cm} 3131.065 & 4.1 & 0.000 & 0.338E+00 \\
3128.653 & 607.0 & 0.510 & 0.249E-04 & \vline \hspace{0.35cm} 3131.070 & 90.1 & 0.000 & 0.276E-01 \\
3128.653 & 607.0 & 0.510 & 0.210E-04 & \vline \hspace{0.35cm} 3131.102 & 26.1 & 9.688 & 0.695E-01 \\
3128.692 & 24.1 & 2.434 & 0.479E+00 & \vline \hspace{0.35cm} 3131.109 & 40.0 & 0.520 & 0.398E+00 \\
3128.692 & 29.0 & 4.974 & 0.195E+00 & \vline \hspace{0.35cm} 3131.115 & 26.0 & 3.047 & 0.194E-05 \\
3128.694 & 23.1 & 2.372 & 0.430E+00 & \vline \hspace{0.35cm} 3131.116 & 76.0 & 1.841 & 0.112E+01 \\
3128.728 & 28.0 & 1.951 & 0.512E-04 & \vline \hspace{0.35cm} 3131.143 & 22.0 & 0.836 & 0.279e-05 \\
3128.737 & 39.1 & 3.376 & 0.646E+01 & \vline \hspace{0.35cm} 3131.194 & 42.0 & 2.499 & 0.441E-01 \\
3128.763 & 72.0 & 0.000 & 0.170E-01 & \vline \hspace{0.35cm} 3131.212 & 24.0 & 3.113 & 0.604E+00 \\
3128.776 & 23.0 & 1.804 & 0.126E-01 & \vline \hspace{0.35cm} 3131.243 & 26.0 & 2.176 & 1.726E-04 \\
3128.782 & 108.0 & 0.897 & 0.102E-01 & \vline \hspace{0.35cm} 3131.255 & 69.1 & 0.000 & 0.240E+00 \\
3128.854 & 23.0 & 1.712 & 0.205E-03 & \vline \hspace{0.35cm} 3131.326 & 27.1 & 2.204 & 0.817E-04 \\
3128.898 & 26.0 & 1.557 & 0.223E-02 & \vline \hspace{0.35cm} 3131.329 & 108.0 & 1.942 & 0.160E-01 \\
3128.938 & 108.0 & 1.939 & 0.254E-03 & \vline \hspace{0.35cm} 3131.338 & 25.0 & 4.679 & 0.117E-01 \\
3128.949 & 75.0 & 2.061 & 0.166E+01 & \vline \hspace{0.35cm} 3131.339 & 21.1 & 7.381 & 0.372E-02 \\
3128.954 & 25.0 & 2.920 & 0.447E-04 & \vline \hspace{0.35cm} 3131.366 & 108.0 & 1.942 & 0.250E-03 \\
3128.975 & 108.0 & 1.939 & 0.169E-01 & \vline \hspace{0.35cm} 3131.384 & 108.0 & 1.680 & 0.604E-03 \\
3129.001 & 607.0 & 0.508 & 0.210E-04 & \vline \hspace{0.35cm} 3131.394 & 108.0 & 1.680 & 0.511E-03 \\
3129.005 & 27.0 & 0.514 & 0.117E-02 & \vline \hspace{0.35cm} 3131.395 & 26.1 & 3.815 & 0.221E-03 \\
3129.009 & 26.1 & 3.968 & 0.202E-02 & \vline \hspace{0.35cm} 3131.423 & 108.0 & 0.960 & 0.789E-02 \\
3129.009 & 26.1 & 12.966 & 0.340E-02 & \vline \hspace{0.35cm} 3131.458 & 25.1 & 4.340 & 0.294E-04 \\
3129.013 & 24.1 & 12.978 & 0.873E+00 & \vline \hspace{0.35cm} 3131.459 & 26.0 & 6.427 & 0.234E-04 \\
3129.017 & 107.0 & 0.740 & 0.202E-04 & \vline \hspace{0.35cm} 3131.486 & 28.1 & 12.409 & 0.755E-02 \\
3129.038 & 26.2 & 10.311 & 0.229E-03 & \vline \hspace{0.35cm} 3131.502 & 108.0 & 0.494 & 0.185E-02 \\
3129.070 & 22.0 & 6.079 & 0.971E+00 & \vline \hspace{0.35cm} 3131.525 & 28.0 & 7.152 & 0.126E-02 \\
3129.095 & 108.0 & 0.897 & 0.574E-03 & \vline \hspace{0.35cm} 3131.533 & 24.1 & 4.168 & 0.782E-02 \\
3129.110 & 20.0 & 4.625 & 0.185E-02 & \vline \hspace{0.35cm} 3131.545 & 80.0 & 4.887 & 0.912E+00 \\
3129.138 & 107.0 & 0.740 & 0.211E-04 & \vline \hspace{0.35cm} 3131.548 & 24.1 & 4.178 & 0.350E-01 \\
3129.144 & 24.1 & 7.332 & 0.498E-01 & \vline \hspace{0.35cm} 3131.583 & 25.1 & 6.495 & 0.124E-01 \\
3129.153 & 40.1 & 0.527 & 0.479E+00 & \vline \hspace{0.35cm} 3131.656 & 107.0 & 0.787 & 0.140E-03 \\
3129.182 & 26.0 & 6.411 & 0.349E-04 & \vline \hspace{0.35cm} 3131.687 & 108.0 & 1.736 & 0.193E-01 \\
3129.183 & 22.0 & 1.046 & 0.160E-03 & \vline \hspace{0.35cm} 3131.702 & 28.0 & 7.264 & 0.247E-01 \\
3129.209 & 107.0 & 0.740 & 0.223E-04 & \vline \hspace{0.35cm} 3131.711 & 108.0 & 1.736 & 0.319E-03 \\
3129.210 & 24.0 & 3.556 & 0.490E-01 & \vline \hspace{0.35cm} 3131.724 & 26.1 & 4.081 & 0.486E-02 \\
3129.228 & 76.0 & 2.191 & 0.537E+00 & \vline \hspace{0.35cm} 3131.754 & 108.0 & 0.960 & 0.550E-03 \\
3129.300 & 28.0 & 0.275 & 0.625E-02 & \vline \hspace{0.35cm} 3131.812 & 72.0 & 1.306 & 0.263E+01 \\
3129.305 & 66.0 & 0.000 & 0.135E-01 & \vline \hspace{0.35cm} 3131.825 & 27.0 & 1.740 & 0.161E-01 \\
3129.333 & 26.0 & 1.485 & 0.114E-01 & \vline \hspace{0.35cm} 3131.838 & 80.0 & 4.887 & 0.912E+00 \\
3129.348 & 25.0 & 3.379 & 0.490E-04 & \vline \hspace{0.35cm} 3131.935 & 107.0 & 0.787 & 0.148E-03 \\
3129.348 & 25.0 & 3.379 & 0.195E-03 & \vline \hspace{0.35cm} 3132.053 & 24.1 & 2.483 & 0.120E+01 \\
3129.376 & 11.1 & 32.944 & 0.102E+01 & \vline \hspace{0.35cm} 3132.062 & 22.0 & 5.975 & 0.530E-03 \\
3129.389 & 23.0 & 2.115 & 0.979E-03 & \vline \hspace{0.35cm} 3132.063 & 40.0 & 0.543 & 0.105E+01 \\
3129.454 & 8.1 & 25.640 & 0.290E+00 & \vline \hspace{0.35cm} 3132.109 & 24.1 & 4.775 & 0.643E-04 \\
3129.478 & 23.1 & 8.574 & 0.152E+00 & \vline \hspace{0.35cm} 3132.142 & 23.0 & 0.262 & 0.589E-05 \\
3129.481 & 27.0 & 1.883 & 0.951E-02 & \vline \hspace{0.35cm} 3132.186 & 108.0 & 0.901 & 0.964E-02 \\
3129.538 & 108.0 & 0.516 & 0.178E-02 & \vline \hspace{0.35cm} 3132.193 & 107.0 & 0.787 & 0.160E-03 \\
3129.548 & 73.0 & 1.147 & 0.107E+00 & \vline \hspace{0.35cm} 3132.212 & 27.0 & 0.101 & 0.377E-01 \\
3129.589 & 72.0 & 0.000 & 0.135E-01 & \vline \hspace{0.35cm} 3132.281 & 106.0 & 0.488 & 0.661E-01 \\
3129.636 & 22.0 & 1.443 & 0.158E-02 & \vline \hspace{0.35cm} 3132.281 & 106.0 & 0.488 & 0.592E-01 \\
3129.652 & 41.1 & 1.321 & 0.115E+00 & \vline \hspace{0.35cm} 3132.288 & 25.0 & 4.332 & 0.316E+00 \\
3129.763 & 40.1 & 0.039 & 0.331E+00 & \vline \hspace{0.35cm} 3132.355 & 23.0 & 1.043 & 0.984E-04 \\
3129.774 & 24.0 & 2.708 & 0.344E-02 & \vline \hspace{0.35cm} 3132.392 & 108.0 & 1.990 & 0.131E-03 \\
3129.857 & 24.0 & 2.968 & 0.845E-02 & \vline \hspace{0.35cm} 3132.405 & 25.0 & 3.373 & 0.818E-02 \\
3129.934 & 39.1 & 3.414 & 0.955E+01 & \vline \hspace{0.35cm} 3132.517 & 68.1 & 1.402 & 0.320E+01 \\
3129.937 & 108.0 & 1.609 & 1.955E-02 & \vline \hspace{0.35cm} 3132.518 & 26.0 & 7.168 & 0.316E+00 \\
3129.943 & 73.0 & 0.697 & 0.724E-01 & \vline \hspace{0.35cm} 3132.532 & 24.1 & 6.805 & 0.387E-03 \\
3129.968 & 64.1 & 1.172 & 0.627E+00 & \vline \hspace{0.35cm} 3132.579 & 26.1 & 7.495 & 0.161E-02 \\
3129.974 & 90.1 & 1.287 & 0.172E+00 & \vline \hspace{0.35cm} 3132.583 & 108.0 & 1.612 & 0.314E-01 \\
3130.056 & 40.0 & 0.519 & 0.200E+00 & \vline \hspace{0.35cm} 3132.591 & 58.1 & 0.295 & 0.288E+00 \\
3130.063 & 26.1 & 13.018 & 0.416E-02 & \vline \hspace{0.35cm} 3132.594 & 42.0 & 0.000 & 0.237E+01 \\
3130.075 & 108.0 & 2.295 & 0.0029174 & \vline \hspace{0.35cm} 3132.596 & 23.1 & 2.900 & 0.859E-01 \\
3130.126 & 108.0 & 0.842 & 0.794E-02 & \vline \hspace{0.35cm} 3132.631 & 607.0 & 0.510 & 0.210E-04 \\
3130.145 & 108.0 & 1.987 & 0.0141579 & \vline \hspace{0.35cm} 3132.656 & 73.0 & 0.491 & 0.110E+00 \\
3130.157 & 22.0 & 5.941 & 0.346E+00 & \vline \hspace{0.35cm} 3132.657 & 27.0 & 2.878 & 0.240E-04 \\
3130.202 & 25.1 & 4.801 & 0.193E+00 & \vline \hspace{0.35cm} 3132.710 & 22.0 & 5.954 & 0.171E+00 \\
3130.254 & 106.0 & 0.521 & 0.0660693 & \vline \hspace{0.35cm} 3132.725 & 25.1 & 6.177 & 0.215E-02 \\
3130.257 & 23.1 & 0.348 & 0.513E+00 & \vline \hspace{0.35cm} 3132.788 & 25.0 & 4.268 & 0.316E+00 \\
3130.281 & 108.0 & 0.250 & 0.134E-01 & \vline \hspace{0.35cm} 3132.794 & 8.2 & 36.895 & 0.933E+00 \\
3130.290 & 106.0 & 0.521 & 0.0731139 & \vline \hspace{0.35cm} 3132.809 & 23.1 & 2.510 & 0.297E-01 \\
3130.340 & 58.1 & 0.529 & 0.705E+00 & \vline \hspace{0.35cm} 3132.816 & 108.0 & 1.947 & 0.313E-03 \\
3130.353 & 27.1 & 2.985 & 0.465E-03 & \vline \hspace{0.35cm} 3132.822 & 24.0 & 3.122 & 0.322E+00 \\
3130.370 & 106.0 & 0.033 & 0.111E-01 & \vline \hspace{0.35cm} 3132.845 & 108.0 & 1.947 & 0.348E-02 \\
3130.376 & 22.0 & 1.430 & 0.275E-01 & \vline \hspace{0.35cm} 3132.864 & 28.1 & 2.865 & 0.223E-03 \\
3130.407 & 108.0 & 1.756 & 0.324E-03 & \vline \hspace{0.35cm} 3132.865 & 108.0 & 0.686 & 0.578E-03 \\
3130.408 & 108.0 & 1.670 & 0.103E-03 & \vline \hspace{0.35cm} 3132.878 & 44.0 & 1.317 & 0.174E+00 \\
\enddata
\tablenotetext{*}{Line list used by \citet{Garcia-Lopez95} and in our previous works. The dissociation
energy for OH, CH, NH and CN molecules is 4.39, 3.46, 3.47 and 7.65 eV respectively.}
\end{deluxetable}
\begin{deluxetable}{lcccccclrr}
\tablecaption{Stars analyzed in this work.\label{tabla}}
\tablewidth{0pt}
\tablehead{
\colhead{Star} & \colhead{T$_{\rm eff}$} & \colhead{log \textit{g}} &
\colhead{$\xi_{t}$} & \colhead{[Fe/H]} & \colhead{V} & \colhead{planet} &
\colhead{Spectral type\tablenotemark{a}} & \colhead{log $\epsilon$(Be)} &
\colhead{log $\epsilon$(Li)}\\
\colhead{} & \colhead{[K]} & \colhead{[cm s$^{-2}$]} & \colhead{[km s$^{-1}$]}}
\startdata
HD2638 & 5198 & 4.43 & 0.74 & 0.12 & 9.44 & yes & G5 & 0.49 & $<$0.16 \\
HD8326 & 4971 & 4.48 & 0.81 & 0.02 & 8.70 & no & K2V & <$-0.16$ & $<$0.09 \\
HD8389A & 5283 & 4.37 & 1.06 & 0.34 & 7.84 & no & K0VCN+2 & 0.16 & $<$0.73 \\
HD9796 & 5179 & 4.38 & 0.66 & -0.25 & 8.81 & no & K0V & 0.27 & $<$0.17 \\
HD11964A\tablenotemark{b} & 5332 & 3.90 & 0.99 & 0.08 & 6.42 & yes& G9VCN+1 & 0.55 & 1.41 \\
HD15337 & 5179 & 4.39 & 0.70 & 0.06 & 9.10 & no & K1V & 0.58 & $<$0.42 \\
HD16270 & 4786 & 4.39 & 0.84 & 0.06 & 8.37 & no & K3.5Vk: & $<$-0.32 & $<$0.03 \\
HD21019\tablenotemark{b} & 5468 & 3.93 & 1.05 & -0.45 & 6.20 & no& G2V & 0.22 & 1.39 \\
HD27894 & 4952 & 4.39 & 0.78 & 0.20 & 9.42 & yes & K2V & $<$-0.38 & $<$0.22 \\
HD35854 & 4928 & 4.46 & 0.54 & -0.13 & 7.74 & no & K2V & $<$-0.31 & $<$-0.22 \\
HD40105\tablenotemark{b} & 5137 & 3.85 & 0.97 & 0.06 & 6.52 & no& K1IV-V & $<$-0.12 & 1.40 \\
HD44573 & 5071 & 4.48 & 0.80 & -0.07 & 8.46 & no & K2.5Vk: & 0.75 & $<$-0.01 \\
HD63454 & 4840 & 4.30 & 0.81 & 0.06 & 9.37 & yes & K3Vk: & $<$-0.32 & $<$-0.03 \\
HD93083 & 5105 & 4.43 & 0.94 & 0.09 & 8.33 & yes & K2IV-V & $<$-0.14 & $<$0.16 \\
HD213042 & 4831 & 4.38 & 0.82 & 0.08 & 7.66 & no & K5V & $<$-0.40 & $<$0.06 \\
\enddata
\tablenotetext{a}{Values taken from Simbad}
\tablenotetext{b}{Evolved stars}
\end{deluxetable}
\begin{figure*}[ht!]
\centering
\includegraphics[width=16cm]{fig6.eps}
\caption{Observed spectra and difference in fluxes for six pairs of planet-host
stars (red lines) and stars without detected planets (blue dashed lines). The
position of Be lines are indicated by the arrows.}
\label{be_comp}
\end{figure*}
\section{Analysis}
In general, Be abundances for stars cooler than 5200 K are probably not reliable
since in this regime Be lines are barely sensitive to changes in the abundance.
In Figure \ref{ajustes_be} we can observe that for the coolest stars the fits are not good.
At those temperatures \ion{Mn}{1} line at 3129.037 \AA{} dominates the feature
and the presence of Be is negligible \citep[see also][]{Garcia-Lopez95}. We make
several tests to try to improve those fits.
In Figure \ref{be_test} we show several spectral syntheses. In panel \textit{a} we have used
different line lists for HD 213042. The orange dashed line is a fit made with original
Kurucz line list\footnote{http://kurucz.harvard.edu/line lists.html}. It is clear that
the Mn-Be feature cannot be well reproduce with the original value of log $gf$ for
Mn line even if we increase Be abundance (light blue dashed-pointed line). Furthermore,
the lambda of the whole feature do not match Be line, so this star do not present so
much Be and another line is required to fit the observed spectrum. \citet{primas} used
an artificial Fe line at 3131.043 which in our case does not help to fit the spectrum
(purple dashed-three pointed line) since this line gets stronger in metallic stars.
Another option is to increase the $gf$ of Mn line at 3131.037 \AA{}. This was first proposed
by \citet{Garcia-Lopez95} to reproduce solar spectrum and we also used this modified line
in our previous works on Be.
The green line represents a fit with Kurucz line list and this modified line. We note that
for our syntheses we have used Mn measured abundances by \citet{neves} with the same models.
This fit is better but the feature is still stronger than observed in its red wing
although the Be abundance used in the syntehsis is negligible.
The synthetic spectra of these cool stars present strong molecular lines. If we decrease
O and C abundances (red line) we can get a better fit though this does not affect Be line
at 3131.06 \AA{}.\\
The different contributions of molecular an atomic lines is shown in panel \textit{b} for star
HD 16270. The Mn-Be feature is practically filled with atomic lines, that is, Mn line.
Therefore we can put an upper limit in Be abundance since increasing Be abundance would
result in a stronger line with a core shifted towards higher $\lambda$ which would not
fit the observed spectrum. We can again get a better fit if we decrease O and C abundances
(red line). However, as the stars get cooler we need to use lower C and O abundances,
$\sim$ -0.6 dex for O and $\sim$ -0.9 dex for C in this star. This is
not a realistic approach but fortunately C and O abundances hardly affect the Be feature.\\
In panel \textit{c} we show three syntheses for HD 63454 with the solar line list derived by
\citet{Garcia-Lopez95} and decreasing C and O abundances -0.3 and -0.2 dex respectively.
The fits with log Be = 0.6 and 0.3 dex presents a stronger line than observed,
so we can adopt an upper limit of 0.0 dex for this star.
Certainly, Be line is not very sensitive at these low temperatures and we think
that we may be overestimating Be depletion. However, it is imposible to fit the spectra using
high Be abundances so we can put an upper limit in Be content though we cannot calculate
accurate abundances. Moreover, the sensitivity of Be line does not seem to be related to
T$_{\rm eff}$ since some cool young objets present strong Be lines with abundances similar
to solar \citep{smiljanic,Randich07}. This is the case of the star R1 in the young cluster
IC 2602 of 46 Myr and solar metallicity. This star was observed by \citet{smiljanic} who
found log Be = 1.25.
We have analyzed this star using the same spectrum and three different line lists (see
panel \textit{d} of Figure \ref{be_test}): \citet{primas} line list, used by these authors, gives
log Be = 1.22; our solar line list gives log Be = 1.37 and Kurucz line list with Mn line
modified and decreasing C and O abundances by 0.3 dex gives log Be = 1.32, all of them
in perfect agreement with the value previously found. Therefore, this test probably
suggests that our line list does work for cool stars which present a measurable quantity of Be,
and our fittings should be valid at least to put an upper limit in Be abundances.\\
\begin{figure*}[ht]
\centering
\includegraphics[width=8cm]{fig7.eps}
\includegraphics[width=8cm]{fig8.eps}
\caption{\textit{Left panel:} Be abundances as a function of effective
temperature for dwarf stars with and without detected planets from this work
(red filled and blue open triangles, respectively) and dwarf stars with (red
filled circles) and without planets (blue open circles) from previous studies
\citep{santos_be1,santos_be3,santos_be2,galvez,Delgado11}. The three evolved
stars from this work are depicted with squares. The Sun is denoted by
the usual symbol. The dashed lines represent 4 Be depletion models of
\citet{Pinsonneault} (Case A) with different initial angular momentum for solar
metallicity and an age of 1.7 Gyr. The solid line represents an assumed
initial Be abundance of 1.26 \cite{santos_be2}. The dotted line represents the
Be depletion isochrone for 4.6 Gyr taken from the models including mixing by
internal waves of \citet{Montalban}. \textit{Right panel:} Be abundances as a
function of Li abundances. Filled and open circles are stars with
and without planets, respectively, from previous surveys. Filled and open
triangles are stars with and without planets, respectively, from this work.
Filled and open squares are evolved stars with and without detected planets,
respectively, with measured Be abundance in this work. Colors denote different
temperature ranges.}
\label{be_teff}
\end{figure*}
\section{Discussion}
In this section we will discuss the different trends of Be abundances
with the effective temperature, the metallicity and the oxygen abundance.
\subsection{Is Be depleted in stars with planets?}
We have seen in the previous section that absolute Be abundances determined with
spectral synthesis are probably not reliable for cool stars, preventing a comparison
between stars with and without detected planets. In order to search for
differences in Be abundances between those stars we proposed to compare
directly their spectra in a previous paper \citep{Delgado11}. We selected pairs
of stars with similar stellar parameters so if there were a difference between
their spectra in the Be region it should be due to a difference in Be abundance.
In that paper we presented two planet-host stars, HD 330075 and HD 13445, with
an extra Be depletion when compared to analogous stars without detected
planets, showing that the effect observed in Li abundances for solar-type stars
with giant planets might also occur for Be in cooler stars.\\
In this work we present new spectra for 15 cool stars (see Table \ref{tabla})
with and without detected planets. Using these stars and others from previous
works \citep{santos_be3,santos_be2,galvez,Delgado11} we made 13 new pairs of
analogous stars with differences in T$_{\rm eff}$, log \textit{g} and
[Fe/H] lower than 50K, 0.15 dex and 0.10 dex respectively. We also note that v
sin\textit{i} values of these stars are very similar. Some examples of these
couples are shown in Figure \ref{be_comp}. We can see in these plots that stars with
and without planets present similar Be features and the differences in flux
between spectra are very low around Be features. In the previous paper we found
three stars from the comparison sample with clearly higher Be abundance
than the planet-host star HD 330075, whose abundance is log $\epsilon$(Be) $<$
-0.25. Now we have two more stars with very similar parameters. HD 44573 has a
higher Be abundance (see Table \ref{tabla}) and the difference in fluxes between
both stars is around 5$\sigma$ in the
position of the two Be lines. However, HD 8326 has a spectrum very
similar to HD 330075 and Be abundances are almost equal. Therefore,
planet host stars and comparison stars can have similar Be abundances.
We note, however, that our sample of cool stars is small and although unlikely,
it might be possible that we are only observing the fraction of comparison stars
with a strong Be depletion. A similar situation is seen for solar-type stars
without detected planets, where 50\% have high Li abundances while the other
50\% do have its Li depleted like planet-hosts.\\
\begin{figure*}[ht]
\centering
\includegraphics[width=8cm]{fig9.eps}
\includegraphics[width=8cm]{fig10.eps}
\caption{\textit{Left panel:} Be abundances as a function of metallicity. Filled
and open circles are dwarf stars with and without planets, respectively, from
previous surveys. Filled and open triangles are dwarf stars with and without
planets, respectively, from this work. Colors denote different temperature
ranges. \textit{Right panel:} Be abundances as a function of metallicity for
dwarf stars with 5600 K $<$ T$_{\rm eff}$ $<$ 6150 K and detections of Be. We
have overplotted two linear least square fittings for planet-host stars (red
filled circles, red line) and stars without detected planets (blue open circles,
blue dashed line).}
\label{be_feh}
\end{figure*}
\begin{figure}[ht]
\centering
\includegraphics[width=8cm]{fig11.eps}
\caption{\textit{Upper panel:} Be abundances as a function of temperature for
dwarf stars with effective temperature between 5600 K and 6150 K with (red
filled circles) and without planets (blue open circles). Green circles are stars
with planets with oxygen abundances measured in this work. \textit{Middle
panel:} Be abundances as a function of [O/H] using O abundances from OH
\citep{Ecuvillon06} lines for stars with effective temperature between 5400 K
and 6300K. Symbols like in top panel. \textit{Lower panel:} Be abundances
corrected of temperature effect for the same stars than middle panel.}
\label{be_oxi}
\end{figure}
\subsection{Be versus T$_{\rm eff}$}
In left panel of Figure \ref{be_teff} we plot the derived Be abundances as a
function of effective temperature for planet-host stars in our sample (see Table
\ref{tabla}) together with previous samples. In this plot we have removed
subgiants and giants (except the three analyzed in this work) to avoid
evolutionary effects in the abundances. In this selection we took the spectral
types of the stars from \citet{galvez}. With this new sample of stars we have
completed the coolest part of the plot and now we can see the behaviour of Be
abundances in a wide range of effective temperatures.
As mentioned in previous papers, the Be abundances decrease from a maximum near
T$_{\rm eff}$ = 6100 K towards higher and lower temperatures, in a similar way
as Li abundances behave. In the high temperature domain, the steep decrease with
increasing temperatures resembles the well known Be gap for F stars
\citep[e.g.][]{Boesgaard02}. The decrease of the Be content towards lower
temperatures is smoother and may show evidence for continuous Be burning during
the main sequence evolution of these stars.\\
When we move to the coolest stars, we observe a strong Be depletion regardless
of the presence of planets. Hence, the planetary formation processes that
affect Li depletion in solar analog stars
\citep{israelian09,gonzalez10} do not seem to take effect in the Be abundances.
On the other hand, we have found several stars with undetectable or very weak Be
lines. The steep decrease of Be abundances for T$_{\rm eff}$ $<$ 5500 K is in
contradiction with models of Be depletion \citep{Pinsonneault}, which predict
either constant or increasing Be abundances as T$_{\rm eff}$ decreases (see
Figure \ref{be_teff}). Even taking into account mixing by internal waves
\citep{Montalban}, Be depletion is still lower than predicted.
Be abundances have also been studied in clusters of different ages.
\citet{Randich07} found that stars in the 150 Myr old cluster NGC 2516 with
5000 $<$ T$_{\rm eff}$ $<$ 5500 K have not depleted Be while Hyades stars
(600 Myr) of the same T$_{\rm eff}$ present some Be depletion, something that
cannot be explained by models with only convective mixing.
\citet{Garcia-Lopez95} also found evidence of Be depletion for
two cool Hyades members.
Although uncertainties in Be abundandes for the coolest stars are large and it might be
some kind of systematic effect due to the \ion{Mn}{1} $\lambda$~3131.037~{\AA}
line, many of them seem to have their Be totally destroyed. Therefore, none of
those models including convective mixing, rotation or gravitational waves
seems to fit the observed Be abundances, at least at these cool temperatures.
Other depletion mechanisms have been proposed to explain Li destruction,
such as tachocline diffusion \citep{Brun,Piau} or episodic accretion of planetary
material during the PMS \citep{Baraffe} though Be has not been analyzed. However,
accretion of planetary material in the ZAMS \citep{theado,theado11} can lead to
an enhanced depletion of Li and Be, as well. These models can explain the
difference in the Li abundances between stars with and without planets and might
also explain the observed Be differences in a pair of cool planet-hosts when
compared to several stars without detected planets \citep{Delgado11}. On the other
hand the depletion produced by episodic accretion does not seem to be enough to explain the
strong Be depletion observed in cool stars and the presence of planetary material would be
necessary in all stars though many of them do not have detected planets.\\
\subsection{Be versus Li}
A beryllium versus lithium diagram can give us information about the depletion
rates in main-sequence stars. In right panel of Figure \ref{be_teff} this
relation is shown for all the stars in our sample. In general Be abundances
increase with increasing Li abundances and stars with and without planets
(filled and open symbols, respectively) behave in a similar way. If we take into
account the effective temperatures we can divide the sample in four groups. The
first one, with the hottest stars (light blue points), present both high Be and
Li abundances except for two stars with T$_{\rm eff}$ around 6300K which fall in
the Be-Li dip region, where the depletion of those elements is atributted to
slow mixing and depends on age and temperature \citep{Boesgaard02,Boesgaard04a},
though one of them, HD 120136 presents a normal Li abundance (log $\epsilon$(Li)
= 2.04). In the second group (orange points), with effective temperatures
between 5600 and 5900 K, Be abundances remain high although Li abundances
present different depletion rates. There are several
objects with low Be abundances and solar temperatures (specially the comparison
star HD 20766 T$_{\rm eff}$ = 5733 K, log $\epsilon$(Be) $<$ -0.09), which could
form some kind of ``Be-gap'' where the depletion of Be may be related to
different pre-main sequence rotational histories \citep{santos_be2}. In a recent
work, \citet{takeda2011} found four stars with solar temperatures and Be line
undetectable. Those stars are also depleted in Li and have low \textit{v}
sin\textit{i} values indicating that slow rotation can be the cause for that
strong Be depletion.
In the third group (blue points), composed by stars with 5150K $<$ T$_{\rm eff}$
$<$ 5600K, Be abundances begin to decrease and Li is severely depleted except
for two objects, HD 11964A and HD 21019, with particularly high Li abundances
considering its temperature. These two objects are evolved stars, therefore a
possible explanation for the high Li content is that they have just left the
main sequence, where they were hotter than they are now, but there has not been
time for their Li and Be to become strongly depleted. Other possible
explanations are a dredge up effect from a ``buffer'' below the former main
sequence convective envelope \citep{Deliyannis90} or accretion of planetary
metal-rich material, although recent models indicate that accretion of planetary
material might destroy Li instead of producing an enhancement
\citep{Baraffe,theado}. Finally, the coolest objects of our sample (red symbols)
are depleted in both Be and Li, although some stars have preserved some Be.
There are two objects with anomalous high Li content. HD 74576 seems to be a
young star, something that probably justifies its high Li content \citep[see
discussion in][]{santos_be2} while HD 40105 is an evolved star that might have
suffered some of the processes proposed for HD 11964A and HD 21019. We note that
we have removed from the plots all the evolved stars from previous works. Only
the three subgiants observed for this work are analyzed here but there are
similar cases in the whole sample \citep[see][]{santos_be2,galvez}.\\
This figure confirms what found in \citet{santos_be3,santos_be2,galvez}, Be and
Li burning seems to follow the same trend. Both elements increase their content
as the temperature rises, although Li depletion begins at 5900 K while strong Be
depletion starts below 5500 K. Stars with and without planets behave in a
similar way.\\
\subsection{Be versus [Fe/H]}
In the left panel of Figure \ref{be_feh} we show Be abundances as a function of
metallicity for different temperature ranges. We can see that the stars are
equally distributed regardless of their temperature and for higher metallicities
the dispersion in Be abundances increases, mainly due to objects with
5150 K $<$ T$_{\rm eff}$ $<$ 5600 K. It is well known that Be abundances
increase with metallicity for [Fe/H] $<$ -1 with a steep slope near 1
\citep[e.g.][]{Rebolo88,Boesgaard09}, but for higher metallicities this relation
is not so well defined.
\citet{Boesgaard04b} found a slope of 0.38 for stars with metallicity between
-0.6 and 0.2, in agreement with \citet{takeda2011} whose slope value is 0.49,
although they take into account only solar analogs with -0.3 $<$ [Fe/H] $<$ 0.3.
On the other hand, \citet{Boesgaard09} argued that stars of one solar mass with
solar metallicities (their most metallic star has [Fe/H] = 0.11) match the slope
of 0.86 for metal-poor stars.
In the right panel of Figure \ref{be_feh} all the dwarf stars of our sample with
5600 K $<$ T$_{\rm eff}$ $<$ 6150 K are plotted. We have chosen this limit in
temperature in order to remove most of the upper limits in Be abundances and
those stars from the Li-Be dip. We have also removed stars from the solar
Be-gap. We find a slope of 0.13 for planet-host stars and 0.07 for comparison
sample stars. These values are lower than previously found by other authors but
this could be possibly due to the lack of stars with [Fe/H] $<$ -0.2 in our
sample in comparison with the high number of metal-rich objects. It seems that
stars with solar and supersolar metallicities do not follow the same trend as
metal-poor stars, and Be has been produced at a lower rate as the Galaxy has
evolved. However, we note that Be abundances present higher uncertainties in
metal-rich stars that might be affecting these trends. If we have a look in the
upper envelope, which might be more consistent with low
er metallicity stars (around [Fe/H] = -0.5), we obtain a fit with a slope of
0.38, more consistent with previous results.
\begin{deluxetable}{lrrrrrr}
\tablecaption{Oxygen abundances for stars with T$_{\rm eff}$ $>$ 5400 K from
\citet{Delgado11}.\label{tabla_oxigeno}}
\tablewidth{0pt}
\tablehead{
\colhead{Star} & \colhead{[O/H]$_{1}$} & \colhead{[O/H]$_{2}$} &
\colhead{[O/H]$_{3}$} & \colhead{[O/H]$_{4}$} & \colhead{[O/H]$_{final}$} &
\colhead{log $\epsilon$(Be)}}
\startdata
HD2039 & 0.17 & 0.12 & 0.17 & 0.29 & 0.18 & 1.44 \\
HD4203 & 0.08 & 0.01 & 0.22 & 0.20 & 0.13 & 1.18 \\
HD73526 & - & - & 0.20 & 0.25 & 0.22 & 1.24 \\
HD76700 & 0.16 & 0.16 & 0.28 & 0.35 & 0.24 & 1.08 \\
HD154857 & -0.23 & -0.20 & -0.18 & -0.18 & -0.20 & 0.54 \\
HD208847 & 0.13 & 0.08 & 0.05 & 0.10 & 0.09 & 1.28 \\
HD216770 & -0.01 & 0.04 & 0.06 & 0.06 & 0.04 & 0.66 \\
\enddata
\end{deluxetable}
\subsection{Be versus [O/H]}
Beryllium is produced by spallation reactions between galactic cosmic rays
(GCRs) and the CNO nuclei in the interstellar medium \citep[see e.g.][and
references therein]{Tan}. Therefore, oxygen abundances can provide us with
complementary information about the galactic evolution of Be. We compile oxygen
abundances from \citet{Ecuvillon06} \citep[see the final lists in][]{galvez},
derived with the OH bands in the near-UV since this region is closed to Be
lines. We use the same temperature ranges than in previous section, 5600 K $<$
T$_{\rm eff}$ $<$ 6150 K. Thus, none of the stars analyzed in this work are
included in this plot because all of them are cooler than 5600 K. However, 6
planet-host stars from our previous work \citep{Delgado11} are hotter and we
have displayed them as green circles in Figure \ref{be_oxi}. Oxygen abundances
for these stars have been obtained in the same way as \citet{Ecuvillon06} and
are presented in Table \ref{tabla_oxigeno}.\\
In the middle panel of Figure \ref{be_oxi}, Be abundances are plotted as a
function of [O/H]. We can see that the slopes of planet hosts and stars without
detected planets are slightly different, possibly due to the low number of
comparison sample stars in this range. We have removed the star HD154857 since
its O abundance is the lowest of this sample and its content in Be is
considerable lower than other stars of similar oxygen abundance, hence it would
increase the slope by 0.3. We make again a fit of the upper envelope for stars
with planets to take into account the lack of stars with low [O/H] in our plot.
These trends might be influenced by the effect of T$_{\rm eff}$ in Be
abundances. To remove this effect, we have made a new Be vs. [O/H] plot (lower
panel of Figure \ref{be_oxi}) with ``corrected'' Be abundances, calculated as
log $\epsilon$(Be)-log $\epsilon$(Be(T$_{\rm eff}$)), where log
$\epsilon$(Be(T$_{\rm eff}$)) is obtained from a linear fit of Be as a function
o
f temperature in the same T$_{\rm eff}$ range (upper panel of Figure
\ref{be_oxi}). However, the slope is still the same for stars with planets. it
seems that there remains a correlation between Be and O abundances, which may
indicate that the sensitivity of Be to the Teff is not affecting at all this
correlation. On the other hand, the comparison sample is so small at this
temperature range, that no strong
conclusion can be extracted from the different slopes obtained before and after
the Be-Teff linear correction.
Nevertheless, these slopes are considerable lower than previously found for
metal-poor stars, even the slope of the upper envelope in the middle panel.
\citet{Boesgaard99} obtained a slope of 1.45, using stars with -3 $<$ [Fe/H] $<$
0.05, while \citet{Tan} reported a slope of 1.49, for stars with -2.3 $<$ [Fe/H]
$<$ -0.6. Therefore, the slope for metal-rich stars is much flatter than for
metal-poor stars as it seems to happen for Be versus [Fe/H]. A slope like this
would not be in agreement with models of Be production that predict quadratic or
linear relations \citep[see e.g.][and references therein]{Tan}. We note again
that oxygen abundances present a high dispersion in metal-rich stars and even at
low metallicities, the use of different O indicators change this trend.
\section{Conclusions}
We present new high-resolution UVES/VLT near-UV spectra of 15 stars with and
without planets in order to find possible differences in Be abundances between
them and confirm our previous results that suggested a greater depletion of Be
in planet-host stars when compared with stars without detected planets. Be needs
higher temperatures than Li to be destroyed so we have to search for these
differences in cooler stars, which have deeper convective envelopes and are able
to carry the material towards Be-burning layers. We have made new couples of
analog stars with and without planets with the purpose
of comparing directly their spectra to see possible differences in Be
abundances. Although in a previous work we found a pair of planet-host stars
with their Be severely depleted, now we have not observed important differences
in Be abundances between both groups of stars. Thus, the
effect caused by protoplanetary disks and rotational history on extra Li
depletion in solar-type stars with planets is not taking effect in Be
abundances, apparently.\\
Furthermore, Be abundances in cool stars (T$_{\rm eff}$ $<$ 5200 K) are much
lower than predicted by the models. We have analyzed for the first time a
considerable number of cool objects and found a strong destruction of Be in most
of the stars. Although we cannot provide absolute abundances we can give
upper values for Be. This gives a steep drop of Be abundances as T$_{\rm eff}$
diminishes in contradiction with current models of Be depletion.\\
Finally, the slopes of Be abundances as a function of [Fe/H] and [O/H] are
considerably lower than found for metal-poor stars, indicating that the
production rate of Be may have diminished with the evolution of the Galaxy.
\acknowledgments
E.D.M, J.I.G.H. and G.I. would like to thank financial
support from the Spanish Ministry project MICINN AYA2008-04874.
J.I.G.H. acknowledges financial support from the Spanish Ministry of
Science and Innovation (MICINN) under the 2009 Juan de la Cierva
Programme.\\
N.C.S. would like to thank the support by the
European Research Council/European Community under the FP7 through a
Starting Grant, as well from Funda\c{c}\~ao para a Ci\^encia e a
Tecnologia (FCT), Portugal, through a Ci\^encia\,2007
contract funded by FCT/MCTES (Portugal) and POPH/FSE (EC),
and in the form of grant reference PTDC/CTE-AST/098528/2008
from FCT/MCTES.\\
This research has made use of the SIMBAD database
operated at CDS, Strasbourg, France. \\
This work has also made use of
the IRAF facility, and the Encyclopaedia of extrasolar planets.
\vspace{10cm}
|
\section{Introduction}
The $Fermi$ Large Area Telescope ($Fermi$-LAT) has discovered over 80 $\gamma$-ray pulsars, and
revealed that the $\gamma$-ray pulsars are a major class of Galactic $\gamma$-ray sources. Of
these, $Fermi$-LAT first detected pulsed $\gamma$-ray emission from 11 millisecond pulsars,
hereafter denoted as MSPs (Abdo et al. 2010a,b,c, 2009a,b; Saz Parkinson et al. 2010; Guillemot et al. 2011). The
$\gamma$-ray emission from pulsars has a long history, having been discussed in the context of
a polar cap accelerator (Ruderman \& Sutherland 1975; Daugherty \& Harding 1982, 1996), a slot gap
(Arons 1983; Muslimov \& Harding 2004; Harding, Usov \& Muslimov 2005; Harding et al. 2008;
Harding \& Muslimov 2011) and an outer gap accelerator (Cheng, Ho \& Ruderman 1986a,b; Hirotani
2008; Takata, Wang \& Cheng 2010b). The cut-off features of the $\gamma$-ray spectra of the
Crab and the Vela pulsars as measured by $Fermi$ imply that the $\gamma$-ray emission site of
canonical pulsars is located in the outer magnetosphere. This is in contrast to a site near the
polar cap region which produces a cut-off feature steeper than observed (Aliu et al. 2009; Abdo et
al. 2009c, 2010d). However, Venter, Harding \& Guillemot (2009) found that the observed
pulse profiles of several MSPs detected by the $Fermi$ cannot be explained by the outer gap and/or
the slot gap models. Hence, they proposed a pair-starved polar cap model in which particles are
continuously accelerated to high altitude because multiplicity of the pairs is insufficient to screen
the electric field. Therefore, the origin of the $\gamma$-ray emission from MSPs remains to be clarified.
We present the spin period vs. the dipole moment for the rotation powered MSPs in Figure~\ref{pb},
in which the $\gamma$-ray emitting MSPs are marked with open circles.
The spin period in the accretion stage may be related to the equilibrium spin period, which is obtained by equating the co-rotation radius $r_{co}=(GMP^2/4\pi^2)^{1/3}$ to the Alfven
radius $r_M=(\mu^4/2GM\dot{M})^{1/7}$, where $M$ is the neutron star mass and $\dot{M}$ is the
accretion rate. The equilibrium spin period implies $\mu\propto P^{7/6}$ (Alpar et al. 1982). In Figure~\ref{pb}, we
illustrate the relation $\mu_{26}=\kappa P^{7/6}_{-3}$, where $P_{-3}=(P/1~\mathrm{ms})$, $\mu_{26}=
(\mu/10^{26}~\mathrm{G~cm^3})$. The normalization factor is chosen as $\kappa=1.5$ (solid
line), which gives an upper limit to the magnetic field for the recently activated radio
MSP PSR~J1023+0038 (see below), and $\kappa=1/3$ (dash line), which reproduces the typical $P-\mu$
relation for the Fermi-LAT MSPs and the radio MSPs. In section~\ref{accretion}, we apply the
above relations to estimate the magnetic moment of accreting MSPs.
In Figure~\ref{pb}, the new radio MSPs, whose locations are coincident with the $Fermi$ unidentified
sources are marked with boxes. In fact, it has been pointed out that over 20 new radio MSPs have been
discovered as $Fermi$ identified sources (e.g. Keith et al. 2011). Since the detection of the spin
period of MSPs (in particular in binary systems) via a blind search is very difficult, it is likely
that $Fermi$ has missed the identification of many MSPs. Thus, some of them may be associated with
the $Fermi$ unidentified sources. For example, Takata, Wang and Cheng (2011a,b,c) performed
population studies of $\gamma$-ray pulsars and argued on statistical grounds that the $\gamma$-ray
emission from over 100 MSPs are missed, possibly contributing to the $Fermi$ unidentified sources.
PSR~J1023+0038 is known to be the first and only rotation powered radio MSP in a quiescent
low-mass X-ray binary (LMXB) FIRST J102347.67+003841.2. The possibility of $\gamma$-ray
emission from the newly born MSP PSR~J1023+0038 has been pointed out by Tam et al. (2010).
This system showed clear evidence of an accretion disk in 2001 and its possible absence in
2002 (Wang et al. 2009), perhaps indicating that the pulsar became activated. To facilitate
the transition from an accreting MSP X-ray pulsars (AMP) to a rotation powered MSP, the accretion
of matter onto the NS must decrease rapidly. Campana et al. (1998) argued that the ``propeller''
effect, in which the Alfven radius exceeds the co-rotation radius, operates during the quiescent
state of an X-ray transient LMXB phase (Romanova et al. 2009) to eject matter from the system.
Takata, Cheng \& Taam (2010a) suggest that the activation of the rotation powered pulsar phase
in a short period LMXB is likely to occur during the quiescent state and that the $\gamma$-ray
emission produced in the outer gap accelerator in the pulsar magnetosphere irradiates the
surrounding disk, thereby, further enhancing this ejection of disk material.
The activation of rotation powered pulsars has been hypothesized based on the orbital
modulations in the optical emission observed in the quiescent state of LMXBs, i.e., First
102347.68+003841.2 (Thorstensen \& Armstrong 2005), SAX J1808.4-3658 (Burderi et al. 2003; Deloye
et al. 2008), XTE~1814-338 (D'Avanzo et al. 2009), and IGR~J00291+5934 (Jonker et al. 2008).
In these systems, the amplitude of the optical modulation can not be explained by irradiation
associated with the X-ray emission from the disk or from the neutron star (NS) surface due to
insufficient luminosity. To provide an explanation of the orbital modulation of the
optical emission, pulsar wind models have been suggested in which heating of the donor is
due to the effect of a relativistic pulsar wind (Burderi et al. 2003).
However there are some unresolved issues on this picture; for example,
(1) heating process of the stellar matter by the pulsar magnetic field if the wind energy near the companion star
is dominated by the magnetic energy
and (2) conversion mechanism from electromagnetic
energy into the particle energy within $r\sim 10^{10-11}$~cm from the pulsar if the wind energy is dominated
by the particle energy. In this paper, because of the large theoretical uncertainties on the pulsar wind model,
we will not pursue this picture in further detail. Alternatively, we will discuss
the $\gamma$-ray irradiation from the outer gap as a possible heating process of the companion star
(section~\ref{accretion}).
In the X-ray band, the emission from rotation powered MSPs can be composed of thermal emission
from a heated polar cap region plus (for some pulsars) non-thermal emission of magnetospheric
origin. Although the number of the non-thermal X-ray emitting MSPs have increased due to the
improved sensitivity of recent X-ray instruments (e.g. Zavlin 2007), the origin of pulsed
non-thermal X-ray emission from MSPs is not well understood.
Since the population of the $\gamma$-ray MSPs has increased due to the
detections by the $Fermi$-LAT, a study of the emission process in the X-ray and $\gamma$-ray
bands is necessary for discriminating between various non-thermal emission processes in
the magnetosphere.
Observationally, X-ray emission from the NS surface has been detected during the quiescent state of
several LMXBs (see Heinke et al. 2009 and reference therein). This emission coupled with estimates for the time averaged mass accretion rate
has been utilized to probe the equation of state of NS matter (e.g. Yakovlev, Levenfish \& Haensel
2003; Campana et al. 2008). For some systems, the thermal X-ray emission may be very weak in the
quiescent state (e.g., SAX~J1808.4-3658) and the X-ray emission may be well fit by a power law
spectrum (Heinke et al. 2009). The finding of non-thermal X-ray emission from SAX~J1808.4-3658 may
provide possible evidence for the activation of the rotation powered activity in the quiescent state,
although the origin of this power law component is unclear at present, as it may be due, for example,
to the emission from the magnetosphere or intra-binary shock.
The X-ray emission from an intra-binary shock in the MSP and low mass star (hereafter LMS) system
has been suggested to explain the observed unresolved non-thermal X-ray emissions from PSR~B1957/LMS
binary system (Arons \& Tavani 1993; Stappers et al. 2003; Huang \& Becker 2007), in which the LMS
eclipses the radio emission from the MSP. Moreover, similar binary (so called ``black widow'')
systems have been discovered at the positions of the $Fermi$ unidentified sources
(Roberts et al. 2011), which raises
questions regarding the origin of the high-energy emissions from MSP/LMS system.
As we have described above, observational evidence for non-thermal X-ray and $\gamma$-ray
emission from the different evolutionary stages of MSPs has been accumulating. In this paper,
we discuss the $\gamma$-ray and X-ray emissions from isolated rotation powered MSPs and
those in binary systems. Specifically, we model the $\gamma$-ray emission from the outer gap
of the MSPs, and explore (1) the origin of the $\gamma$-ray emission from rotation powered MSPs
detected by $Fermi$-LAT and (2) the possibility that irradiation by the $\gamma$-rays
from the outer gap can explain the optical modulation of quiescent LMXBs. We also discuss
the non-thermal X-ray emission associated with the outer gap activities and
the high-energy emission from an intra-binary shock. In section~\ref{model},
we review the $\gamma$-ray emission from the outer gap accelerator in a pulsar magnetosphere
and suggest that the thermal X-ray emission from the polar cap region can arise from the
heating due to incoming particles, which were accelerated in the outer gap. In
section~\ref{rotation}, we apply the model to the rotation powered MSPs and compare the
predicted $\gamma$-ray luminosity with the $Fermi$ observations. In section~\ref{accretion},
the various processes determining the NS surface temperature which depend on the
NS model of the accreting millisecond pulsars (AMPs) in quiescent LMXBs are discussed. In
addition, we estimate the $\gamma$-ray
luminosity from the outer gap accelerator. In section~\ref{nonx}, the non-thermal X-ray
emission is discussed within the context of the outer gap model, and we predict the relation
between the non-thermal X-ray luminosity and the spin down power of rotation powered MSPs. In
addition, the non-thermal X-ray luminosity from the AMP in the quiescent LMXB is discussed as
a function of the time averaged accretion rate.
In section~\ref{intra}, we discuss high-energy emission from an intra-binary shock in MSP/LMS
binary, and describe the difference in the properties between the intra-binary
shock emission and the magnetospheric emission. Finally, a brief summary is presented in
the last section.
\section{X-ray and $\gamma$-ray radiation processes}
\label{model}
\subsection{$\gamma$-ray emissions from the outer gap}
In the outer gap accelerator model for rotation powered pulsars, electrons and/or positrons
can be accelerated by the electric field along the magnetic field lines in the region where the
local charge density deviates from the Goldreich-Julian charge density. The typical strength of the
accelerating field in the gap is expressed as (Zhang \& Cheng 1997)
\begin{equation}
E_{||}\sim \frac{f^2V_{a}}{s}\sim 9\times 10^5f^2 \left(\frac{P}{10^{-3}~\mathrm{s}}\right)^{-3}
\left(\frac{\mu}{10^{26}~\mathrm{G~cm^3}}\right)\left(\frac{s}{R_{lc}}\right)^{-1}\mathrm{(c.g.s.)},
\label{electric}
\end{equation}
where $V_a=\mu /R_{lc}^2$ is the electrical potential drop across the polar cap, $P$ is the spin period, $\mu$ is the dipole moment of the NS, $s$ is the curvature radius of the magnetic
field line and $R_{lc}=cP/2\pi$ is the light cylinder radius. In addition, $f$ is the fractional
gap thickness, which is defined as the ratio of the gap thickness at the light cylinder
to the light cylinder radius $R_{lc}$. From the energy balance between the dipole radiation and the spin down of the pulsar, the dipole
moment of the NS is estimated from
\begin{equation}
\mu \sim \sqrt{3\pi}I^{1/2}R_{lc}^{3/2}P^{-1}\dot{P}^{1/2}\sim
3\times 10^{26}\sqrt{P_{-3}\dot{P}_{-19}}~\mathrm{G~\mathrm{cm^3}}
\end{equation}
where $I$ is NS moment of inertia assumed to be $I=10^{45}~\mathrm{g~cm^2}$. In addition,
$\dot{P}_{-19}$ is the time derivative of the spin period in units of $10^{-19}~\mathrm{s/s}$.
By assuming force balance between the acceleration and radiation reaction of the curvature radiation
process, the typical Lorentz factor of the particles in the gap is
\begin{equation}
\Gamma=\left(\frac{3s^2}{2e}E_{||}\right)^{1/4}\sim 2\times 10^7 f^{1/2}
P^{-1/4}_{-3}\mu_{26}^{1/4}s_*^{1/4},
\label{gamma}
\end{equation}
where $s_*=(s/R_{lc})$. The typical energy of the curvature
photons is estimated to be
\begin{equation}
E_{c}=\frac{3hc\Gamma^3}{4\pi s}\sim 25f^{3/2}P_{-3}^{-7/4}\mu_{26}^{3/4}s_{*}^{-1/4}~\mathrm{GeV}.
\label{lcut}
\end{equation}
The $\gamma$-ray luminosity from the outer gap is typically
\begin{equation}
L_{\gamma}\sim f^3L_{sd}=3.8\times 10^{35}f^3P_{-3}^{-4}\mu_{26}^2~\mathrm{erg~s^{-1}},
\label{lgamma}
\end{equation}
where we used the spin down power $L_{sd}=2(2\pi)^4\mu^2/3P^4c^3$.
We note that
the electron and positron pairs are mainly created around
the inner boundary of the outer gap (Cheng, Ruderman \& Zhang 2000;
Takata, Chang \& Shibata 2008; Takata et al. 2010b),
implying that the outgoing particles are accelerated with
whole gap potential drop, whereas the incoming particles can only receive
$\le 10\%$ of the gap potential drop.
Hence, we expect that the power carried by the outgoing particles
is, at least, one order of magnitude greater than
that carried by the incoming particles. In the gap, furthermore, because
the curvature radiation process is
occurring under the force balance
between the acceleration force and the radiation
back reaction force, as assumed in equation~(\ref{gamma}),
the particles lose most of the energy gain via the
curvature radiation process. Hence, the luminosity of the outgoing
$\gamma$-rays is, at least, one order of magnitude greater
than that of the incoming $\gamma$-rays. As we will discuss
in section~\ref{heating}, the incoming particles eventually reach the
stellar surface and heat up the polar cap region, producing
the observed thermal X-ray emissions. The Lorentz factor of
the incoming particles is reduced from $\sim 10^7$ at the inner boundary
of the outer gap to
$\sim 8\times 10^{5}$ at the stellar surface, indicating the X-ray emissions
from the heated polar cap region is at least two order of magnitude
less than the outgoing $\gamma$-rays radiated by the outgoing particles.
As equation~(\ref{lgamma}) indicates,
the $\gamma$-ray luminosity from the pulsar
depends on its spin down luminosity. On the other hand, the spin
down luminosity depends on the structure of magnetosphere
and the inclination angle between
the magnetic axis and the rotation axis, indicating the predicted
\ $\gamma$-ray luminosity from the outer gap depends on the structure of the
pulsar magnetosphere. For example, Spitkovsky (2006)
represented the dipole field structure in the ideal MHD limit, in which
there is no accelerating field,
and found that the inferred spin down luminosity, $L^{MHD}_{sd}\sim (2\pi)^4
\mu^2/(P^{4}c^3)\times (1+\sin^{2}\theta_o^2)$, can be up to 3 times larger than
the standard vacuum formula $L^{va}_{sd}\sim 2(2\pi)^4
\mu^2/(3P^{4}c^3)$, where $\theta_o$ is the inclination angle. Although
the magnetospheric structure has not been understood well,
the real magnetosphere with the accelerating region will be between
the vacuum limit and the ideal MHD limit (e.g. Constantinos et al. 2011;
Li, Spitkovsky \& Tchekhovskoy, 2011; Wada and Shibata 2011), and therefore
the spin down luminosity will be only
a factor of 1-3 different from that of standard formula ($L_{sd}^{va}$).
Furthermore, the detailed structure of the magnetosphere,
such as the magnetic field configurations, will be more important to
the pulse profiles and the phase-resolved spectra, which
are beyond the scope of this paper.
In this paper, therefore, we apply the standard formula
as a typical magnitude of the spin down power.
\subsection{Heated polar cap by incoming current}
\label{heating}
Half of the particles accelerated in the outer gap will return to the polar cap region and heat
the stellar surface. Several rotation powered MSPs exhibit thermal X-ray radiation characterized by
a temperature of $T\sim 10^6$~K (Zavlin 2007), which is much higher than $T\sim 10^5$~K expected
from the standard neutrino cooling scenario (Yakovlev \& Pethick 2004 for the review). Furthermore,
the size of the inferred emitting region corresponds to an effective radius of $10^{4-5}$~cm, which
is much smaller than the NS radius. For some MSPs the observed X-ray spectra can be fit by black body
radiation with two components, that is, a ``core'' component described by a higher temperature
($T_c>10^6$~K), but smaller effective radius ($R_c\sim 10^{3-4}$~cm), and a
``rim'' component with a lower temperature ($T_r\sim 5\times 10^5$~K) and larger effective radius ($R_r\sim 10^{5}$~cm). The observed temperatures and the effective radii of the 11 $\gamma$-ray emitting
MSPs are summarized in Table~1. These observations provide indirect evidence for additional
heating of the polar cap region.
It has been hypothesized that near the stellar surface, the magnetic field configuration is not dominated by a dipole field (Ruderman 1991; Chen, Ruderman \& Zhu 1998). Higher order multipole field configurations are likely and the strength of these components can be 1-3 orders of magnitude
greater than the global dipole field. The distance $\delta R_{eq}$ from the star for which the
local magnetic field is comparable to the dipole field is estimated from the relation
\begin{equation}
B_s\left(\frac{l+\delta R_{eq}}{l}\right)^{-3}=B_d\left(\frac{R_s+\delta R_{eq}}
{R_s}\right)^{-3}
\end{equation}
where $B_s$ is the strength of the local magnetic field at the stellar surface, $R_s$ is the stellar
radius, and $l\sim (1-3)\times 10^{5}$~cm is the thickness of the NS crust. With $B_s=10-1000B_d$,
we find $\delta R_{eq}\sim (1-3)R_s$. We expect that because of the bending of the local magnetic
field with a smaller curvature radius ($\sim 10^{5-6}$~cm) than $\sim 10^{7}$~cm of the dipole field,
the curvature photons emitted between $R_s<r<R_s+\delta R_{eq}$ will illuminate a wider area on the
stellar surface than that connected to the outer gap via the global dipolar magnetic field lines.
In Figure~\ref{localst1}, we illustrate this picture for the structure
of the polar cap region. The incoming relativistic particles, which
were accelerated in the gap, lose their energy via curvature radiation
between the stellar surface and the inner boundary of the gap, and
its Lorentz factor decreases
to $\Gamma\sim 2\times 10^{6}P^{1/3}_{-3}$ [equation~(\ref{gamma1})]
near the stellar surface ($r\sim 2R_s$). Near the stellar surface,
the incoming particles emit $\gamma$-ray photons
with an energy $m_ec^2/\alpha_f~\sim 70$~MeV, where $m_ec^2$ is the electron
rest mass energy and $\alpha_f$
is the fine structure constant (Wang et al. 1998; Takata et al. 2010b).
These $~70-100$MeV photons may be converted
into pairs via the local magnetic field, whose strength
may reach $B\sim 10^{11}$~Gauss. The local
magnetic field will bend the trajectory of the incoming particles so that
the $\sim 100$~MeV photons will illuminate and heat a significant
part of the polar cap region. The incoming particles eventually impact on
the stellar surface with a Lorentz factor
$\sim 8\times 10^5$ [equation~(\ref{gammas})], and their remaining energy
will heat an area much smaller than the polar cap region. We expect that
the former and latter components are observed as the rim and core components
respectively. With a strong local magnetic field, the typical size of
the core component is estimated as $R_c\sim fR_p(B_d/B_s)^{1/2}$
where $R_p$ is the size of the polar cap region of the dipole field.
This yields $R_c\sim
0.2(f/0.5)P^{-1/2}_{-3}$~km with $B_s=10^2B_d$, which is consistent with
the observations; for example $R_c\sim 0.4$~km for PSR J~0030+4051 (Table~1).
Between the NS star surface and the inner boundary of the gap, the evolution of the Lorentz factor
is approximately described by
\begin{equation}
m_ec^3\frac{d\Gamma}{dr}=\frac{2}{3}\frac{\Gamma^4e^2c}{s^2}.
\end{equation}
Near the surface, $s\sim\sqrt{rR_{ls}}$ provides a good approximation for the curvature radius of
the dipole field. Assuming that $\delta R_{eq}\sim R_s$ and the Lorentz factor at $r=R_s+
\delta R_{eq}$ is much smaller than that at the inner boundary of the gap, we obtain
\begin{equation}
\Gamma(R_s+\delta R_{eq})\sim \Gamma_{1}
[\mathrm{ln}(r_0/(R_s+\delta R_{eq}))]^{-1/3}\sim \Gamma_{1},
\label{gamma1}
\end{equation}
where $r_0$ is the radial distance to the inner boundary of the gap and
\[
\Gamma_{1}\equiv \left(\frac{R_{lc}m_ec^2} {2e^2}\right)^{1/3}\sim 2\times 10^6P_{-3}^{1/3}.
\]
In the region between $R_s<r<R_s+\delta R_{eq}$, the trajectories of the particles are described by
the local magnetic field lines, which have a curvature radius of $\zeta \sim 10^{6}$~cm.
In such a case, the Lorentz factor of the particles at the stellar surface can be estimated from
\begin{equation}
\Gamma_s\sim \left(\frac{m_ec^2\zeta}{2\pi e^2}\right)^{1/3}
\sim 8\times 10^5\zeta_6^{1/3}
\label{gammas}
\end{equation}
where $\zeta_6=\zeta/10^6$~cm. The X-ray luminosity and temperatures for the rim $(L_{X,r},T_r)$
and core ($L_{X,c}, T_c)$ components are calculated from
\begin{equation}
L_{r}=(\Gamma_{1}-\Gamma_s)\dot{N}m_ec^2,~T_r=\left(\frac{L_{X,r}}{4\pi R^2_{r}
\sigma_{s}}\right)^{1/4},
\end{equation}
and
\begin{equation}
L_{c}=\Gamma_s\dot{N}m_ec^2,~T_c=\left(\frac{L_{X,c}}{4\pi R^2_{c}
\sigma_s}\right)^{1/4},
\end{equation}
respectively, where $\dot{N}$ is the number of incoming particles per unit time and $\sigma_s$
is the Stefan-Boltzmann constant. In addition, $R_{r}$ and $R_{c}$ are the effective radii of the
heated region of the rim and core components respectively.
Wang, Takata and Cheng (2010) fit the phase-averaged spectrum of $\gamma$-ray pulsars using the
outer gap model and suggested that the averaged current density
is about 50~\% of the Goldreich-Julian value.
Hence, we take the rate of the incoming particles as
\begin{equation}
\dot{N}=f\frac{c\mu}{2eR_{lc}^2}.
\label{primary}
\end{equation}
As a result, the X-ray luminosity of the rim and core components are described by
\begin{equation}
L_{r}=(\Gamma_1-\Gamma_s)m_ec^2\dot{N}\sim 10^{32}fP_{-3}^{-5/3}\mu_{26}Q~\mathrm{erg~s^{-1}},
\label{xlr}
\end{equation}
and
\begin{equation}
L_{c}=\Gamma_sm_ec^2\dot{N}\sim 5\times 10^{31}fP^{-2}_{-3}\mu_{26}
\zeta^{1/3}_6~\mathrm{erg~s^{-1}},
\end{equation}
where $Q=1-[(\zeta/\pi R_{lc})^{1/3}]$. The temperature of the heated surface becomes
\begin{equation}
T_{r}\sim 2\times 10^6f^{1/4}P_{-3}^{-5/12}\mu_{26}^{1/4}R^{-1/2}_{r,5}
Q^{1/4}~\mathrm{K},
\label{trim}
\end{equation}
and
\begin{equation}
T_{c}\sim5\times 10^6f^{1/4}P_{-3}^{-1/2}\mu_{26}^{1/4}R^{-1/2}_{c,4}
\zeta_6^{1/12}~K,
\label{tc}
\end{equation}
respectively, where $R_{r,5}=R_r/(10^{5}$~cm) and $R_{c,4}=R_c/(10^{4}$~cm).
For the canonical pulsar, the core component
is observed with a luminosity
one or two order of magnitude fainter than that of the
rim component; for example, $L_c\sim 10^{30}~\mathrm{erg~s^{-1}}$ and
$L_r\sim 10^{32}~\mathrm{erg~s^{-1}}$ for the Geminga pulsar
(Kargaltsev et al. 2005). Halpern \& Ruderman (1993) studied
the thermal X-ray emissions from the heated polar cap due to the bombardment of the incoming particles, which were accelerated
in the outer gap, and found that
the predicted luminosity $\sim 10^{32}~\mathrm{erg~s^{-1}}$ with a temperature $T>
10^{6}$~K for the Geminga pulsar is too bright compared with the observed luminosity of
the core component $L_c\sim 10^{30}~\mathrm{erg s^{-1}}$.
Hence, they proposed that most of the X-ray photons from the
heated polar cap region is scattered by the resonant Compton scattering
and are eventually redistributed as a
thermal emission from almost entire surface with a temperature $T<10^{6}$~K, which is observed
as the rim component(Wang et al. 1998; Cheng \& Zhang 1999).
For the MSPs, the observed luminosity of the core
component ($L_{c}\sim 10^{31}~\mathrm{erg~s^{-1}}$)
is comparable to that of the rim component ($L_{r}$), as Table~1 indicates.
In fact, our model predicts that
most of thermal emissions from the
heated polar cap regions directly emerge without the resonant scattering (section~\ref{rotation}).
Because of the weaker magnetic field of the MSPs, the resonant Compton scattering may be less effective as
compared with the case of the canonical pulsar.
\section{Application to the Rotation Powered Millisecond Pulsars}
\label{rotation}
As argued by Takata et al. (2010b) the outer gap can be controlled by either the photon-photon
pair-creation process or the magnetic pair-creation process. In this section, we first review
the outer gap model controlled by the photon-photon pair-creation process in section~\ref{photon}
and the magnetic pair-creation process in section~\ref{magnetic}. For reference, Table~1 summarizes
the observed parameters of the MSPs detected by the $Fermi$-LAT.
\subsection{Outer gap controlled by the photon-photon pair-creation process}
\label{photon}
The $\gamma$-rays emitted in the outer gap can collide with the X-rays from the heated polar cap and
convert into electron-positron pairs. It is possible that this process, itself, controls the
thickness of the outer gap (Zhang \& Cheng 1997, 2003). However, given the existence of the rim and core components, it is not
clear theoretically which component controls the outer gap. If both components illuminate the gap,
the core component, which is of a higher temperature than the rim component, is more likely to
control the size of the outer gap. On the other hand, with a small effective radius of $R_c\sim
10^{3-4}$~cm, it is possible that the core component does not illuminate the outer gap. In this
respect, the X-ray emission of the rim component, whose effective radius $R_{r}\sim 10^{5-6}$~cm,
has a greater likelihood of illuminating the outer gap. To avoid complexity in the theoretical
argument, therefore, we only present the case that thermal X-rays from the rim component control the
size of the outer gap. We remark that if the $\gamma$-ray emission is
controlled by the core component, the predicted $\gamma$-ray luminosity is several times smaller
than presented here. However, the main conclusion in this paper will remain unchanged.
The use of the pair-creation condition $E_{\gamma}E_{X}\sim (m_ec^2)^2$ together with
equations~(\ref{lcut}) and~(\ref{trim}) leads to the fractional gap thickness controlled by the
photon-photon pair-creation process, $f\equiv f_{p}$, as
\begin{equation}
f_p\sim0.1P_{-3}^{26/21}\mu_{26}^{-4/7}s_*^{1/7}R_{r,5}^{2/7}Q^{-1/7},
\label{fp}
\end{equation}
where the typical X-ray photon energy $E_{X}=3kT_{r}$ is used. The $\gamma$-ray luminosity
(\ref{lgamma}) and the typical radiation energy~(\ref{lcut}) can be
rewritten as
\begin{equation}
L^p_{\gamma}\sim3.8\times 10^{32}P^{-2/7}_{-3}\mu_{26}^{2/7}s_*^{3/7}
R_{r,5}^{6/7}Q^{-3/7}
~\mathrm{erg~s^{-1}},
\label{plgamma}
\end{equation}
and
\begin{equation}
E^p_{c}\sim0.8P^{3/28}_{-3}\mu_{26}^{-3/28}s_*^{-1/28}
R_{r,5}^{3/7}Q^{-3/14}~\mathrm{GeV}.
\end{equation}
The total X-ray luminosity~$L_{X}=L_{r}+L_{c}$ and the temperature of the
heated surface are now given by
\begin{equation}
L^p_{X}\sim10^{31}P^{-3/7}_{-3}\mu^{3/7} s_*^{1/7}
R_{r,5}^{2/7}Q^{-1/7}~\mathrm{erg~s^{-1}},
\end{equation}
\begin{equation}
T^p_{r}\sim 10^{6}P^{-3/28}_{-3}\mu_{26}^{3/28}s_*^{1/28}
R^{-3/7}_{r,5}Q^{3/14}~\mathrm{K},
\label{ptr}
\end{equation}
and
\begin{equation}
T^p_{c}\sim3\times 10^{6}P^{-4/21}_{-3}\mu_{26}^{3/28}s_*^{1/28}R_{r,5}^{1/14}
R_{c,4}^{-1/2}\zeta_6^{1/12}Q^{-1/7}~\mathrm{K},
\label{ptc}
\end{equation}
respectively.
In Table~2, we compare between the observed (second-fourth columns)
and predicted (fifth-eighth columns) X-ray and $\gamma$-ray emission properties.
For the effective radii of the rim and core components, (1) we apply the
observational value if it is available or (2) we apply
the typical value $R_r=3\times 10^5$~cm and $R_c=10^4$~cm, respectively,
if observational results are unavailable. In addition,
we use the curvature radii corresponding to $s_*=0.5$ and $\zeta_6=1$.
We find in Table~2 that the predicted surface
temperatures of the rim ($T^p_r\sim 0.5-0.7\times 10^6$~K) and
core components ($T^p_c\sim 1-4\times 10^6$~K) are approximately consistent
with the observations. The predicted $\gamma$-ray luminosity
$L^p_{\gamma}\sim 5\times 10^{32}~\mathrm{erg~s^{-1}}$ is consistent
with the observations for older MSPs with
$\tau_c> 5\times 10^{9}$~yrs, where $ \tau_c\equiv
\dot{P}/(2P)\sim 1.5\times 10^9 P_{-3}^2/\mu_{26}^2$~yrs (Table~1).
For the younger MSPs (PSRs J0218+4232, B1937+21 and B1957+20),
we find that the outer gap model predicts a $\gamma$-ray
luminosity one order of magnitude less than that from the $Fermi$ observation.
\subsection{Outer gap controlled by the magnetic pair-creation process}
\label{magnetic}
Takata et al. (2010b) argued that the incoming particles emit photons with an energy
$m_ec^2/\alpha_f\sim 70 \rm MeV$ by curvature radiation near the stellar surface. These photons
can become pairs via the magnetic pair creation process and the secondary pairs can continue to
radiate several MeV photons via synchrotron radiation. In this case, the photon multiplicity can
easily exceed $10^4$ per incoming particle. For a simple dipole field structure, all pairs move
inward and cannot affect the outer gap accelerator. However if the local field lines near the
surface are bent sideward due to the strong multipole field (e.g. shown in Figure~\ref{localst1}),
the pairs created in these local magnetic field lines can have an angle greater than 90$^{\circ}$,
which results in an outgoing flow of pairs. Only a very tiny fraction (1-10) out of $10^4$ photons
is required to create pairs in these field lines, which are sufficient to provide screening in
the outer gap when they migrate to the outer magnetosphere. In this model, the fractional gap
thickness in this circumstance is
\begin{equation}
f_m\sim0.025KP^{1/2}_{-3},
\label{fm}
\end{equation}
where $K\sim B_{m,12}^{-2}s_7$ characterizes the local parameters. Here, $B_{m,12}$ and $s_7$ are
the local magnetic field in units of $10^{12}$G and the local curvature radius in units of $10^7$cm,
respectively. For the case of MSPs, the local parameter can be in the range $B_{m,12}\sim 0.01-0.1$
and $s_7\sim 0.01-0.1$, which yields $K$ of the order of ten. Substituting equation~(\ref{fm}) into
equations~(\ref{lgamma}) and~(\ref{lcut}), we obtain the expected $\gamma$-ray luminosity and the typical radiation energy as
\begin{equation}
L^m_{\gamma}\sim 6\times 10^{33}K^{3}_{1}P^{-5/2}_{-3}\mu_{26}^2~\mathrm{erg~s^{-1}},
\label{mlgamma}
\end{equation}
and
\begin{equation}
E^m_c\sim 3K^{3/2}_{1}P^{-1}_{-3}\mu_{26}^{3/4}s_*^{-1/4}~\mathrm{GeV},
\label{ecm}
\end{equation}
respectively, where $K_1=K/10$. The total X-ray luminosity and the temperatures of the heated
surface are described as
\begin{equation}
L^m_X\sim 3\times 10^{31}K_1P^{-7/6}_{-3}\mu_{26}~\mathrm{erg~s^{-1}},
\end{equation}
\begin{equation}
T^m_r\sim 10^6K_1^{1/4}P_{-3}^{-7/24}\mu_{26}^{1/4}R_{r,5}^{-1/2}
Q^{1/4}~\mathrm{K},
\label{mtr}
\end{equation}
and
\begin{equation}
T^m_c\sim4\times 10^{6}K_1^{1/4}P^{-3/8}_{-3}\mu_{26}^{1/4} R_{c,4}^{-1/2}\zeta_6^{1/12}~\mathrm{K},
\label{mtc}
\end{equation}
respectively.
The X-ray and
$\gamma$-ray emission properties predicted by
the outer gap model controlled by the magnetic pair-creation
process near the stellar surface are summarized in ninth-twelfth columns of
Table~2. By comparing between the emission properties predicted by the
photon-photon pair-creation model and the magnetic pair-creation model
in Table~2, we find that the X-ray/$\gamma$-ray emission properties
for the older MSPs do not depend strongly
on the specific pair-creation process controlling
the outer gap. For younger MSPs (PSRs J0218+4232, B1937+21 and B1957+20),
however, the magnetic pair-creation model predicts $\gamma$-ray
emission about one order of magnitude
brighter than that of the photon-photon pair-creation
model in closer agreement with the $Fermi$ observation.
For PSR~J1614-223 (or J0437-4718), the theoretical predictions
in Tables~2 are found to lie below (exceed) the measured value.
This discrepancy may be affected by the local structure,
since the $\gamma$-ray luminosity described by equation~(\ref{mlgamma})
is sensitive to the local structure as $L_{\gamma}\propto K^3$. Furthermore
the uncertainties of the solid angle and the distance may also
contribute to these discrepancies. As indicated in Table~1,
the observed flux of
PSR~J1614-223 implies a $\gamma$-ray emission efficiency
of $L_{\gamma}/L_{sd}\sim 1$
assuming a solid angle $\Delta\Omega_{\gamma}=4\pi$ and distance $d=1.3$~kpc. Comparing with
the expected efficiency $L_{\gamma}/L_{sd}\sim 0.1$ for the outer gap model, the
actual solid angle and/or distance may be several factors smaller than
$\Delta\Omega_{\gamma}=4\pi$ and/or $d=1.3$~kpc. In addition to the local effect and
the observational parameters, the observed inefficient $\gamma$-ray luminosity of PSR~J0437-4718
may be a result of an unfavorable viewing angle.
Takata et al. (20011a,b,c) showed that the observed $\gamma$-ray flux from
the outer gap decreases with decreasing viewing angle as measured from
the spin axis. Hence, $Fermi$ has preferentially discovered pulsars
with larger viewing angle $\xi\sim 90^{\circ}$. For the
$\gamma$-ray pulsars with $\xi\sim 90^{\circ}$, the
luminosity inferred from the flux can be characterized
by equation~(\ref{mlgamma}). On the other hand, if the viewing angle
is much smaller than $90^{\circ}$, the
$\gamma$-ray luminosity inferred from the observed
flux will lie below the prediction~(\ref{mlgamma}).
In fact, a smaller Earth viewing geometry is preferred to reproduce
the observed single pulse profile of PSR~J0437-4718 (Abdo et al. 2009b)
using the outer gap model (Takata et al. 2011c).
Based on statistical grounds, Takata et al. (2010b) suggest that the outer gap controlled by the
pair-creation model may provide a preferable explanation for the possible observational correlation
between the characteristics of the $\gamma$-ray emission and the pulsar characteristics. For the
MSPs, the $\gamma$-ray luminosity $L_{\gamma}$ (\ref{plgamma}) and~(\ref{mlgamma}) can be cast in
terms of the spin down power $L_{sd}= 2(2\pi)^4 \mu /(3c^3P^4)$ or the
characteristic age
$\tau_c=P/2\dot{P}$ yielding
\begin{equation}
L^p_{\gamma}\sim 3\times 10^{32}L_{sd,34}^{1/14}\mu_{26}^{1/7}
s_*^{1/7}R_{r,5}^{2/7}Q^{-1/7}
~\mathrm{erg~s^{-1}},
\label{lglsp}
\end{equation}
or
\begin{equation}
L^p_{\gamma}\sim 4\times 10^{32}\tau_{c,9}^{-1/7}
s_*^{1/7}R_{r,5}^{2/7}Q^{-1/7}~\mathrm{erg~s^{-1}}
~\mathrm{erg~s^{-1}},
\label{lgtaup}
\end{equation}
for the outer gap controlled by the photon-photon pair-creation process, and
\begin{equation}
L^m_{\gamma}\sim 6\times 10^{32}L_{sd,34}^{5/8}K_1^{3}\mu_{26}^{3/4}
~\mathrm{erg~s^{-1}},
\label{lglsm}
\end{equation}
or
\begin{equation}
L^m_{\gamma}\sim 10^{34}\tau_{c,9}^{-5/4}K_1^{3}\mu_{26}^{-1/2}
~\mathrm{erg~s^{-1}},
\label{lgtaum}
\end{equation}
by the magnetic pair-creation process. Here, $L_{sd, 34}=(L_{sd}/10^{34}~
\mathrm{erg~s^{-1}})$ and $\tau_{c,9}=(\tau_c/10^{9}~\mathrm{yrs})$. In Figures~\ref{lgls}
and~\ref{lgtau}, the model predictions given by equations~(\ref{lglsp})-(\ref{lgtaum}) are plotted
with the solid lines (for the photon-photon pair-creation) or dashed lines (for the magnetic
pair-creation). The filled circles represent the MSPs detected by the $Fermi$-LAT. Notwithstanding
the large observational errors, the data points at large $L_{\gamma}$ in Figures~\ref{lgls}
and~\ref{lgtau} may suggest that the magnetic pair-creation model is preferred over the
photon-photon pair-creation model for the $L_{\gamma} - L_{sd}$ and $L_{\gamma}- \tau_c$ relations.
Given that the X-rays from the heated polar cap may be prevented from illuminating the outer gap by
the resonant cyclotron scattering process, the magnetic pair-creation process may be the more
important process to control the outer gap. The cross section for Thomson scattering can be
represented as (Halpern \& Ruderman 1993; Zhang \& Cheng 1997),
\begin{equation}
\sigma=\sigma_T(\hat{\varepsilon}\cdot\hat{B})^2+\frac{2\pi^2e^2}{m_ec}
(\hat{\varepsilon}\times\hat{B})^2\delta(\omega_B-\omega),
\end{equation}
where $\hat{\varepsilon}$ is the (electric field) polarization of the X-ray photons, $\hat{B}$ is
the unit vector of the direction of the background magnetic field and $\omega_B=eB(r)/mc$. In
the case where $\hat{\varepsilon}\cdot\hat{B}=0$ and for a local dipole magnetic field $B_s\sim
10^{11}$~G, the resonant scattering will be efficient at $\delta R\sim l(B_se\hbar /E_Xm_ec)^{1/3}
\sim 7\times 10^{5}\mathrm{cm} (l/3\cdot 10^5~\mathrm{cm})(B_s/3\times 10^{11}~\mathrm{G})^{1/3}
(T/10^{6}~ \mathrm{K})^{-1/3}$ from the stellar surface, where $l$ is the thickness of the crust.
Since we obtain
\begin{equation}
\int \sigma dr\sim \frac{2\pi^2 e^2}{m_ec\omega}\left(\frac{eB_sl^3}{m_ec\omega}\right)^{1/3}\sim 10^{-13}\left(\frac{l}{3\cdot 10^{5}\mathrm{cm}}\right)
\left(\frac{B_s}{3\cdot 10^{11}\mathrm{G}}\right)^{1/3}
\left(\frac{T}{10^{6}\mathrm{K}}\right)^{4/3}~\mathrm{cm^{3}},
\end{equation}
a number density $n_{\pm}>10^{13}~\mathrm{cm^{-3}}$ at $\delta R\sim 10^{6}$~cm leads to an optically
thick cyclotron resonant scattering layer. The number density of the incoming primary particles near
the stellar surface, which may be about 50~\% of the Goldreich-Julian value (e.g.
equation~\ref{primary}), becomes $n\sim B/(2Pce)\sim 3\times 10^{9} (B/10^8~\mathrm{G})P^{-1}_{-3}
~\mathrm{cm^{3}}$. This value implies that the cyclotron resonant scattering can become optically
thick for the multiplicity of the incoming particles due to the magnetic pair-creation ($\ge 10^{4}$,
see Takata et al. 2010b). Because the radial distance to the inner boundary of the outer gap from the
stellar surface is only $\sim 5\times 10^{6}-10^{7}~$cm for MSPs, it is possible that the scattering
layer prevents the illumination of the inner part of the outer gap by X-ray photons.
In such a case, the magnetic pair-creation process can control the outer gap.
\section{Application to Millisecond Pulsar in Quiescent LMXBs}
\label{accretion}
\subsection{Thermal X-ray emissions from MSPs}
Observations of the optical modulation of LMXBs in the quiescent state have provided indirect evidence
for additional heating of the companion star, possibly due to the rotational energy loss of the NS
associated with pulsar activity, e.g. J102347.68+003841.2 (Thorstensen \& Armstrong 2005), SAX
J1808.4-3658 (Burderi et al. 2003; Deloye et al. 2008), XTE 1814-338 (D'Avanzo et al. 2009) and IGR
J00291+5934 (Jonker, Torres \& Steeghs 2008). Specifically, the
irradiation luminosity required to produce the amplitude of the modulation
is $\sim 10^{33-34}~\mathrm{erg~s^{-1}}$ for the isotropic radiation.
This is significantly greater than that associated with the X-ray
emission from the disk or neutron star, indicating the need for the
operation of an additional heating source. If the pulsar magnetosphere is sufficiently clear of matter during the quiescent
state of the LMXB, the outer gap accelerator can be activated and
the emitted $\gamma$-rays may irradiate and heat the companion star.
If the optical modulation is a result of irradiation from the outer gap,
the actual irradiated luminosity may be several factors less than that for the
isotropic case because the outer gap emission is beamed with a solid angle
$\Delta\Omega_{\gamma}\sim2-3$~radian (Takata et al. 2010b). The amplitude of the optical
modulation is estimated as $L_{o}\sim (\pi \theta^2/\Delta\Omega_{\gamma})
L_{\gamma}$, where $\theta$ is the angle of the size of the
companion star measured from the pulsar and
$\Delta\Omega_{\gamma}$ is the solid angle of the $\gamma$-ray beam.
If the companion fills its Roche lobe,
$\theta\sim 0.462[q/(1+q)]^{1/3}$ for typical LMXB $0.1<q<0.8$, where $q$
is the mass ratio of the system (Frank, King \& Raine, 2002).
Deep X-ray observations have been carried out during the quiescent state to search for the thermal
emission from the NS star surface. The detected emissions indicate that the NS is hotter in
comparison to expectations based on traditional cooling curves of NSs. An explanation for this
difference is a consequence of heating associated with nuclear fusion in the crust. In this
picture, the base of the accreted matter in the crust is sufficiently compressed by the overlying
weight of newly accreted matter, leading to pycnonuclear reactions at $\rho\sim 10^{12-13}~\mathrm{g~
cm^{-3}}$. These reactions release about 1-2~MeV per accreted baryon, resulting in heating of the
crust and the core (Brown, Lars \& Rutledge 1998; Haensel \& Zdunik 1990, 2003). On a time scale of
$10^4$~yr, thermal equilibrium is established between heating during the accretion stage and
cooling during the quiescent stage (Colpi et al. 2001). Accordingly, the NS core and surface
temperatures can reach $\sim 10^{8-9}$~K in the interior and $\sim 10^{6}$~K at the surface. Given this
surface thermal emission, the outer gap in the quiescent stage may be controlled by the photon-photon
pair-creation between the $\gamma$-rays and the X-rays from the NS surface.
To explore the $\gamma$-ray emission from AMPs in the quiescent state, the X-ray emission in this state
is calculated following the model description by Yakovlev et al. (2003). The NS core
temperature, $T_i$, surface temperature, $T_s$, and long-term time-average ($\sim 10^4$~yr)
mass accretion rate, $<\dot{M}>$, are related by
\begin{equation}
L_{h}(<\dot{M}>)=L_{\nu}(T_i)+L_{th}(T_s),
\end{equation}
where $L_h$ is heating term due to the pycnonuclear reactions, $L_{\nu}$ is cooling term associated
with neutrino emission, and $L_{th}=4\pi R_s^2\sigma_s cT_s^4$ is the thermal emission from the NS
surface. The heating rate is expressed as
\begin{equation}
L_h(<\dot{M}>)=\frac{<\dot{M}>}{m_u}Q_{\nu}\sim 8.7\times 10^{33}
\frac{<\dot{M}>}{10^{-10} {\rm ~M}_{\odot} \mathrm{yr^{-1}}}~\mathrm{erg~s^{-1}},
\end{equation}
where $m_u$ is the atomic mass unit and $Q_{\nu}\sim 1.45$~MeV is the nuclear energy release per baryon.
The neutrino emission $L_{\nu}$ is calculated using equation~(4) in Yakovlev et al. (2003). A $T_i-T_s$
relation was obtained by Gudmundsson, Pethick, \& Epstein (1983) as
\begin{equation}
T_i\sim1.288\times 10^8(T^4_{s,6}/g_{14})^{0.455}~K,
\end{equation}
where $g_{14}$ is the surface gravity $g=GM\mathrm{e}^{-\Phi}/R^2$ in units of $10^{14}~\mathrm
{cm~s^{-2}}$. Here, $\Phi= \sqrt{1-2GM/c^2R_s^2}$ and $T_{s,6}= T_s/10^6$~K.
Figures~\ref{axplum} and \ref{axptemp} display the X-ray luminosity and the surface temperature in
the quiescent state as a function of the averaged accretion rate respectively. The different lines
correspond to cooling of a low mass NS (solid line) and various enhanced cooling mechanisms for
high mass NSs. In Figures~\ref{axplum} and \ref{axptemp}, the observational data for various quiescent
LMXBs are given for reference, based on the work by Heinke et al. (2009).
\subsection{$\gamma$-ray emissions from the outer gap}
\subsubsection{$\gamma$-ray luminosity}
Using the pair-creation condition $E_{\gamma}E_X=(m_ec^2)$ with $E_X=3kT$, the relation between the
fractional gap thickness and the surface temperature is given as
\begin{equation}
f\sim 0.12P_{-3}^{7/6}\mu_{26}^{-1/2}s_*^{1/6}T_{s,6}^{-2/3},
\end{equation}
yielding a $\gamma$-ray luminosity in the quiescent state corresponding to
\begin{equation}
L_{\gamma}\sim 6\times 10^{32}P_{-3}^{-1/2}\mu_{-26}^{1/2}s_*^{1/2}T_{s.6}^{-2}
~\mathrm{erg ~s^{-1}}.
\end{equation}
Figure~\ref{glum} represents the predicted $\gamma$-ray luminosity for four LMXBs (SAX~J1808.4-3658,
XTE J0929-314, XTE J1814-328 and IGR J00291-5934) as a function of the averaged accretion rate. The
solid line (for the low mass NS) and dashed line (for the high-mass NS with
neucleon matter) represent the results for the outer gap model controlled by
the photon-photon pair-creation process between the $\gamma$-rays
and the X-ray from the full surface cooling emissions.
Furthermore, the dotted and dotted-dashed horizontal lines represent
results for the outer gap model controlled by the photon-photon pair-creation process between
the $\gamma$-ray and the X-rays from the heated polar cap region and by the magnetic pair-creation
process near the stellar surface respectively. Since the magnetic fields for the MSPs in quiescent
LMXBs have not been constrained, we present the results for two extreme cases as thick and thin
lines. For the thick lines, we assume that $\mu_{26}=1.5P_{-3}^{7/6}$, which gives an upper limit
of the magnetic field for the recently turned on the radio millisecond pulsar PSR~J1023+0038 (see
Figure~\ref{pb}). For the thin-lines, on the other hand, we estimate the dipole magnetic field from
$\mu_{26}=P_{-3}^{7/6}/3$, which describes the relation for the MSPs detected by $Fermi$-LAT.
It can be seen in Figure~\ref{glum} that the predicted $\gamma$-ray luminosity given by the solid and
dashed lines increases with decreasing time averaged accretion rates. This dependence reflects the
fact that the surface temperature decreases with a decrease of the averaged accretion rate, as
Figure~\ref{axptemp} reveals. In the present case, the fractional gap thickness is related with the
surface temperature as $f\propto T_{s,6}^{-2/3}$, indicating that the gap is thicker for lower
accretion rates. Since the $\gamma$-ray luminosity is expressed by $L_{\gamma}\sim f^3L_{sd}$, the
predicted $\gamma$-ray luminosity increases with a decrease of the averaged accretion rate. We note
that the predicted $\gamma$-ray luminosity presented in Figure~\ref{glum} is insensitive to the
spin periods of the known MSPs in the quiescent LMXBs because it is assumed that $\mu\propto
P^{7/6}$, which results in $L_{\gamma}\propto P^{1/12}$.
\subsubsection{Irradiation of $\gamma$-rays to companion star}
\label{irrad}
To explain the observed optical modulation of the companion star in quiescent state,
Takata et al. (2010a) discussed the irradiation of
$\gamma$-rays from the outer gap to the companion star. The magnetospheric $\gamma$-ray irradiating the companion star may be absorbed via the so-called pair-creation process in the Coulomb field by the nuclei in the stellar matter. Further absorption will occur as the relativistic pairs created with a Lorentz factor $\Gamma\sim 10^3$ will transfer their energy and momentum to the stellar matter via the ionization and/or the Coulomb scattering processes. The cross section of the above pair-creation process for the photon with
energy $E_{\gamma}$ is given by Lang (1999) as
\begin{equation}
\sigma(E_{\gamma})=3.5\times 10^{-3}Z^2\sigma_T
\left[\frac{7}{9}\log \left(\frac{183}{Z^{1/2}}\right)-\frac{1}{53}\right]
~~\mathrm{for}~~\frac{E_{\gamma}}{2}>>\frac{m_ec^2}{\alpha Z^{1/3}},
\end{equation}
where $Z$ is the atomic number, $\sigma_T$ is the Thomson cross section.
All $\gamma$-rays irradiating the star can be absorbed
if the column density of
the star exceeds
\begin{equation}
\Sigma>\frac{m_p}{\sigma}\sim 60~\mathrm{g~cm^{-2}},
\end{equation}
where we use $Z^2=3$ appropriate for the solar abundance. We can see that
the above condition is easily satisfied for the typical low mass companion star,
which has $\sim 0.1 {\rm ~M}_{\odot}$ and the radius $\sim 10^{10}$~cm.
The created pars will transfer their energy to the stellar matter via
exciting and ionizing atoms in the matter. For
atomic hydrogen, the energy loss rate of the pairs per unit length is given
by Lang (1999) as
\begin{equation}
\frac{dE}{dx}\sim -2.54\times 10^{-19}N_e W~\mathrm{eV~cm^{-1}}
\end{equation}
where $N_e$ is the number density of the electrons in the matter, and $W$
is a factor of 10-100. All energy of the created pairs will transfer to
the stellar material if the column density of the star exceeds
\begin{equation}
\Sigma> 70\left(\frac{E}{1\mathrm{GeV}}\right)\left(\frac{W}{100}\right)^{-1}
~\mathrm{g~cm^{-2}},
\end{equation}
which is easily satisfied for the typical companion star. Hence, all of
the irradiation energy from the outer gap will transfer to the stellar matter.
In Figure~\ref{glum}, the observational data represents the
lower limit of the irradiating luminosity $L_{irr}$ required to explain the optical modulation. We find that if the optical modulation
originates from the irradiation of
the $\gamma$-rays from the outer gap of the pulsar magnetosphere, the observations can constrain
the theoretical model. For example, the level of the inferred irradiation luminosity $L_{irr}\sim
10^{34}~\mathrm{erg s^{-1}}$ of SAX~J1808.4-3658, XTE~J1814-325 and IGR~J00291+5934 suggests that the outer gap with the high-mass NS model is preferable. The outer gap model with low mass NS
predicts, on the other hand, a lower $\gamma$-ray luminosity by an order of magnitude due to the
higher NS surface temperatures and may be applicable to J0929-314.
It should be noted that $L_{irr}\sim 10^{34}~\mathrm{erg~s^{-1}}$ cannot be produced by the outer
gap model controlled by the photon-photon pair-creation process between
the $\gamma$-rays and the X-rays from the heated polar cap region.
Within the context of the present treatment for the magnetic field determination, that is $\mu_{26}=\kappa P_{-3}^{7/6}$, where $\kappa$
is the proportional factor, the $\gamma$-ray luminosity described by equation~(\ref{plgamma}) is less
dependent on the proportional factor $\kappa$. That is, $L_{\gamma}\propto \kappa^{2/7}$ and unrealistic
values, say $\kappa\sim 1000$
are required to produce $L_{\gamma}\sim 10^{34} ~\mathrm{erg~s^{-1}}$. However, such a model would not be consistent with the
properties of the quiescent X-ray emission.
On the other hand, $L_{\gamma}\propto \kappa^2$ for the outer
gap with the magnetic pair-creation process implies that such a model can produce $L_{\gamma}\sim 10^{34}~\mathrm{erg~s^{-1}}$ with a
reasonable value of the NS magnetic field.
As indicated by the dashed-dotted lines in Figure~\ref{glum}
(also Figure~\ref{xlum}), the model prediction is sensitive to the specific relation between the spin
period and magnetic moment. Hence a measurement the possible $P-\mu$ relation for the accreting
MSPs is desired for further discussion on the origin of $\gamma$-ray emission from the MSPs in quiescent LMXBs.
\section{Magnetospheric Non-thermal X-ray emissions from MSPs }
\label{nonx}
\subsection{Rotation powered MSPs}
The $\gamma$-rays ($>$GeV) from the outer gap may be converted into pairs via the photon-photon
pair-creation process involving the thermal X-rays from the NS surface before escaping the
magnetosphere. The secondary pairs produced near the light cylinder, where the magnetic field
$B\sim 10^{5-6}$~G, will emit non-thermal X-rays via the synchrotron radiation process with a
typical energy of $E_{n,X}\sim 2(\Gamma/10^3)^2(B/10^{6}~\mathrm{G})^2
(\sin\theta/0.1)$~keV. Here,
$\Gamma_s\sim (1~\mathrm{GeV}/2m_ec^2)$ is the typical Lorentz factor of the secondary pairs, and
$\theta$ is the pitch angle. The synchrotron damping length of the secondary pairs
is $l_d\sim 1.5\times 10^{6}~(\Gamma_s/10^{3})^{-1} (B\sin\theta/10^{5})^{-2}$~cm, implying the
secondary pairs quickly lose their perpendicular momentum inside the light cylinder.
In such a case, the luminosity of the non-thermal X-rays is estimated as
\begin{equation}
L_{n,X}\sim \tau_{X\gamma}L_{\gamma},
\label{nxl}
\end{equation}
where $\tau_{X\gamma} (<1)$ is the optical depth of the photon-photon pair-creation process. Note that
because the local cyclotron energy near the light cylinder is less than the X-ray energy, the resonant
scattering is inefficient near the light cylinder, implying the non-thermal X-rays can freely escape
from the magnetosphere.
It can be seen from equations~(\ref{xlr})-(\ref{tc}) that the ratio of the photon
number density, which is proportional to $L_{r}/T_r$, of the rim component is about a factor of ten
larger than that of the core component, indicating the optical depth $\tau_{X\gamma}$ near the light
cylinder can be estimated as
\begin{equation}
\tau_{X\gamma}\sim \frac{L_{r}}{4\pi R_{lc}^2cE_{r}}\sigma_{X\gamma}
R_{lc}\sim0.014f^{3/4}P_{-3}^{-9/4}\mu^{3/4}_{26}R_{r,5}^{1/2}Q^{3/4}.
\end{equation}
Here, $E_{r}=3kT_r$ and $\sigma_{X\gamma}\sim \sigma_T/3$ with $\sigma_T$ corresponding to the
Thomson cross section. Inserting the gap fractions described by equations~(\ref{fp}) and~(\ref{fm})
into equation~(\ref{nxl}), the non-thermal X-ray luminosity can be expressed as a function of the
spin-down power as
\begin{equation}
L_{n,X}\sim 5\times 10^{29}L_{sd,35}^{45/112}\mu_{26}^{-11/56}s_*^{15/28}R^{11/7}_{r,5}
Q^{3/14}~\mathrm{erg~s^{-1}}
\label{nxp}
\end{equation}
for the outer gap model controlled by the photon-photon pair-creation process and
\begin{equation}
L_{n,X}\sim 6\times 10^{30}L_{sd,35}^{35/32}K_1^{15/4}\mu_{26}^{3/8}R^{1/2}_{r,5}
Q^{3/4}~\mathrm{erg~s^{-1}},
\label{nxm}
\end{equation}
by the magnetic pair-creation process. The model results described by equations~(\ref{nxp})
and~(\ref{nxm}) are plotted in Figure~\ref{millix} with the solid and dashed lines respectively. For the
observational data in Zavlin (2007), we represent the $Fermi$ MSPs with the filled circles and the radio MSPs with
filled triangles. We find in Figure~\ref{millix} that the outer gap model controlled by the magnetic
pair-creation process can explain the observed possible correlation between the non-thermal X-ray
luminosity and the spin down power slightly better than the model controlled by the photon-photon pair
creation process.
For the synchrotron emission from the secondary pairs, it is expected that the spectrum is described
by a photon index of $\alpha\sim 1.5$ below $E<E_{n,X}$ (Takata, Chang \& Cheng 2007). On the other
hand, the secondary particles will escape from the light cylinder with a Lorentz factor $\Gamma_{c}
\sim 5\times 10^{2}$, for which the synchrotron damping length is comparable to the light cylinder radius.
We expect that the emission of the secondary pairs beyond the light cylinder is not observed
as pulsed emission because the co-rotation motion with
central star can not be retained beyond the light cylinder.
In such a case, the observed pulsed emissions has a break
at the energy of $E_{c}\sim 400~(\Gamma_c/5\cdot
10^{2})^2(B/10^{6})(\sin\alpha/0.1)$~eV, below which the spectrum is characterized by
a photon index $\alpha=2/3$, which corresponds to the
spectrum described by the emission from the single particle. Hence,
the present model predicts that the non-pulsed X-ray emissions
from rotation powered MSPs are described by spectra with a photon index
$\alpha=2/3-1.5$, which may not be in conflict with the observed range
$\alpha\sim 1-2$ (Zavlin 2007). Furthermore, the present model predicts that
a spectral cut-off of the secondary emissions appears at
$E_{n,X}\sim 2(\Gamma/10^3)^2(B/10^{6}~\mathrm{G})^2
(\sin\theta/0.1)$~keV. If the cut-off energy position is
located in the observation energy range, the photon index fit to the observed spectrum
with the single power law function can be larger than $\alpha=1.5$.
\subsection{Accretion powered pulsars in quiescent LMXBs}
\label{lmxbnx}
The predicted relation between the non-thermal X-ray
luminosity in the quiescent state and the averaged accretion rate for LMXBs is summarized in
Figure~\ref{xlum}. The non-thermal X-ray luminosity
is calculated from $L_{n,X}\sim \tau_{X\gamma} L_{\gamma}$, where the optical depth $\tau_{X\gamma}$
is given by $\tau_{X\gamma}=L_{th}(T_s)\sigma_{X\gamma}/(4\pi R_{lc}cE_{s})$ with $E_s=3kT_s$.
The thermal X-ray luminosity $L_{th}$ and the surface temperature $T_s$ correspond to the results
represented in Figures~\ref{axplum} and~\ref{axptemp}, respectively. In addition, we calculate the
expected $\gamma$-ray luminosity $L_{\gamma}$ with $P=2.5$~ms of SAX~J1808.4-3658 for reference. Note
that the calculated X-ray luminosity is not sensitive to the spin periods in the observed range
of presently known MSPs in quiescent LMXBs. In Figure~\ref{xlum}, the solid and dashed lines
represent the non-thermal X-ray luminosity for the low-mass NS and the high mass NS models with
nucleon matter cooling processes respectively. The thick and thin lines represent the results for
magnetic fields determined from $\mu_{26}=1.5P_{-3}^{7/6}$ and $\mu_{26}=P_{-3}^{7/6}/3$ respectively.
For comparison, we plot the model predictions if the outer gap is controlled by the photon-photon
pair-creation with the X-rays from the heated polar cap (dotted lines) and by the magnetic
pair-creation process near the NS surface (dashed-dotted lines) respectively.
For the case of the outer gap controlled by the surface X-ray emission (solid and dashed lines) in
Figure~\ref{xlum}, we find that the predicted non-thermal X-ray luminosity decreases with decreasing
time averaged mass accretion rates, although the $\gamma$-ray luminosity increases as Figure~\ref{glum}
shows. This results from the fact that the number of the thermal X-ray photons, which absorb the
$\gamma$-ray photons in the outer magnetosphere, decreases with a decrease of the accretion rate, and
the decrease of the thermal X-ray luminosity is more rapid than the increase of the $\gamma$-ray
luminosity, as Figures~\ref{glum} and~\ref{xlum} show. As a result, the present non-thermal X-ray
luminosity, which is proportional to $L_{n,X}\sim \tau_{X\gamma}L_{\gamma}\propto L_{th}(<\dot{M}>)
L_{\gamma}$, decreases with the decrease of the accretion rate. It is also found in Figure~\ref{xlum}
that the non-thermal X-ray luminosity is less dependent on the accretion rate in comparison with the
$\gamma$-ray luminosity. For example, if the accretion rate changes between $10^{-13}~{\rm ~M}_{\odot}~\mathrm{
yr^{-1}}$ and $10^{-8}~{\rm ~M}_{\odot} \mathrm{yr^{-1}}$, the X-ray luminosity varies less than factor of ten
(see Figure~\ref{xlum}), whereas the $\gamma$-ray luminosity varies more than factor of ten (see
Figure~\ref{glum}).
Finally, it has been known that X-ray emission (0.5-10~keV)
from SAX~J1808.4-3658 in quiescence is well fit by
a power law spectrum with a photon index of $\alpha\sim 1.4-2$ and
with a luminosity of $L_{n,X}\sim 10^{31-32}~\mathrm{erg~s^{-1}}$ (Heinke et al.
2009). Because the pulsed period in the non-thermal
X-ray emissions has not been confirmed yet,
its origin is still unclear. As Figure~\ref{xlum} indicates,
if the non-thermal X-ray emission
originates from the magnetosphere, the outer gap model with the
low-mass X-ray NS or with the magnetic pair-creation process
for the gap closing may be more favorable, although the observational
error is large. As we discussed in section~\ref{irrad}
(c.f. Figure~\ref{glum}), the optical modulation of the companion star
observed during the quiescent state can be explained
by $\gamma$-ray irradiation from the outer gap with the high-mass
X-ray NS or with the magnetic pair-creation process for the gap closing.
For the high-mass NS model, the predicted non-thermal X-ray
luminosity associated with the outer gap may be insufficient to explain
the observational result, as Figure~\ref{xlum} shows, indicating the need for
an additional component. In this case, the non-thermal emission
originating from the interaction between the pulsar wind and the stellar wind
via an intra-binary shock may be observable,
as we will discuss in section~\ref{intra}.
\section{High-energy emissions from intra-binary shock}
\label{intra}
Stappers et al. (2003) detected the unresolved X-ray emission around a binary system
composed of the millisecond pulsar (PSR~B1957+20) and low mass star (LMS),
in which the wind of the companion star eclipses
the pulsed radio emission for $\sim 10$\% of
every orbit. They suggested that
the pulsar wind is ablating the low-mass companion star, and that
the observed unpulsed X-ray emission
from the PSR~B1957+20 binary system
originates from the interaction between the pulsar
wind and the stellar wind via an intra-binary shock
(Arons \& Tavani 1993; Cheng, Taam \& Wang 2006).
Similar binary systems, so called ``black widow'' systems, in which
the MSP is destroying a low mass companion star, has also been discovered as $Fermi$ $\gamma$-ray sources
in the Galactic field (Table~3), suggesting the existence of high energy particles in
the ``black widow'' system (Roberts et al. 2011). For AMP, the intra-binary shock in quiescent
state is also suggested to explain the observed non-thermal X-ray emission from SAX~J1808.4-3658
(e.g. Campana et al. 1998).
The high-energy emission from an intra-binary shock between
a pulsar and a high-mass companion star has been established
for so called $\gamma$-ray binary PSR~B1256/LS~2883 system,
which is composed of a canonical
radio pulsar with a period of $P\sim 48$~ms and
high mass Be star~(Tam et al. 2011; Abdo et al. 2011; Aharonian et al. 2009).
In the $\gamma$-ray binary, it has been proposed that
the shock stands at the interface between the
pulsar wind and the Be wind/disk, and unpulsed
radio to TeV radiations are produced
via the synchrotron and inverse Compton processes of the accelerated
particles at the shock (Tavani \& Arons 1997; Takata \& Taam 2009;
Kong et al. 2011).
The observed photon index in the X-ray bands from the $\gamma$-ray binary
varies with the orbital phase in the range $\alpha=1-2$.
For a pulsar/low mass star binary, Arons \& Tavani (1993) proposed the
intra-binary shock model to explain the unpulsed X-ray emissions
from the PSR~B1957+20/LMS system (see also Cheng et al. 2006). In particular,
the separation between the pulsar and the companion star
is order of $a_o\sim 10^{10-11}$~cm, which is about
2-3 orders of magnitude smaller than $a_o\sim 1$~AU
of pulsar/high mass star system, PSR~B1256/LS~2883.
Therefore, if the observed non-thermal X-ray emission from PSR~B1957+20, is produced by
an intra-binary shock, the MSP/LMS system (i.e., black widow system)
provides a unique laboratory to probe the physics of
the pulsar wind in the vicinity of the neutron star. In this section,
therefore, we discuss the emission from the intra-binary shock in
pulsar/LMS systems, and examine the predicted $\gamma$-ray and
X-ray luminosity for known black widow systems.
It has been pointed out that the synchrotron
emission from the shock caused by the interaction between
the pulsar wind and ISM contributes to the resolved X-ray nebula
around PSR~1957+20 (Stappers et al. 2003; Huang \& Becker 2007;
Cheng et al. 2006). On the other hand, we expect that the emission
from the pulsar wind nebula does not extend to $\gamma$-ray bands,
and therefore is not a candidate for the origin of
the $\gamma$-ray emission detected by the $Fermi$ observations.
The Lorentz factor of the accelerated particles at
the shock between pulsar wind
and ISM may be limited below the critical value, at which the gyroradius,
$r_g=\Gamma m_ec^2/eB$, is equal to the size of the shock. Assuming
the shock size $r_s\sim 5\times 10^{16}~$cm and the magnetic
field $B\sim 10^{-5}$~G (Cheng el al. 2006), the Lorentz factor is
limited below $\Gamma=3\times 10^{8} (B/10^{-5}\mathrm{G})
(r_s/5\cdot 10^{16}\mathrm{cm})$. This indicates that
the synchrotron spectrum extends up to the hard X-ray bands, $E_c\sim
(3h\Gamma^2 eB/4\pi m_ec)\sim 15(B/10^{-5}\mathrm{G})^3(r_s/5\cdot10^{16}
\mathrm{cm})^2$keV, implying the spectrum can not extend in $\gamma$-ray energy bands. We, henceforth, examine the high-energy emission
from an intra-binary shock and from the magnetosphere in black widow systems as possible sites for the origin
of the $\gamma$-ray emissions indicated by the $Fermi$ observations.
The distance ($r_s$) from the pulsar to the intra-binary shock
may be estimated by the pressure
balance as
\begin{equation}
\frac{L_{sd}}{4\pi r_s^2 c}=\frac{\dot{M}v_g}{4\pi f_{\omega}(a_o-r_s)^2},
\label{shock}
\end{equation}
where we assumed that the pulsar wind is emitted with the solid angle $4\pi$,
$\dot{M}$ is the mass loss rate from the star, $v_g$ is the velocity for
gaseous material, and $f_{w}$ is the outflow fraction in units of $4\pi$.
For the black widow system, PSR~B1957+20, the mass loss rate from LMS
is expected as $\dot{M}_c\sim M_{c}\dot{P}_{b}P_b^{-1}
\sim 10^{-10}M_{\odot}~\mathrm{yr^{-1}}$, where $M_{c}=0.02M_{\odot}$ is the mass of
LMS, $P_{b}=33001$~s is the orbital period and $\dot{P}_b\sim 10^{-11}$
is the orbital period derivative (Fruchter et al. 1990).
For the quiescent LMXBs, the mass loss rate will be smaller than
$\dot{M}\sim 10^{-11}-10^{-12}M_{\odot}~\mathrm{yr^{-1}}$, based on the study
by Heinke et al. (2009).
With the mass-loss rate of
$\dot{M}\sim 10^{-10}-10^{-12}M_{\odot}~\mathrm{yr^{-1}}$,
we find that the shock distance from the pulsar
is of the order of the orbital separation. For example,
if we assume the stellar
wind velocity is the escape velocity from the system,
we obtain $v_g\sim \sqrt{2GM_{\odot}/a_o}\sim 10^{8}~\mathrm{cm~s^{-1}}$, where
$a_0\sim5\times 10^{10}$~cm is the typical separation between two components.
Using $L_{sd}=10^{34}~\mathrm{erg~s^{-1}}$,
$\dot{M}=10^{-11}M_{\odot}\mathrm{yr^{-1}}$, $a_0=5\times 10^{10}~$cm, and
$f_{\omega}=1$, the equation~(\ref{shock}) implies
the shock distance of $r_s\sim 3.5\times 10^{10}~$cm. Hence,
the shock radius from the pulsar will be
of order of the separation between two components.
We assume that the kinetic energy dominated flow of the pulsar wind
is formed within the distance $r\sim a_0$ from the pulsar, that is,
the so called magnetization parameter $\sigma$, which is the
ratio of the magnetic energy to kinetic energy of the unshocked flow,
is smaller than unity, although the formation of the kinetic energy
dominated flow within $r\sim 10^{10-11}$~cm from the MSP is still of matter of
debate (Lyubarsky \& Kirk 2001). The magnetic field upstream and
behind the shock are estimated as $B_1(r)=(L_{sd}\sigma/r^2c)^{1/2}$ and
$B_2(r_s)=3B_1(r_s)$, respectively.
We assume a power law distribution of the accelerated particles at the shock,
that is, $f(\Gamma)\propto \Gamma^{-p}$ for $\Gamma_{min}\le\Gamma
\le\Gamma_{max}$. The photon index in the 0.5-10~keV band
is observed as $\alpha\sim 2$ for
PSR~B1957+20 (Huang \& Becker 2007) and
$\alpha=1.4-2$ for SAX~J1808.4-3658 (Heinke et al. 2009), implying
a power law index $p=1.8-3$, which is found in the range predicted by
the shock acceleration model (Baring 2004). We assume
that the minimum Lorentz factor is comparable to the Lorentz factor
of the bulk motion of the un-shocked flow, which is
$\Gamma_{min}\sim 10^{5}$ for $\sigma\ll 1$ (Takata \& Taam 2009).
We determine the maximum Lorentz factor of the accelerated particles as
$\Gamma_{max}=\mathrm{Min}(\Gamma_{syn},\Gamma_g)$, where
$\Gamma_{syn}$ is the Lorentz factor at which
the synchrotron cooling time scale
$\tau_s\sim9m_e^3c^5/4e^4B^2\Gamma$ is equal
to the acceleration time scale, $\tau_{acc}\sim \Gamma m_ec/eB_2$, that is,
$\Gamma_{syn}\sim 5\times 10^7L_{sd,34}^{-1/4}\sigma_{0.1}^{-1/4}r^{1/2}_{s,11}$,
where $L_{sd,34}=(L_{sd}/10^{34}~\mathrm{erg~s^{-1}})$,
$\sigma_{0.1}=(\sigma/0.1)$ and $r_{s,11}=(r_s/10^{11}\mathrm{cm})$. In addition,
$\Gamma_g$ is the Lorentz factor at which the gyroradius is equal
to the size of the system ($\sim r_s$), that is, $\Gamma_g\sim 3\times 10^{8}
L_{sd,34}^{1/2}\sigma_{0.1}^{1/2}$.
The synchrotron spectrum extends from
$E_{min}\sim 260L_{sd,34}^{1/2}\sigma_{0.1}^{1/2}r_{s,11}^{-1}$~eV to soft-$\gamma$-ray bands $E_{max}\sim 2
m_ec^2\sin\theta/8\alpha_f\sim 200~$MeV if $\Gamma_{max}=\Gamma_{syn}$,
where $\alpha_f$ is the fine structure constant, and $E_{max}\sim
35~L_{34,sd}^{3/2}\sigma_{0.1}^{3/2}r_{s,11}^{-1}$~GeV if $\Gamma_{max}=\Gamma_{g}$.
The steady
continuity equation, $\partial \dot{\Gamma}N_{tot}(\Gamma)/\partial
\Gamma=\dot{Q}$, implies
that the distribution of the total number within the radiation cavity is
expressed as $N_{tot}(\Gamma)\propto \Gamma^{-p}$ for $\Gamma_{min}
\le \Gamma\le \Gamma_c $ in the slow cooling regime
and $N_{tot}(\Gamma)\propto \Gamma^{-1-p}$ for $\Gamma_{c}\le \Gamma\le
\Gamma_{max}$ in the fast cooling regime, where $\Gamma_c$ is the Lorentz factor
at which the synchrotron cooling scale $\tau_{s}$ is equal to the dynamical
time scale $\tau_{d}\sim r_s/c$ (see Kong et al. 2011).
In addition, the normalization is determined
from the energy conservation that $\int \Gamma N(\Gamma)d\Gamma=\eta L_{sd}
\tau_d/m_ec^2$ (Arons \& Tavani 1993), where $\eta$
is the solid angle that the pulsar wind is stopped by the injected material
from the low mass star. For the case of PSR~B1957+20, the eclipse of
the pulsar ($\sim$10-20\% of the orbital period) suggests that the angle of the
mass flow from the LMS measured from the pulsar is $\theta_a\sim 0.2$.
If the pulsar wind is emitted spherically,
then $\eta\sim \pi\theta_a^2/4\sim 0.03$.
The photon index of the emission
is characterized by $\alpha=(p+1)/2$ for $E_{min}<E<E_c$ and
by $\alpha=(p+2)/2$ for $E_c<E<E_{max}$, where $E_c\sim 1.6L_{sd,34}^{-3/2}\sigma_{0.1}^{-3/2}r_{s,11}$~MeV is the characteristic
photon energy emitted by the particles with the Lorentz factor $\Gamma_c$.
The predicted luminosity in the $E_1<E<E_2$ energy band is obtained by
\begin{eqnarray}
&&L_{s}(E_1<E<E_2))\sim\int_{\Gamma_{min}}^{\Gamma_{max}}\int_{E_1}^{E_2}N_{tot}P_{syn}dEd\Gamma \nonumber \\
&\sim&2\times 10^{30}\left(\frac{\eta}{0.03}\right)
\left(\frac{\Gamma_{c}}{10^5}\right)L^2_{sd,34}
\sigma_{0.1}r_{s,11}^{-1}Q
~\mathrm{erg~s^{-1}} \mathrm{~~for~~} p>1,
\label{inx-ray}
\end{eqnarray}
where $P_{syn}$ is the synchrotron power per unit energy. In addition,
the factor $Q$ is expressed as
\[
Q=(1-p)(2-p)\frac{[\int_{\Gamma_{min}/\Gamma_c}^{1}\Gamma'^{2-p}d\Gamma'+
\int_{1}^{\Gamma_{max}/\Gamma_{c}}\Gamma'^{1-p}d\Gamma']
\int_{x_{min}}^{x_{max}} F(x)dx}{(1-p)[1-(\Gamma_{min}/\Gamma_{c})^{2-p}]-(2-p)
[(\Gamma_{max}/\Gamma_{c})^{1-p}-1]}
\]
where $\Gamma'=\Gamma/\Gamma_{min}$ and $x=E/E_s$ with
$E_s=3h\Gamma^2eB\sin\theta/4\pi m_ec$, and $x_{min}=E_1/E_s$
and $x_{max}=E_2/E_s$ and $F(x)=x\int_x^{\infty}K_{5/3}(y)dy$ with
$K_{5/3}$ being the modified Bessel function of order 5/3.
For SAX~J1808.4+3365, the spin down power is expected to be
$L_{sd}\ge 10^{34}~\mathrm{erg~s^{-1}}$ from the enhancement of
the optical emissions, and the separation between two components is
inferred as $a_0\sim 3\times 10^{10}$~cm (Chakrabarty \& Morgan, 1998).
Assuming $L_{sd}=5\times 10^{34}~\mathrm{erg~s^{-1}}$, for example,
we find that
the expected X-ray luminosity in the 0.5-10~keV energy band from equation
~(\ref{inx-ray}) can be consistent with the observation,
$L_{X}\sim 10^{31-32}~\mathrm{erg~s^{-1}}$ (Heinke et al. 2009), when the
magnetization parameter is larger than $\sigma\sim10^{-2}$.
For the black widow systems, the predicted luminosity~(\ref{inx-ray})
in the 0.5-10~keV and 0.1-10~GeV energy bands are summarized
in the seventh and eighth columns in Table~3, respectively, where the first
and the second values corresponds to the results for the
power law index $p=1.5$ and $p=3$, respectively. In addition, we assume
the separation between two components (sixth column)
as the shock distance ($r_s$). In Table~3,
we also present the $\gamma$-ray (ninth column)
and X-ray (tenth column) luminosity predicted by the outer gap model
controlled by the magnetic pair-creation process.
The predicted luminosity of the intra-binary shock emission
depends on the power law index ($p$) and magnetization parameter ($\sigma$).
In Table~3, we see that the $\gamma$-ray luminosity predicted by
the intra-binary shock changes by at least two orders of magnitude with
the power law index between $p=1.5$ and $p=3$. This dependence of the $\gamma$-ray
luminosity on the power law index is more sensitive than
that for the X-ray luminosity, since
the number of particles that emit GeV photons is very sensitive to
the photon index, while most particles emit X-rays.
Figure~\ref{sigma} summarizes the dependence of
the predicted luminosity of the shock emission in
X-ray (solid line) and $\gamma$-ray (dashed line) bands
on the magnetization parameter. Here, the results are for
PSR B1957+20 and for the power law index of $p\sim2$.
For the X-ray bands, we can see
in Figure~\ref{sigma} that the luminosity (solid line)
decreases with decreasing magnetization parameter.
This results from the fact that the X-rays are emitted by
the particles in the slow cooling regime and because
the magnetic field decreases with decreasing $\sigma$.
In the 0.1-10~GeV band, the luminosity (dashed-line) is not sensitive to
the magnetization parameter if $\sigma> 10^{-3}$.
For $\sigma> 10^{-3}$, the $\gamma$-ray photons are emitted by
both particles in the slow and fast cooling regimes.
As the magnetization parameter decreases, the magnetic field
decreases, while more particles remain in the slow cooling regime.
Since the former and latter effects tend to decrease and to
increases the emissivity, respectively,
two effects compensate each other. Below $\sigma\sim 10^{-3}$,
we can see in Figure~\ref{sigma} that the $\gamma$-ray
luminosity quickly decreases with decreasing magnetization parameter.
This is because the maximum Lorentz factor
for $\sigma <10^{-3}$ is limited
by $\Gamma_g$, at which the gyroradius is equal to size of the system,
and because the spectral cut-off of the
synchrotron radiation,
$E_{max}\sim 35~L_{34,sd}^{3/2}(\sigma/0.001)^{3/2}r_{s,11}^{-1}$~MeV,
appears below 100~MeV.
For PSR~B1957+20,
the non-pulsed X-ray luminosity is observed with
a luminosity level of $L_{X}(0.3-10\mathrm{keV})\sim 2.2\times
10^{31}~\mathrm{erg~s^{-1}}$ (Huang \& Becker 2007), which can be explained by
the intra-binary shock model if the magnetization parameter
$\sigma> 0.01$. Guillemot et al. (2011) find that
the X-ray emissions from PSR~B1957+20 are composed of the
pulsed and non-pulsed components, although detail spectral
properties of the pulsed component is still unknown.
Guillemot et al. (2011) also report the detection of the
pulsed $\gamma$-ray radiation from PSR~B1957+20 using the $Fermi$ data,
whose observed flux level implies the luminosity of order of
$L_{\gamma}\sim 1 \times 10^{34}(d/2~\mathrm{kpc})^2
(\Delta\Omega_{\gamma}/4\pi)~\mathrm{erg~s^{-1}}$ (also see Kerr et al. 2010;
Ray \& Saz Parkinson 2011). As Table~3 indicates, the predicted luminosity
level of the outer gap model (ninth column in Table~3) is found to be
consistent with the $Fermi$ observations with a typical solid angle
$\Delta\Omega_{\gamma}\sim 3$.
For other systems,
Table~3 indicates that the magnitude of the the predicted luminosity by
the intra-binary shock model lies below the results of the $Fermi$
observations, unless the power law index $p\sim 1.5$.
For the $\gamma$-ray luminosity of PSR~J1810+17,
the predictions of both the outer gap and intra-binary shock models
lie below the observation,
implying that the true spin down power may be larger than
that assumed using the relation $\mu_{26}=P^{7/6}_{-3}/3$.
Due to the uncertainties of the magnetization parameter $\sigma$ and the
power law index $p$, it is difficult to discriminate between
the intra-binary shock model and the outer gap model for the unresolved X-ray
emission from the ``black widow'' pulsar, unless the pulsed period
is detected in the data. In addition to the pulsation search,
future observations may be able to discriminate between models. For example,
a measurement of the spectral shape
in soft/hard X-ray bands and $\gamma$-ray bands
can discriminate the emission models.
The synchrotron spectrum from the intra-binary shock will
extend from $E_{min}\sim 200$~eV to $E_{max}\sim 200$~MeV with a break
at $E_c\sim 1.6L_{sd,34}^{-3/2}\sigma_{0.1}^{-3/2}r_{s,11}$~MeV,
implying the observed spectrum
can be fit by a single power law function in the soft/hard X-ray
bands and by a large photon index above $\sim100$~MeV.
For the outer gap model, on the other hand,
the spectral break will appear at $\sim$keV corresponding to the synchrotron
radiation from the secondary pairs,
as was discussed in section~\ref{nonx}, and at $\sim$GeV corresponding to the
curvature radiation in the outer gap.
The observed soft X-ray spectrum with a index $\alpha\sim 2$, such
as the X-ray emission of PSR~B1957+20 (Huang \& Becker 2007),
may support the intra-binary shock, although the possibility that
the spectrum with a cut-off energy $\sim$
keV predicted by the outer gap model can not be excluded.
Finally, we would like to remark that (i) many new
Black Widow systems will be associated with the $Fermi$ un-identified sources
and (ii) those systems will exhibit an eclipse of the radio emission.
First, because of the radio eclipse, it is likely that many Black Widow systems have been missed
by the previous radio surveys with the shorter observations.
On the other hand, the MSPs in the Black Widow systems are
younger and have higher spin-down power.
This indicates that the MSPs in the Black Window
systems have larger $\gamma$-ray luminosity,
$L_{\gamma}\sim f^3L_{sd}$ [equation~(\ref{lgamma})], than ordinary
MSPs. Accumulating data of the $Fermi$ observation will
enable us to detect the $\gamma$-ray emissions from the Black Widow systems.
In particular, the population studies (e.g. Kaaret \& Philip 1996; Faucher-Gigu$\grave{\mathrm{e}}$re \& Loeb 2010; Takata et al. 2011a,b,c)
have pointed out that unidentified MSPs will be associated
with the $\gamma$-ray sources located at higher Galactic latitude; for example, Takata et al. (2011a,b,c) argued statistically that
the distribution of the high galactic latitude of the MSPs of
the $Fermi$ un-identified sources that manifest
the spectral properties similar to the pulsars can be explained by the
distribution of the MSPs. Second, because the $\gamma$-ray emission from
the outer gap is greater in the direction perpendicular to the spin axis,
$Fermi$ is more likely to discover a greater number of MSPs
with the Earth viewing angle $\sim 90^{\circ}$ measured
from the rotation axis (Takata, et al. 2010b; Takata, et al.2011c).
If the angular momentum transferred from the accreting
matter to the neutron star in the accreting
stage produces the pulsar's spin axis perpendicular to the orbital plane,
the $\gamma$-ray emissions from MSPs in Black Widow will be
greater in the orbital plane. Hence, $Fermi$ will find the
Black Widow systems with the Earth viewing angle described by
edge-on rather than by face-on with respect
to the orbital plane. In such a case,
a greater number of the $Fermi$ Black Window systems will reveal
eclipses of the radio emissions by the matter ejected
from the companion star.
\section{Summary and conclusion}
\label{summary}
With the recent accumulation of evidence for non-thermal X-ray and $\gamma$-ray emissions
from different evolutionary stages of MSPs, we have investigated
the high-energy emission processes of isolated rotation powered
MSPs and those in binary systems. To understand their observational
properties, the high-energy emission associated with
the outer gap accelerator and with the intra-binary shock in the binary
system has been investigated.
For the $\gamma$-ray emitting MSPs, the polar cap region is heated
by incoming particles accelerated in the outer gap. These particles emit $\sim100$MeV photons
near the stellar surface which irradiate the polar cap region,
eventually impacting on the stellar surface. The former and latter heating processes are
identified with the rim component characterized by $(T_r,~r_r)\sim (7\times 10^5\mathrm{K},
~1\mathrm{km})$ and the core component $(T_c,~r_c)\sim (2\times 10^6\mathrm{K},~0.1\mathrm{km})$,
respectively. For the outer gap model, the emission properties are controlled by either the
photon-photon pair-creation process between X-rays from the heated polar cap and
the $\gamma$-rays or the magnetic pair-creation near the stellar surface. It has been found,
based on the statistical grounds, that the outer gap controlled by the magnetic pair-creation
process near the stellar surface (Takata et al. 2010b) is preferable in explaining the possible
correlations in $L_{\gamma}$ vs. $L_{sd}$ (or $\tau)$ and $L_{n, X}$ vs. $L_{sd}$.
For the AMP in the quiescent state of LMXBs, the observed modulation of the optical
emissions and/or the non-thermal X-ray emission suggests the presence of
rotation powered activities during this state. The thermal
X-ray emission at the neutron star surface resulting from deep crustal heating
can control the gap and, hence, the $\gamma$-ray emission properties. We find
that if the optical modulation originates from the irradiation of $\gamma$-rays
from the outer gap, the observed amplitude can constrain the
NS model. For example, the level of the inferred irradiation luminosity
$L_{irr}\sim 10^{34}~\mathrm{erg~s^{-1}}$ of SAX~J1808.4-3658 suggests
that the outer gap with a high-mass NS model is preferable.
As argued by Takata et al. (2010a), the presence of the outer gap emission
would be responsible for the transition of
the system from the LMXB phase to the rotation powered MSP phase.
Finally, we have discussed the high-energy emission from an intra-binary shock
in the black widow systems, which are frequently found from radio searches of
$Fermi$ unidentified sources. Within the context of the simple one-zone model,
we calculate the synchrotron emission from the accelerated particles at
the shock. For PSR~B1957+20, the observed non-pulsed X-ray
emission ($L_{X}\sim 2.2\times 10^{31}\mathrm{erg~s^{-1}}$) can
be explained by the intra-binary shock model if the magnetization
parameter is $\sigma>0.01$. In addition, it is found that the observed luminosity
of the pulsed $\gamma$-ray emission from PSR~B1957+20 can be explained
by the outer gap model. On the other hand, the 0.1-10~GeV emissions from the
intra-binary shock is several orders of magnitude smaller than that from the outer gap
emission, unless the magnetization parameter and the power law index
of the accelerated particles are $\sigma\ge 10^{-3}$ and
$p\sim 1.5$, respectively. For the other black widow systems detected
from $Fermi$ unidentified sources, the predicted
luminosity from the intra-binary shock model is consistent with
the $Fermi$ observations only
if the power law index $p\sim 1.5$. In addition to the pulse search,
the origin of the high-energy emissions from black widow systems
will be constrained by a measurement of the spectral shape
by new studies in the soft/hard X-ray bands, for example, by the Astro-H satellite, (Takahashi
et al. 2010) ) and/or by $Fermi$.
\acknowledgements
We thank A.H.Kong, C.Y.Hui, P.H.T.Tam, R.H.H.Huang,
Lupin-C.C.Lin, M.Ruderman, and S.Shibata for the useful discussions.
We express our appreciation to an anonymous referee for useful comments.
J.T. and K.S.C. are supported by a GRF grant of Hong Kong Government under
HKU700911P and R.E.T This are supported in part by
the Theoretical Institute for Advanced
Research in Astrophysics (TIARA) operated under the Academia Sinica Institute of Astronomy \&
Astrophysics in Taipei, Taiwan.
|
\section{Introduction}
Pair-instability supernovae (PISNe) are thought to occur for stars with helium cores between $\sim$64 and 133 $M_{\odot}$ \citep{Heger2002}. At zero metallicity, this corresponds to initial stellar masses between $\sim 140$ and 260 $M_{\odot}$. These enormous stellar masses may have been reached by Pop III stars, predicted to have a top-heavy mass distribution \citep{Bromm2004}. However, at lower redshifts, as the universe was enriched, Pop III stars ceased to form once the local metallicity exceeded a critical threshold $Z_{crit} \sim 10^{-3} Z_{\odot}$ \citep{Bromm2003}. Since it is almost impossible to raise the intergalactic medium metallicity in a homogeneous way \citep{Furlanetto2003, Scannapieco2003}, pristine metal-free stars will still be formed past the end of the reionization epoch $z \lesssim 6$ \citep{Trenti2009}, conceivably all the way down to $z=2.5$ \citep{Tornatore2007}. Observations have confirmed the existence of extremely metal-poor star formation at moderate redshifts of $z=3.357$ and $z=5.563$ \citep{Fosbury2003, Raiter2010}. The detectability of PISNe from Pop III stars at these moderate redshifts was investigated by \citet{Scannapieco2005a}.
Outside these surviving pristine regions, there is a wide range of observations that support an upper limit to stellar mass at $\sim 150M_{\odot}$ in our Galactic neighborhood\citep{Figer2005, Weidner2010}, preventing the formation of PISNe from Pop II/I stars (but see \citet{Crowther2010} for stars determined to be above $150M_{\odot}$ in the R136 cluster.) Nevertheless, even if very massive stars can form in metal rich regions, these radiatively supported stars are loosely bound and have strong winds driven mainly by radiation pressure through spectral lines, scaling as $\dot{M}\propto Z^{0.5\sim 0.7}$ \citep{Vink2001, Kudritzki2002}. So even Pop II/I stars with initial masses between $\sim 140$ and 260 $M_{\odot}$ will suffer copious mass loss during both the hydrogen and helium burning stages, and may not end their lives with enough mass remaining to die as PISNe; this prediction could be contested, as there are still large uncertainties in mass loss models from hot massive stars \citep{Puls2008}. Nevertheless, due to mass loss the possibility of PISNe is usually not considered for solar composition stars, even though the pair instability arises irrespective of the progenitor's metallicity.
Regardless, PISNe have very likely already been observed at low redshifts, most convincingly in the case of the very luminous and long duration event SN 2007bi \citep{Gal-Yam2009}. More recently, the Palomar Transient Factory observed a new presumed PISN, PTF 10nmn (Gal-Yam 2011, submitted; Yaron et al., in preparation), and another PISN candidate was reported by the Pan-STARRS1 survey, PS1-11ap (Kotak et al., in preparation.) Although it may be possible to explain bright events like SN 2007bi with alternative models (e.g. \citet{Woosley2007}, \citet{Moriya2010}, \citet{Kasen2010}) on the whole the observations seem to favor a scenario in which a large total mass and radioactive mass were ejected, as in a PISN explosion. The observations therefore suggest that very massive stars above the Galactic limit are formed in the local universe. The metallicities of the host galaxies of both supernovae are low but well above the maximum metallicity required to form Pop III stars \citep{Young2010}. Either pockets of pristine gas survived in the dwarf host galaxy of SN 2007bi and PTF 10nmn, or the initial mass function (IMF) of Pop II/I stars merely steepens at the very high end (instead of a hard upper limit), or there are other exotic ways to form a very massive star.
In theory, mergers of stars can form massive SN progenitors at any metallicity and circumvent the upper mass limit for Pop II/I stars at $\sim 150M_{\odot}$. The most likely environment for such mergers is a dense, young star cluster undergoing core-collapse, in which a runaway collision product can become massive enough to die as an ultra-luminous supernova. \citet{PortegiesZwart2007} first investigated this scenario for a collapsar, and \citet{Yungelson2008, Glebbeek2009, Vanbeveren2009} discussed the conditions under which the runaway collision product will end its life as a PISN.
In this paper, we calculate the number of the collision runaway merger products within dense young star clusters that lie in the PISN progenitor mass range, and show that the predicted event rate is roughly equal to the inferred rate of PISNe from the detection of SN 2007bi and PTF 10nmn in existing surveys, without requiring the supernova progenitor to be metal free. We further investigate the observability and rate of these events in the low redshift universe with the \emph{Large Synoptic Survey Telescope} (LSST).
\section{Rates from Runaway Collisions}
An appreciable fraction of stars are born in clusters; \citet{Bastian2008} found the fraction of mass that forms in clusters $>100M_{\odot}$ out of the total star formation rate to be $\Gamma \sim 8\pm 3\%$. As soon as the cluster forms, the massive stars start to sink to the cluster center due to dynamic friction, driving the cluster into a state of core collapse on a timescale of $t_{cc} \sim 0.2 t_{rh}$, where $t_{rh}$ is the relaxation time:
\begin{eqnarray}
\nonumber t_{rh} &\approx& 2 {\rm Myr} \left(\frac{r}{1 {\rm pc}}\right)^{\frac{3}{2}}\left(\frac{m}{1 M_{\odot}}\right)^{-\frac{1}{2}}\frac{N}{\log\lambda}\\
&\approx& 200 {\rm Myr} \left(\frac{r}{1 {\rm pc}}\right)^{\frac{3}{2}}\left(\frac{m}{10^6 M_{\odot}}\right)^{\frac{1}{2}}\frac{\langle m \rangle}{M_{\odot}}.
\label{RelaxationTimeEq}
\end{eqnarray}
Here $m$ is the cluster mass, $r$ is its half mass radius, $N$ is the number of stars, $\langle m \rangle = N/m \approx 0.5 M_{\odot}$ is the average stellar mass, and $\log\lambda \approx \log(0.1 N) \sim 10$ \citep{PortegiesZwart2010}. In sufficiently compact clusters, the formation of a dense central subsystem of massive stars may lead to a collision runaway, where multiple stellar mergers result in the formation of an unusually massive object \citep{Gurkan2004, Freitag2006}. This prescription is often invoked to form intermediate-mass black holes via the photodisintegration instability that collapses a super-massive star directly into a black hole.
For a successful collision runaway to occur, the star cluster must experience core collapse before the most massive stars explode as a SN ($\sim$3 Myr). For compact clusters ($t_{rh} \lesssim 100$ Myr), basically all massive stars sink to the cluster core during the runaway, and the final merged object's mass scales with the cluster mass, $m_{r} \approx 8 \times 10^{-4} m \log\lambda$ \citep{PortegiesZwart2002}. For clusters with longer relaxation times, only a portion of massive stars sink to the core in time and the merged object's mass scales as $m t_{rh}^{-1/2}$ \citep{McMillan2004}. A fitting formula for combining these scalings is given by \citet{PortegiesZwart2006}, calibrated by N-body simulations for Salpeter-like mass functions:
\begin{equation}
m_r \sim 0.01m(1+\frac{t_{rh}}{100 {\rm Myr}})^{-\frac{1}{2}}.
\label{RunawayMassEq}
\end{equation}
To get statistics on the final runaway mass $m_r$ from equations (\ref{RelaxationTimeEq}) \& (\ref{RunawayMassEq}), we need to specify the number distribution of clusters as a function of their mass $m$ and radius $r$.
The functional form of the cluster initial mass function is well represented by a \citet{Schechter1976} distribution,
\begin{equation}
\Phi(m)=\frac{dN}{dm}=Am^{-\beta}e^{-m/m_{*}},
\end{equation}
where observationally $\beta \sim 2$ \citep{Zhang1999, McCrady2007}. For Milky Way-type spiral galaxies the break mass $m_{*} \approx 2\times 10^5 M_{\odot}$ \citep{Gieles2006, Larsen2009}, whereas for interacting galaxies and luminous IR galaxies $m_{*}\gtrsim 10^6 M_{\odot}$ \citep{Bastian2008}. Our results are not sensitive to the choice of $m_{*}$.
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{PISNRateCollisionRunawayMassDistribution-New-tlimit5.eps}
\caption{Differential number distribution of the final runaway mass formed, \emph{per $1 M_{\odot}$ of stellar mass formed in all clusters}. The calculated distribution is not perfectly smooth owing to the finite number of samples in the observed radius distribution.}
\label{PISNRateCollisionRunawayMassDistribution}
\end{figure}
Several studies have discussed the lack of any clear correlation between the size of a cluster and its mass or luminosity \citep{Larsen2004, Scheepmaker2007}. Lacking a functional distribution of cluster radii, we use the empirical distribution of radii for each cluster mass bin, for observed clusters younger than $5$ Myr, compiled in Tables 2-4 of \citet{PortegiesZwart2010}. The restriction on cluster age is important, as clusters expand considerably during the first 10 Myr of their evolution. Note that this empirical construction underestimates the number of super-massive collision runaway objects ($>10^3 M_{\odot}$), as there happens to be no observed $>10^6 M_{\odot}$ clusters younger than $5$ Myr in the current sample, but this does not drastically affect our PISN rate estimates. With the joint number distribution of clusters as a function of their mass and radii, we can find the number distribution of the final mass of the runaway collision merged object, see Figure \ref{PISNRateCollisionRunawayMassDistribution}.
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{PISNRateCollisionRunawayVolumetricRate-New-tlimit5.eps}
\caption{ Predicted rate of PISN events per comoving Mpc$^3$ per year. The pair-instability SNe progenitor mass range is a major uncertainty for non-pristine stars. Here we use 140-260 $M_{\odot}$ (blue line) for metal-free Pop III stars from the models of \citet{Heger2002}, whereas the mass range of $\sim$250-800 $M_{\odot}$ (red line) is taken from \citet{Yungelson2008}, who account for increased mass loss at solar metallicity. The environments of low redshift PISNe will likely lie in between these two cases. A stronger mass loss scenario is presented by \citet{Belkus2007}, who found that PISNe progenitors can only be created at metallicities below 0.02 $Z_{\odot}$, with a mass range $\sim$300-1000 $M_{\odot}$ (black line); as the fraction of matter in $Z<0.02$ $Z_{\odot}$ galaxies is not well constrained past low redshifts, we do not plot this rate past $z=0.5$. The strongest mass loss scenarios presented by \citet{Glebbeek2009} and \citet{Vanbeveren2009} predict a PISNe rate of practically zero, so we do not plot it here. The dashed lines are the PISNe rates of \citet{Yungelson2008} and \citet{Belkus2007} adjusted for mass loss from stellar collisions. The black star shows the inferred PISN rate from current surveys.}
\label{PISNRateCollisionRunawayVolumetricRate}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{PISNRateCollisionRunawaySqDegreeRate-New-tlimit5.eps}
\caption{Number of new PISNe per deg$^2$ per unit redshift per year, for the same models as Figure \ref{PISNRateCollisionRunawayVolumetricRate}. Note that LSST is expected to cover over 20,000 deg$^2$ of sky.}
\label{PISNRateCollisionRunawaySqDegreeRate}
\end{figure}
However, as we have not taken mass loss into account in our estimate of the final runaway mass in equation (\ref{RunawayMassEq}), we artificially inflate the mass range for PISN progenitors required at the end of the last merger event, to compensate for the mass lost during the collision runaway merger sequence. For zero-metallicity Pop III stars, mass loss via line-driven winds should be negligible, and \citet{Heger2002} found the progenitor mass range to be 140-260 $M_{\odot}$; this should set the upper limit of the PISNe rate from runaway collision products.
As for Pop II/I PISNe progenitors, we caution that mass loss via
stellar winds for massive stars $M>100M_{\odot}$ is not well
understood. In fact, the observations of PISNe at low redshifts
\citep{Gal-Yam2009}, and of Type IIn SNe whose progenitors are found
to sometimes retain their hydrogen envelopes until shortly before
their explosion \citep{Smith2011}, suggest that most commonly-used
stellar mass loss models are inaccurate for very massive stars, and
likely overestimate the total mass loss, as the models do not allow
such SNe to exist at the measured metallicities. Therefore, to
account for this uncertainty, we present here various PISN progenitor
mass range scenarios described in literature, dependent on the assumed
metallicity and mass loss prescription.
\citet{Yungelson2008} studied the evolution and fate of super-massive
stars with solar metallicity from the zero-age main sequence using
detailed stellar structure models. However, instead of extrapolating
commonly used mass loss models, e.g. \citet{deJager1988, Vink2001,
Kudritzki2002}, \citet{Yungelson2008} used an ad-hoc mass-loss
prescription consistent with existing models in their relevant regimes
and more consistent with the observed Hertzsprung-Russell diagram
location and mass loss ranges found for young massive stars in
clusters in the Milky Way and the Magellanic Clouds. Notably, their
time-averaged Wolf-Rayet (WR) mass loss rate $\dot{M}_{WR}$ hardly
exceeds $10^{-4}$ $M_{\odot}$ yr$^{-1}$, which better fits
observations of hydrogen-rich WR stars that account for iron-line
blanketing and clumping in determining $\dot{M}_{WR}$
\citep{Hamann2006}, and also agrees well with $\dot{M}_{WR}$ estimates based on radio observations \citep{Cappa2004}.
On the contrary, the extrapolation of Wolf-Rayet mass loss formulas to high stellar masses given by \citet{Langer1989},
\citet{Nugis2000}, \citet{Nelemans2001}, and \citet{deDonder2003}
overestimate the mass loss rates compared with these observations.
Therefore, we use the results of \citet{Yungelson2008} as our fiducial
model. They allow the creation of PISNe progenitors at $Z\sim
Z_{\odot}$ in the initial mass range of $\sim$250-800 $M_{\odot}$;
however, they do not account for the mass loss from stellar
collisions.
Alternatively, by extrapolating theoretical mass-loss rates for
radiation-driven wind, \citet{Belkus2007} found that when the
metallicity $Z$ is between 0.001 and 0.02 $Z_{\odot}$, one may expect
PISN candidates for stars with masses from $\sim$300-1000 $M_{\odot}$;
however, at $Z>0.02$ $Z_{\odot}$ no PISNe are expected. Using the
observed galaxy luminosity-metallicity relationship
\citep{Kirby2008,Guseva2009} and the galaxy luminosity function at low
redshifts $z\sim 0.1$ \citep{Blanton2003}, we find that $\sim 0.3\%$
of stellar mass is formed in $Z \lesssim 0.02$ $Z_{\odot}$ galaxies,
and fold this factor into the predicted PISNe rate for this scenario.
In addition, \citet{Glebbeek2009} follows the evolution of the
collision product for a few merger sequences for a $m \sim 5\times
10^5 M_{\odot}$ cluster, including mass loss along the course of the
collision sequence by using the prescription of \citet{Vink2001}, and
found that above $Z=0.001$ $Z_{\odot}$, the collision runaway product
cannot die with sufficient mass to undergo a PISN. The main sequence
stellar wind mass loss rate between this work and \citet{Yungelson2008} are
similar, however, \citet{Glebbeek2009} also calculates the mass loss
from stellar collisions to be roughly $\sim 20\%$ of the total merger
product mass before mass loss. Nevertheless, the main source of
discrepancy between their conclusions is due to their very different
Wolf-Rayet mass loss rates. \citet{Glebbeek2009} implements a strong
Wolf-Rayet mass loss rate from \citet{Nugis2000} (up to $3.6\times
10^{-3}$ $M_{\odot}$ yr$^{-1}$ at $Z=0.02$), bringing the collision
product down to only $m_r \sim 10 M_{\odot}$ by the end of core
helium-burning. Using a comparable mass loss rate,
\citet{Vanbeveren2009} reaches the same conclusion that PISNe cannot
occur above $Z=0.001$ $Z_{\odot}$. Note that with these mass loss
rates, essentially no star in the low redshift universe below
$M\sim$1000 $M_{\odot}$ will end their lives as a PISN, irrespective
of the collision runaway mechanism.
To account for the $\sim 20\%$ mass loss due to unbound ejecta from
the stellar collision, we can further increase the required PISN
progenitor mass range. The new adjusted mass range for PISN
progenitors would be $\sim$313-1000 $M_{\odot}$ in the
\citet{Yungelson2008} scenario, and $\sim$375-1250 $M_{\odot}$ in the
\citet{Belkus2007} scenario.
Combining the above, we can estimate the number of collision runaway
products that have a final mass $m_r$ in the various PISN progenitor
mass range scenarios. Using the global comoving star formation rate
from \citet{Reddy2009}, we estimate the PISN rate as a function of
redshift in Figures \ref{PISNRateCollisionRunawayVolumetricRate} and
\ref{PISNRateCollisionRunawaySqDegreeRate}. If the collision runaway
mechanism is indeed responsible for creating PISNe progenitors at Pop
II/I environments in the local universe, we find that only the mass
loss prescription described by \citet{Yungelson2008} fits the current
rate of PISNe inferred from observation.
\section{Observability with LSST}
The \emph{Large Synoptic Survey Telescope} is a planned wide-field survey telescope that should begin operations at the end of this decade. It has a very wide field of view of $9.6$ deg$^2$, and 6 bands: $u, g, r, i, z$, and $y$, covering 320-1080 nm. For the most sensitive bands $g, r$, and $i$, a single visit will reach $M_{AB}=25.0$, 24.7, and 24.0 (5$\sigma$) sensitivity, respectively. These bands will be visited 10, 23, and 23 times every year during the 10 years of operation, reaching a coadded depth of $M_{AB}=$26.3, 26.4, and 25.7 per year by stacking multiple images. Note that for objects much dimmer than $\sim 22$ mag/arcsec$^2$, or $\sim 25.5$ mag/pixel for LSST, the signal will be dominated by the sky background (e.g. airglow and zodiacal light), so in this regime the limiting signal flux needed to reach a fixed signal-to-noise ratio is inversely proportional to the square root of the integration time.
We use simulated PISN light curves and spectra from \citet{Kasen2011}, who improved radiative transfer calculations by using a multi-wavelength Monte Carlo code which includes detailed line opacities. In particular, we use models He130, He100, and He80, which represent pair-instability explosions of non-rotating bare helium cores with masses 130, 100, and 80 $M_{\odot}$, respectively, as non-pristine massive stars formed via runaway collision will likely lose most of their hydrogen envelope by the end of their life. The brightest helium core model He130 peaks at around $M_{AB} \sim -22$ in the rest frame r band, and stays above $M_{AB}=-21$ for half a year, and above $M_{AB}=-20$ for almost one year. Such an event in the local universe will be easily detectable; however, the rates for PISNe from both Pop III and Pop II/I progenitors is predicted to be very low at $z\sim 0$. These rates increase at higher redshifts, but since the higher wavelength $z, y$ LSST bands are much less sensitive, the best strategy to find PISNe is to continue using the $g, r$, and $i$ bands and observe at the rest frame UV and optical luminosity of the supernovae.
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{LSST_max_redshift.eps}
\caption{Maximum redshift observable by LSST, as a function of the number of stacked images, for various PISN progenitor models. Here we use the co-added $r$ band 5$\sigma$ sensitivities, for which LSST will visit the same location 23 times every year, or once every $\sim 16$ days on average. We consider a PISN at a certain redshift as observable if it stays brighter than the limiting co-added depth for a duration longer than the time it takes to observe that number of images. $z_{max}$ eventually drops with increasing image count, as the PISN flux falls off but the sky background remains, reducing the integrated signal-to-noise. The brightest PISN will be observable with LSST out to a redshift of $\sim 1.8$.}
\label{LSST_max_redshift}
\end{figure}
Using the coadded depth sensitivities, we find that using the $r$ band is optimal for the helium core PISN models, and that we can observe the brightest He130 model out to a redshift of $z\sim 1.8$ by stacking $\sim 10$ images (see Figure \ref{LSST_max_redshift}). Below $z<2$, the PISN is visible in the $r$ band for over 1 year in the observer frame; however, at $z\geq 2$, the supernova will be too dim in the rest frame UV wavelengths being effectively probed, even though the $(1+z)$ time dilation allows more stacked images. Even if one combines data from the $g, r$, and $i$ bands over one year, and reaches a coadded depth of $M_{AB}\sim 27$, the supernova will still be too dim to be observable beyond $z=3$. Alternative PISN models where the progenitors are red supergiants which retain their hydrogen envelopes have a longer plateau in their light curves, and thus stay visible slightly longer than the helium core models. However, the conclusions are similar - in terms of instrument capability, redshifts $z<2$ are most suitable for detecting PISNe in the normal operation mode of LSST. The smaller He100 model is only visible out to $z\sim 1.2$, while even smaller progenitors are too dim to be seen beyond $z<0.4$. Combined with Figure \ref{PISNRateCollisionRunawaySqDegreeRate}, we estimate that LSST will see on the order of $\sim 10^2$ new PISNe per year that originated from the final collision runaway object in young, dense clusters.
These conclusions differ from those of \citet{Trenti2009} as well as the LSST Science Book \citep{LSSTScienceCollaborations2009}, which concluded that PISN at $z\sim 4$ will be within the capability of LSST. The difference arises because \citet{Trenti2009} approximated the PISN with a blackbody spectrum with $T_{eff}=1.5\times 10^4$K, which overestimates the rest frame UV flux compared to the spectrum obtained by the radiation hydrodynamics simulations of \citet{Kasen2011}. Also, in the LSST Science Book, when calculating that hundreds of $z=2-4$ PISNe will be detected by LSST (Chapter 11.14), the authors used $z$ and $y$ band sensitivies of $\sim 26.2$. This is unrealistic as $M_{AB} \sim 26.2$ can only be reached in the $z$ band by stacking all images over the entire 10 year lifetime of the survey, but no PISN will stay bright enough that long even with time dilation; the $y$ band is even less sensitive. Our findings suggest that, to find PISNe at $z>2$, an instrument with better infrared capabilities such as the \emph{James Webb Space Telescope}\footnote{http://www.jwst.nasa.gov/} is required.
Although stacking multiple images averages out the time variation in the supernova light curve, LSST also allows a secondary survey over a smaller area of sky, going substantially deeper in a single epoch. However, due to the steep luminosity function of PISNe, we will preferentially see only the massive PISN events beyond the local universe, so narrow, deep exposures by LSST are more useful for improving light curve coverage, instead of supernova discovery.
\section{Discussion}
Runaway collisions were explored most seriously in massive, dense clusters, so equation (\ref{RunawayMassEq}) may not be accurate for $m<10^4 M_{\odot}$. However, only more massive clusters can make runaway masses $m_r$ in the PISN progenitor mass range, so this does not affect the predicted PISN rate. In addition, initial mass segregation of stars within young clusters observed by \citet{Degrijs2002} and \citet{Stolte2006} will shorten the time to runaway collisions and increase $m_r$, but we do not take this into account.
For $z\lesssim 6$, the rate and detectability of PISN from Pop III progenitors born in surviving pockets of metal-free gas was investigated by \citet{Scannapieco2005a}. To model the PISN light curves, they used an implicit hydrodynamics code which only implements gray diffusive radiation transport; for spectra and colors they assumed a blackbody distribution. Depending on the intergalactic medium metal enrichment history, their predicted rates span two orders of magnitude, with their lower end roughly equal to our collision runaway rates at $z=1-2$. However, a PISN with a pristine host galaxy has yet to be observed.
A pilot search done using the \emph{Spitzer}/IRAC dark field found no candidates above the sensitivity limit of $M_{AB}(3.6\mu m)\sim 24$, placing an 95$\%$ confidence upper limit of 23 per deg$^2$ per year for $>1$$\mu$Jy sources with plateau timescales less than $400/(1+z)$ \citep{Frost2009}, which does not contradict the predicted rate of $<0.1$ PISN per deg$^2$ per year for our collision runaway model.
More recently, observers have discovered a class of ultra-luminous supernova, with luminosities exceeding those of the brightest pair-instability events, and rates of order $\sim 10^{-8}$ Mpc$^{-3}$ yr$^{-1}$at $z\approx 0.3$. These events do not appear be standard radioactively powered PISNe, as their luminosities are too high and their light curve durations too short \citep[e.g.,][]{Quimby2011, Chomiuk2011}. Comparing the rate of those events to that of the two putative observed PISNe, Gal-Yam found that PISNe are roughly $\sim$5 times rarer than the \citet{Quimby2011} ultra-luminous supernovae (Gal-Yam 2011, Science, submitted). This gives a PISN rate of $\sim 2\times 10^{-9}$ Mpc$^{-3}$ yr$^{-1}$ in the local universe, roughly consistent with the collision runaway rates found in Figure \ref{PISNRateCollisionRunawayVolumetricRate}.
If the collision runaway of massive stars in young, massive stellar clusters do give rise to PISNe at Pop II/I metallicities, we expect to see such a young, massive cluster at the same location, after the light of the supernova has faded away. However, even a $10^5 M_{\odot}$ cluster only has an absolute magnitude of about -8.2 mag, so the PISN will have to occur close by ($z<0.05$) for its host cluster to be observed with current telescopes. Also, these PISNe should follow the distribution of clusters, and appear in the luminous parts of their host galaxies, analogous to the position of long duration gamma-ray bursts. Note that due to the steep distribution of collision runaway masses (see Figure \ref{PISNRateCollisionRunawayMassDistribution}), the rates of PISNe from collision runaways will still be higher in environments with low metallicities, as long as mass loss for massive stars is proportional to metallicity.
Alternatively, if mass loss models are wrong and the Galactic stellar mass limit is violated, we need not invoke stellar mergers to create the massive progenitors required for the observed non-pristine PISNe. \citet{Langer2007} found that hydrogen-rich PISNe could occur at metallicities as high as $Z_{\odot}/3$, resulting in a rate of about 1 PISN per $10^3$ SNe in the $z\approx 0$ universe. For a more conservative metallicity threshold of $Z_{\odot}/10$, the rate would be about 1 PISN per $10^4$ SNe. However, even the latter is a few times higher than the current inferred rate of PISNe.
\section{Conclusion}
We have shown that the runaway collision and merger of stars in a young, dense star cluster may form the massive progenitor of a pair-instability supernova at non-zero metallicity. The volumetric rate of such events is a few times $10^{-9}$ Mpc$^{-3}$ yr$^{-1}$ in the local universe, roughly matching the inferred rate of pair-instability supernova events SN 2007bi and PTF 10nmn in ongoing surveys, both of which have a metal-poor but not metal-free host galaxy. We expect that the primary survey of the Large Synoptic Survey Telescope would see $\sim 10^2$ such events per year.
\section*{Acknowledgments.}
We thank Charlie Conroy for helpful comments. TP was supported by the Hertz Foundation. This work was supported in part by NSF grant AST-0907890 and NASA grants NNX08AL43G and NNA09DB30A.
DK is supported in part by the Director, Office of Energy
Research, Office of High Energy and Nuclear Physics, Divisions of
Nuclear Physics, of the U.S. Department of Energy under Contract No.
DE-AC02-05CH11231, and by an NSF Astronomy and Astrophysics Grant NSF-AST-1109896.
This research has been supported by the DOE
SciDAC Program (DE-FC02-06ER41438). We are grateful for computer time provided by ORNL through an INCITE award and by NERSC.
\bibliographystyle{mn2e}
|
\section{Introduction}
It may be argued that the foundation of financial mathematics consists in giving a mathematical characterization of market models satisfying certain financial axioms. This leads to so-called \emph{fundamental theorems of asset pricing}. Harrison and Pliska \cite{Harrison1981} were the first to observe that, on finite probability spaces, the absence of arbitrage opportunities (condition \emph{no arbitrage, (NA)}) is equivalent to the existence of an equivalent martingale measure. A definite version was shown by Delbaen and Schachermayer \cite{Delbaen1994}. Their result, commonly referred to as \emph{the} Fundamental Theorem of Asset Pricing, states that for locally bounded semimartingale models there exists an equivalent probability measure under which the price process is a local martingale, if and only if the market satisfies the condition \emph{no free lunch with vanishing risk (NFLVR)}. Delbaen and Schachermayer also observed that (NFLVR) is satisfied if and only if there are no arbitrage opportunities (i.e. (NA) holds), and if further it is not possible to make an unbounded profit with bounded risk (we say there are \emph{no arbitrage opportunities of the first kind}, condition \emph{(NA1)} holds). Since in finite discrete time, (NA) is equivalent to the existence of an equivalent martingale measure, it was then a natural question how to characterize continuous time market models satisfying only (NA) and not necessarily (NA1). For continuous price processes, this was achieved by Delbaen and Schachermayer \cite{Delbaen1995a}, who show that (NA) implies the existence of an \emph{absolutely continuous} local martingale measure.
Here we complement this program, by proving that for locally bounded processes, (NA1) is equivalent to the existence of a \emph{dominating} local martingale measure. Constructing dominating probability measures is rather delicate, and F\"ollmer's measure (\cite{Follmer1972}) associated to a nonnegative supermartingale appears naturally in this context.
Let us give a more precise description of the notions of arbitrage considered in this work, and of the obtained results.
Let $S$ be a $d$-dimensional stochastic process on a filtered probability space $(\Omega, \mathcal{F}, (\mathcal{F}_t)_{t \ge 0}, P)$. We assume throughout the paper that the filtration $(\mathcal{F}_t)$ is right-continuous, and that $\mathcal{F} = \mathcal{F}_\infty = \vee_{t \ge 0} \mathcal{F}_t$. We think of $S$ as the (discounted) price process of $d$ financial assets.
We should point out that the filtration $(\mathcal{F}_t)$ will \emph{not} be complete with respect to $P$. Our aim is to construct dominating measures which may charge $P$-null sets. Therefore we cannot complete the filtration by the $P$-null sets, hence we only assume that $(\mathcal{F}_t)$ is right-continuous. This means that we have to slightly deviate from the usual definition of a semimartingale. For us, a semimartingale is the sum of a local martingale and a process of finite variation on bounded intervals. However we only assume that semimartingales are almost surely (a.s.) c\`adl\`ag. A semimartingale does \emph{not} need to be c\`adl\`ag for \emph{every} $\omega \in \Omega$. Our definition follows Jacod and Shiryaev \cite{Jacod2003}, who also work with non-complete filtrations. We argue in Section \ref{sec: supermartingale densities} and Appendix \ref{app:complete filtration} that the non-completeness of our filtration will not pose any problem.
A \emph{strategy} is a predictable process $(H_t)_{t\ge 0}$ with values in $\mathbb{R}^d$. If $S$ is a semimartingale and $\lambda > 0$, then a strategy $H$ is called \emph{$\lambda$-admissible} (for $S$) if the stochastic integral $H\cdot S = \int_0^\cdot H_s \cdot dS_s$ exists and satisfies $P((H\cdot S)_t \ge -\lambda)=1$ for all $t \ge 0$. Here we write $a \cdot b = \sum_{i=1}^d a_i b_i$ for the usual inner product on $\mathbb{R}^d$.
Similarly, a \emph{simple strategy} is a process of the form $H_t = \sum_{j=0}^{n-1} F_j 1_{(T_j, T_{j+1}]}(t)$ for stopping times $0 \le T_0 < T_1 < \dots < T_n < \infty$ and bounded $\mathcal{F}_{T_j}$-measurable random variables $F_j$ with values in $\mathbb{R}^d$. If $S$ is a right-continuous adapted process, then the integral $H\cdot S$ is defined as
\begin{align*}
(H\cdot S)_t = \sum_{j \ge 0} F_j (S_{T_{j+1}\wedge t} - S_{T_j\wedge t}),
\end{align*}
and $\lambda$-admissible strategies are defined analogously to the semimartingale case.
We denote by
\begin{align}\label{eq:W1}
\mathcal{W}_1 = \{1 + (H\cdot S)_\cdot : H \text{ is a 1-admissible strategy and } (H\cdot S)_t \text{ a.s. converges as } t \rightarrow \infty \}
\end{align}
all wealth processes obtained by using 1-admissible strategies under initial wealth 1, and such that the terminal wealth is well defined. Similarly we define
\begin{align*}
\mathcal{W}_{1,s} = \{1 + (H\cdot S)_\cdot : H \text{ is a 1-admissible simple strategy} \}.
\end{align*}
Note that the convergence condition in \eqref{eq:W1} is trivially satisfied for simple strategies. We will also need
\begin{align}\label{eq:K1}
\mathcal{K}_1 = \{X_\infty : X \in \mathcal{W}_1 \} \qquad \text{and} \qquad \mathcal{K}_{1,s} = \{X_\infty : X \in \mathcal{W}_{1,s} \},
\end{align}
i.e. all terminal wealths that are attainable with initial wealth 1 and using 1-admissible strategies.
We write $L^0 = L^0(\Omega, \mathcal{F}, P)$ for the space of real-valued random variables on $(\Omega, \mathcal{F})$, where we identify random variables that are $P$-almost surely equal. We equip $L^0$ with the distance $d(X,Y) = E(|X-Y|\wedge 1)$, under which it becomes a complete metric space.
Recall that a family of random variables $\mathcal{X}$ is called \emph{bounded in probability}, or \emph{bounded in $L^0$}, if
\begin{align*}
\lim_{M \rightarrow \infty} \sup_{X \in \mathcal{X}} P(|X| \ge M) = 0.
\end{align*}
\begin{defn}
We say that a semimartingale $S$ satisfies \emph{no arbitrage of the first kind (NA1)} if $\mathcal{K}_1$ is bounded in probability. We say that $S$ satisfies \emph{no arbitrage (NA)} if there is no $X \in \mathcal{K}_1$ with $X \ge 1$ and $P(X > 1) > 0$, . If both (NA1) and (NA) hold, we say that $S$ satisfies \emph{no free lunch with vanishing risk (NFLVR)}.
Similarly we say that a right-continuous adapted process $S$ satisfies \emph{no arbitrage of the first kind with simple strategies (NA1$_s$)}, \emph{no arbitrage with simple strategies (NA$_s$)}, or \emph{no free lunch with vanishing risk with simple strategies (NFLVR$_s$)}, if $\mathcal{K}_{1,s}$ satisfies the corresponding conditions.
\end{defn}
Heuristically, (NA) says that it is not possible to make a profit without taking a risk. (NA1) says that is not possible to make an unbounded profit if the risk remains bounded. This is why (NA1) is also referred to as ``no unbounded profit with bounded risk'' (NUPBR), see for example Karatzas and Kardaras \cite{Karatzas2007}.
The main result of this paper is that for locally bounded semimartingales $S$, (NA1) is equivalent to the existence of a dominating local martingale measure. As a byproduct of the proof, we obtain that a locally bounded, right-continuous, and adapted process $S$ that satisfies (NA1$_s$) is already a semimartingale, and in this case $S$ also satisfies (NA1).
When constructing absolutely continuous probability measures, it suffices to work with random variables. In Section \ref{sec: motivation} below, we argue that dominating measures correspond to nonnegative supermartingales with strictly positive terminal values. We also show that a dominating local martingale measure corresponds to a supermartingale density in the following sense.
\begin{defn}\label{def:supermartingale density}
Let $\mathcal{Y}$ be a family of stochastic processes. A \emph{supermartingale density} for $\mathcal{Y}$ is an almost surely c\`adl\`ag and nonnegative supermartingale $Z$ with $Z_\infty = \lim_{t \rightarrow \infty} Z_t > 0$ a.s., such that $YZ$ is a supermartingale for every $Y\in \mathcal{Y}$.
If all processes in $\mathcal{Y}$ are of the form $1 + (H\cdot S)$ for suitable integrands $H$, and if $Z$ is a supermartingale density for $\mathcal{Y}$, then we will sometimes call $Z$ a supermartingale density for $S$.
\end{defn}
In the literature, supermartingale densities are usually referred to as \emph{supermartingale deflators}. We think of a supermartingale density as the ``Radon-Nikodym derivative'' $dQ/dP$ of a dominating measure $Q \gg P$. This is why we prefer the term supermartingale density.
First we give an alternative proof of a well-known result.
\begin{thm}\label{thm: ex supermartingale density}
Let $S$ be a $d$-dimensional adapted process, a.s. right-continuous (resp. a $d$-dimensional semimartingale). Then (NA1$_s$) (resp. (NA1)) holds if and only if there exists a supermartingale density for $\mathcal{W}_{1,s}$ (resp. for $\mathcal{W}_1$).
\end{thm}
As a consequence, (NA1$_s$) implies the semimartingale property for locally bounded processes.
\begin{cor}\label{cor:NA1 implies semimartingale}
Let $S$ be a $d$-dimensional adapted process, a.s. right-continuous. If every component $S^i$ of $S=(S^1,\dots, S^d)$ is locally bounded from below and if $S$ satisfies (NA1$_s$), then $S$ is a semimartingale that satisfies (NA1), and any supermartingale density for $\mathcal{W}_{1,s}$ is also a supermartingale density for $\mathcal{W}_{1}$.
\end{cor}
Given a supermartingale density $Z$ for $S$, we then apply Yoeurp's \cite{Yoeurp1985} results on F\"ollmer's measure \cite{Follmer1972}, to construct a dominating measure $Q \gg P$ associated to $Z$. We define $\gamma$ to be a right-continuous version of the density process $\gamma_t = dP/dQ|_{\mathcal{F}_t}$, and $T$ to be the first time that $\gamma$ hits zero, $T = \inf \{t \ge 0: \gamma_t = 0\}$. Set
\begin{align*}
S^{T-}_t = S_t 1_{\{t<T\}} + S_{T-} 1_{\{t \ge T\}} = S_t 1_{\{t<T\}} + \lim_{s \rightarrow T-} S_s 1_{\{t \ge T\}}.
\end{align*}
Note that $S$ and $S^{T-}$ are $P$-indistinguishable.
In the predictable case we then obtain the following result.
\begin{thm}\label{thm:predictable supermartingale density}
Let $S$ be a predictable semimartingale. If $Z$ is a supermartingale density for $\mathcal{W}_1$, then $Z$ determines a probability measure $Q \gg P$ such that $S^{T-}$ is a $Q$-local martingale. Conversely, if $Q \gg P$ is a dominating local martingale measure for $S^{T-}$, then $\mathcal{W}_1$ admits a supermartingale density.
\end{thm}
Theorem \ref{thm:predictable supermartingale density} is false if $S$ is not predictable, as we will demonstrate on a simple example. But in the non-predictable case we are able to exhibit a subclass of supermartingale densities that \emph{do} give rise to dominating local martingale measures. Conversely every dominating local martingale measure corresponds to a supermartingale density, even for processes that are not predictable. Therefore the following theorem, the main result of this paper, is valid for all locally bounded processes that are adapted and a.s. right-continuous. In the non-predictable case we build on results of \cite{Takaoka2012} that are only formulated for processes on finite time intervals. So in the theorem we let $T_\infty = \infty$ if $S$ is predictable, and $T_\infty \in (0,\infty)$ otherwise.
\begin{thm}\label{thm:main result}
Let $(S_t)_{t \in [0,T_\infty]}$ be a locally bounded, adapted process, a.s. right-continuous. Then $S$ satisfies (NA1$_s$) if and only if there exists a dominating $Q\gg P$, such that $S^{T-}$ is a $Q$-local martingale.
\end{thm}
This work is motivated by insights from the theory of filtrations enlargements. A filtration $(\mathcal{G}_t)$ is called \emph{filtration enlargement} of $(\mathcal{F}_t)$ if $\mathcal{G}_t \supseteq \mathcal{F}_t$ for all $t \ge 0$. A basic question is then under which conditions all members of a given family of $(\mathcal{F}_t)$-semimartingales are $(\mathcal{G}_t)$-semimartingales. We say that \emph{Hypoth\`{e}se $(H')$} is satisfied if \emph{all} $(\mathcal{F}_t)$-semimartingales are $(\mathcal{G}_t)$-semimartingales. Given a $(\mathcal{F}_t)$-semimartingale that satisfies (NFLVR), i.e. for which there exists an equivalent local martingale measure, one might also ask under which conditions it still satisfies (NFLVR) under~$(\mathcal{G}_t)$. It is well known, and we llustrate this in an example below, that the (NFLVR) condition is usually violated after filtration enlargements.
However it turns out that (NA1) is relatively stable under filtration enlargements. If $(\mathcal{G}_t)$ is an initial enlargement of $(\mathcal{F}_t)$, i.e. $\mathcal{G}_t = \mathcal{F}_t \vee \sigma(L)$ for some random variable $L$, then Jacod's criterion \cite{Jacod1985} is a celebrated condition that guarantees Hypoth\'ese $(H')$ to hold. We show that Jacod's criterion implies in fact the existence of a \emph{universal supermartingale density}. A strictly positive process $Z$ is called universal supermartingale density if $Z M$ is a $(\mathcal{G}_t)$-supermartingale for every nonnegative $(\mathcal{F}_t)$-supermartingale $M$. This is of course a much stronger than Hypoth\`{e}se $(H')$, and in particular it implies that every process satisfying (NA1) under $(\mathcal{F}_t)$ also satisfies (NA1) under $(\mathcal{G}_t)$.
We also show that if $(\mathcal{G}_t)$ is a general (not necessarily initial) filtration enlargement of $(\mathcal{F}_t)$, and if there exists a universal supermartingale density for $(\mathcal{G}_t)$, then a generalized version of Jacod's criterion is necessarily satisfied.
\subsection{Structure of the paper}
Section \ref{sec: motivation} describes our motivation from filtration enlargements in more detail. In Section \ref{sec: motivation} we also argue that dominating local martingale measures should correspond to supermartingale densities. In Section \ref{sec: supermartingale densities} we prove that the existence of supermartingale densities is equivalent to (NA1$_s$). In Section \ref{sec: dominating measures} we prove if $S$ is predictable, then $Z$ is a supermartingale density for $S$ if and only if $S^{T-}$ is a local martingale under the F\"ollmer measure $P^Z$. We also prove our main result, Theorem \ref{thm:main result}, for general locally bounded processes (not necessarily predictable). In Section \ref{sec: jacod criterion} we return to filtration enlargements and examine how Jacod's criterion relates to our results.
\subsection{Relevant literature}
Supermartingale densities were first considered by Kramkov and Schachermayer \cite{Kramkov1999} and Becherer \cite{Becherer2001}.
The semimartingale case of Theorem \ref{thm: ex supermartingale density} was shown by Karatzas and Kardaras \cite{Karatzas2007}. Their proof extensively uses the semimartingale characteristics of $S$, and can therefore not be applied to general processes satisfying (NA1$_s$). Note that Corollary \ref{cor:NA1 implies semimartingale} states that any locally bounded process satisfying (NA1$_s$) is a semimartingale. But for unbounded processes this is no longer true, as we shall demonstrate in a simple counterexample below. A more general result than Theorem \ref{thm: ex supermartingale density} is shown in Rokhlin \cite{Rokhlin2010}, using arguments that are related to our proof. In fact our arguments are powerful enough to imply the results of \cite{Rokhlin2010}. We were not aware of either of these works before completing our proof, and decided to keep it in the paper because we believe that it gives a nice application of \emph{convex compactness}, as introduced by Zitkovic \cite{Zitkovic2010}. Oversimplifying things a bit, one can understand convex compactness as an elegant way of formalizing convergence and compactness results that are usually shown by ad-hoc considerations based on results like Lemma A1.1 of \cite{Delbaen1994}. We also believe that our techniques may be interesting in more complicated contexts, say under transaction costs, where arbitrage considerations no longer imply the the semimartingale property of the price process.
It is well known that a locally bounded process satisfying (NA1$_s$) must be a semimartingale, see Ankirchner's thesis \cite{Ankirchner2005}, Theorem 7.4.3, and also Kardaras and Platen \cite{Kardaras2011a}. See also \cite{Delbaen1994} for a first result in this direction. This part of Corollary \ref{cor:NA1 implies semimartingale} is an immediate consequence of Theorem \ref{thm: ex supermartingale density}. We rely on \cite{Kardaras2011a} to obtain that (NA1$_s$) implies (NA1) for locally bounded processes, and that in that case supermartingale densities for $\mathcal{W}_{1,s}$ are supermartingale densities for $\mathcal{W}_1$.
Recently there has been an increased interest in F\"ollmer's measure, motivated by problems from mathematical finance. F\"ollmer's measure appears naturally in the construction and study of \emph{strict local martingales}, i.e. local martingales that are not martingales. These are used to model bubbles in financial markets, see Jarrow, Protter and Shimbo \cite{Jarrow2010}. A pioneering work on the relation between F\"ollmer's measure and strict local martingales is Delbaen and Schachermayer \cite{Delbaen1995}. Other references are Pal and Protter \cite{Pal2010} and Kardaras, Kreher and Nikeghbali \cite{Kardaras2011}. The work most related to ours is Ruf \cite{Ruf2010}, where it is shown that in a diffusion setting, (NA1) implies the existence of a dominating local martingale measure. All these works have in common that they study F\"ollmer measures of strictly positive local martingales. Carr, Fisher and Ruf \cite{Carr2011} study the F\"ollmer measure of a local martingale which is not strictly positive.
To the best of our knowledge, the current work is the first time that the F\"ollmer measure of an actual supermartingale (i.e. a supermartingale which is not a local martingale) is used as a local martingale measure. In F\"ollmer and Gundel \cite{Follmer2006}, supermartingales $Z$ are associated to ``extended martingale measures'' $P^Z$. But they define $P^Z$ to be an extended martingale measure if and only if $Z$ is a supermartingale density. This does not obiously imply that $S^{T-}$ or $S$ is a local martingale under $P^Z$ (and in general this is not true). Here we show that if $S$ is predictable, then any supermartingale density $Z$ corresponds to a dominating local martingale measure $P^Z$ - meaning that $S^{T-}$ is a local martingale under $P^Z$. For non-predictable $S$ we give a counterexample. In that case we identify a subclass of supermartingale densities that correspond to local martingale measures.
Another related work is Kardaras \cite{Kardaras2010}, where it is shown that (NA1) is equivalent to the existence of a finitely additive equivalent local martingale measure. Here we construct countably additive measures, that are however not equivalent but only dominating.
The main motivation for this work comes from the theory of filtrations enlargements, see for example Amendinger, Imkeller and Schweizer \cite{Amendinger1998}, Ankirchner's thesis \cite{Ankirchner2005}, and Ankirchner, Dereich and Imkeller \cite{Ankirchner2006}. In these works it is shown that if $M$ is a continuous local martingale in a given filtration $(\mathcal{F}_t)$, then under an enlarged filtration $(\mathcal{G}_t)$, assuming suitable conditions, $M$ is of the form $M = \tilde{M} + \int_0^\cdot \alpha_s d \langle \tilde{M} \rangle_s$, where $\tilde{M}$ is a $(\mathcal{G}_t)$-local martingale. It is then a natural question to ask whether there exists an equivalent measure $Q$ that ``eliminates'' the drift, i.e. under which $M$ is a $(\mathcal{G}_t)$-local martingale. In general the answer to this question is negative. However Ankirchner \cite{Ankirchner2005}, Theorem 9.2.7, observed that if there exists a well-posed utility maximization problem in the large filtration, then the \emph{information drift} $\alpha$ must be locally square integrable with respect to $\tilde{M}$. Here we show that for continuous processes, the square integrability of the information drift is \emph{equivalent} to the well-posedness of a utility maximization problem in the large filtration, we relate these conditions to (NA1), and we show that this allows to construct dominating local martingale measures. We also give the corresponding results for discontinuous processes.
\section{Motivation}
\label{sec: motivation}
In this section we show that under filtration enlargements, generally there exists no longer an equivalent local martingale measure.
Then we recall that as long as Jacod's condition holds, there still is a dominating local martingale measure. Finally we argue that under Jacod's condition, (NA1) is often satisfied in the large filtration. We hope that this convinces the reader that (NA1) resp. (NA1$_s$) should be in some relation to the existence of dominating local martingale measures. Assuming that a dominating local martingale measure exists, we examine its Kunita-Yoeurp decomposition under $P$, and we see that it corresponds to a supermartingale density.
\subsection*{Equivalent local martingale measures and filtration enlargements}
Consider a filtered probability space $(\Omega, \mathcal{F}, (\mathcal{F}_t)_{t\ge 0}, P)$ with $P(A) \in \{0,1\}$ for every $A \in \mathcal{F}_0$. Define $\mathcal{F}_\infty = \vee_{t\ge 0} \mathcal{F}_t$. Let $S$ be a one-dimensional semimartingale that describes a complete market (i.e. for every $F \in L^\infty(\mathcal{F}_\infty)$ there exists a predictable process $H$, integrable with respect to $S$, such that $F = F_0 + \int_0^\infty H_s dS_s$ for some constant $F_0 \in \mathbb{R}$). Let $L$ be a random variable that is $\mathcal{F}_\infty$-measurable. Assume that $L$ is not $P$-a.s. constant. Define the initially enlarged filtration
\begin{align*}
(\mathcal{G}_t = \mathcal{F}_t \vee \sigma (L): t \ge 0).
\end{align*}
This is a toy model for insider trading. At time 0, the insider has the additional knowledge of the value of $L$. Since $L$ is not constant, there exists $A \in \sigma(L)$ such that $P(A) \in (0,1)$. Assume $Q$ is an equivalent $(\mathcal{G}_t)$-local martingale measure for $S$. Consider the $(Q,(\mathcal{F}_t))$-martingale $N_t = E_Q(1_A|\mathcal{F}_t)$, $t \ge 0$. Since the market is complete, $1_A$ can be replicated. That is, there exists a $(\mathcal{F}_t)$-predictable strategy $H$ such that $N_\cdot = Q(A) + \int_0^\cdot H_s dS_s$. But then $\int_0^\cdot H_s dS_s$ is a bounded $(Q, (\mathcal{G}_t))$-local martingale. Hence it is a martingale, and since $A^c \in \mathcal{G}_0$, we obtain
\begin{align*}
0 = E_Q(1_{A^c} 1_A) = E_Q\left(1_{A^c} \left( Q(A) + \int_0^\infty H_s dS_s\right)\right) = Q(A^c) Q(A) > 0,
\end{align*}
which is absurd. The last step follows because $Q$ was assumed to be equivalent to $P$.
So already in the simplest insider trading models there may not exist an equivalent local martingale measure. If $S$ is locally bounded, then by the Fundamental Theorem of Asset Pricing at least one of the conditions (NA) or (NA1) has to be violated.
\subsection*{Jacod's criterion and dominating local martingale measures}
Let $(\mathcal{G}_t)$ be a filtration enlargement of $(\mathcal{F}_t)$, i.e. $\mathcal{F}_t \subseteq \mathcal{G}_t$ for every $t \ge 0$. Let $\mathcal{S}$ be a family of $(\mathcal{F}_t)$-semimartingales. One of the typical questions in filtration enlargements is under which conditions all $S \in \mathcal{S}$ are $(\mathcal{G}_t)$-semimartingales. \emph{Hypoth\`{e}se $(H')$} is said to be satisfied if \emph{all} $(\mathcal{F}_t)$-semimartingales are $(\mathcal{G}_t)$-semimartingales.
Jacod's criterion \cite{Jacod1985} is a famous condition that implies Hypoth\`{e}se $(H')$. Here we give an equivalent formulation, first found by F\"ollmer and Imkeller \cite{Follmer1993} and later generalized and carefully studied by Ankirchner, Dereich and Imkeller \cite{Ankirchner2007}. Let $L$ be a random variable and consider the initial enlargement $\mathcal{G}_t = \mathcal{F}_t \vee \sigma(L)$. Define the product space
\begin{align*}
\overline{\Omega} = \Omega \times \Omega, \qquad \overline{\mathcal{G}} = \mathcal{F}_\infty \otimes \sigma(L), \qquad \overline{\mathcal{G}}_t = \mathcal{F}_t \otimes \sigma(L).
\end{align*}
We define two measures on $\overline{\Omega}$. The decoupling measure $\overline{Q} = P|_{\mathcal{F}_\infty} \otimes P|_{\sigma(L)}$, and $\overline{P} = P \circ \psi^{-1}$, where $\psi: \Omega \rightarrow \overline{\Omega}$, $\psi(\omega) = (\omega, \omega)$. We then have the following result, which in this setting is just a reformulation of Jacod's criterion.
\begin{thm*}[Theorem 1 in \cite{Ankirchner2007}]
If $\overline{P} \ll \overline{Q}$, then Hypoth\`ese $(H')$ holds, i.e. any $(\mathcal{F}_t)$-semimartingale is a $(\mathcal{G}_t)$-semimartingale.
\end{thm*}
In this formulation it is quite obvious why Jacod's criterion works. Under the measure $\overline{Q}$, the additional information from $L$ is independent of $\mathcal{F}_\infty$. Therefore any $(\mathcal{F}_t)$-martingale $M$ will stay a $(\overline{\mathcal{G}}_t)$-martingale under $\overline{Q}$ (if we embed $M$ from $\Omega$ to $\overline{\Omega}$ by setting $\overline{M}_t(\omega, \omega') = M_t(\omega)$). By assumption, $\overline{Q}$ \emph{dominates} $\overline{P}$. So an application of Girsanov's theorem implies that $\overline{M}$ is a $\overline{P}$-semimartingale. But it is possible to show that any $(\overline{P}, (\overline{\mathcal{G}}_t))$-semimartingale is a $(P, (\mathcal{G}_t))$-semimartingale, which completes the argument. The message is that Jacod's criterion implies the existence of a dominating measure under which any $(\mathcal{F}_t)$-martingale is a $(\mathcal{G}_t)$-martingale.
It is not hard to see that Jacod's criterion is always satisfied if $L$ takes its values in a countable set, regardless of the structure of $(\Omega, \mathcal{F}, (\mathcal{F}_t), P)$ and $S$. So if we recall our example of an initial filtration enlargement in a complete market from above, then we observe that Jacod's criterion may be satisfied even though there is no equivalent local martingale measure in the large filtration.
\subsection*{Utility maximization and filtration enlargements}
There are many articles devoted to calculating the additional utility of an insider. Assume $S$ is a semimartingale in the large filtration $(\mathcal{G}_t)$. Then we define the set of attainable terminal wealths $\mathcal{K}_1(\mathcal{F}_t)$ and $\mathcal{K}_1(\mathcal{G}_t)$ as in \eqref{eq:K1}, using $(\mathcal{F}_t)$-predictable and $(\mathcal{G}_t)$-predictable strategies respectively.
If $S$ describes a complete market under $(\mathcal{F}_t)$, and if $(\mathcal{G}_t)$ is an initial enlargement satisfying Jacod's criterion, then it is shown in Ankirchner's thesis (\cite{Ankirchner2005}, Theorem 12.6.1, see also \cite{Ankirchner2006}), that the maximal expected logarithmic utility under $(\mathcal{G}_t)$ is given by
\begin{align*}
\sup_{X \in \mathcal{K}_1(\mathcal{G}_t)} E(\log(X)) = \sup_{X \in \mathcal{K}_1(\mathcal{F}_t)} E(\log(X)) + I(L, \mathcal{F}_\infty),
\end{align*}
where $I(L, \mathcal{F}_\infty)$ is the mutual information between $L$ and $\mathcal{F}_\infty$. This mutual information may be finite, and therefore the maximal expected utility under $(\mathcal{G}_t)$ may be finite. But we show in Proposition \ref{prop: l0-bded iff finite utility} below that finite utility and (NA1) are equivalent.
\begin{lem}\label{lem:utility concrete}
$S$ satisfies (NA1) under $(\mathcal{G}_t)$ if and only if there exists an unbounded increasing function $U$ such that the maximal expected utility is finite, i.e. such that
\begin{align*}
\sup_{X \in \mathcal{K}_1(\mathcal{G}_t)} E(U(X)) < \infty.
\end{align*}
\end{lem}
\begin{proof}
This is an immediate consequence of Proposition \ref{prop: l0-bded iff finite utility} below.
\end{proof}
In conclusion we showed that the (NFLVR) condition is not very robust with respect to filtration enlargements. Since (NFLVR) is equivalent to (NA) and (NA1), either (NA) or (NA1) must be violated after a typical filtration enlargement. We observed that the maximal expected logarithmic utility under an enlarged filtration may well be finite, and that this implies (NA1). Therefore we conclude that (NA) is the part of (NFLVR) that is less robust with respect to filtration enlargements (see Remark \ref{rmk:filtration discussion} for a more thorough discussion). Moreover in the examples where (NA1) holds, Jacod's criterion is satisfied as well. As we saw above, Jacod's criterion implies the existence of a dominating local martingale measure. Hence (NA1) seems to be related to the existence of a dominating local martingale measure. In this paper we prove that the two conditions are equivalent.
\subsection*{Supermartingale densities}
Now let us assume that we are given a dominating local martingale measure $Q \gg P$, and let us examine what type of object this gives us under $P$. We consider a fixed right-continuous filtration $(\mathcal{F}_t)$, and we assume that $S$ is a local martingale under $Q$ with $P \ll Q$. Define $\gamma$ to be the right-continuous density process, $\gamma_t = dP/dQ|_{\mathcal{F}_t}$. Then $T = \inf\{t \ge 0: \gamma_t = 0\}$ is a stopping time, and we can define the adapted process $Z_t = 1_{\{ t < T\}}/\gamma_t$. Let $H$ be 1-admissible for $S$ under $Q$, that is $Q(\int_0^{t} H_s dS_s \ge -1)=1$ for all $t \ge 0$. Let $s,t\ge 0$ and let $A \in \mathcal{F}_t$. We have
\begin{align}\label{eq:supermartingale density derivation}
E_P(1_{A} Z_{t+s}(1 + (H\cdot S)_{t+s})) & = E_Q\left( \gamma_{t+s} 1_{A} \frac{1_{\{t+s < T\}}}{\gamma_{t+s}}(1 + (H\cdot S)_{t+s})\right) \\ \nonumber
& \le E_Q\left( 1_{A} 1_{\{t < T\}}(1 + (H\cdot S)_{t+s}) \right) \\ \nonumber
& \le E_Q\left( 1_{A} 1_{\{t < T\}}(1 + (H\cdot S)_{t}) \right) \\ \nonumber
& = E_P(1_{A} Z_t (1 + (H\cdot S)_t))
\end{align}
using in the second line that $1_{A}(1+ (H\cdot S)_{t+s})$ is nonnegative, and in the third line that $1 + (H\cdot S)$ is a nonnegative $Q$-local martingale and therefore a $Q$-supermartingale. This indicates that $Z$ should be a supermartingale density. Of course here we only considered strategies that are 1-admissible under $Q$, and there might be strategies that are 1-admissible under $P$ but not under $Q$. The way to deal with this problem is to consider $S^{T-}$ rather than $S$. We will make this rigorous later.
Note that the couple $(Z,T)$ is the \emph{Kunita-Yoeurp decomposition} of $Q$ with respect to $P$. The Kunita-Yoeurp decomposition is a progressive Lebesgue decomposition on filtered probability spaces. It was introduced in Kunita \cite{Kunita1976} in a Markovian context, and generalized to arbitrary filtered probability spaces in Yoeurp \cite{Yoeurp1985}. Namely we have for every $t \ge 0$
\begin{enumerate}
\item $P(T = \infty) = 1$,
\item $Q(\cdot \cap \{T \le t\})$ and $P$ are mutually singular on $\mathcal{F}_t$,
\item $Q(\cdot \cap \{T > t\})|_{\mathcal{F}_t} \ll P|_{\mathcal{F}_t}$, and for $A \in \mathcal{F}_t$ we have $Q(A \cap \{T>t\}) = E_P(1_A Z_t)$.
\end{enumerate}
Hence our program will be to find a supermartingale density $Z$, and to construct a measure $Q$ and a stopping time $T$, such that $(Z,T)$ is the Kunita-Yoeurp decomposition of $Q$ with respect to $P$. But the second part was already solved by \cite{Yoeurp1985}, and $Q$ will be the F\"ollmer measure of $Z$. After studying the relation between $S$ and $Z$, we will see that $S^{T-}$ is a local martingale under $Q$.
Before doing so, let us prove Lemma \ref{lem:utility concrete}. This is an immediate consequence of the following de la Vall\'{e}e-Poussin type theorem for families of random variables that are bounded in $L^0$.
\begin{prop} \label{prop: l0-bded iff finite utility}
A family of random variables $\mathcal{X}$ is bounded in probability if and only if there exists a nondecreasing and unbounded function $U$ on $[0, \infty)$, such that
\begin{align*}
\sup_{X \in \mathcal{X}} E(U(|X|)) < \infty.
\end{align*}
In this case $U$ can be chosen concave and such that $U(0) = 0$.
\end{prop}
\begin{proof}
First, assume that such a $U$ exists. Then
\begin{align*}
\sup_{X \in \mathcal{X}} P(|X| \ge M) \le \sup_{X \in \mathcal{X}} P(U(|X|) \ge U(M)) \le \frac{\sup_{X \in \mathcal{X}} E(U(|X|))}{U(M)}.
\end{align*}
Since $U$ is unbounded, the right hand side converges to zero as $M$ tends to $\infty$.
Conversely, assume that $\mathcal{X}$ is bounded in probability. We need to construct a nondecreasing, unbounded, and concave function $U$ with $U(0)=0$, such that $E(U(|X|))$ is bounded for $X$ running through $\mathcal{X}$. Our construction is inspired by the proof of de la Vall\'{e}-Poussin's theorem. That is, we will construct a function $U$ of the form
\begin{align*}
U(x) = \int_0^x g(y) dy \qquad \text{where} \qquad g(y) = g_n, y \in [n-1, n)
\end{align*}
for a decreasing sequence of positive numbers $g_n$. This $U$ will be increasing, concave, $U(0) = 0$. It will be unbounded if and only if $\sum_{n=1}^\infty g_n = \infty$.
For $U$ of this form we have by monotone convergence and Fubini (all terms are nonnegative)
\begin{align*}
E(U(|X|)) & = \sum_{n=1}^\infty E(U(|X|) 1_{\{ |X| \in [n-1, n)\}} ) \le \sum_{n=1}^\infty U(n) P( |X| \in [n-1, n) ) \\
& = \sum_{n=1}^\infty \sum_{k=1}^n g_k P( |X| \in [n-1, n) ) = \sum_{k=1}^\infty \sum_{n=k}^\infty g_k P(|X| \in [n-1,n)) \\
& = \sum_{k=1}^\infty g_k P(|X| \ge k-1) \le \sum_{k=1}^\infty g_k F_\mathcal{X}(k-1),
\end{align*}
where $F_\mathcal{X}(k-1) = \sup_{X \in \mathcal{X}} P(|X| \ge k-1)$.
So the proof is complete if we can find a decreasing sequence $(g_k)$ of positive numbers, such that $\sum_{k=1}^\infty g_k = \infty$ but $\sum_{k=1}^\infty g_k F_\mathcal{X}(k-1) < \infty$. Let $n \in \mathbb{N}$. By assumption $(F_\mathcal{X}(k))$ converges to zero as $k \rightarrow \infty$, and therefore it also converges to zero in the Ces\`{a}ro sense. So we obtain for large enough $K_n$
\begin{align}\label{eq:cesaro-conv}
\frac{1}{K_n}\sum_{k=1}^{K_n} F_\mathcal{X}(k-1) \le \frac{1}{n}.
\end{align}
We choose an increasing sequence of numbers $K_n \ge n$, such that every $K_n$ satisfies \eqref{eq:cesaro-conv}. Define
\begin{align*}
g^n_k = \begin{cases}
\frac{1}{nK_n}, & k \le K_n \\
0, & k > K_n,
\end{cases}
\end{align*}
and let $n_k$ denotes the smallest $n$ for which $g^n_k \neq 0$, i.e. the smallest $n$ for which $K_n \ge k$. The sequence $(K_n)$ is increasing, and therefore $n_k \le n_{k+1}$ for all $k$. Then the sequence $(g_k)$, where
\begin{align*}
g_k = \sum_{n=1}^\infty g^n_k = \sum_{n = n_k}^\infty \frac{1}{n K_n} \le \sum_{n=n_k}^\infty \frac{1}{n^2} < \infty,
\end{align*}
is decreasing in $k$. Moreover we have by Fubini
\begin{align*}
\sum_{k=1}^\infty g_k = \sum_{k=1}^\infty \sum_{n=1}^\infty g^n_k = \sum_{n=1}^\infty \sum_{k=1}^\infty g^n_k = \sum_{n=1}^\infty \sum_{k=1}^{K_n} \frac{1}{nK_n} = \sum_{n=1}^\infty \frac{1}{n} = \infty,
\end{align*}
and at the same time we get from \eqref{eq:cesaro-conv}
\begin{align*}
\sum_{k=1}^\infty g_k F_\mathcal{X}(k-1) = \sum_{n=1}^\infty \sum_{k=1}^{K_n} \frac{F_\mathcal{X}(k-1)}{nK_n} \le \sum_{n=1}^\infty \frac{1}{n^2} < \infty,
\end{align*}
which completes the proof.
\end{proof}
\begin{rmk}
In Loewenstein and Willard \cite{Loewenstein2000}, Theorem 1, it is shown that the utility maximization problem for It\^{o} processes is well posed if and only if there is absence of a certain notion of arbitrage. They describe the critical arbitrage opportunities very precisely, and they consider more general utility maximization problems, allowing for intermediate consumption. Proposition~\ref{prop: l0-bded iff finite utility} is much simpler and more obvious, but therefore also more robust. It is applicable in virtually any context, say to discontinuous price processes that are not semimartingales, with transaction costs, and under trading constraints. The family of portfolios need not even be convex.
\end{rmk}
\begin{rmk}
Note that supermartingale densities are the dual variables in the utility maximization problem, see \cite{Kramkov1999}. So taking Proposition \ref{prop: l0-bded iff finite utility} into account, Theorem \ref{thm: ex supermartingale density} states that the utility maximization problem is non degenerate if and only if the space of dual minimizers is nonempty. This insight might also be useful in more complicated contexts, say in market with transaction costs. As a sort of meta-theorem holding for many utility maximization problems, we expect that the space of dual variables is nonempty if and only if the space of primal variables is bounded in probability.
\end{rmk}
A first corollary is that any locally bounded process satisfying (NA1$_s$) is a semimartingale.
\begin{cor}
Let $S$ be a locally bounded, c\`adl\`ag process satisfying (NA1$_s$). Then $S$ is a semimartingale.
\end{cor}
\begin{proof}
Since $\mathcal{K}_{1,s}$ is bounded in probability, Proposition \ref{prop: l0-bded iff finite utility} implies that there exists an unbounded utility function $U$ for which $\sup_{X \in \mathcal{K}_{1,s}} E(U(X))<\infty$. Theorem 7.4.3 of \cite{Ankirchner2005} then implies that $S$ is a semimartingale.
\end{proof}
This result will also be an immediate consequence of Theorem \ref{thm: ex supermartingale density}.
\section{Existence of supermartingale densities}
\label{sec: supermartingale densities}
Now let us prove Theorem \ref{thm: ex supermartingale density} Let $(\Omega, \mathcal{F}, (\mathcal{F}_t)_{t \ge 0}, P)$ be a filtered probability space with a right-continuous filtration. We do not require $(\mathcal{F}_t)$ to be complete. This goes against a long tradition in probability theory to only work with filtrations satisfying the usual conditions. The most important reasons to consider complete filtrations are that the cross-section theorem (\cite{Dellacherie1980}, 44) only holds in complete $\sigma$-algebras, and as a consequence entrance times into Borel sets are generally only stopping times with respect to complete filtrations, and that supermartingales only have c\`adl\`ag modifications in complete filtrations.
However there are at least two classical books in stochastic analysis that avoid using complete filtrations as much as possible, Jacod \cite{Jacod1979} and Jacod and Shiryaev \cite{Jacod2003}. With non-complete filtrations one can obtain results that are nearly as powerful as the ones for complete filtrations. For example every stopping time $T$ in the completed filtration $(\mathcal{F}^P_t)$ is $P$-a.s. equal to a $(\mathcal{F}_t)$-stopping time $\tilde{T}$. And it is easy to see that entrance times of right-continuous processes into open or closed sets are hitting times as long as $(\mathcal{F}_t)$ is right-continuous. If $(\mathcal{F}_t)$ is right-continuous, then any supermartingale $Z$ with right-continuous expectation $t \mapsto E(Z_t)$ has a modification that is right-continuous for \emph{every} $\omega \in \Omega$, and which $P$-a.s. has left limits, see Remark I.1.37 of \cite{Jacod2003}. Note also that in \cite{Jacod2003} stochastic integration is done for non-complete filtrations. In Appendix \ref{app:complete filtration} we moreover recall that for every $(\mathcal{F}^P_t)$-adapted process that is a.s. c\`adl\`ag there exists an indistinguishable $(\mathcal{F}_t)$-adapted process, and similar results hold for $(\mathcal{F}^P_t)$-predictable and -optional processes.
We hope that this convinces the reader that there are no problems with using non-complete filtrations. Whenever we apply a result not from \cite{Jacod1979} or \cite{Jacod2003}, we point out why it also holds under non-complete filtrations.
After becoming aware of Rokhlin's work \cite{Rokhlin2010}, we noticed that our arguments prove in fact the main result of \cite{Rokhlin2010}, which is stronger than Theorem \ref{thm: ex supermartingale density}.
A family of nonnegative stochastic processes $\mathcal{Y}$ is called \emph{fork-convex}, see \cite{Zitkovic2002} or \cite{Rokhlin2010}, if every $Y \in \mathcal{Y}$ stays in zero once it hits zero, i.e. $Y_s = 0$ implies $Y_t = 0$ for all $0\le s \le t < \infty$, and if further for all $Y^1, Y^2, Y^3 \in \mathcal{Y}$, for all $s > 0$, and for all $\mathcal{F}_s$-measurable random variables $\lambda_s$ with values in $[0,1]$, we have that
\begin{align}\label{eq:fork-convex}
Y_\cdot = 1_{[0,s)}(\cdot) Y^1_s + 1_{[s,\infty)}(\cdot) Y^1_s \left( \lambda_s \frac{Y^2_\cdot}{Y^2_s} + (1 - \lambda_s) \frac{Y^3_\cdot}{Y^3_s}\right) \in \mathcal{Y}.
\end{align}
Here and throughout the paper we interpret $0/0=0$. Note that a fork-convex family of processes with $Y_0 = 1$ for all $Y \in \mathcal{Y}$ is convex. If moreover $\mathcal{Y}$ contains the constant process 1, then $\mathcal{Y}$ is stable under stopping at deterministic times, i.e. for all $Y \in \mathcal{Y}$ and for all $t \ge 0$ also $Y_{\cdot \wedge t} \in \mathcal{Y}$.
Rokhlin's \cite{Rokhlin2010} main result is the following.
\begin{thm}\label{thm:abstract ex supermartingale density}
Let $\mathcal{Y}$ be a fork-convex family of right-continuous and nonnegative processes containing the constant process 1 and such that $Y_0=1$ for all $Y \in \mathcal{Y}$. Let
\begin{align*}
\mathcal{K} = \left\{ Y_\infty: Y \in \mathcal{Y}, Y_\infty = \lim_{t\rightarrow \infty} Y_t \text{ exists}\right\}.
\end{align*}
Then $\mathcal{K}$ is bounded in probability if and only if there exists a supermartingale density for $\mathcal{Y}$.
\end{thm}
We split up the proof in several lemmas.
\begin{lem}\label{prop: l0-bded iff Z ex}
Let $\mathcal{X}$ be a convex family of nonnegative random variables. Then $\mathcal{X}$ is bounded in probability if and only if there exists a strictly positive random variable $Z$ such that
\begin{align*}
\sup_{X \in \mathcal{X}} E(X Z) < \infty.
\end{align*}
\end{lem}
\begin{proof}
The sufficiency is Theorem 1 of \cite{Yan1980}. Note that Yan does not require the $\sigma$-algebra to be complete. Yan makes the additional assumption that $\mathcal{X}$ is contained in $L^1$. But since we are considering nonnegative random variables, this can be avoided by applying Theorem 1 of \cite{Yan1980} to the convex hull of the bounded random variables $\{X \wedge n\}$ for $n \in \mathbb{N}$, as suggested in Remark (c) of \cite{Dellacherie1980}, VIII-84.
So let us assume that $Z$ exists. Normalizing by $E(Z)$, we obtain an equivalent measure $Q$ such that $\mathcal{X}$ is norm bounded in $L^1(Q)$ and therefore bounded in $Q$-probability. Since $P\ll Q$, it is easy to see that $\mathcal{X}$ is also bounded in $P$-probability.
\end{proof}
\begin{rmk}
Convexity is necessary. Let $\{A^n_k: 1 \le k \le 2^n, n \in \mathbb{N}\}$ be an increasing sequence of partitions of $\Omega$, such that for every $n,k$ we have $P(A^n_k) = 2^{-n}$. Define the nonnegative random variables $X^n_k = 1_{A^n_k} 2^{2n}$. Then $(X^n_k: n,k)$ is bounded in probability. Let $Z$ be a nonnegative random variable such that $E(Z X^n_k) \le C$ for some $C>0$ and all $n, k$. Then
\begin{align*}
E(1_{A^n_k} Z) = E(Z X^n_k) 2^{-2n} \le C 2^{-2n}.
\end{align*}
Summing over $k$, we obtain $E(Z) \le C 2^{-n}$ for all $n$, and therefore $E(Z) = 0$. Since $Z\ge 0$, this implies $Z=0$.
\end{rmk}
We call a family of random variables \emph{$L^p$-bounded} for $p\ge 1$ if it is norm bounded in $L^p$.
\begin{rmk}
Lemma \ref{prop: l0-bded iff Z ex} states that a convex family of nonnegative random variables $\mathcal{X}$ is bounded in probability if and only if there exists a measure $Q \sim P$, such that $\mathcal{X}$ is $L^1(Q)$-bounded. One might ask if this can be improved. For example there could exist $Q\sim P$ such that $\mathcal{X}$ is $L^p(Q)$-bounded for some $p>1$. However this is not true in general. Even for a continuous martingale $M$ there might not be an absolutely continuous $Q \ll P$, such that $\mathcal{K}_1(M)$ (defined in terms of $M$ as in \eqref{eq:K1}) is uniformly integrable under $Q$. To see this, choose an increasing sequence of partitions $(A^n_k: 1 \le k \le 2^n, n \in \mathbb{N})$ of $\mathbb{R}$, such that $\nu(A^n_k) = 2^{-n}$ for all $n,k$, where $\nu$ denotes the standard normal distribution. Let $M$ be a Brownian motion. Define the random variables $X^n_k = 1_{A^n_k}(M_1) 2^n$. Then $X^n_k \in L^\infty$, and $E(X^n_k) = 1$ for all $n,k$. By the predictable representation property of Brownian motion, $X^n_k \in \mathcal{K}_1(M)$ for all $n, k$. Now let $Q \ll P$, and let $g \ge 0$ be such that $\lim_{x \rightarrow \infty} g(x) / x = \infty$. If we show that $(g(X^n_k))_{n,k}$ is unbounded in $L^1(Q)$, then de la Vall\'ee-Poussin's theorem implies that $\mathcal{K}_1(M)$ cannot be uniformly integrable under $Q$. Let $C > 0$ and let $n \in \mathbb{N}$ be such that $g(2^n) \ge C 2^{n}$. Choose $k$ for which $Q( M_1 \in A^n_k) \ge 2^{-n}$. Such a $k$ must exist because $Q$ has total mass 1. Then
\begin{align*}
E_Q(g(X^n_k)) \ge E_Q(1_{A^n_k}(M_1) 2^n C) \ge 2^{-n} 2^n C = C.
\end{align*}
Since $C>0$ was arbitrary, this shows that $E_Q(g(\cdot))$ is unbounded on $\mathcal{K}_1(M)$.
\end{rmk}
The following Lemma establishes Theorem \ref{thm:abstract ex supermartingale density} in the case of two time steps. The general case then follows easily.
\begin{lem}\label{lem: one period supermartingale}
Let $\mathcal{Y}$ be a $L^1$-bounded family of nonnegative processes indexed by $\{0,1\}$, adapted to a filtration $(\mathcal{F}_0, \mathcal{F}_1)$. Assume that $\mathcal{Y}$ is fork-convex and that $\mathcal{Y}$ contains a process of the form $(1, Y^*_1)$ for a strictly positive $Y^*_1$.
Then there exists a strictly positive $\mathcal{F}_0$-measurable random variable $Z$, such that $(Y_0 Z, Y_1)$ is a supermartingale for every $Y \in \mathcal{Y}$. $Z$ can be chosen such that for every $Y \in \mathcal{Y}$
\begin{align}\label{eq: bound on one period supermartingale}
\qquad E(Y_0 Z) \le \sup_{Y\in \mathcal{Y}} \max_{i = 0,1} E(Y_i).
\end{align}
\end{lem}
\begin{proof}
We define a nonnegative set function $\mu$ on $\mathcal{F}_0$ by setting
$\mu(A) := \sup_{Y \in \mathcal{Y}} E( 1_A Y_1/Y_0)$.
Let us apply the fork-convexity of $\mathcal{Y}$ to show that for every $Y \in \mathcal{Y}$ there exists $\tilde{Y} \in \mathcal{Y}$, such that $Y_1/Y_0 = \tilde{Y}_1$. We take $s=0$ and $Y^1 = (1, Y^*_1)$ and $Y^2 = Y$ and $\lambda_s = 1$ in \eqref{eq:fork-convex}. Then $\tilde{Y} \in \mathcal{Y}$, where
\begin{align*}
\tilde{Y}_\cdot = 1_{\{0\}}(\cdot) + 1_{\{1\}}(\cdot) \frac{Y_1}{Y_0}.
\end{align*}
In particular we can use the $L^1$-boundedness of $\mathcal{Y}$ to obtain
\begin{align*}
\mu(A) = \sup_{Y \in \mathcal{Y}} E\left( 1_A \frac{Y_1}{Y_0}\right) \le \sup_{\tilde{Y} \in \mathcal{Y}} E(1_A \tilde{Y}_1) < \infty
\end{align*}
for all $A$, i.e. we obtain that $\mu$ is finite. In fact $\mu$ is a finite measure. Let $A,B \in \mathcal{F}_0$ be two disjoint sets and let $Y^A, Y^B \in \mathcal{Y}$. We take $s=0$, $Y^1 = (1, Y^*_1)$, $Y^2 = Y^A$, $Y^3 = Y^B$, and $\lambda_s = 1_A$ in \eqref{eq:fork-convex}, which implies $\tilde{Y} \in \mathcal{Y}$, where
\begin{align*}
\tilde{Y} = 1_{\{0\}}(\cdot) + 1_{\{1\}}(\cdot) \left(1_A \frac{Y^A_1}{Y^A_0} + 1_{A^c} \frac{Y^B_1}{Y^B_0}\right).
\end{align*}
Because $A$ and $B$ are disjoint, this $\tilde{Y}$ satisfies
\begin{align*}
1_{A \cup B} \frac{\tilde{Y}_1}{\tilde{Y}_0} = 1_A \frac{Y^A_1}{Y^A_0} + 1_{B} \frac{Y^B_1}{Y^B_0}.
\end{align*}
As a consequence we obtain
\begin{align*}
\mu(A) + \mu(B) & = \sup_{(Y^A, Y^B) \in \mathcal{Y}^2} E\left(1_A \frac{Y^A_1}{Y^A_0} + 1_B \frac{Y^B_1}{Y^B_0}\right) \le \sup_{\tilde{Y} \in \mathcal{Y}} E\left(1_{A \cup B} \frac{\tilde{Y}_1}{\tilde{Y}_0}\right) = \mu(A \cup B).
\end{align*}
But $\mu(A \cup B) \le \mu(A) + \mu(B)$ is obvious, and therefore $\mu$ is finitely additive.
Now let $(A_n)$ be a sequence of disjoint sets in $\mathcal{F}_0$. Then
\begin{align*}
\mu(\cup_{n=1}^\infty A_n) & = \sup_{Y \in \mathcal{Y}} \sum_{n=1}^\infty E\left( 1_{A_n} \frac{Y_1}{Y_0}\right) \le \sum_{n=1}^\infty \sup_{Y^n \in \mathcal{Y}} E\left( 1_{A_n} \frac{Y^n_1}{Y^n_0}\right) = \sum_{n=1}^\infty \mu(A_n).
\end{align*}
The opposite inequality is easily seen to be true for any finitely additive nonnegative set function. Thus $\mu$ is a finite measure on $\mathcal{F}_0$, which is absolutely continuous with respect to $P$. Therefore there exists a nonnegative $Z \in L^1(\mathcal{F}_0, P)$, such that
\begin{align}\label{eq:radon proof1}
\mu(A) = E(1_A Z) = \sup_{Y \in \mathcal{Y}} E\left( 1_A \frac{Y_1}{Y_0} \right).
\end{align}
It is easy to see that we can replace $1_A$ in \eqref{eq:radon proof1} by any nonnegative $\mathcal{F}_0$-measurable random variable. In particular for any $Y \in \mathcal{Y}$ and $A \in \mathcal{F}_0$
\begin{align*}
E(1_A Y_0 Z) = \sup_{\tilde{Y} \in \mathcal{Y}} E\left(1_A Y_0 \frac{\tilde{Y}_1}{\tilde{Y}_0}\right) \ge E\left(1_A Y_0 \frac{Y_1}{Y_0}\right) = E(1_A Y_1),
\end{align*}
proving that $(Y_0 Z, Y_1)$ is a supermartingale as long as $E(Y_0 Z) < \infty$. But the bound stated in \eqref{eq: bound on one period supermartingale} follows immediately from the fork-convexity of $\mathcal{Y}$, because the process $\bar{Y} = (Y_0,Y_0 \tilde{Y}_1/\tilde{Y}_0)$ is in $\mathcal{Y}$ for any $\tilde{Y} \in \mathcal{Y}$, and thus
\begin{align*}
E(Y_0 Z) = \sup_{\tilde{Y} \in \mathcal{Y}} E\left(Y_0 \frac{\tilde{Y}_1}{\tilde{Y}_0}\right) \le \sup_{\bar{Y} \in \mathcal{Y}} E\left(\bar{Y}_1\right).
\end{align*}
It remains to show that $Z$ is strictly positive. But this is easy, because $(1,Y^*_1)$ is in $\mathcal{Y}$, and $Y^*_1$ is strictly positive. Therefore $(Z,Y^*_1)$ is a supermartingale with strictly positive terminal value, which is only positive if also $Z$ is strictly positive.
\end{proof}
\begin{rmk}
Some sort of stability assumption is necessary for Lemma \ref{lem: one period supermartingale} to hold. Even for a uniformly integrable and convex family of processes $\mathcal{Y}$, the proposition may fail without assuming fork convexity or a similar stability property. Let again $\{A^n_k: 1 \le k \le 2^n, n \in \mathbb{N}\}$ be an increasing sequence of partitions of $\Omega$, such that for every $n$ and $k$ we have $P(A^n_k) = 2^{-n}$. Define the random variables $Y^n_k = 1_{A^n_k} 2^n/n$. Let $M > 1$ be such that $2^{n_0-1} < M \le 2^{n_0}$. Then
\begin{align*}
\sup_{n,k} E(|Y^n_k| 1_{\{|Y^n_k| \ge M\}}) \le E(|Y^{n_0}_1|) = \frac{1}{n_0},
\end{align*}
proving that $(Y^n_k)_{n,k}$ is uniformly integrable. From de la Vall\'{e}e-Poussin's theorem and Jensen's inequality we obtain that also the convex hull $\mathcal{C}$ of the $Y^n_k$ is uniformly integrable. Define $\mathcal{F}_0 = \mathcal{F}_1 = \sigma(A^n_k: 1 \le k \le 2^n, n \in \mathbb{N})$, and $\mathcal{Y} = \{(1,Y): Y \in \mathcal{C}\}$. Assume there exists $Z > 0$ such that $E(1_A Y) \le E(1_A Z)$ for all $A \in \mathcal{F}_0$ and $Y \in \mathcal{C}$. Then for every $n \in \mathbb{N}$
\begin{align*}
E(Z) = \sum_{k=1}^{2^n} E(1_{A^n_k} Z) \ge \sum_{k=1}^{2^n} E(1_{A^n_k} Y^n_k) = \frac{2^n}{n}.
\end{align*}
so that $E(Z) = \infty$. Therefore $(Z, Y)$ cannot be a supermartingale for any $Y \in \mathcal{C}$. In fact it is possible to show that $E(1_{A^n_k} Z) = \infty$ for all $k,n$, and since $\mathcal{F}_0$ is generated by $(A^n_k)_{n,k}$, this implies $P(Z = \infty) = 1$.
\end{rmk}
\begin{cor}\label{cor: discrete time supermartingale}
Let $\mathcal{Y}$ be a $L^1$-bounded family of nonnegative processes indexed by $\{0, \dots, n\}$, adapted to a filtration $(\mathcal{F}_k: 0 \le k \le n)$. Assume that $\mathcal{Y}$ is fork-convex, and that it contains the constant process $(1, \dots, 1)$.
Then there exists a strictly positive and adapted process $(Z_k: 0 \le k \le n)$, with $Z_n=1$ and such that $ZY$ is a supermartingale for every $Y \in \mathcal{Y}$. $Z$ can be chosen such that that every $Y \in \mathcal{Y}$
\begin{align}\label{eq: bound on multiperiod supermartingale}
\max_{k=0\dots, n} E(Z_k Y_k) \le \sup_{Y\in \mathcal{Y}} \max_{k = 0,\dots, n} E(Y_k).
\end{align}
\end{cor}
\begin{proof}
We prove the result by induction. For $n=1$, this is just Lemma \ref{lem: one period supermartingale}. Take $Z_0 = Z, Z_1 = 1$.
Now assume the result holds for $n$. Let $\mathcal{Y}$ be a family of processes indexed by $\{0,\dots, n+1\}$, and assume that $\mathcal{Y}$ satisfies all requirements stated above. Then also $(Y_1, \dots, Y_{n+1})$ satisfies all those requirements. By induction hypothesis, there exists a strictly positive and adapted process $Z = (Z_1, \dots, Z_{n},1)$ such that $ZY$ is a supermartingale for all $Y\in \mathcal{Y}$, and such that
\begin{align}\label{eq:multiperiod proof1}
\max_{k=1\dots, n+1} E(Z_k Y_k) \le \sup_{Y\in \mathcal{Y}} \max_{k = 1,\dots, n+1} E(Y_k).
\end{align}
Therefore it suffices to construct a suitable $Z_0$. This is achieved by applying Lemma \ref{lem: one period supermartingale} to the family of processes $\tilde{\mathcal{Y}}=\{(Y_0, Z_1 Y_1): Y \in\mathcal{Y}\}$. Since $\mathcal{Y}$ contains the constant process $(1,\dots,1)$, $\tilde{\mathcal{Y}}$ contains the process $(1, Z_1)$, and $Z_1$ is strictly positive. Furthermore it is straightforward to check that $\tilde{Y}$ is fork-convex. $L^1$-boundedness of $\tilde{Y}$ follows from \eqref{eq:multiperiod proof1}. Hence we can apply Lemma \ref{lem: one period supermartingale} to $\tilde{Y}$, and the result follows.
\end{proof}
To prove Theorem \ref{thm:abstract ex supermartingale density}, we have to go from finite discrete time to continuous time. This is achieved by means of a compactness argument. Compactness for right-continuous functions is not very easy to show, and it would require us to use some form of the Arzel\`a-Ascoli theorem. However we want to construct a \emph{supermartingale} $Z$, and therefore it will be sufficient to construct its ``skeleton'' $(Z_q: q \in \mathbb{Q}_+)$. Using standard results for supermartingales, we can then use this skeleton to construct a right-continuous supermartingale density.
We will need the notion of convex compactness as introduced by Zitkovic \cite{Zitkovic2010}.
\begin{defn}
Let $X$ be a topological vector space. A closed convex subset $C \subseteq X$ is called \emph{convexly compact} if for any family $\{F_\alpha: \alpha \in A\}$ of closed convex subsets of $C$, we can only have $\cap_{\alpha \in A} F_\alpha = \emptyset$ if there exist already finitely many $\alpha_1,\dots, \alpha_n \in A$ for which $\cap_{i=1}^n F_{\alpha_i} = \emptyset$.
\end{defn}
Recall that $L^0$ is the space of real valued random variables, equipped with the topology of convergence in probability. Zitkovic \cite{Zitkovic2010} then characterizes convexly compact sets of nonnegative elements of $L^0$.
\begin{lem}[Theorem 3.1 of \cite{Zitkovic2010}]\label{prop: zitkovic}
Let $\mathcal{X}$ be a convex set of nonnegative random variables, closed with respect to convergence in probability. Then $\mathcal{X}$ is convexly compact in $L^0$ if and only if it is bounded in probability.
\end{lem}
Note that Zitkovic works on a complete probability space. But completeness is not used in the proof of Theorem 3.1. There is only one point in the proof where it is not immediately clear whether completeness of the $\sigma$-algebra is needed: when Lemma A1.2 of \cite{Delbaen1994} is applied. However this lemma is formulated for general probability spaces.
In Proposition \ref{prop: tychonoff result} we prove a Tychonoff theorem for countable families of convexly compact subsets of metric spaces. This will be used in the following proof.
\begin{lem}\label{lem: Q indexed supermartingale}
Let $\mathcal{Y}$ and $\mathcal{K}$ be as in Theorem \ref{thm:abstract ex supermartingale density}. Then there exists a nonnegative supermartingale $(\bar{Z}_q)_{q \in \mathbb{Q}_+ \cup\{\infty\}}$ with $\bar{Z}_\infty > 0$, such that $(\bar{Z}_q Y_q)_{q \in \mathbb{Q}_+}$ is a supermartingale for all $Y \in \mathcal{X}$.
\end{lem}
\begin{proof}
Recall that $\mathcal{Y}$ is convex, and therefore $\mathcal{K}$ is convex as well.
By Lemma \ref{prop: l0-bded iff Z ex} there exists $Q \sim P$ such that $\mathcal{K}$ is $L^1(Q)$-bounded. We define $C = \sup_{X \in \mathcal{K}} E_Q(X)$ and introduce the family of processes
\begin{align*}
\mathcal{C} = \{(Z_q)_{q \in \mathbb{Q}_+ \cup\{\infty\}}: Z_\infty = 1, Z_q \ge 0, Z_q \in \mathcal{F}_q \text{ and } E_Q(Z_q)\le C \text{ for all } q\}.
\end{align*}
By Lemma \ref{prop: zitkovic} and Proposition \ref{prop: tychonoff result}, this is a convexly compact set in $\prod_{q \in \mathbb{Q}_+ \cup \{ \infty \}} L^0(\mathcal{F}_q, Q)$ equipped with the product topology (where every $L^0(\mathcal{F}_q, Q)$ is equipped with its usual topology). Define for given $q,r \in \mathbb{Q}_+ \cup\{\infty\}$
\begin{align*}
\mathcal{C}(q,r) = \{Z \in \mathcal{C}: E_Q(Z_{q+r} Y_{q+r}/Y_q|\mathcal{F}_q) \le Z_q \text{ for all } Y \in \mathcal{Y}\},
\end{align*}
where for $r=\infty$ it is understood that we only consider those $Y \in \mathcal{Y}$ for which $Y_\infty = \lim_{t\rightarrow \infty} Y_t$ exists. The sets $\mathcal{C}(q,r)$ are convex, and by Fatou's lemma they are also closed. Furthermore they are subsets of the convexly compact set $\mathcal{C}$.
So if
\begin{align*}
\cap_{q \in \mathbb{Q}_+, r \in \mathbb{Q}_+ \cup \{\infty\}}\mathcal{C}(q,r)
\end{align*}
was empty, then already a finite intersection would have to be empty. But if a finite intersection was empty, then there would be some $0 \le t_0< \dots < t_n \le \infty$ for which it is impossible to find $(Z_{t_i}: i=0,\dots, n)$ with $Z_{t_n}=1$ and such that $(Y_{t_i} Z_{t_i}: i = 0,\dots, n)$ is a $Q$-supermartingale for every $Y \in \mathcal{Y}$ (respectively for every $Y \in \mathcal{Y}$ for which $\lim_{t \rightarrow \infty} Y_t$ exists in case $t_n = \infty$). This would contradict Corollary \ref{cor: discrete time supermartingale}.
So let $Z$ be in the intersection of all $\mathcal{C}(q,r)$ and let $Y \in \mathcal{Y}$. Then for any $q, r \in \mathbb{Q}_+$
\begin{align*}
E_Q\left(\left.Z_{q+r} \frac{Y_{q+r}}{Y_q}\right|\mathcal{F}_q\right) \le Z_q,
\end{align*}
which shows that $ZY$ is a $Q$-supermartingale on $\mathbb{Q}_+$. Taking $Y\equiv 1$, we also see that $(Z_q: q \in \mathbb{Q}_+ \cup \{\infty\})$ is a supermartingale. To complete the proof it suffices now to define for $q \in \mathbb{Q}_+ \cup \{\infty\}$
\begin{align*}
\bar{Z}_q = Z_q \left.\frac{dP}{dQ}\right|_{\mathcal{F}_q}.
\end{align*}
\end{proof}
We are now ready to prove Rokhlin's result.
\begin{proof}[Proof of Theorem \ref{thm:abstract ex supermartingale density}]
It remains to show that given the skeleton $(\bar{Z}_q: q \in \mathbb{Q}_+ \cup \{\infty\})$ we can construct a right-continuous supermartingale density with left limits almost everywhere. This is a standard result on supermartingales. For the reader's convenience and to dispel possible concerns about the non-completeness of our filtration, we give the arguments below.
Since $(\mathcal{F}_t)$ is right-continuous, for every $t\ge 0$ there exists a null set $N_t \in \mathcal{F}_t$, such that for $\omega \in \Omega \backslash N_t$
\begin{align*}
\lim_{\substack{r\rightarrow s-\\r \in \mathbb{Q}}} \bar{Z}(\omega)_r \text{ and } \lim_{\substack{r\rightarrow s+\\r \in \mathbb{Q}}} \bar{Z}(\omega)_r
\end{align*}
exist for every $s \le t$. See for example Ethier and Kurtz \cite{Ethier1986}, right before Proposition 2.2.9. We define for $t \in [0,\infty)$
\begin{align*}
Z_t(\omega) = \begin{cases}
\lim_{\substack{s\rightarrow t+\\s \in \mathbb{Q}}} \bar{Z}_s(\omega), & \omega \in \Omega\backslash N_t \\
0, & \text{otherwise}.
\end{cases}
\end{align*}
Then $Z$ is adapted because $(\mathcal{F}_t)$ is right-continuous, and $Z$ is right-continuous by definition. However it may not have left limits everywhere. But outside of the null set $N = \cup_{n \in \mathbb{N}} N_n$ it has left limits at every $t > 0$.
Let us show that $ZY$ is a supermartingale for every $Y \in \mathcal{Y}$. Recall that the processes in $\mathcal{Y}$ are right-continuous. Using Fatou's Lemma in the first step and Corollary 2.2.10 of \cite{Ethier1986} in the second step, we obtain
\begin{align*}
E_Q(Z_{t+s} Y_{t+s}|\mathcal{F}_t) &\le \liminf_{\substack{r\rightarrow (t+s)+\\r \in \mathbb{Q}}} E_Q(\bar{Z}_r Y_r |\mathcal{F}_t) = \liminf_{\substack{r\rightarrow (t+s)+\\r \in \mathbb{Q}}} \liminf_{\substack{u\rightarrow t+\\u \in \mathbb{Q}}} E_Q(\bar{Z}_r Y_r |\mathcal{F}_u) \\
&\le \liminf_{\substack{r\rightarrow (t+s)+\\r \in \mathbb{Q}}} \liminf_{\substack{u\rightarrow t+\\u \in \mathbb{Q}}} \bar{Z}_u Y_u = Z_t Y_t.
\end{align*}
The same arguments with $s=\infty$ and $Y=1$ show that if we set $Z_\infty = \bar{Z}_\infty$, then $(Z_t)_{t\in [0,\infty]}$ is a nonnegative supermartingale with strictly positive terminal value. By Theorem I.1.39 of \cite{Jacod2003}, $Z_t$ almost surely converges to a limit $\tilde{Z}_\infty$ as $t\rightarrow \infty$. Theorem VI-6 of \cite{Dellacherie1980} now implies $\tilde{Z}_\infty \ge Z_\infty > 0$. Since \cite{Dellacherie1980} work with complete filtrations, let us provide the argument. Define $M_t = E(Z_\infty |\mathcal{F}_t)$ for $t \in [0,\infty]$. This is a uniformly integrable martingale which almost surely converges to $Z_\infty$ as $t \rightarrow \infty$. By the supermartingale property of $Z$ we have $M_t \le Z_t$ for all $t \ge 0$. Thus
\begin{align*}
Z_\infty = \lim_{t\rightarrow \infty} M_t \le \lim_{t \rightarrow \infty} Z_t = \tilde{Z}_\infty.
\end{align*}
It remains to show necessity. If $Z$ is a supermartingale density for $\mathcal{Y}$, then for any $Y \in \mathcal{Y}$, $Z_t Y_t$ converges as $t\rightarrow \infty$, see Theorem I.1.39 of \cite{Jacod2003}. Since $Z_t$ converges to a strictly positive limit, $Y_t$ must converge as well, and we have $E(Z_\infty Y_\infty) \le E(Z_0 Y_0) = E(Z_0)$. Now Lemma \ref{prop: l0-bded iff Z ex} implies that $\mathcal{Y}$ is bounded in probability.
\end{proof}
\begin{cor}\label{cor:abstract semimartingale}
If $\mathcal{Y}$ is as in Theorem \ref{thm:abstract ex supermartingale density}, then every $Y \in \mathcal{Y}$ is a semimartingale for which $Y_t$ almost surely converges as $t \rightarrow \infty$.
\end{cor}
\begin{proof}
Convergence was shown in the proof of Theorem \ref{thm:abstract ex supermartingale density}. The semimartingale property follows from It\^{o}'s formula. Let $Z$ be a supermartingale density for $\mathcal{Y}$. Then $Z$ is strictly positive, and therefore $1/Z$ is a semimartingale, implying that $Y = (1/Z) (ZY)$ is a semimartingale.
\end{proof}
In case $\mathcal{Y} = \mathcal{W}_1$ and under the stronger assumption (NFLVR$_s$), Corollary \ref{cor:abstract semimartingale} was already shown in \cite{Delbaen1994}.
\begin{proof}[Proof of Theorem \ref{thm: ex supermartingale density}]
It suffices to note that $\mathcal{W}_1$ and $\mathcal{W}_{1,s}$ satisfy the assumptions of Theorem \ref{thm:abstract ex supermartingale density}. This is easy and shown for example in Rokhlin \cite{Rokhlin2010}, in the proof of Theorem 2. Rokhlin only treats the case of $\mathcal{W}_1$ and $\mathcal{K}_1$, but the same arguments also work for $\mathcal{W}_{1,s}$ and $\mathcal{K}_{1,s}$.
\end{proof}
\begin{proof}[Proof of Corollary \ref{cor:NA1 implies semimartingale}]
Let $S$ have components that are locally bounded from below and assume that $S$ satisfies (NA1$_s$). Recall that local semimartingales are semimartingales, see for example Protter \cite{Protter2004}, Theorem II.6. Protter works with complete filtrations, but it is an immediate consequence of Lemma \ref{lem:complete cadlag} in Appendix \ref{app:complete filtration} that for every $(\mathcal{F}^P_t)$-semimartingale there exists an indistinguishable $(\mathcal{F}_t)$-semimartingale. Let $1 \le i \le d$. Since $S^i$ is locally bounded from below, there exists an increasing sequence of stopping times $(T_n)$ with $\lim_{n \rightarrow \infty} T_n = \infty$, and a sequence of strictly positive numbers $\alpha_n$, such that $(1 + \alpha_n S^i_{t \wedge T_n})_{t\ge0} \in \mathcal{W}_{1,s}$. By Corollary \ref{cor:abstract semimartingale} the stopped process $S^i_{\cdot\wedge T_n}$ is a semimartingale, and therefore $S^i$ is a semimartingale.
It remains to show that in the locally bounded from below case, any supermartingale density for $\mathcal{W}_{1,s}$ is a supermartingale density for $\mathcal{W}_1$. But this is exactly Kardaras and Platen \cite{Kardaras2011a}, Section 2.2. (And it will also follow from our considerations in Section \ref{sec: dominating measures}.)
\end{proof}
Of course we could also assume that every component of $S$ is either locally bounded from below or locally bounded from above, and we would still obtain the semimartingale property of $S$ under (NA1$_s$). But in the totally unbounded case, $S$ is not necessarily a semimartingale. A simple counterexample is given by a one-dimensional L\'{e}vy-process with jumps that are unbounded both from above and from below, to which we add an independent fractional Brownian motion with Hurst index $H \neq 1/2$. Their sum is not a semimartingale. But there are no 1-admissible strategies other than 0, so that $\mathcal{K}_{1,s} = \{1\}$, which is obviously bounded in probability.
\section{Construction of dominating local martingale measures}
\label{sec: dominating measures}
\subsection{The Kunita-Yoeurp problem and F\"ollmer's measure}\label{subsec: Kunita-Yoeurp}
Now let $Z$ be a strictly positive supermartingale with $Z_\infty >0$ and $E_P(Z_0) = 1$. Here we construct a dominating measure associated to $Z$. Naturally it is more delicate to construct a dominating measures than to construct an absolutely continuous measure
Our aim is to construct a dominating measure $Q$ and a stopping time $T$, such that $(Z,T)$ is the Kunita-Yoeurp decomposition of $Q$ with respect to $P$. We call this the ``Kunita-Yoeurp problem''. Recall that $(Z,T)$ is the Kunita-Yoeurp decomposition of $Q$ with respect to $P$ if
\begin{enumerate}
\item $P(T = \infty) = 1$,
\item $Q(\cdot \cap \{T \le t\})$ and $P$ are mutually singular on $\mathcal{F}_t$,
\item $Q(\cdot \cap \{T > t\})$ is absolutely continuous with respect to $P$ on $\mathcal{F}_t$, and for $A \in \mathcal{F}_t$
\begin{align}\label{eq: kunita decomposition}
Q(A \cap \{T>t\}) = E_P(1_A Z_t).
\end{align}
\end{enumerate}
In this case one can show that for any stopping time $\tau$ and any $A \in \mathcal{F}_\tau$
\begin{align}\label{eq: kunita decomposition stopping time}
Q(A \cap \{T > \tau\}) = E_P\left(1_{A \cap \{\tau < \infty\}} Z_\tau\right),
\end{align}
see for example \cite{Yoeurp1985}, Proposition 4.
In general it is impossible to construct $Q$ and $T$ without making further assumptions on the underlying filtered probability space. For example the space could be too small. Take an $\Omega$ that consists of one element, and define $\mathcal{F} = \mathcal{F}_t = \{\emptyset, \Omega\}$ for all $t \ge 0$. Then
\begin{align*}
Z_t = \frac{1}{2}(1+e^{-t}), \qquad t \ge 0,
\end{align*}
is a continuous and positive supermartingale with $Z_\infty >0$. But there exists only one probability measure on $\Omega$, and therefore any $Q$ would have Kunita-Yoeurp decomposition $(1,\infty)$ with respect to $P$, and not $(Z,T)$. This is reminiscent of the Dambis Dubins-Schwarz Theorem without the assumption $\langle M \rangle_\infty = \infty$ (see Revuz and Yor \cite{Revuz1999}, Theorem V.1.7). This problem can be solved by enlarging $\Omega$.
But even if the space is large enough, it might still not be possible to find $Q$ and $T$, because the filtration might be too large. Assume that the filtration $(\mathcal{F}_t)$ is complete with respect to $P$ and that $E_P(Z_0) = 1$. Then \eqref{eq: kunita decomposition} implies that $Q$ is absolutely continuous with respect to $P$ on $\mathcal{F}_0$. Since $\mathcal{F}_0$ contains all $P$-null sets, this means that $Q$ is absolutely continuous with respect to $P$, and therefore by \eqref{eq: kunita decomposition} again $Z_t = E_P(dQ/dP|\mathcal{F}_t)$. That is, $Z$ has to be a uniformly integrable martingale under $P$. So if $Z$ is a supermartingale, then the filtration $(\mathcal{F}_t)$ should not be completed. This problem can be avoided by assuming that $(\mathcal{F}_t)$ is the right-continuous modification of a standard system, to be defined below.
If we are allowed to enlarge $\Omega$ and if $(\mathcal{F}_t)$ is the right-continuous modification of a standard system, then the problem of constructing $Q$ and $T$ has been solved by Yoeurp \cite{Yoeurp1985} with the help of F\"ollmer's measure. Let us describe Yoeurp's solution.
First we remove the second problem in constructing $Q$ and $T$ by assuming that the filtration $(\mathcal{F}_t)$ is the right-continuous modification of a standard system $(\mathcal{F}^0_t)$. A filtration $(\mathcal{F}^0_t)$ is called \emph{standard system} if
\begin{enumerate}
\item for all $t \ge 0$, the $\sigma$-algebra $\mathcal{F}^0_t$ is $\sigma$-isomorphic to the Borel $\sigma$-algebra of a Polish space; that is, there exists a Polish space $(\mathcal{X}_t, \mathcal{B}_t)$, and a bijective map $\pi: \mathcal{F}^0_t \rightarrow \mathcal{B}_t$, such that $\pi^{-1}$ preserves countable set operations;
\item if $(t_i)_{i \ge 0}$ is an increasing sequence of positive times, and if $(A_i)_{i \ge 0}$ is a decreasing sequence of atoms of $\mathcal{F}^0_{t_i}$, then $\cap_{i \ge 0} A_i \neq \emptyset$.
\end{enumerate}
Then $(\mathcal{F}_t)$ is defined by setting $\mathcal{F}_t = \cap_{s > t} \mathcal{F}^0_{s}$.
Path spaces equipped with the canonical filtration are only standard systems if we allow for ``explosion'' to a cemetery state in finite time, see F\"ollmer \cite{Follmer1972}, Example 6.3, 2), or Meyer \cite{Meyer1972}.
Next we enlarge $\Omega$ in order to solve the possible problem of $\Omega$ being too small. Define $\overline{\Omega} := \Omega \times [0, \infty]$ and $\overline{\mathcal{F}} = \mathcal{F} \otimes \mathcal{B}[0,\infty]$, where $\mathcal{B}[0,\infty]$ denotes the Borel $\sigma$-algebra of $[0,\infty]$. Also define $\overline{P} = P \otimes \delta_\infty$, where $\delta_\infty$ is the Dirac measure at $\infty$. The filtration $(\overline{\mathcal{F}}_t)$ is defined by
\begin{align*}
\overline{\mathcal{F}}_t = \cap_{s > t} \mathcal{F}_s \otimes \sigma([0,r]: r \le s).
\end{align*}
Note that if $(\mathcal{F}_t)$ is the right-continuous modification of the standard system $(\mathcal{F}^0_t)$, then $(\overline{\mathcal{F}}_t)$ is the right-continuous modification of the standard system $\overline{\mathcal{F}}^0_0 = \mathcal{F}^0_0 \otimes \sigma(\{0\})$,
\begin{align*}
\overline{\mathcal{F}}^0_t = \mathcal{F}^0_t \otimes \sigma([0,s]: s \le t), \qquad t > 0.
\end{align*}
Random variables $X$ on $\Omega$ are embedded into $\overline{\Omega}$ by setting $\overline{X}(\omega, \zeta) = X(\omega)$.
Let us remark that $(\overline{\Omega}, \overline{\mathcal{F}}, (\overline{\mathcal{F}}_t), \overline{P})$ is an enlargement of $(\Omega, \mathcal{F}, (\mathcal{F}_t), P)$ in the sense of \cite{Revuz1999}.
\begin{defn}[\cite{Revuz1999}, p. 182]
A filtered probability space $(\tilde{\Omega}, \tilde{\mathcal{F}}, (\tilde{\mathcal{F}}_t), \tilde{P})$ is an \emph{enlargement} of $(\Omega, \mathcal{F}, (\mathcal{F}_t), P)$ if there exists a measurable map $\pi: \tilde{\Omega} \rightarrow \Omega$, such that $\pi^{-1}(\mathcal{F}_t) \subseteq \tilde{\mathcal{F}}_t$ and such that $\tilde{P}\circ \pi^{-1} = P$. In this case, random variables are embedded from $(\Omega, \mathcal{F})$ into $(\tilde{\Omega}, \tilde{\mathcal{F}})$ by setting $\tilde{X}(\tilde{\omega}) = X(\pi(\tilde{\omega}))$.
\end{defn}
Define $\pi(\omega, \zeta) = \omega$. Then we have $\pi^{-1}(A) = A \times [0, \infty] \in \overline{\mathcal{F}_t}$ for every $A \in \mathcal{F}_t$, and therefore $\pi^{-1}(\mathcal{F}_t) \subseteq \overline{\mathcal{F}_t}$. For any set $A \in \mathcal{F}$ we have $\overline{P}\circ \pi^{-1}(A) = P \otimes \delta_\infty (A \times [0, \infty]) = P(A)$. And if $X$ is a random variable on $(\Omega, \mathcal{F})$, then $\overline{X}(\omega, \zeta) = X(\omega) = X(\pi(\omega, \zeta))$.
Now we can proceed to construct $(\overline{Q},\overline{T})$ on $(\overline{\Omega}, \overline{\mathcal{F}}, (\overline{\mathcal{F}}_t))$. In fact it suffices to construct $\overline{Q}$, because we will take $\overline{T}(\omega, \zeta) = \zeta$, so that $\overline{P}(\overline{T} = \infty) = 1$. However there is one last remaining problem. In general $\overline{Q}$ will not be uniquely determined by $(\overline{Z},\overline{T})$. The measure $\overline{Q}$ must satisfy
\begin{align}\label{eq:martingale iff T infinite}
1 = \overline{Q}(\overline{\Omega}) = \overline{Q}(\overline{\Omega} \cap\{ t < \overline{T}\}) + \overline{Q}(\overline{\Omega} \cap\{ t \ge \overline{T}\}) = E_{\overline{P}}(\overline{Z}_t) + \overline{Q}( t \ge \overline{T})
\end{align}
for all $t \ge 0$. But $\overline{Q}$ is supposed to solve the Kunita-Yoeurp problem associated with $(\overline{Z},\overline{T})$, and therefore at time $\overline{T}$, the measure $\overline{Q}$ should stop being absolutely continuous with respect to $\overline{P}$, and \eqref{eq:martingale iff T infinite} implies that $\overline{Q}(\overline{T}<\infty) > 0$ if $\overline{Z}$ is not a martingale. So knowing $\overline{P}$, $\overline{Z}$, and $\overline{T}$, in general we can only hope to determine $\overline{Q}$ uniquely on the $\sigma$-field
\begin{align}\label{eq:F T minus predictable}
\overline{\mathcal{F}}_{\overline{T}-} & = \sigma(\overline{\mathcal{F}}_0,\overline{A}_t \cap \{{\overline{T} > t}\}: \overline{A}_t \in \overline{\mathcal{F}}_t, t > 0) = \sigma(A_t \times (t, \infty]: A_t \in \mathcal{F}_t, t \ge 0).
\end{align}
For the second equality we refer to \cite{Follmer1972}. Note that \eqref{eq:F T minus predictable} implies that $\overline{\mathcal{F}}_{\overline{T}-}$ is the predictable sigma-algebra on $\Omega \times [0,\infty]$.
In conclusion, in order to construct $T$ and $Q$, we need to assume that $(\mathcal{F}_t)$ is the the right-continuous modification of a standard system, we need to enlarge $(\Omega, \mathcal{F}, (\mathcal{F}_t), P)$ as described above, and we have to accept that $Q$ will only be defined on $\overline{\mathcal{F}}_{\overline{T}-}$. Under these conditions, we can take $\overline{Q}$ as the F\"ollmer measure of $Z$. Given a nonnegative supermartingale $Z$ with $E_P(Z_0) = 1$, F\"ollmer \cite{Follmer1972} constructs a measure $P^Z$ on $(\overline{\Omega}, \overline{\mathcal{F}}_{T-})$, which satisfies $P^Z(\overline{A}_t \cap \{ \overline{T} > t\}) = E_{\overline{P}}(\overline{Z}_t 1_{\overline{A}_t})$ for all $t \ge 0$ and all $\overline{A}_t \in \overline{\mathcal{F}}_t$. This is exactly the relation \eqref{eq: kunita decomposition}, and therefore $(\overline{Z}, \overline{T})$ is the Kunita-Yoeurp decomposition of $\overline{Q}:=P^Z$ with respect to $\overline{P}$. Note that it is possible to extend $\overline{Q}$ from $\overline{\mathcal{F}}_{\overline{T}-}$ to $\overline{\mathcal{F}}$, generally in a non-unique way. See for example the discussion on p. 9 of \cite{Kardaras2011}. So from now on we make the convention that the measure $\overline{Q}$ is one of these extensions, i.e. $\overline{Q}$ denotes a probability measure on $(\overline{\Omega}, \overline{\mathcal{F}})$ that satisfies $\overline{Q}|_{\overline{\mathcal{F}}_{\overline{T}-}} = P^Z$.
It remains to show that $\overline{Q}$ dominates $\overline{P}$. But this is a general result.
\begin{lem}\label{lem:absolute continuity}
Let $\mu$ and $\nu$ be two probability measures on a filtered probability space $(\Theta, \mathcal{G}, (\mathcal{G}_t))$. Let $(\zeta, \tau)$ be the Kunita-Yoeurp decomposition of $\nu$ with respect to $\mu$. Define $\mathcal{G}_\infty = \vee_{t\ge 0} \mathcal{G}_t$ and assume that $\mu(\zeta_\infty > 0)=1$. Then $\mu|_{\mathcal{G}_{\infty}} \ll \nu|_{\mathcal{G}_{\infty}}$.
\end{lem}
\begin{proof}
Let $A \in \cup_{t\ge0}\mathcal{G}_t$. We use the $\sigma$-continuity of $\nu$, \eqref{eq: kunita decomposition}, and Fatou's lemma, to obtain
\begin{align*}
\nu(A\cap\{\tau=\infty\})=\lim_{t\rightarrow\infty} \nu(A\cap\{\tau>t\}) = \lim_{t\rightarrow \infty} E_\mu(\zeta_t 1_{A}) \ge E_\mu(\zeta_\infty 1_{A}).
\end{align*}
By the monotone class theorem, this inequality extends to all $A \in \mathcal{G}_\infty$. Since $\mu(\zeta_\infty>0)=1$, we obtain $\mu|_{\mathcal{G}_\infty} \ll \nu|_{\mathcal{G}_\infty}$.
\end{proof}
Lemma \ref{lem:absolute continuity} applied to $\overline{P}$, $\overline{Q}$, and $(\overline{\Omega}, \overline{\mathcal{F}}, (\overline{\mathcal{F}}_t))$ implies that $\overline{Q}\gg \overline{P}$.
In the following we make the standing assumption that we work on a probability space $(\Omega, \mathcal{F}, (\mathcal{F}_t), P)$ where it is possible to solve the Kunita-Yoeurp problem of associating a probability measure and a stopping time to a given supermartingale, and we omit the notation $\overline{(
\cdot)}$.
\subsubsection{Calculating expectations under $Q$}
Here we present important results of Yoeurp that allow to rewrite certain expectations under $Q$ as expectations under $P$. More precisely, let $Z$ be a nonnegative supermartingale with $E(Z_0)=1$ and with Doob-Meyer decomposition $Z = Z_0 + M - A$, where $M$ is a local martingale starting in zero, and $A$ is an adapted process, a.s. increasing and c\`adl\`ag. Let $T$ and $Q$ be a stopping time and a probability measure, such that $(Z,T)$ is the Kunita-Yoeurp decomposition of $Q$ wrt $P$.
\begin{lem}\label{lem:yoeurps formula}
Let $Z = Z_0 + M - A$, and let $T, Q$ be as described above. Let $(\tau_n)$ be a localizing sequence for $M$. Then we have for every bounded predictable process $Y$ and for every $n \in \mathbb{N}$
\begin{align}\label{eq:explicit foellmer}
E_Q(Y^{\tau_n}_T) = E_P\left(Y_{\tau_n} Z_{\tau_n} + \int_0^{\tau_n} Y_s dA_s\right).
\end{align}
\end{lem}
\begin{proof}
This is part of Proposition 9 of \cite{Yoeurp1985}. For the convenience of the reader, we provide a proof. First consider a simple processes of the form $Y_r(\omega) = X(\omega) 1_{(t,\infty]}(r)$ for some bounded $X \in \mathcal{F}_t$. For such $Y$ we get from \eqref{eq: kunita decomposition}
\begin{align*}
E_Q(Y^{\tau_n}_T) & = E_Q( X 1_{(t,\infty]} (\tau_n \wedge T)) = E_Q( X 1_{\{t < \tau_n\}} 1_{\{t < T\}}) = E_P( X 1_{\{t < \tau_n\}} Z_t) \\
& = E_P( X 1_{\{t < \tau_n\}} (M_t - A_t)) = E_P( X 1_{\{t < \tau_n\}} (M^{\tau_n}_t - A_t)).
\end{align*}
Now we use that $M^{\tau_n}$ is a uniformly integrable martingale, and that $X 1_{\{t<\tau_n\}}$ is $\mathcal{F}_{t}$-measurable, to replace $M^{\tau_n}_t$ by $M^{\tau_n}_\infty = M_{\tau_n}$. Moreover we have
\begin{align*}
X 1_{\{t < \tau_n\}} A_t & = X 1_{\{t < \tau_n\}} A_{\tau_n} - X 1_{\{t < \tau_n\}}(A_{\tau_n} - A_t) = X 1_{\{t < \tau_n\}} A_{\tau_n} - \int_0^{\tau_n} Y_s dA_s,
\end{align*}
which proves \eqref{eq:explicit foellmer} for such simple $Y$. The general case now follows from the monotone class theorem.
\end{proof}
\begin{cor}\label{cor:yoeurps formula 2}
Let $Y$ be a bounded adapted process that is $P$-a.s. c\`adl\`ag. Define
\begin{align*}
Y^{T-}_t(\omega) = Y_t(\omega) 1_{\{t < T(\omega)\}} + \limsup_{s \rightarrow T(\omega)-} Y_s(\omega) 1_{\{t \ge T(\omega) \}}.
\end{align*}
Let $Z$ and $(\tau_n)$ be as in Lemma \ref{lem:yoeurps formula}. Then
\begin{align*}
E_Q(Y^{T-}_{\tau_n}) = E_P\left( Y_{\tau_n} Z_{\tau_n} + \int_0^{\tau_n} Y_{s-} dA_s\right).
\end{align*}
\end{cor}
\begin{proof}
Define $Y^{-}_t(\omega) = Y_{t-}(\omega) = \limsup_{s \rightarrow t-} Y_s$ for $t > 0$, and $Y^-_0 = Y_0$. Then $Y^-$ is a predictable process, because it is the point-wise limit of the step functions
\begin{align*}
Y^n_t = Y_0 1_{\{0\}}(t) + \sum_{k \ge 0} \limsup_{s \rightarrow k2^{-n} -} 1_{(k2^{-n}, (k+1)2^{-n}]}(t).
\end{align*}
Therefore we can apply Lemma \ref{lem:yoeurps formula} to $Y^-$. Observe that
\begin{align*}
Y^{T-}_{\tau_n} = Y_{\tau_n} 1_{\{T > \tau_n\}} + Y_{T-} 1_{\{T \le \tau_n\}} = Y_{\tau_n} 1_{\{T > \tau_n\}} + (Y^-)_T 1_{\{T \le \tau_n\}}.
\end{align*}
Now \eqref{eq: kunita decomposition} implies that $E_Q(Y_{\tau_n} 1_{\{T > \tau_n\}}) = E_P(Y_{\tau_n} Z_{\tau_n})$, whereas \eqref{eq:explicit foellmer} and then again \eqref{eq: kunita decomposition} applied to the second term give
\begin{align*}
E_Q((Y^-)_T 1_{\{T \le \tau_n\}}) &= E_Q((Y^-)^{\tau_n}_T ) - E_Q((Y^-)_{\tau_n} 1_{\{T > \tau_n\}}) \\
&= E_P\left(Y_{\tau_n-} Z_{\tau_n} + \int_0^{\tau_n} Y_{s-} dA_s\right) - E_P\left(Y_{\tau_n-} Z_{\tau_n}\right) \\
& = E_P\left( \int_0^{\tau_n} Y_{s-} dA_s\right).
\end{align*}
\end{proof}
\subsection{The predictable case}
We still assume that $(\Omega, \mathcal{F}, (\mathcal{F}_t), P)$ is a probability space on which it is possible to solve the Kunita-Yoeurp problem. Let $S$ be a $d$-dimensional predictable semimartingale, let $\mathcal{W}_{1}$ be defined as in \eqref{eq:W1}, and let $Z$ be a supermartingale density for $\mathcal{W}_1$. Here we examine the structure of $S$ and $Z$ closer. This will allow us to apply Lemma \ref{lem:yoeurps formula} to deduce that $S^{T-}$ is a local martingale under the dominating measure associated to $Z$. Note that Yoeurp \cite{Yoeurp1985} also establishes a generalized Girsanov formula, which we could apply directly rather than using Lemma \ref{lem:yoeurps formula}. However the use of Lemma \ref{lem:yoeurps formula} turns out to be rather instructive, and it allows us to obtain some insight into why the non-predictable case is more complicated.
\begin{rmk}
Observe that thanks to predictability, $S-S_0$ is almost surely locally bounded. This follows immediately from I.2.16 of \cite{Jacod2003}, which says that for $C>0$ there exists an announcing sequence for the entrance time of $S$ into $\{x \in \mathbb{R}^d: |x|\ge C\}$. By Corollary \ref{cor:NA1 implies semimartingale} it would in particular suffice to assume that $S$ satisfies (NA1$_s$) and that $Z$ is a supermartingale density for $\mathcal{W}_{1,s}$. Then $S$ is a semimartingale, satisfies (NA1), and $Z$ is a supermartingale density for $\mathcal{W}_1$.
\end{rmk}
Since $S-S_0$ is locally bounded, it is even a special semimartingale (see \cite{Jacod2003}, I.4.23 (iv)). That is, there exists a unique decomposition
\begin{align}\label{eq: semimartingale decomposition}
S = S_0 + M + A,
\end{align}
where $M$ is a local martingale with $M_0 = 0$, and $A$ is a predictable process of finite variation with $A_0 = 0$. This implies that $M = S - S_0 - A$ is predictable. But any predictable right-continuous local martingale is continuous (\cite{Jacod2003}, Corollary I.2.31). Therefore $S$ is of the form \eqref{eq: semimartingale decomposition} with continuous $M$.
But then also $A$ must be continuous, because (NA1) implies $dA^i \ll d\langle M^i \rangle$ for every $i=1,\dots,d$, where $M=(M^1, \dots, M^d)$ and $A=(A^1,\dots, A^d)$. This is a well known fact, see for example Ankirchner's thesis \cite{Ankirchner2005}, Lemma 9.1.2. Otherwise one could find a predictable process $H^i$ which satisfies $H^i \cdot M^i \equiv 0$, but for which $H^i \cdot A^i$ is increasing;
this would clearly contradict $\mathcal{K}_1$ being bounded in probability. Therefore $A$ and then also $S$ must be continuous.
It turns out that $S$ must satisfy the \emph{structure condition} as defined by Schweizer \cite{Schweizer1995}. Recall that $L^2_{\text{loc}}(M)$ is the space of progressively measurable processes $(\lambda_t)_{t \ge 0}$ that are locally square integrable with respect to $M$, i.e. such that
\begin{align*}
\int_0^t \sum_{i,j=1}^d \lambda^i_s \lambda^j_s d \langle M^i, M^j \rangle_s < \infty
\end{align*}
for every $t >0$. For details see \cite{Jacod2003}, III.4.3.
\begin{defn}\label{def:structure condition}
Let $S = S_0 + M + A$ be a $d$-dimensional special semimartingale with locally square-integrable $M$. Define
\begin{align*}
C_t = \sum_{i=1}^d \langle M^i \rangle_t \qquad \text{and for } 1 \le i, j \le d: \qquad \sigma^{ij}_t = \frac{d \langle M^i, M^j \rangle_t }{d C_t}.
\end{align*}
Note that $\sigma$ exists by the Kunita-Watanabe inequality. Then $S$ satisfies the \emph{structure condition} if $dA^i \ll d \langle M^i \rangle$ for every $1 \le i \le d$, with predictable derivative $\alpha^i_t = d A^i_t/d\langle M^i \rangle_t$, and if there exists a predictable process $\lambda_t = (\lambda^1_t, \dots, \lambda^d_t) \in L^2_{\text{loc}}(M)$, such that
\begin{align} \label{eq:structure condition}
(\sigma_t \lambda_t)^i = \alpha^i_t \sigma^{ii}_t
\end{align}
for every $i = 1, \dots, d$. Note that $\lambda$ might not be uniquely determined, but the stochastic integral $\int \lambda dM$ does not depend on the choice of $\lambda$, see \cite{Schweizer1995}. If
\begin{align}\label{eq:structure condition until infty}
\int_0^\infty \sum_{i,j=1}^d \lambda^i_t \sigma^{ij}_t \lambda^j_t dC_t < \infty,
\end{align}
then we say that S satisfies the \emph{structure condition until $\infty$}.
\end{defn}
Recall that two one-dimensional local martingales $L$ and $N$ are called \emph{strongly orthogonal} if $LN$ is a local martingale.
Also recall that the \emph{stochastic exponential} of a semimartingale $X$ is defined by
\begin{align*}
\mathcal{E}(X)_t = 1 + \int_0^t \mathcal{E}(X)_{s-} dX_s, \qquad t \ge 0.
\end{align*}
Finally we recall that every nonnegative supermartingale $Y$ satisfies $Y_{\tau+t} \equiv 0$ for all $t \ge 0$, where $\tau = \inf\{t \ge 0: Y_{t-}=0 \text{ or } Y_t = 0\}$.
\begin{lem}\label{lem: structure condition for predictable S}
Suppose that $Z$ is a supermartingale density for the predictable semimartingale $S$. Then $S$ satisfies the structure condition until $\infty$, and
\begin{align}\label{eq: explicit supermartingale density}
dZ_t = Z_{t-}( - \lambda_t dM_t + dN_t - dB_t)
\end{align}
where $\lambda$ satisfies \eqref{eq:structure condition} and \eqref{eq:structure condition until infty}, $N$ is a local martingale that is strongly orthogonal to $M$, $B$ is increasing,
and $\mathcal{E}(N-B)_\infty > 0$.
Conversely, if a predictable process $S$ satisfies the structure condition until $\infty$, and if $Z$ is defined by \eqref{eq: explicit supermartingale density}, then $Z$ is a supermartingale density for $S$.
In particular, for predictable $S$, the structure condition until $\infty$ is equivalent to (NA1).
\end{lem}
\begin{proof}
This is essentially Proposition 3.2 of \cite{Larsen2007} in infinite time. We provide a slightly simplified version of their proof, because later we will need some results obtained during the proof.
Let $Z$ be a supermartingale density. Since $Z$ is strictly positive, it is of the form $dZ_t = Z_{t-}(dL_t - dB_t)$ for a local martingale $L$ and a predictable increasing process $B$. Since $M$ is continuous, there exists a predictable process $\lambda \in L^2_{\text{loc}}(M)$, such that $dL_t = \lambda_t dM_t + dN_t$, where $N$ is a local martingale that is strongly orthogonal to all components of $M$, see \cite{Jacod2003}, Theorem III.4.11. In particular $[H\cdot M, N]$ is a local martingale for every integrand $H$. Moreover
\begin{align*}
0 < Z_\infty = Z_0 \mathcal{E}( \lambda \cdot M + N - B)_\infty = Z_0 \mathcal{E}( \lambda \cdot M)_\infty \mathcal{E}(N - B)_\infty,
\end{align*}
which is only possible if $\lambda$ satisfies \eqref{eq:structure condition until infty} and if $\mathcal{E}(N-B)_\infty > 0$. It only remains to show that $\lambda$ also satisfies \eqref{eq:structure condition}.
Let $H$ be a 1-admissible strategy. Write $W^H_t = 1 + (H \cdot S)_t$ for the wealth process generated by $H$. Then $W^H Z$ is a nonnegative supermartingale. Since $Z$ is strictly positive, we must have $W^H_t \equiv 0$ for $t \ge \tau^H = \inf\{ s \ge 0: W^H_{s-} = 0 \text{ or } W^H_s = 0\}$. Therefore we may replace $H$ by $H 1_{\{t < \tau^H\}}$ without loss of generality. Define $\pi_t = H_t/W^H_{t-}$, where we interpret $0/0=0$ as before. Then finally
\begin{align*}
W^H_t = 1 + (H \cdot S)_t = 1 + \int_0^t \pi_s W^H_{s-} dS_s.
\end{align*}
So if we slightly abuse notation and define $W^\pi_t = W^H_t$, then $W^\pi_t = 1 + \int_0^t \pi_s W^\pi_{s-} dS_s$, and every wealth process is of this form.
We write $dX_t \sim dY_t$ if $d(X - Y)_t$ is the differential of a local martingale. Integration by parts applied to $Z W^\pi$ gives
\begin{align}\label{eq:ibp supermartingale density} \nonumber
d(Z W^\pi)_t & = W^\pi_{t-} dZ_t + Z_{t-} \pi_t W^\pi_{t-} dS_t + d[(\pi W^\pi_{-} dS), Z]_t \\ \nonumber
& = W^\pi_{t-} Z_{t-} (\lambda_t dM_t + dN_t - dB_t) + Z_{t-} \pi_t W^\pi_{t-} (dM_t + dA_t) \\ \nonumber
& \qquad + W^\pi_{t-} Z_{t-} d[\pi\cdot (M + A), \lambda \cdot M + N - B]_t \\
& \sim - W^\pi_{t-} Z_{t-} dB_t + Z_{t-} \pi_t W^\pi_{t-} dA_t + W^\pi_{t-} Z_{t-} (d[ \pi \cdot M, \lambda \cdot M]_t - d[\pi \cdot A, B]_t).
\end{align}
Of course $[\pi \cdot A, B] \equiv 0$, because $A$ is continuous. But thanks to Proposition I.4.49 c) of \cite{Jacod2003}, \eqref{eq:ibp supermartingale density} has the advantage that it is also valid if $M$ and $A$ are not necessarily continuous (although in that case $\lambda \cdot M$ should be replaced by a general local martingale $N$).
Let now $C$ and $\sigma$ be as described in Definition \ref{def:structure condition}. Theorem III.4.5 of \cite{Jacod2003} implies
\begin{align}\label{eq:second ibp supermartingale density} \nonumber
d(Z W^\pi)_t &\sim W^\pi_{t-} Z_{t-} \left( - dB_t + \pi_t dA_t + d\langle \pi \cdot M, \lambda \cdot M\rangle_t \right) \\
& = W^\pi_{t-} Z_{t-}\left( -dB_t + \sum_{i = 1}^d \pi^i_t \left( dA^i_t + \sum_{j=1}^d \sigma^{ij}_t \lambda^j_t dC_t \right)\right).
\end{align}
Assume that there exists an $i \in \{1,\dots,d\}$ for which the almost surely continuous process
\begin{align*}
D^i_t = \int_0^t \left( Z_{s-} dA^i_s + \sum_{j=1}^d \sigma^{ij}_s \beta^j_s dB_s \right)
\end{align*}
is not evanescent. We claim that then there exists a 1-admissible strategy $\pi$ for which the finite variation part of $(Z W^\pi)$ is increasing on a small time interval. This is a contradiction to $Z W^\pi$ being a supermartingale. By the predictable Radon-Nikodym theorem of Delbaen and Schachermayer (\cite{Delbaen1995a}, Theorem 2.1 b)), there exists a predictable $\gamma^i$ with values in $\{-1, 1 \}$, such that $\int_0^\cdot \gamma^i_s dD^i_s = V^i$, where $V^i$ denotes the total variation process of $D^i$. Note that \cite{Delbaen1995a} work with complete filtrations, but given the $(\mathcal{F}^P_t)$-predictable $\gamma^i$ that they construct, we can apply Lemma \ref{lem:complete predictable optional} to obtain a $(\mathcal{F}_t)$-predictable $\tilde{\gamma}^i$ that is indistinguishable from $\gamma^i$.
Let now $n\in \mathbb{N}$ and take $\pi_t = n \gamma^i_t$ in \eqref{eq:second ibp supermartingale density}. Then $d(Z W^{\pi})_t \sim W^{\pi}_{t-} Z_{t-}\left( -dB_t + n dV^i_t \right)$, and $V^i$ is an increasing process. Since $\pi$ is bounded, $W^{\pi}_{t-} > 0$ for all $t \ge 0$, and of course also $Z_{t-} > 0$ for all $t \ge 0$. But if $H$ is strictly positive, and $F$ is a finite variation process, then $\int_0^\cdot H_s dF_s$ is a decreasing process if and only if $F$ is a decreasing process. And clearly $- B + nV^i$ can only be decreasing for all $n \in \mathbb{N}$ if $V^i \equiv 0$, a contradiction.
Therefore $D^i$ is evanescent, and thus for some predictable $\alpha^i$,
\begin{align*}
0 \equiv \left( dA^i_t + \sum_{j=1}^d \sigma^{ij}_t \lambda^j_t dC_t \right) = \left( \alpha^i_t d\langle M^i_t \rangle + \sum_{j=1}^d \sigma^{ij}_t \lambda^j_t dC_t \right) = \left( \alpha^i_t \sigma^{ii}_t + (\sigma_t \lambda_t)^i\right) dC_t
\end{align*}
so that
\begin{align}\label{eq:predictable proof1}
\alpha^i \sigma^{ii} = -(\sigma \lambda)^i \quad dC(\omega)\otimes P(d\omega)-\text{almost everywhere,}
\end{align}
i.e. \eqref{eq:structure condition} is satisfied, and the proof is complete.
The converse direction is easy and follows directly from \eqref{eq:ibp supermartingale density}.
\end{proof}
Next we will show that $S^{T-}$ is a local martingale under the measure $Q$ that is associated to $Z$. But first we observe that if $Z$ is a supermartingale density, then $S Z$ is \emph{not} necessarily a local martingale.
\begin{cor}
Let $Z$ and $S$ be as in Lemma \ref{lem: structure condition for predictable S}. Then $ZS^i$ is a local supermartingale if and only if $S^i \ge 0$ on the support of the measure $dC$. If $S^i \ge 0$ identically, then $ZS^i$ is a supermartingale.
$ZS^i$ is a local martingale if and only if $S^i = 0$ on the support of the measure $dC$.
\end{cor}
\begin{proof}
\eqref{eq:structure condition} and \eqref{eq: explicit supermartingale density} imply that
\begin{align*}
d(ZS^i)_t & = Z_{t-}dS^i_t + S^i_{t-} dZ_t + d[S^i,Z]_t \sim - Z_{t-} S^i_{t-} dB_t.
\end{align*}
Nonnegative local supermartingales are supermartingales by Fatou's lemma, and therefore the proof is complete.
\end{proof}
Another immediate consequence of Lemma \ref{lem: structure condition for predictable S} is that in the predictable case, the maximal elements among the supermartingale densities are always local martingales. This is important in the duality approach to utility maximization. For details we refer the reader to \cite{Larsen2007}.
We are now ready to prove Theorem \ref{thm:predictable supermartingale density}.
\begin{cor*}[Theorem \ref{thm:predictable supermartingale density}]
Let $S$ be a predictable semimartingale, and let $Z$ be a supermartingale density for $S$. Let $T$ be a stopping time and $Q$ be a probability measure, such that $(Z/E_P(Z_0),T)$ is the Kunita-Yoeurp decomposition of $Q$ wrt $P$. Then $S^{T-}$ is a $Q$-local martingale.
Conversely if $Q \gg P$ with Kunita-Yoeurp decomposition $(Z,T)$ wrt $P$, and if $S^{T-}$ is a local martingale under $Q$, then $Z$ is a supermartingale density for $S$.
\end{cor*}
\begin{proof}
We first show that $S^{T-}$ is $Q$-a.s. locally bounded. Let $\tilde{\rho}_n = \inf \{ t \ge 0: |S^{T-}_t| \ge n\}$. Since $S^{T-}$ was only required to be right-continuous $P$-a.s. and not identically, $\tilde{\rho}_n$ is not necessarily a stopping time. It is however a $(\mathcal{F}^Q_t)$-stopping time. By Lemma \ref{lem:complete stopping time}, we can find a stopping time $\rho_n$ such that $Q(\rho_n = \tilde{\rho}_n)=1$. Then $\sup_n \rho_n$ is a stopping time, and \eqref{eq: kunita decomposition stopping time} implies that
\begin{align}\label{eq:predictable supermartingale density pr1}
Q\left(\sup_n \rho_n < T\right) = E_P\left( Z_{\sup_n \rho_n} 1_{\{\sup_n \rho_n < \infty\}} \right) = 0,
\end{align}
because $P(\sup_n \rho_n < \infty) = 0$. But $S^{T-}_t$ is constant for $t \ge T$, and therefore $\{\sup_n \rho_n \ge T\}$ is $Q$-a.s. contained in $\{\sup_n \rho_n = \infty\}$, showing that $S^{T-}$ is $Q$-a.s. locally bounded.
Let $(\sigma_n)$ be a localizing sequence for $M$, where $Z = Z_0 + M - A$. We define $\tau_n = \rho_n \wedge \sigma_n$. Let $H$ be a strategy that is 1-admissible under $Q$. Then we can apply Corollary \ref{cor:yoeurps formula 2} (which of course extends from bounded $Y$ to nonnegative $Y$), to obtain
\begin{align*}
E_Q(1+(H\cdot S^{T-})_{\tau_n}) = E_P\left( (1+(H\cdot S)_{\tau_n}) Z_{\tau_n} + \int_0^{\tau_n} (1+(H\cdot S)_{s-})dAs\right).
\end{align*}
But now \eqref{eq:second ibp supermartingale density} and \eqref{eq:structure condition} imply that $(1+(H\cdot S)) Z + \int (1+(H\cdot S)_{s-})dAs$ is a nonnegative $P$-local martingale starting in 1, and therefore $E_Q(1+(H\cdot S^{T-})_{\tau_n}) \le 1$. Since $(S^{T-})^{\tau_n}$ is bounded, it must be a martingale.
The only remaining problem is that we only know $Q(\sup_n \tau_n \ge T) = 1$ and not $Q(\sup_n \tau_n = \infty) = 1$. But in fact the same arguments also show that $(S^{T-})^{\rho_n\wedge \tau_m}$ is a martingale for all $n,m \in \mathbb{N}$. Therefore we can apply bounded convergence to obtain for all $s,t \ge 0$
\begin{align*}
E_Q( (S^{T-})^{\rho_n}_{t+s} | \mathcal{F}_t) = \lim_{m \rightarrow \infty} E_Q( (S^{T-})^{\rho_n \wedge \tau_m}_{t+s} | \mathcal{F}_t) = \lim_{m \rightarrow \infty} (S^{T-})^{\rho_n \wedge \tau_m}_{t} = (S^{T-})^{\rho_n}_{t}
\end{align*}
As we argued above, $Q(\sup_n \rho_n = \infty) = 1$, and therefore $S^{T-}$ is a $Q$-local martingale.
Conversely, let $S^{T-}$ be a $Q$-local martingale, and let $H$ be a 1-admissible strategy for $S$ under $P$. Define $\tau = \inf\{t \ge 0: (H \cdot S^{T-})_t < -1 \}$. Then $P(\tau < \infty) = 0$ and therefore $Q(\tau < T) = 0$ by the same argument as in \eqref{eq:predictable supermartingale density pr1}. In particular $H$ is 1-admissible for $S^{T-}$ under $Q$. Now we can repeat the arguments in \eqref{eq:supermartingale density derivation}, to obtain that ${Z}_t = 1_{\{ t < T\}}/\gamma_t$ is a supermartingale density for $S$, where we denoted ${\gamma}_t = (dP/dQ)|_{\mathcal{F}_t}$.
\end{proof}
\begin{rmk}
Note that we only used the predictability of $S$ once: it was only needed to obtain
\begin{align*}
E_P\left( (1+(H\cdot S)_{\tau_n}) Z_{\tau_n} + \int_0^{\tau_n} (1+(H\cdot S)_{s-})dAs\right) \le 1,
\end{align*}
for which we applied (results from the proof of) Lemma \ref{lem: structure condition for predictable S}.
\end{rmk}
\begin{cor}[``Predictable weak fundamental theorem of asset pricing'']\label{thm:predictable ftap}
Let $(\mathcal{F}_t)$ be the right-continuous modification of a standard system. Let $S$ be a predictable stochastic process that is a.s. right-continuous. Then $S$ satisfies (NA1$_s$) if and and only there exists an enlarged probability space $(\overline{\Omega}, \overline{\mathcal{F}}, (\overline{\mathcal{F}}_t), \overline{P})$ and a dominating measure $\overline{Q} \gg \overline{P}$ with Kunita-Yoeurp decomposition $(\overline{Z}, \overline{T})$ with respect to $\overline{P}$, such that $\overline{S}^{\overline{T}-}$ is a $\overline{Q}$-local martingale.
\end{cor}
\begin{proof}
It remains to be shown that the existence of $\overline{Q}$ implies that $S$ satisfies (NA1$_s$). But if $\overline{Q}$ exists, then Theorem \ref{thm:predictable supermartingale density} and Theorem \ref{thm: ex supermartingale density} imply that $\overline{S}$ satisfies (NA1) on $(\overline{\Omega}, \overline{\mathcal{F}},$ $(\overline{\mathcal{F}}_t), \overline{P})$. Since this is an enlargement of $(\Omega, \mathcal{F}, (\mathcal{F}_t), P)$, the process $S$ must also satisfies (NA1).
\end{proof}
\begin{rmk}
We argued above that a predictable process satisfying (NA1) must be continuous. Therefore Corollary \ref{thm:predictable ftap} is not much more general than Ruf \cite{Ruf2010}, where it is shown that a diffusion $S$ that satisfies (NA1) admits a dominating measure $Q$ under which $S^{T-}$ is a local martingale. However one difference is that \cite{Ruf2010} only shows that supermartingale densities that are local martingales correspond to dominating local martingale measures. Here we show that in the predictable case this is in fact true for all supermartingale densities. Also, we show equivalence between (NA1) and the existence of a dominating local martingale measure, and not only that (NA1) implies the existence of $Q$. Of course, as is usually the case for this type of result, the reverse direction is much easier.
\end{rmk}
\subsection{The general case}
We start the treatment of the non-predictable case with two examples that illustrate why it is natural to consider dominating local martingale measures for $S^{T-}$ rather than for $S$.
\begin{ex}
If $S$ is optional and if $Q$ is a dominating local martingale measure for $S$ rather than for $S^{T-}$, then $S$ does not need to satisfy (NA1): Let $T$ be exponentially distributed with parameter 1 under $Q$. Define $S_t = e^t 1_{\{t < T\}}$ for $t \in [0,1]$. Since time is finite, $S$ is a uniformly integrable martingale. Therefore $dP = S_1 dQ$ is absolutely continuous with respect to $Q$. But under $P$ we have $S_t = e^t$ for all $t \in [0,1]$. So clearly $S$ does not satisfy (NA1) under $P$, although $Q$ is a dominating martingale measure for $S$. Of course $S^{T-}$ is \emph{not} a local martingale under $Q$, because $S^{T-}_t = e^t$.
\end{ex}
Recall that a stopping time $\tau$ is called \emph{foretellable} under a probability measure $P$ if there exists an increasing sequence $(\tau_n)$ of stopping times, such that $P(\tau_n < \tau)=1$ for every $n$, and such that $P(\sup_n \tau_n = \tau)=1$. In this case $(\tau_n)$ is called an \emph{announcing sequence} for $\tau$. Every predictable time is foretellable under any probability measure, see Theorem I.2.15 and Remark I.2.16 of \cite{Jacod2003}.
\begin{ex}
Let $S$ be a semimartingale under $P$ and let $Q \gg P$ be a dominating measure with Kunita-Yoeurp decomposition $(Z,T)$ wrt $P$. Assume that $T$ is not foretellable under $Q$. Then there exists an adapted process $\tilde{S}$ which is $P$-indistinguishable from $S$, such that $\tilde{S}$ is not a $Q$-local martingale: Let $x \in \mathbb{R}^d$ and define $\tilde{S}^x_t = S_t 1_{\{t < T\}} + x 1_{\{t \ge T\}}$, which is $P$-indistinguishable from $S$ since $P(T=\infty)=1$. If $\tilde{S}^x$ is a $Q$-local martingale, then we can take the localizing sequence $\tau^x_n = \inf\{t \ge 0: |\tilde{S}^x_t| \ge n\}$.
Since $T$ is not foretellable under $Q$ and since by the same argument as in \eqref{eq: kunita decomposition stopping time} we have $Q(\lim_{n \rightarrow \infty} \tau^x_n \ge T)=1$, there must exist $n \in \mathbb{N}$ for which $Q(\tau^x_n = T) > 0$. Moreover $\tau^x_n = \tau^y_n$ for all $|x|<n$, $|y| < n$, and therefore
\begin{align*}
E_Q(S_0) = E_Q(\tilde{S}^x_{\tau^x_n}) = E_Q(S_{\tau^x_n} 1_{\{\tau^x_n < T\}}) + x Q(\tau^x_n \ge T).
\end{align*}
We obtain a contraction by letting $x$ vary through the ball of radius $n-1$.
\end{ex}
These two examples show that given $Q \gg P$, it is important to choose a good version of $S$ if we want to obtain a $Q$-local martingale. All the results obtained so far indicate that this good version should be $S^{T-}$. Maybe somewhat surprisingly, this is not true in general, as we demonstrate in the following example.
\begin{ex}
Let $(L_t)_{t \in [0,1]}$ be a L\'evy process under $Q$, with jump measure $\nu = \delta_1 + \delta_{-1}$ and drift $b \in \mathbb{R}$. To wit, $L_t = N^1_t - N^2_t + b t$, where $N^1$ and $N^2$ are independent Poisson processes. Let $a > |b|$ and let $\tau$ be an exponential random variable with parameter $a$, such that $\tau$ is independent from $L$. Define $T = \tau$ if $\tau \le 1$, and $T = \infty$ otherwise. Then $(e^{a t} 1_{\{t < T\}})_{t \in [0,1]}$ is a uniformly integrable martingale, and therefore it defines a probability measure $dP = e^a 1_{\{1 < T\}}dQ$. Since $T$ and $L$ are independent, $L$ has the same distribution under $P$ as under $Q$. The Kunita-Yoeurp decomposition of $Q$ wrt $P$ is given by $((e^{-a t})_{t \in [0,1]}, T)$.
We claim that $Z = e^{-a \cdot}$ is a supermartingale density for $L$. Let $(\pi_t W^\pi_{t-})$ be a strategy for $L$, where $W^\pi$ is the wealth process obtained by investing in this strategy. This strategy is $1$-admissible if and only if $|\pi_t| \le 1$ for all $t \in [0,1]$. Moreover we get from \eqref{eq:ibp supermartingale density} that
\begin{align*}
d(Z W^\pi)_t \sim - W^\pi_{t-} Z_{t-} a dt + Z_{t-} \pi_t W^\pi_{t-} b dt = W^\pi_{t-} Z_{t-} (\pi_t b - a) dt.
\end{align*}
Since $W^\pi Z \ge 0$ and since $\pi_t b - a < 0$ (recall that $a > |b|$), the drift rate is negative. Therefore $Z W^\pi$ is a local supermartingale, and since it is a nonnegative process, it is a supermartingale.
Now $T$ is independent from $L$ under $Q$, and $L$ has no fixed jump times. Hence
\begin{align*}
Q(\Delta L_T \neq 0, T < \infty) = \int_{[0,\infty)} Q(\Delta L_t \neq 0) (Q\circ T^{-1})(dt) = 0,
\end{align*}
which implies that $L^{T-} = L^T$, and this is clearly no $Q$-local martingale.
\end{ex}
\begin{rmk}
In the preceding example it is possible to show that the modified process
\begin{align}\label{eq:Levy counterexample}
\tilde{L}_t = L^{T-}_t - \frac{b}{a} 1_{\{t \ge T\}}
\end{align}
is a $Q$-martingale.
More generally we expect that given a semimartingale $S$, a supermartingale density $Z$ for $S$, and a measure $Q \gg P$ with Kunita-Yoeurp decomposition $(Z,T)$ wrt $P$, there should always exist a version $\tilde{S}$ that is $P$-indistinguishable from $S$, such that $\tilde{S}$ is a $Q$-local martingale. But as \eqref{eq:Levy counterexample} shows, we will need to take different $\tilde{S}$ for different supermartingale densities. Therefore this approach seems somewhat unnatural, and we will not pursue it further. We rather note that there exists a subclass of supermartingale densities that turn $S^{T-}$ into a local martingale.
\end{rmk}
Note that all three examples had one thing in common: $T$ was not foretellable under $Q$. It turns out that if we assume $T$ to be foretellable, then things get much simpler. But it is well known, and easy to see, that $T$ is foretellable under $Q$ if and only if $Z$ is a $P$-local martingale, see \cite{Follmer1972}, Proposition~(2.1) or \cite{Yoeurp1985}, Theorem 6.
Therefore we should look for supermartingale densities that are local martingales. We call these supermartingale densities \emph{local martingale densities}. In the case of a one-dimensional $(S_t)_{t \in [0,T_\infty]}$ with finite terminal time $T_\infty$, it is shown by Kardaras \cite{Kardaras2009}, Theorem 1.1, that local martingale densities exist if and only if (NA1) is satisfied. The proof is in the spirit of the article \cite{Karatzas2007}. Takaoka \cite{Takaoka2012} solves the $d$-dimensional case with finite terminal time. More precisely it is easily deduced from Remark 7 of \cite{Takaoka2012} that for locally bounded $d$-dimensional semimartingale $(S_t)_{t \in [0, T_\infty]}$, (NA1) is satisfied if and only if there exists a local martingale density. Takoaka's proof is based on the insight of Delbaen and Schachermayer \cite{Delbaen1995b}, that a change of num\'{e}raire can induce the (NA) property, even if previously there were arbitrage opportunities in the market. \cite{Takaoka2012} continues to show that a clever choice of num\'{e}raire preserves the (NA1) property, so that then the condition (NA) + (NA1) = (NFLVR) is satisfied, which permits to apply the fundamental theorem of asset pricing \cite{Delbaen1994}.
Of course both \cite{Kardaras2009} and \cite{Takaoka2012} work with complete filtrations, but given a local martingale density $\tilde{Z}$ that is $(\mathcal{F}^P_t)$-adapted, there exists an indistinguishable process $Z$ that is $(\mathcal{F}_t)$-adapted, see Lemma \ref{lem:complete predictable optional}.
\begin{lem}
Let $(S_t)_{t \in [0,T_\infty]}$ be a locally bounded semimartingale on a finite time horizon $T_\infty < \infty$, and let $Z$ be a local martingale density for $S$. Let $T$ be a stopping time and $Q$ be a probability measure, such that $(Z/E_P(Z_0),T)$ is the Kunita-Yoeurp decomposition of $Q$ wrt $P$. Then $S^{T-}$ is a $Q$-local martingale.
Conversely if $Q \gg P$ with Kunita-Yoeurp decomposition $(Z,T)$ wrt $P$, and if $S^{T-}$ is a local martingale under $Q$, then $Z$ is a supermartingale density for $S$.
\end{lem}
\begin{proof}
The proof is very similar to the one of Theorem \ref{thm:predictable supermartingale density}. In that proof we only used the predictability of $S$ once, to obtain $E_Q( (H \cdot S)_{\sigma_n}) \le 0$ for all strategies $H$ that are 1-admissible under $Q$. Here $(\sigma_n)$ was a localizing sequence for $M$ under $P$, where $Z = Z_0 + M - A$.
So let $(\sigma_n)$ be a localizing sequence for the local martingale $Z$ under $P$, and let $H$ be a strategy that is $1$-admissible for $S$ under $Q$ (and then also under $P$). We can apply Lemma \ref{lem:yoeurps formula} with $A=0$, to obtain
\begin{align*}
E_Q(1+(H\cdot S)_{\sigma_n}) = E_P((1 + (H\cdot S)_{\sigma_n}) Z_{\sigma_n}) \le 1,
\end{align*}
because $Z$ is a supermartingale density. Now we can just copy the proof of Theorem \ref{thm:predictable supermartingale density}.
\end{proof}
We obtained our main result, a weak fundamental theorem of asset pricing.
\begin{cor*}[Theorem \ref{thm:main result}]
Let $(\mathcal{F}_t)$ be the right-continuous modification of a standard system. Let $S$ be an a.s. locally bounded stochastic process that is a.s. right-continuous. Then $S$ satisfies (NA1$_s$) if and and only there exists an enlarged probability space $(\overline{\Omega}, \overline{\mathcal{F}}, (\overline{\mathcal{F}}_t), \overline{P})$ and a dominating measure $\overline{Q} \gg \overline{P}$ with Kunita-Yoeurp decomposition $(\overline{Z}, \overline{T})$ with respect to $\overline{P}$, such that $\overline{S}^{\overline{T}-}$ is a $\overline{Q}$-local martingale.
\end{cor*}
\begin{rmk}
There is another subclass of supermartingale densities of which one might expect that they correspond to local martingale measures for $S^{T-}$: the maximal elements among the supermartingale densities. A supermartingale density $Z$ is called maximal if it is indistinguishable from any supermartingale density $Y$ that satisfies $Y_t \ge Z_t$ for all $t \ge 0$. If $S$ is not continuous, then some maximal supermartingale densities are supermartingales and not local martingales, see Example 5.1' of \cite{Kramkov1999}. It turns out that such $Z$ will usually \emph{not} correspond to local martingale measures for $S$. Assume for example that we are in the situation described in Theorem 2.2 of \cite{Kramkov1999}, i.e. we have a dual optimizer $Z$ and a primal optimizer $H$ for a certain utility maximization problem. Then point iii) of this theorem states that $(1+ (H\cdot S))Z$ is a uniformly integrable martingale. If we assume now that $Z$ is not a local martingale, as is the case in Example 5.1' of \cite{Kramkov1999}, and if $(\tau_n)$ is a localizing sequence for the martingale part $M$ of $Z = Z_0 + M - A$, then we obtain from Corollary \ref{cor:yoeurps formula 2}
\begin{align}\label{eq:rmk after main result}\nonumber
E_Q((1+(H\cdot S^{T-})_{\tau_n}) Z_{\tau_n}) &= E_P((1+(H\cdot S)_{\tau_n})Z_{\tau_n}) + E_P\left( \int_0^{\tau_n} (1+(H\cdot S)_{s-}) dA_s\right) \\
& = 1 + E_P\left( \int_0^{\tau_n} (1+(H\cdot S)_{s-}) dA_s\right),
\end{align}
where we used that $(1 + (H\cdot S))Z$ is a uniformly integrable martingale. Now, since $H$ is optimal, the wealth process $(1+(H\cdot S)_{s-})$ will be strictly positive with positive probability. Since also $dA \neq 0$ with positive probability, the expectation in \eqref{eq:rmk after main result} is strictly positive for large $n$, and therefore $H \cdot S^{T-}$ cannot be a $Q$-supermartingale, i.e. $S^{T-}$ cannot be a $Q$-local martingale.
\end{rmk}
\section{Relation to filtration enlargements}
\label{sec: jacod criterion}
Here we show that Jacod's criterion for initial filtration enlargements is in fact a criterion for the existence of a universal supermartingale density (to be defined below). We also treat general filtration enlargements. We show that if there exists a universal supermartingale density in an enlarged filtration, then a generalized version of Jacod's criterion is satisfied.
\subsection{Jacod's criterion and universal supermartingale densities}
Let $(\Omega, \mathcal{F}, (\mathcal{F}_t), P)$ be a filtered probability space, and let $(\mathcal{G}^0_t)$ be an initial filtration enlargement of $(\mathcal{F}_t)$, meaning that there exists a random variable $L$ such that $\mathcal{G}^0_t = \mathcal{F}_t \vee \sigma(L)$ for every $t \ge 0$. We define the right-continuous regularization of $(\mathcal{G}^0_t)$ by setting $\mathcal{G}_t = \cap_{s > t} \mathcal{G}^0_s$ for all $t \ge 0$.
Recall that Hypoth\`{e}se $(H')$ is satisfied if all $(\mathcal{F}_t)$-semimartingales are $(\mathcal{G}_t)$-semimartingales.
We now give the classical formulation of Jacod's criterion, see \cite{Jacod1985}. For this purpose we need to assume that $L$ takes its values in a Lusin space, which we denote by $(\mathbb{L}, \mathcal{L})$, where $\mathcal{L}$ is the Borel $\sigma$-algebra on $\mathbb{L}$. In particular the regular conditional distributions
\begin{align*}
P_t(\omega, d\ell) = P(L \in d\ell | \mathcal{F}_t) (\omega)
\end{align*}
exist for all $t \ge 0$. We write $P_L$ for the distribution of $L$. Jacod's criterion states that Hypoth\`{e}se $(H')$ is satisfied as long as almost surely
\begin{align}\label{eq:jacod}
P_t(\omega, dx) \ll P_L(dx)
\end{align}
for every $t \ge 0$.
Below we give an alternative proof of this result and we relate it to the existence of a \emph{universal supermartingale density}.
First observe that Hypoth\`{e}se $(H')$ is satisfied if and only if all nonnegative $(\mathcal{F}_t)$-martingales are $(\mathcal{G}_t)$-semimartingales: This follows by decomposing every $(\mathcal{F}_t)$-local martingale into a sum of a locally bounded local martingale and a local martingale of finite variation, and by observing that every bounded process can be made nonnegative by adding a deterministic constant.
\begin{defn}
Let $(\mathcal{G}_t)$ be a filtration enlargement of $(\mathcal{F}_t)$. Let $Z$ be an adapted process that is a.s. c\`adl\`ag, such that $P(Z_t > 0)=1$ for all $t \ge 0$. Then $Z$ is called \emph{universal supermartingale density} for $(\mathcal{G}_t)$ if $ZM$ is a $(\mathcal{G}_t)$-supermartingale for every nonnegative $(\mathcal{F}_t)$-supermartingale $M$.
\end{defn}
Note that here we do not require $Z_\infty$ to be positive. This is because local semimartingales are semimartingales, and therefore it suffices to verify the $(\mathcal{G}_t)$-semimartingale property of $M$ on $[0,t]$ for every $t \ge 0$. Hence it suffices if $Z_t > 0$ for every $t \ge 0$.
Also note that $ZM$ must be a $(\mathcal{G}_t)$-supermartingale for every nonnegative $(\mathcal{F}_t)$-supermartingale $M$, and not just for nonnegative $(\mathcal{F}_t)$-martingales.
This has the advantage that now we see immediately that in finite time every process satisfying (NA1) under $(\mathcal{F}_t)$ satisfies also (NA1) under $(\mathcal{G}_t)$: If $Y$ is a $(\mathcal{F}_t)$-supermartingale density for $S$, then $ZY$ is a $(\mathcal{G}_t)$-supermartingale density for $S$.
The first result of this section shows that Jacod's criterion is not so much a criterion for Hypoth\`ese $(H')$ to hold, but rather a criterion for the existence of a universal supermartingale density.
\begin{prop}\label{prop:jacod}
Let $(\mathcal{G}_t)$ be an initial enlargement of $(\mathcal{F}_t)$ with a random variable $L$ taking its values in a Lusin space. Assume Jacod's criterion \eqref{eq:jacod} is satisfied. Then there exists a universal supermartingale density for $(\mathcal{G}_t)$
\end{prop}
\begin{proof}
\begin{enumerate}
\item Define for every $t \ge 0$
\begin{align*}
Y_t(\omega, \ell) = \frac{dP_t(\omega, \cdot)}{dP_L}(\ell).
\end{align*}
By Doob's disintegration theorem there exists a version of $Y_t$ that is $\mathcal{F}_t \otimes \mathcal{L}$-measurable, see also the proof of Theorem VI.2.10 in \cite{Protter2004}. Let $t, s \ge 0$. We first show that $P \otimes P_L$-almost surely
\begin{align}\label{eq:Yt zero then Yt+s zero}
\{(\omega, \ell): Y_t(\omega,\ell) = 0\} \subseteq \{(\omega, \ell): Y_{t+s}(\omega,\ell) = 0\}.
\end{align}
Note that $Y_{t+s} \ge 0$, and therefore by Fubini's Theorem and the tower property
\begin{align*}
\int_{\Omega \times \mathbb{L}} &1_{\{Y_t(\omega, \ell) = 0\}} Y_{t+s}(\omega, \ell) P\otimes P_L(d\omega, d\ell) = \int_\Omega \int_\mathbb{L} 1_{\{Y_t(\omega, \ell) = 0\}} P_{t+s}(\omega, d\ell) P(d\omega) \\
& = \int_\Omega \int_\mathbb{L} 1_{\{Y_t(\omega, L(\omega)) = 0\}} P(d\omega) = \int_\Omega \int_\mathbb{L} 1_{\{Y_t(\omega, \ell) = 0\}} P_t(\omega, d\ell) P(d\omega) = 0,
\end{align*}
since $P_t(\omega, \cdot)$-almost surely $Y_t(\omega, \cdot) > 0$.
\item Define $\tilde{Z}_t(\omega,\ell) = 1_{\{Y_t(\omega, \ell) > 0\}}/Y_t(\omega, \ell)$ and
\begin{align*}
Z_t (\omega) = \tilde{Z}_t(\omega, L(\omega)).
\end{align*}
This $Z$ is $(\mathcal{G}_t)$-adapted. Let $M$ be a nonnegative $(\mathcal{F}_t)$-supermartingale.
Let $t, s \ge 0$, let $A \in \mathcal{F}_t$, and $B \in \mathcal{L}$. Then we can apply the tower property to obtain
\begin{align*}
E&\left( 1_{A}1_B(L)M_{t+s}Z_{t+s}\right) = \int_\Omega 1_{A}(\omega) \int_\mathbb{L} 1_B(\ell)M_{t+s}\tilde{Z}_{t+s}(\omega, \ell) P_{t+s}(\omega, d\ell) P(d\omega) \\
& = \int_\Omega 1_{A}(\omega) \int_\mathbb{L} 1_B(\ell)M_{t+s}(\omega) \frac{Y_{t+s}(\omega, x)}{Y_{t+s}(\omega, x)}1_{\{Y_{t+s}(\omega, x) > 0\}} P_L(d\ell) P(d\omega) \\
& \le \int_\Omega 1_{A}(\omega) \int_\mathbb{L} 1_B(\ell) M_{t+s}(\omega) 1_{\{Y_{t}(\omega, x) > 0\}} P_L(d\ell) P(d\omega).
\end{align*}
In the last step we used \eqref{eq:Yt zero then Yt+s zero} and that $1_{A}(\omega) 1_B(\ell)M_{t+s}(\omega)$ is $P_L \otimes P$-a.s. nonnegative. Using the $(\mathcal{F}_t)$-supermartingale property of $M$ in conjunction with Fubini's theorem, we obtain
\begin{align*}
\int_\Omega &1_{A}(\omega) \int_\mathbb{L} 1_B(\ell) M_{t+s}(\omega) 1_{\{Y_{t}(\omega, \ell) > 0\}} P_L(d\ell) P(d\omega) \\
& \le \int_\mathbb{L} 1_B(\ell) \int_\Omega 1_{A}(\omega) M_{t}(\omega) 1_{\{Y_{t}(\omega, \ell) > 0\}} P(d\omega)P_L(d\ell) \\
& = \int_\Omega 1_{A}(\omega) \int_\mathbb{L} 1_B(\ell) M_{t}(\omega) \frac{Y_t(\omega, \ell)}{Y_t(\omega, \ell)}1_{\{Y_{t}(\omega, \ell) > 0\}} P_L(d\ell) P(d\omega) \\
& = \int_\Omega 1_{A}(\omega) \int_\mathbb{L} 1_B(\ell) M_{t}(\omega) \tilde{Z}_t(\omega, \ell) P_t(\omega, d\ell) P(d\omega) \\
& = E\left[ 1_{A} 1_B(L) (1 + (H \cdot S)_t) Z_t\right]
\end{align*}
The monotone class theorem allows to pass from sets of the form $A \cap L^{-1}(B)$ to general sets in $(\mathcal{G}^0_t)$, and therefore $MZ$ is a $(\mathcal{G}^0_t)$-supermartingale. Taking $M \equiv 1$, we see that also $Z$ is a $(\mathcal{G}^0_t)$-supermartingale.
\item Let us show that $Z_t$ is $P$-a.s. strictly positive for every $t \ge 0$. For this purpose it suffices to show that $P(\omega: Y_t(\omega, L(\omega))=0) = 0$. By the tower property
\begin{align*}
E(1_{\{Y_t(\cdot, L(\cdot)) = 0\}}) & = \int_\Omega \int_\mathbb{L} 1_{\{Y_t(\omega, \ell) = 0\}} P_t(\omega, d\ell) P(d\omega) = 0.
\end{align*}
\item $Z$ is not necessarily right-continuous, and also we did not show yet that $ZM$ is a $(\mathcal{G}_t)$-supermartingale and not just a $(\mathcal{G}^0_t)$-supermartingale. But the construction of a right-continuous universal supermartingale density is now done exactly as in the proof of Theorem \ref{thm:abstract ex supermartingale density}. The supermartingale property under $(\mathcal{G}_t)$ follows from Corollary 2.2.10 of \cite{Ethier1986}, which states that $E(X|\mathcal{G}_t) = \lim_{s \downarrow t} E(X|\mathcal{G}^0_s)$ for every $L^1$-random variable $X$.
\end{enumerate}
\end{proof}
\begin{rmk}
If we are only interested whether Hypoth\`ese $(H')$ holds and not whether there exists a universal supermartingale density, then we can also work with the unregularized filtration $(\mathcal{G}^0_t)$. Since Hypoth\`ese $(H')$ holds for $(\mathcal{G}_t)$ and since $(G^0_t)$ is a filtration shrinkage of $(\mathcal{G}_t)$, Stricker's theorem implies that Hypoth\`ese $(H')$ is also satisfied for $(\mathcal{G}^0_t)$.
\end{rmk}
\begin{rmk}\label{rmk:filtration discussion}
We may replace assumption \eqref{eq:jacod} by $P_t(\omega, d\ell) \gg P_L(d\ell)$ or $P_t(\omega, d\ell) \sim P_L(d\ell)$. In the first case we could use the same proof as for Proposition \ref{prop:jacod} to obtain the existence of a nonnegative martingale $Z$, not necessarily strictly positive, such that $Z M$ is a $(\mathcal{G}_t)$-supermartingale for every nonnegative $(\mathcal{F}_t)$-supermartingale $M$. In particular there exists an absolutely continuous measure $Q \ll P$, such that every locally bounded $(P,(\mathcal{F}_t))$-local martingale is a $(Q,(\mathcal{G}_t))$-local martingale. Since (NA) is related to the existence of absolutely continuous local martingale measures, see \cite{Delbaen1995a}, this indicates that the (NA) property may be stable under initial filtration enlargements that satisfy this ``reverse Jacod condition''. Of course it is much harder to satisfy this assumption, for example it will never be satisfied if $L$ is $\mathcal{F}_t$-measurable for some $t \ge 0$.
In case $P_t(\omega, d\ell) \sim P_L(d\ell)$, the same proof as for Proposition \ref{prop:jacod} yields the existence of an equivalent measure $Q \sim P$, such that every nonnegative $(P,(\mathcal{F}_t))$-supermartingale is a nonnegative $(Q,(\mathcal{G}_t))$-supermartingale. In particular, every locally bounded $(P,(\mathcal{F}_t))$-local martingale is a $(Q,(\mathcal{G}_t))$-local martingale. This condition has been intensely studied by Amendinger, Imkeller and Schweizer \cite{Amendinger1998} as well as Amendinger \cite{Amendinger2000}. Obviously it is harder to satisfy than Jacod's condition or the reverse Jacod condition. In financial applications one may however assume that the knowledge of the ``insider'' is perturbed by a small Gaussian noise that is independent of $\mathcal{F}_\infty$ (or more generally by an independent noise with strictly positive density wrt Lebesgue measure). Then $P_t(\omega, d\ell) \sim P_L(d\ell)$ is always satisfied. If the density of the noise is not strictly positive, then we only obtain $P_t(\omega, d\ell) \ll P_L(d\ell)$.
\end{rmk}
\subsection{Universal supermartingale densities and the generalized Jacod criterion}
Let $(\Omega, \mathcal{F}, (\mathcal{F}_t), P)$ be a filtered probability space and let $\mathcal{G}_t \supseteq \mathcal{F}_t$ be a filtration enlargement. Assume that $\mathcal{G}_t$ is countably generated for every $t \ge 0$, to wit, $\mathcal{G}_t = \sigma(B^t_1, B^t_2, \dots)$. In particular the regular conditional probabilities
\begin{align*}
P_t(\omega, \cdot)|_{\mathcal{G}_t} = P(\cdot |\mathcal{F}_t)|_{\mathcal{G}_t}(\omega)
\end{align*}
exist. We say that the \emph{generalized Jacod condition} is satisfied if for all $t, s \ge 0$ a.s.
\begin{align*}
P_{t+s}|_{\mathcal{G}_t}(\omega, \cdot) \ll P_t|_{\mathcal{G}_t}(\omega, \cdot).
\end{align*}
It is known that neither Jacod's condition nor the generalized Jacod condition are necessary for Hypoth\`{e}se $(H')$ to hold. But if we assume that there exists a universal supermartingale density for $(\mathcal{G}_t)$, then the generalized Jacod condition necessarily holds.
\begin{prop}
Assume that there exists a universal supermartingale density for $(\mathcal{G}_t)$. Then the generalized Jacod condition is satisfied.
\end{prop}
\begin{proof}
\begin{enumerate}
\item For every $A \in \mathcal{F}$ the process $M^A_t:=E_P(1_A|\mathcal{F}_t)$, $t \ge 0$, is a nonnegative $(\mathcal{F}_t)$-martingale. Therefore $M^AZ$ is a $(\mathcal{G}_t)$-supermartingale. Fix $t, s \ge 0$. Let $A \in \mathcal{F}_{t+s}$ and $B \in \mathcal{G}_t$. Then for every $n \in \mathbb{N}$
\begin{align*}
E\left(1_A 1_B \frac{Z_{t+s}}{Z_t} 1_{\{Z_t \ge 1/n\}}\right) & = E\left( \frac{1_B 1_{\{Z_t \ge 1/n\}}}{Z_t} M^A_{t+s} Z_{t+s} \right) \\
& \le E\left( \frac{1_B 1_{\{Z_t \ge 1/n\}}}{Z_t} M_{t}^A Z_{t}\right) \\
& = E(1_A E(1_B 1_{\{Z_t \ge 1/n\}}|\mathcal{F}_t)).
\end{align*}
Applying monotone convergence on both sides, we obtain that
\begin{align*}
E\left(1_A 1_B \frac{Z_{t+s}}{Z_t}\right) \le E(1_A E_P(1_B |\mathcal{F}_t)).
\end{align*}
The same inequality holds if we replace $Z_{t+s}/Z_t$ by a version $\tilde{Z}_{t+s}/\tilde{Z}_t$ that is strictly positive for \emph{every} $\omega \in \Omega$. Since the inequality holds for all $A \in \mathcal{F}_{t+s}$, this implies
\begin{align*}
\int 1_B(\omega') \frac{\tilde{Z}_{t+s}}{\tilde{Z}_t}(\omega') P_{t+s}(\omega, d\omega') \le P_t(\omega, B) \text{ for almost every } \omega \in \Omega.
\end{align*}
This looks promising. The only problem is that the null set outside of which this inequality holds may depend on $B$.
\item Now we use the assumption that $\mathcal{G}_t$ is countably generated. This means that we can find an increasing sequence of finite partitions
\begin{align*}
\mathcal{P}^{n} = \cup_{k=1}^{K_n} B_k^n
\end{align*}
of $\Omega$ such that $\mathcal{G}_t = \sigma(\mathcal{P}^n: n \ge 0)$.
Since $\cup_{n \ge 0} \sigma(\mathcal{P}^n)$ is countable, we can choose a null set $N$ such that for all $\omega \in \Omega \backslash N$ and all $B \in \cup_n \sigma(\mathcal{P}^n)$ we have
\begin{align}\label{eq:generalized jacod proof}
\int 1_B(\omega') \frac{\tilde{Z}_{t+s}}{\tilde{Z}_t}(\omega') P_{t+s}(\omega, d\omega') \le P_t(\omega, B).
\end{align}
Now $B \in \cup_n \sigma(\mathcal{P}^n)$ is stable under finite intersections (it even is an algebra). The monotone class theorem then implies that \eqref{eq:generalized jacod proof} holds for all $B \in \sigma(\mathcal{P}^n: n \ge 0) = \mathcal{G}_t$. Since $\tilde{Z}_{t+s}(\omega')/\tilde{Z}_t(\omega') > 0$ for every $\omega' \in \Omega$, the proof is complete.
\end{enumerate}
\end{proof}
\begin{cor}\label{cor:generalized jacod complete}
Let $(\mathcal{F}_t)$ be a filtration for which a continuous local martingale $M$ has the predictable representation property. Assume that under $(\mathcal{G}_t)$, $M$ is of the form
\begin{align*}
M_t = \tilde{M}_t + \int_0^t \alpha_s d\langle \tilde{M}\rangle_s
\end{align*}
for a $(\mathcal{G}_t)$-local martingale $\tilde{M}$ and a predictable integrand $\alpha \in L^2_{\text{loc}}(\tilde{M})$. Then the generalized Jacod condition holds.
\end{cor}
\begin{proof}
In this case the stochastic exponential
\begin{align*}
Z_t = \exp\left( -\int_0^t \alpha_s d \tilde{M}_s - \frac{1}{2} \int_0^t \alpha^2_s d\langle \tilde{M} \rangle_s \right)
\end{align*}
is a universal supermartingale density.
\end{proof}
Corollary \ref{cor:generalized jacod complete} was previously shown by Imkeller, Pontier and Weisz \cite{Imkeller2001} for the case of initial enlargements and under the stronger assumption
\begin{align*}
E\left(\int_0^\infty \alpha^2_s d\langle \tilde{M} \rangle_s\right) < \infty.
\end{align*}
For simplicity we gave the one-dimensional formulation of Corollary \ref{cor:generalized jacod complete}. Of course the same argument works in the multidimensional setting: if $M = (M^1, \dots, M^d)$ has the predictable representation under $(\mathcal{F}_t)$, and if $M$ satisfies the structure condition under $(\mathcal{G}_t)$, then the generalized Jacod condition is satisfied.
\begin{appendix}
\section{Convex compactness and Tychonoff's theorem}
In Section \ref{sec: supermartingale densities} we needed Tychonoff's theorem for products of convexly compact spaces. Since we only applied the result to a countable product of metric spaces, we will only prove this special case. Let us recall the following definitions.
\begin{defn}
\begin{enumerate}
\item A set $A$ is called \emph{directed} if it is partially ordered and if for every $a, b \in A$ there exists $c \in A$ such that $a \le c$ and $b \le c$.
\item Let $X$ be a topological space. A \emph{net} in $X$ is a map from some directed set $A$ to $X$.
\item A net $\{x_\alpha\}_{\alpha \in A}$ in $X$ \emph{converges} to a point $x \in X$ if for every open neighborhood $U$ of $x$ there exists $\alpha \in A$, such that $x_{\alpha'} \in U$ for every $\alpha' \ge \alpha$.
\end{enumerate}
\end{defn}
\begin{ex}
If $A = \mathbb{N}$, then a net in $X$ is just a sequence with values in $X$.
\end{ex}
Zitkovic \cite{Zitkovic2010} introduces the notation $\mathrm{Fin}(A)$, which denotes all non-empty finite subsets of a given set $A$. If $B$ is a subset of a vector space, then $\mathrm{conv}(B)$ denotes the convex hull of $B$. Zitkovic then gives the following definition.
\begin{defn}\label{def:subnet convex}
Let $\{x_\alpha\}_{\alpha \in A}$ be a net in a topological vector space $X$. A net $\{y_\beta\}_{\beta \in B}$ is called a \emph{subnet of convex combinations of $\{x_\alpha\}_{\alpha \in A}$} if there exists a map $D: B \rightarrow \mathrm{Fin}(A)$ such that
\begin{enumerate}
\item $y_\beta \in \mathrm{conv}\{x_\alpha: \alpha \in D(\beta)\}$ for every $\beta \in B$, and
\item for every $\alpha \in A$ there exists $\beta \in B$ such that $\alpha' \ge \alpha$ for all $\alpha' \in \cup_{\beta' \ge \beta} D(\beta')$.
\end{enumerate}
\end{defn}
\begin{lem}\label{rmk:subnet of subnet}
Let $\{y_\beta\}_{\beta \in B}$ be a subnet of convex combinations of $\{x_\alpha\}_{\alpha \in A}$, and let $\{z_\gamma\}_{\gamma \in C}$ be a subnet of convex combinations of $\{y_\beta\}_{\beta \in B}$. Then $\{z_\gamma\}_{\gamma \in C}$ is a subnet of convex combinations of $\{x_\alpha\}_{\alpha \in A}$.
\end{lem}
\begin{proof}
Let $D_B: B \rightarrow \mathrm{Fin}(A)$ and $D_C: C \rightarrow \mathrm{Fin}(B)$ be two maps as described in Definition \ref{def:subnet convex}, $D_B$ for $\{y_\beta\}_{\beta \in B}$ and $D_C$ for $\{z_\gamma\}_{\gamma \in C}$. Define
\begin{align*}
D: C \rightarrow \mathrm{Fin}(A), \qquad D(\gamma) = \cup_{\beta \in D_C(\gamma)} D_B(\beta).
\end{align*}
Then we have for all $\gamma \in C$
\begin{align*}
z_\gamma \in \mathrm{conv}\{y_\beta: \beta \in D_C(\gamma)\} \subseteq \mathrm{conv}\{x_\alpha: \alpha \in \cup_{\beta \in D_C(\gamma)} D_B(\beta)\} = \mathrm{conv} \{x_\alpha: \alpha \in D(\gamma)\},
\end{align*}
and therefore condition (1) of Definition \ref{def:subnet convex} is satisfied. As for condition (2), let $\alpha \in A$. Then there exists $\beta \in B$, such that $\alpha'\ge \alpha$ for all $\alpha' \in \cup_{\beta'\ge \beta} D_B(\beta')$. For this given $\beta$, there exists $\gamma \in C$, such that $\beta' \ge \beta$ for all $\beta' \in \cup_{\gamma'\ge \gamma} D_C(\gamma')$. Hence $\alpha' \ge \alpha$ for all
\begin{align*}
\alpha' \in \cup_{\gamma'\ge \gamma} D(\gamma') = \cup_{\gamma'\ge \gamma} \cup_{\beta' \in D_C(\gamma')} D_B(\beta') \subseteq \cup_{\beta' \ge \beta} D_B(\beta').
\end{align*}
\end{proof}
One of Zitkovic's \cite{Zitkovic2010} main results is the following Proposition.
\begin{prop}\label{prop:characterization convex compactness}
A closed and convex subset $C$ of a topological vector space $X$ is convexly compact if and only if for any net $\{x_\alpha: \alpha \in A\}$ in $C$ there exists a subnet of convex combinations $\{y_\beta: \beta \in B\}$, such that $\{y_\beta\}$ converges to some $y \in X$.
\end{prop}
We will use this insight to prove a weak version of Tychonoff's theorem for convexly compact sets.
\begin{prop}\label{prop: tychonoff result}
Let $\{X_n: n \in \mathbb{N}\}$ be a \emph{countable} family of convexly compact \emph{metric} spaces. Then $\prod_{n \in \mathbb{N}} X_n$ is convexly compact in the product topology.
\end{prop}
\begin{proof}
Let $\{(x_\alpha(n))_{n\in\mathbb{N}}: \alpha \in A\}$ be a net in $\prod_{n \in \mathbb{N}} X_n$. Then $\{x_\alpha(1): \alpha \in A\}$ is a net in $X_1$. By Proposition \ref{prop:characterization convex compactness}, there exists a subnet of convex combinations $\{(y^1_\beta(n))_{n \in \mathbb{N}}: \beta \in B_1\}$, such that $\{y^1_\beta(1): \beta \in B_1\}$ converges to some $y(1) \in X_1$. We can now inductively construct for every $k > 1$ a subnet of convex combinations
\begin{align*}
\{(y^k_\beta(n))_{n \in \mathbb{N}}: \beta \in B_k\} \text{ of } \{(y^{k-1}_\beta(n))_{n \in \mathbb{N}}: \beta \in B_{k-1}\},
\end{align*}
such that $\{y^k_\beta(k): \beta \in B_k\}$ converges to some $y(k) \in X_k$. By Remark \ref{rmk:subnet of subnet}, every $\{(y^k_\beta(n))_{n \in \mathbb{N}}: \beta \in B_k\}$ is a subnet of convex combinations of $\{x_\alpha: \alpha \in A\}$. We denote the corresponding maps from $B_k$ to $\mathrm{Fin}(A)$ by $D_k$. Note that by construction, $\{y^k_\beta(l): \beta \in B_k\}$ converges to $y(l)$ for all $1 \le l \le k$. Now take the directed set $\mathbb{N} \times A$ with the partial order $(k, \alpha) \le (k', \alpha')$ if $k \le k'$ and $\alpha \le \alpha'$. We write $d_l$ for the distance on $X_l$. Define for $(k,\alpha) \in \mathbb{N}\times A$ the set of ``admissible indices'' as
\begin{align*}
B(k, \alpha) = \left\{\beta \in B_k:\alpha' \ge \alpha \text{ for all } \alpha' \in D_k(\beta) \text{ and } d_l(y^k_\beta(l), y(l)) \le \frac{1}{k}\text{ for } l=1,\dots, k \right\}.
\end{align*}
By construction of the $(y^k_\beta(n))_{n \in \mathbb{N}}$, every $B(k,\alpha)$ is non-empty. For every $(k, \alpha) \in \mathbb{N} \times A$ choose $\beta(k, \alpha)\in B(k,\alpha)$. Note that here we explicitly apply the Axiom of Choice! Set $z_{(k,\alpha)} = y^k_{\beta(k, \alpha)}$ and $D((k,\alpha)) = D_k(\beta(k,\alpha))$. Then by construction, $\{z_{(k, \alpha)}: (k, \alpha) \in \mathbb{N} \times A\}$ is a subnet of convex combinations of $\{x_\alpha: \alpha \in A\}$, which converges to some $(y(n))_{n \in \mathbb{N}} \in \prod_{n \in \mathbb{N}} X_n$ in the product topology. Therefore $\prod_n X_n$ is convexly compact in the product topology.
\end{proof}
\begin{rmk}
The proof is surprisingly technical considering that we have a countable product of metric spaces. In this case compactness is equivalent to sequential compactness, and therefore the proof of Tychonoff's theorem is a triviality, based on a diagonal sequence argument. But so far there seems to be no characterization of convex compactness in terms of sequential compactness, and therefore we had to work with nets rather than with sequences.
\end{rmk}
\section{The question of complete filtrations}\label{app:complete filtration}
Here we collect some classical observations that allow to transfer results of other authors that were obtained under complete filtrations to our setting. We follow \cite{Jacod2003}.
Let $(\Omega, \mathcal{F}, (\mathcal{F}_t)_{t \ge 0}, P)$ be a filtered probability space with a right-continuous filtration $(\mathcal{F}_t)$. Write $\mathcal{F}^P$ for the $P$-completion of $\mathcal{F}$, and $\mathcal{N}^P$ for the $P$-null sets of $\mathcal{F}^P$. Then $\mathcal{F}_t^P = \mathcal{F}_t \vee \mathcal{N}^P$ is the $\sigma$-algebra generated by $\mathcal{F}_t$ and $\mathcal{N}^P$, and $(\mathcal{F}^P_t)$ satisfies the usual conditions. We call $(\Omega, \mathcal{F}^P, (\mathcal{F}^P_t), P)$ the \emph{completion} of $(\Omega, \mathcal{F}, (\mathcal{F}_t), P)$.
Recall that the optional $\sigma$-algebra over $(\mathcal{F}_t)$ is the $\sigma$-algebra on $\Omega \times [0,\infty)$ that is generated by all processes of the form $X_t(\omega) = 1_A(\omega) 1_{[r,s)}(t)$ for some $0\le r < s < \infty$ and $A \in \mathcal{F}_r$. The predictable $\sigma$-algebra over $(\mathcal{F}_t)$ is the $\sigma$-algebra on $\Omega \times [0,\infty)$ that is generated by all processes of the form $X_t(\omega) = 1_A(\omega) 1_{\{0\}}(t) + 1_B(\omega) 1_{(r,s]}(t)$ for some $0\le r < s < \infty$, for $A \in \mathcal{F}_0$, and $B \in \mathcal{F}_r$. Similarly we define the predictable and optional $\sigma$-algebra over $(\mathcal{F}^P_t)$.
The first result relates stopping times under $(\mathcal{F}_t)$ and under $(\mathcal{F}^P_t)$.
\begin{lem}[Lemma I.1.19 of \cite{Jacod2003}]\label{lem:complete stopping time}
Any stopping time on the completion $(\Omega, (\mathcal{F}^P_t))$ is a.s. equal to a stopping time on $(\Omega, (\mathcal{F}_t))$.
\end{lem}
Next, we show a similar result on the level of processes.
\begin{lem}\label{lem:complete predictable optional}
Any predictable (resp. optional) process on the completion $(\Omega, (\mathcal{F}^P_t))$ is indistinguishable from a predictable (resp. optional) process on $(\Omega, (\mathcal{F}_t))$.
\end{lem}
\begin{proof}
The predictable case is Lemma I.2.17 of \cite{Jacod2003}. The proof of the optional case works exactly in the same way: the claim is trivial for the generating processes described above, and we can use the monotone class theorem to pass to indicator functions of general optional sets. Then we use monotone convergence to pass to general optional processes.
\end{proof}
This allows us to obtain a similar result for c\`adl\`ag processes.
\begin{lem}\label{lem:complete cadlag}
Let $S$ be a $(\mathcal{F}^P_t)$-adapted process that it a.s. c\`adl\`ag. Then $S$ is indistinguishable from a $(\mathcal{F}_t)$-adapted process (which is then of course
a.s. c\`adl\`ag as well).
\end{lem}
\begin{proof}
Since $(\mathcal{F}^P_t)$ is complete, $S$ admits an indistinguishable version $\tilde{S}$ that is $(\mathcal{F}_t^P)$-adapted and c\`adl\`ag for \emph{every} $\omega \in \Omega$. This $\tilde{S}$ is optional, so now the result follows from Lemma \ref{lem:complete predictable optional}.
\end{proof}
\end{appendix}
\textbf{Acknowledgement:} N.P. thanks Asgar Jamneshan for the introduction to filtration enlargements. We are grateful to Stefan Ankirchner, Kostas Kardaras, and Johannes Ruf for their helpful comments on an earlier version of this paper. We thank Alexander Gushchin for pointing out the reference \cite{Rokhlin2010}. Part of the research was carried out while the authors were visiting the Department of Aerospace Engineering at the University of Illinois at Urbana-Champaign. We are grateful for the hospitality at UIUC. N.P. is supported by a Ph.D. scholarship of the Berlin Mathematical School.
|
\section{Introduction}
The effect of spatial graph coupling, first discovered for low-density error correction codes, attracts increasing interest from the communications research community. Iterative decoding of block low-density parity-check (LDPC) codes, despite its known strength, fails to achieve the same limits as the optimum maximum likelihood (ML) decoding. It has been shown, however~\cite{lscz10}, and then rigorously proved~\cite{kru10}, that iterative decoding thresholds of convolutional (or spatially coupled) LDPC codes coincide with the ML decoding thresholds. A coupled code is constructed by copying the protograph of an initial block code a number of times and then connecting the neighboring copies to form a chain. The graphs located at the ends of the chain contain variable nodes with fewer constraints, since these graphs are connected to their neighbors from only one side. As a result, iterative decoding progress initiates at the ends of the chain, due to the effect of the slight irregularity, and then propagates through the entire chain. The principle of spatial graph coupling has proven to be applicable to several other areas, including multiple user detection, compressive sensing, and quantum coding.
In this paper, we focus our attention on spatial graph coupling in relation to multiple access communications. The capacity and the achievable rate region for the multiple access channel (MAC) are well known. The corner points of the rate region polytope can be achieved by successive (onion peeling) decoding~\cite{CovTho06} while the middle part is achieved by rate splitting and/or time sharing added to the onion peeling~\cite{RimUrb01}. However, communication approaches that allow for more robust and less complex joint parallel detection/decoding, such as code-division multiple access (CDMA), have had limited success in achieving the inner points of the MAC rate region. The equal power user case is typically the most difficult, since all users happen to be operating under same conditions, and there is no structural irregularity to initiate the decoding convergence. It has been shown that regular random CDMA can only support a fixed system load\footnote{The system load $\alpha$ is defined as the ratio of the number of supported users $K$ to the number of available signal dimensions $N$ (or chips in CDMA), i.e, $\alpha=K/N$.} equal to $\alpha= 1.49$~\cite{Tanaka2002} with equal power users, while sparse synchronous CDMA can only support $\alpha=2.07$~\cite{TruSchKrz2007_IT}.
In this work we consider a multiple access multiuser situation where each user modulates his signal as a sum of redundant, independent, equal power, data streams, as described in \cite{TruSchKrz2007_IT}\cite{GM2010}. The modulated signal allows for sparse graph representation similar to the LDPC code's Tanner graph. The sparse graphs of the transmitted signals couple together ``in the air,'' and the receiver observes a coupled sequence of graphs which is later processed by an iterative detector/decoder. We focus on a communication scenario in which all users transmit an equal number of equal power BPSK data streams that are encoded by error correction codes optimal for the binary input additive white Gaussian noise (AWGN) channel. We then prove that the (real-valued) AWGN channel capacity can be achieved exactly by such a system.
\section{System Model}
We consider the generalized modulation type format~\cite{GM2010}. Each user modulates a number of independent data streams and transmits their sum. One data stream is modulated as follows. First the data is encoded by a binary error control encoder to produce a binary data stream $\{u_{l}\}^L_{l=1}$. The stream is first duplicated $M$ times resulting in $M$ identical data sub-streams, $\{u_{m,l}\}^{L}_{l=1}$, $m=1,2,\ldots,M$. Each sub-stream $\{u_{m,l}\}^{L}_{l=1}$ is permuted by an interleaver $\pi_{m}$ to produce $\{\tilde{u}_{m,l}\}^{L}_{l=1}$, $m \in \{1,2,\ldots,M\}$. At the next step each bit of $\{\tilde{u}_{m,l}\}^{L}_{l=1}$ is multiplied by an $N$-dimensional signature vector $\vec{s}_{m}$. Signature vector maps the data into the available resource space. This operation is similar to spreading in a CDMA system.
Each of the $M$ sub-streams is multiplied by an amplitude $\sqrt{P/M}$ and they are all added up to produce one modulated signal stream
\begin{equation}
\vec{v}_{l} = \sum_{m=1}^{M} \sqrt{\frac{P}{M}} \tilde{u}_{m,l} \vec{s}_{m} \quad l=1,2,\ldots,L\ .
\label{eq:userkTx}
\end{equation}
Each signature sequence $\vec{s}_{m} = \{s_{m,1},s_{m,2},\ldots,s_{m,N}\}$ is real and energy normalized such that $\mathbb{E} |s_{m,i}|^2 = \frac{1}{N}$, $i=1,2,\ldots,N$.
We also require~\footnote{The following property can be easily guaranteed by
choosing $s_{m,i} \in \left\{-\frac{1}{\sqrt{N}},\frac{1}{\sqrt{N}}\right\}$ iid Bernoulli with probability $1/2$. However, other signature sequence choices are also possible.} $\mathbb{E} \vec{s}_{m_1} \vec{s}\ensuremath{^{\mathrm{T}}}_{m_2} = 1/N$, $m_1 \neq m_2$.
We consider the case in which $2W+1$ users transmit $K/(2W+1)$ data streams each, producing total system load of $\alpha=K/N$. The power of each modulated stream is $P=1$ and communication happens over the real-valued AWGN channel with noise power $\sigma^2$. We focus on a scenario in which the packets of the users are received asynchronously as shown in Fig.~\ref{Fig:Packets}. Each packet is divided into $2W+1$ sections of length $(2W+1)/L$ bits each. At time slot $t=-W$ the first user starts to transmit his packet and transmits one section. At time $t=-W+1$ the second user joins and transmits one section while the first user transmits his second section. At time $t=-W+2$ the third user joins and so on. Once a user finishes transmitting his packet he immediately starts to transmit his next packet.
\begin{figure}[t]
\setlength{\unitlength}{1mm}
\begin{picture}(85,37)
\put(0,-2){\includegraphics{CoupledStreamsFig.pdf}
}
\end{picture}
\caption{Transmitted packets coupling in the channel, $W=2$.}
\label{Fig:Packets}
\end{figure}
We consider two types of receivers. The first receiver is a {\em two-stage receiver} where the first stage is an iterative parallel interference cancellation (PIC) which layers the received data streams and the second stage is an error control decoding performed for all layered streams independently and in parallel. The first, PIC stage, can be described as follows. At every iteration the received signals are filtered using the signature sequences $\vec{s}_{k,m}$. Assuming a simple matched filter, the filtered signal of data bit $u_{k,l}$ from data stream $k \in \{1,2,\ldots,K\}$, and sub-stream $m \in \{1,2,\ldots,M\}$ equals
\begin{align}
q_{k,m,l} &= \frac{1}{\sqrt{M}} \vec{s}\ensuremath{^{\mathrm{T}}}_{k,m} \bigl( u_{k,m,l} \vec{s}_{k,m} + \hspace*{-4mm}\sum_{(k_1,m_1) \neq (k,m)} \tilde{u}_{k_1,m_1,l} \vec{s}_{k_1,m_1} \bigr) \nonumber \\
&= \frac{1}{\sqrt{M}} u_{k,m,l} + I_{k,m,l}
\end{align}
where $I_{k,m,t}$ is the interference, suppressed by the filtering. Each bit $u_{k,l}$ has $M$ replicas $u_{k,m,l}$ which are dispersed along the packet. The power $x_m$ of the interference term $I_{k,m,l}$ depends on the section of the packet where replica $u_{k,m,l}$ belongs to. The signal power of each replica is $1/M$. We estimate the transmitted bits by summing up weighted received replicas $q_{k,m,l}$ and applying the conditional expectation estimate
\begin{equation}
\hat{u}_{k,m,l} = \tanh \left(\sum_{m_1 \neq m} \xi_{m_1} q_{k,m_1,l} \right)\ .
\label{eq:bitest}
\end{equation}
The optimal weighting coefficients
\[
\xi_m = \frac{1}{x_m}\left(\sum_{m'=1}^M \frac{1}{x_{m'}} \right)^{-1}
\]
ensure that the signal-to-noise-and-interference ratio (SINR) of the weighted sum inside the parenthesis in (\ref{eq:bitest}) is maximized and equals
\begin{equation}
z_1 = \frac{1}{M} \sum_{m_1 \neq m} \frac{1}{x_{m_1}}\ .
\label{eq:SINR}
\end{equation}
The bit estimates (\ref{eq:bitest}) are used to reconstruct the entire transmitted signals (\ref{eq:userkTx}) and perform interference cancellation by subtraction of the reconstructed signals from the received signal. The iterative process is repeated $i$ times until the SINR $z_i$ of the data bits at iteration $i$ becomes greater than the decoding threshold $\theta$ of the error control code used for their encoding. The error control decoding (second stage) can than be applied to all the packets (received at given time) for the final error correction.
The second receiver, which we call a {\em modified successive interference cancellation (SIC) receiver} is working as follows. It starts with the first stage of the two-stage receiver described above. However, the error correction decoding of the data streams for which $z_i > \theta$ is performed after each interference cancellation iteration. These decoded streams are then removed from the interference pool. The modified SIC receiver is slightly more complex since it requires a feedback loop between the decoder and the interference canceller. However, the modified SIC allows for simpler analysis.
\section{Analysis}
We start with an analysis of the iterative interference cancellation, i.e., the first stage of the two-stage receiver. We track the evolution of the noise-and-interference variance which we denote by $x_i^t$, and the SINR of the data bits which we denote by $z_i^t$. The upper index $t$ indicates the time slot of interest and $i$ is the iteration number. A discrete system of equations describing the evolution of $x_i^t$ and $z_i^t$ have been derived in~\cite{SchTruISIT11}
\begin{align}
x_i^t &= \frac{\alpha}{2W+1} \sum_{j=-W}^{W} g\left(z_i^{t+j}\right) + \sigma^2 &i\geq 0, t \geq 1\label{eq:xdis} \\
z_i^t &= \frac{1}{2W+1}\sum_{j=-W}^{W} \frac{1}{x_{i-1}^{t+j}} &i> 0, t \geq W+1 \label{eq:zdis}
\end{align}
where the function $g(\cdot)$ given by
\[
g(a) = \mathbb{E}\left[\left(1-\tanh\left(a+\xi\sqrt{a}\right)\right)^2\right]\ , \quad\xi \sim \mathcal{N}(0,1)\ ,
\]
determines the mean squared error of the data bits with SINR $a$ and $\mathcal{N}(0,1)$ denotes a standard normal random variable. Function $g(\cdot)$ is a continuous differentiable function that is strictly decreasing from $g(0)=1$ to $g(+\infty)=0$.
Variable $z_i^t$ defines the SINR of the packet starting at time $t-W$ and finishing at $t+W$, i.e., centered at time $t$.
SINR $z_i^t$ is computed from the SINRs of the $M$ partitions uniformly distributed along
the packet. The SINRs $\frac{1}{x_{i-1}^{t+l}}$ of the partitions are averaged in (\ref{eq:zdis}), just like in (\ref{eq:SINR}). Packets that start between time $t-2W$ and $t$ contribute to the noise-and-interference variance at time $t$ in (\ref{eq:xdis}).
The iterative system is initialized by
\begin{equation}
z_0^t = \left\{ \begin{array}{ll}
+\infty, & \quad t \leq W \\
0, & \quad t \geq W+1\ ,
\end{array} \right.
\label{eq:zdis_ini}
\end{equation}
since the packets that start before time $t=0$ are completely known (or absent) and, therefore, their SINR is $+\infty$. These packets are centered at time before $W+1$. The packets
centered at time $t=W+1$ and after are completely unknown. Hence, their SINR is $0$.
Considering $W \rightarrow \infty$ and normalizing the packet length to $1$ the equations (\ref{eq:xdis}) and (\ref{eq:zdis})
can be transformed to
\begin{align}
x_i^t &= \alpha \int_{-1/2}^{1/2} g(z_i^{t+\tau}) d \tau + \sigma^2 &t \geq -\frac{1}{2}, i \geq 0 \label{eq:xcon} \\
z_i^t &= \int_{-1/2}^{1/2} \frac{1}{x_{i-1}^{t+\tau}} d \tau &t \geq 0, i > 0 \label{eq:zcon}
\end{align}
with a shift of the time origin from $1/2$ to $0$ which is done for notation simplicity.
The initialization is done using a step function
\begin{equation}
z_0^t = \left\{ \begin{array}{ll}
+\infty, & \quad t < 0 \\
0, & \quad t \geq 0\ .
\end{array} \right.
\label{eq:zcon_ini}
\end{equation}
The SINR of the data bits at iteration $I$ of the interference cancellation stage equals $z_I^t$. The second, error correction decoding stage, of the two-stage
reception is completed successfully iff $z_I^t > \theta$ where $\theta$ is the error correction code threshold.
Contrary to the two-stage receiver, the modified SIC receiver is performing decoding after every interference cancellation iteration. SIC can be described by (\ref{eq:xcon}) together with
\begin{equation}
z_{\mathrm{SIC},i}^t = \left\{ \begin{array}{ll}
+\infty, & \quad \int_{-1/2}^{1/2} \frac{1}{x_{i-1}^{t+\tau}} d \tau > \theta \\
\int_{-1/2}^{1/2} \frac{1}{x_{i-1}^{t+\tau}} d \tau, & \quad \int_{-1/2}^{1/2} \frac{1}{x_{i-1}^{t+\tau}} d \tau \leq \theta\ .
\end{array} \right.
\label{eq:zconSIC}
\end{equation}
replacing equation (\ref{eq:zcon}). We notice that the operator $F(\cdot)$ defining the evolution of the SINR throughout the iterations
via (\ref{eq:xcon}) and (\ref{eq:zconSIC}),
\begin{equation}
z_{i+1}^t = F(z_i^t)\ ,
\label{eq:F}
\end{equation}
is monotone, i.e., if $z_1(t) \leq z_2(t)$ for $t \in (-\infty,+\infty)$, then $F(z_1(t)) \leq F(z_2(t))$ for $t \in (-\infty,+\infty)$.
\begin{theorem}
\label{thm:speff}
Spectral efficiency of
\begin{equation}
\alpha \mathcal{C}_{\mathrm{BIAWGN}}\left(\frac{1}{\alpha} \ln \frac{\alpha+\sigma^2}{\sigma^2}-\epsilon\right)
\label{eq:siccapeq}
\end{equation}
is achievable by the modified SIC receiver for any $\alpha \in [0,\infty)$ and $\epsilon>0$.
\end{theorem}
\begin{proof}
We proceed with (\ref{eq:zcon_ini}) and (\ref{eq:xcon}) calculating
\begin{align}
x_0^t &= \int\limits_{t-1/2}^{t+1/2} \alpha g\left(z_0^{\tau}\right) d \tau + \sigma^2 = \left\{ \begin{array}{ll}
\alpha t+\frac{\alpha}{2} + \sigma^2, &t \in \left[-\frac{1}{2},\frac{1}{2}\right] \\
\alpha + \sigma^2, &t \in \left(\frac{1}{2},\infty\right)
\end{array} \right. \nonumber
\end{align}
that after the application of (\ref{eq:zcon}) implies
\begin{align}
z_1^0 &= \int\limits_{-1/2}^{1/2} \frac{1}{x_{0}^{\tau}} d \tau = \int\limits_{-1/2}^{1/2} \frac{1}{\alpha \left(\tau+\frac{1}{2}\right) + \sigma^2} d \tau = \frac{1}{\alpha} \ln \frac{\alpha+\sigma^2}{\sigma^2}, \nonumber\\
z_1^\tau &= \frac{1}{\alpha} \ln \frac{\alpha+\sigma^2}{\sigma^2} - \frac{\alpha}{\sigma^2(\alpha+\sigma^2)}\delta+ o(\delta), \quad \tau \in [0,\delta]
\end{align}
for small $\delta >0$.
Let us assume that the binary error correction codes for all data streams are of rate
\begin{equation}
\mathcal{C}\stxt{BIAWGN}\left(\frac{1}{\alpha} \ln \frac{\alpha+\sigma^2}{\sigma^2} - \frac{\alpha}{\sigma^2(\alpha+\sigma^2)}\delta+ o(\delta) \right),
\label{eq:rate}
\end{equation}
i.e., these codes have threshold
\[
\theta = \frac{1}{\alpha} \ln \frac{\alpha+\sigma^2}{\sigma^2} - \frac{\alpha}{\sigma^2(\alpha+\sigma^2)}\delta+ o(\delta)
\]
In this case (\ref{eq:zconSIC}) implies
\[
z_{\mathrm{SIC},1}^t = +\infty, \quad t < \delta\ ,
\]
indicating that the packets centered at $t \in [0,\delta)$ are successfully decoded and cancelled.
The situation will repeat at the next iteration since $z_{\mathrm{SIC},1}^t \geq z_0^{t-\delta}$, $t\in(-\infty,+\infty)$ implies $z_{\mathrm{SIC},2}^t = +\infty$ for $t \in [\delta, 2\delta)$
from the monotonicity of the SINR evolution operator (\ref{eq:F}). Thus, the convergence of the SINR to $+\infty$ will continue at the speed of at least $\delta$ per iteration. The total achievable sum rate is (see (\ref{eq:rate}))
\[
\alpha \left( \mathcal{C}\stxt{BIAWGN} \left(\frac{1}{\alpha} \ln \frac{\alpha+\sigma^2}{\sigma^2} - \frac{\alpha}{\sigma^2(\alpha+\sigma^2)}\delta+ o(\delta) \right)\right)\ .
\]
By choosing appropriately small $\delta$ we obtain the statement of the Theorem.
\end{proof}
Let us define the total system SNR parameter $s= \frac{\alpha}{\sigma^2}$. The next theorem states that for any fixed $s$ the AWGN channel capacity is achievable by the
modified SIC decoder. Let us consider $\sigma^2 = \alpha/s$, i.e., we are keeping $s$ fixed. We denote the limiting spectral efficiency of the system considered in Theorem~\ref{thm:speff}
and the corresponding SNR per bit by
\begin{align}
\mathcal{C}\stxt{eff}(\alpha,s) &= \alpha \mathcal{C}_{\mathrm{BIAWGN}}\left(\frac{1}{\alpha} \ln (1+s)\right) \label{eq:ceffalphs}\\
\frac{E\stxt{b}}{N_0}(\alpha,s) &= \frac{1}{2 \mathcal{C}_{\mathrm{BIAWGN}}\left(\frac{1}{\alpha} \ln (1+s)\right) \sigma^2}\ . \label{eq:snralphs}
\end{align}
We also define the real-valued AWGN channel capacity $\mathcal{C}\stxt{AWGN}\left( \frac{E\stxt{b}}{N_0} \right)$ for given $\frac{E\stxt{b}}{N_0}$ as the root of the equation
\begin{equation}
\mathcal{C}\stxt{AWGN}\left( \frac{E\stxt{b}}{N_0} \right) = \frac{1}{2} \log_2\left(1 + 2\mathcal{C}\stxt{AWGN}\left( \frac{E\stxt{b}}{N_0} \right) \frac{E\stxt{b}}{N_0}\right)\ .
\label{eq:cawgn}
\end{equation}
Finally, we state the main result of the paper.
\begin{theorem}
\label{thm:capach}
\[
\lim_{\alpha \rightarrow \infty} \mathcal{C}\stxm{eff}(\alpha,s) = \lim_{\alpha \rightarrow \infty} \mathcal{C}\stxm{AWGN}\left( \frac{E\stxm{b}}{N_0}(\alpha,s) \right) = \frac{1}{2}\log_2 (1+s)
\]
\end{theorem}
\begin{proof}
The capacity of the binary input AWGN channel for SNRs $\gamma<1$ can be bounded as follows~\cite{Gal06}
\begin{equation}
\frac{1}{2\ln 2} \gamma + A_1 \gamma^2 \leq \mathcal{C}_{\mathrm{BIAWGN}}(\gamma) \leq \frac{1}{2\ln 2} \gamma + A_2 \gamma^2 \label{eq:Galbnd}
\end{equation}
where $A_1 < A_2$ are constants. Therefore, application of (\ref{eq:Galbnd}) to (\ref{eq:ceffalphs}) implies
\begin{multline}
\frac{1}{2} \log_2 (1+s) + A_1 \frac{\ln^2 (1+s)}{\alpha} \leq \mathcal{C}\stxt{eff}(\alpha,s) \\ \leq \frac{1}{2} \log_2 (1+s) + A_2 \frac{\ln^2 (1+s)}{\alpha}
\end{multline}
while application of (\ref{eq:Galbnd}) to (\ref{eq:snralphs}) implies
\begin{multline}
\frac{s}{ \log_2 (1+s) + 2 A_2 \frac{\ln^2 (1+s)}{\alpha}} \leq \frac{E\stxt{b}}{N_0}(\alpha,s) \\ \leq \frac{s}{ \log_2 (1+s) + 2 A_1 \frac{\ln^2 (1+s)}{\alpha}}\ .
\end{multline}
Taking the limit in the above inequalities we obtain
\begin{align}
&\lim_{\alpha \rightarrow \infty} \mathcal{C}\stxt{eff}(\alpha,s) = \frac{1}{2}\log_2 (1+s) \label{eq:ceffs}\\
&\lim_{\alpha \rightarrow \infty} \frac{E\stxt{b}}{N_0}(\alpha,s) = \frac{s}{ \log_2 (1+s)} \label{eq:ebn0s}
\end{align}
Equations~\ref{eq:ceffs} and \ref{eq:ebn0s} imply equality
\[
\lim_{\alpha \rightarrow \infty} \frac{E\stxt{b}}{N_0}(\alpha,s) = \frac{2^{2\lim_{\alpha \rightarrow \infty} \mathcal{C}\stxt{eff}(\alpha,s)}-1}{2\lim_{\alpha \rightarrow \infty} \mathcal{C}\stxt{eff}(\alpha,s)}
\]
which ensures (see (\ref{eq:cawgn}))
\[
\lim_{\alpha \rightarrow \infty} \mathcal{C}\stxm{eff}(\alpha,s) = \lim_{\alpha \rightarrow \infty} \mathcal{C}\stxm{AWGN}\left( \frac{E\stxm{b}}{N_0}(\alpha,s) \right)\ .
\]
The Theorem is proved.
\end{proof}
Spectral efficiency $\mathcal{C}\stxm{eff}(\alpha,s)$ achieved by the modified SIC receiver (see Theorem~\ref{thm:speff}) is plotted in Fig.~\ref{Fig:CapCeff}. The three magenta curves correspond to $\alpha =10,100,$ and $500$ (from bottom to top). For each curve $\alpha$ is kept constant while $s$ is varying. The channel capacity $\mathcal{C}\stxm{AWGN}$ is given by the blue curve. Finally, the brown curve plots spectral efficiency achieved by the two-stage receiver.
\pagebreak
\begin{figure}[t]
\setlength{\unitlength}{1mm}
\begin{picture}(90,56)
\put(0,-3){\includegraphics{CapAchFig2.pdf}}
\end{picture}
\caption{Achievable spectral efficiency for the modified SIC receiver (magenta), two-stage receiver (brown) and the AWGN channel capacity (blue).}
\label{Fig:CapCeff}
\end{figure}
\vspace*{-7mm}
\section{Conclusions}
In this paper, it has been proven that sparse graph modulation with spatial coupling can achieve AWGN channel capacity under modified SIC reception. The sparse graph modulation format is based on superposition of low-rate redundant data streams which can be easily designed to operate near binary input AWGN channel capacity. Numerical results also demonstrate that the two-stage reception where parallel interference cancellation is followed by parallel error control decoding is capable of operating within 1dB of the channel capacity.
|
\section{Introduction }
A capital letter (such as $T$) means a bounded linear operator on a Hilbert space $\mathcal{H}$. $T$ is said to be positive (denoted by $T\geq 0$) if $(Tx, x) \geq 0$ for all $x\in \mathcal{H} $, and $T$ is said to be strictly positive (denoted by $T>0$) if $T$ is positive and invertible. The usual order $S\geq T$ among selfadjoint operators on $\mathcal{H}$ is defined by $(Sx, x)\geq (Tx, x)$ for all $x\in \mathcal{H}$. Let $I$ denote the indentity operator.
As an essential and historical extension of the famous L\"{o}wner-Heinz inequality: $A\geq B\geq 0 \Rightarrow A^{\alpha}\geq B^{\alpha}$ if $\alpha \in [0, 1]$, T. Furuta proved the following operator inequality in 1987.
\noindent{\bf Theorem 1.1.} (Furuta Inequality, \cite{Furuta_1987}) If $A\geq B\geq 0$, then for each $r\geq 0$,
\begin{eqnarray}
(A^{\frac r 2}A^{p}A^{\frac r 2})^{\frac 1 q}\geq (A^{\frac r 2}B^{p}A^{\frac r 2})^{\frac 1 q},\\
(B^{\frac r 2}A^{p}B^{\frac r 2})^{\frac 1 q}\geq (B^{\frac r 2}B^{p}B^{\frac r 2})^{\frac 1 q}
\end{eqnarray}
hold for $p\geq 0$ and $q\geq 1$ with $(1+r)q\geq p+r$.
\begin{figure}[h] \centering \includegraphics{FIFig.eps} \caption{ Domain of Furuta inequality} \label{fige1} \end{figure}
K. Tanahashi showed that the conditions $p$ and $q$ in Figure 1 are best possible for each $r\geq 0$. See \cite{Tanahashi_1996}. It is well known that Furuta inequality has many applications. See \cite{Aluthge_1990, Aluthge_Powers, Fujii_1993, F_F, Huruya, Ito, Yuan_Gao_2008, Yuan_2010MIA, Yuan_2010OM}.
In 1995, T. Furuta showed the following theorem which interpolates Furuta inequality.
\noindent{\bf Theorem 1.2.} (Grand Furuta Inequality, \cite{Furuta_1995}) If $A\geq B\geq 0$ with $A>0$, then for each $t\in [0, 1]$ and $p\geq 1$,
\begin{eqnarray}
A^{1-t+r}\geq\{A^{\frac r 2}(A^{-{\frac t 2}}B^{p}A^{-{\frac t 2}})A^{\frac r 2}\}^{\frac {1-t+r}{(p-t)s+r}}
\end{eqnarray}
holds for $s\geq 1$ and $r\geq t$.
K. Tanahashi proved that the exponent value ${\frac {1-t+r}{(p-t)s+r}}$ of grand Furuta inequality is the best possible in \cite{Tanahashi_2000}. Afterwards, the proof was improved by T. Yamazaki and M. Fujii et al., respectively. See \cite{Yamazaki_1999} and \cite{Fujii_1999}.
In 2003, grand Furuta inequality was extended by M. Uchiyama in \cite{Uchiyama_2003} as follows:
\noindent{\bf Theorem 1.3.} (Extended Grand Furuta Inequality, \cite{Uchiyama_2003}) If $A\geq B\geq C\geq 0$ with $B>0$, then for each $t\in [0, 1]$ and $p\geq 1$,
\begin{eqnarray}
A^{1-t+r}\geq {A^{\frac r 2}(B^{-{\frac t 2}}C^{p}B^{-{\frac t 2}})^{s}A^{\frac r 2}}^{\frac {1-t+r}{(p-t)s+r}}
\end{eqnarray}
holds for $s\geq 1$ and $r\geq t$.
In 2008, grand Furuta inequality was given another extension in \cite{Furuta_2008} as follows:
\noindent{\bf Theorem 1.4.} (Extension of Furuta Inequality, \cite{Furuta_2008}) If $A\geq B\geq 0$ with $A>0$, $t\in [0, 1]$ and $p_{1}, p_{2}, \ldots, p_{2n}\geq 1$ for any natural number $n$, then the following inequality
\begin{equation}
\begin{split}
A^{1-t+r}\geq &\ \big \{A^{\frac r 2}\big[A^{-{\frac t 2}}\cdots \big[A^{-{\frac t 2}}\{A^{\frac t 2}(A^{-{\frac t 2}}B^{p_{1}}A^{-{\frac t 2}})^{p_{2}}\\ &\ A^{\frac t 2}\}^{p_{3}}A^{-{\frac t 2}}\big]^{p_{4}}
\cdots A^{-{\frac t 2}}\big]^{p_{2n}}A^{\frac r 2}\big\}^{\frac {1-t+r}{\phi (2n)-t+r}}
\end{split}
\end{equation}
holds for $r\geq t$, where $\phi (2n)=\{\cdots[\{[(p_{1}-t)p_{2}+t]p_{3}-t\}p_{4}+t]p_{5}-\cdots -t\}p_{2n}+t$.
In 2010, C. Yang and Y. Wang showed the following theorem which interpolates extended grand Furuta inequality in \cite{Yang_2010}.
\noindent{\bf Theorem 1.5.} (Further Extension of Furuta Inequality, \cite{Yang_2010}) If $A_{2n+1}\geq A_{2n}\geq A_{2n-1}\geq \cdots\geq A_{3}\geq A_{2}\geq A_{1}\geq 0$ with $A_{2}>0$, $t_{1}, t_{2}, \ldots, t_{n-1}, t_{n} \in [0, 1]$ and $p_{1}, p_{2}, \ldots, p_{2n-1}, p_{2n}\geq 1$ for a natural number $n$, then the following inequality
\begin{equation}
\begin{split}
A^{1-t_{n}+r}_{2n+1}\geq &\ \{A^{\frac r 2}_{2n+1}[A^{-{\frac {t_{n}}{2}}}_{2n}\{A^{\frac {t_{n-1}}{2}}_{2n-1} \cdots A^{\frac {t_{2}}{2}}_{5}[A^{-{\frac {t_{2}}{2}}}_{4}\{A^{\frac {t_{1}}{2}}_{3}(A^{-{\frac {t_{1}}{2}}}_{2}A^{p_{1}}_{1}A^{-{\frac {t_{1}}{2}}}_{2})^{p_{2}}\\
&\ A^{\frac {t_{1}}{2}}_{3}\}^{p_{3}}A^{-{\frac {t_{2}}{2}}}_{4}]^{p_{4}}A^{\frac {t_{2}}{2}}_{5}
\cdots A^{\frac {t_{n-1}}{2}}_{2n-1}\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n}]^{p_{2n}}A^{\frac r 2}_{2n+1}\}^{\frac {1-t_{n}+r}{\psi [2n]-t_{n}+r}}
\end{split}
\end{equation}
holds for $r\geq t_{n}$, where $\psi [2n]=\{\cdots [\{[(p_{1}-t_{1})p_{2}+t_{1}]p_{3}-t_{2}\}p_{4}+t_{2}]p_{5}-\cdots -t_{n}\}p_{2n}+t_{n}$.
Recently, some beautiful results on characterizations of operator order have been shown, such as \cite{Fujii_2002}, \cite{Lin_2007} and \cite{Lin_2011}. C.-S. Lin, by using Furuta inequality, showed the characterizations of operator order for two strictly positive operators in \cite{Lin_2011}. Afterwards, he and Y. J. Cho, by using extended grand Furuta inequality, showed the characterizations of operator order for three strictly positive operators. The aim of the present paper is to show the characterizations of operator order $A_{k}\geq A_{k-1}\geq \cdots \geq A_{2}\geq A_{1}>0$ for any positive integer $k$ in terms of operator inequality via further extension of Furuta inequality. An application of the characterizations is given to operator equalities due to Douglas's majorization and factorization theorem.
\section{Main results and proofs}
\indent In this section, we show the characterizations of operator order for $k$ strictly positive operators. First, we assume that $k$ is an odd integer ($k=2n+1$).\\
\noindent{\bf Theorem 2.1.} Let $A_{1}, A_{2}, A_{3}, \cdots, A_{2n-1}, A_{2n}, A_{2n+1}$ be strictly positive operators. Then the following two assertions are equivalent.\\
(I) $A_{2n+1}\geq A_{2n}\geq A_{2n-1}\geq \cdots \geq A_{3}\geq A_{2}\geq A_{1}$.\\
(II) If $t_{1}, t_{2}, \cdots, t_{n}\in [0, 1]$, $p_{1}, p_{2}, \cdots, p_{2n-1}, p_{2n}\geq 1$, $\psi [2n]=\{\cdots[\{[(p_{1}-t_{1})p_{2}+t_{1}]p_{3}-t_{2}\}p_{4}+t_{2}]p_{5}-\cdots -t_{n}\}p_{2n}+t_{n}$, then the following inequalities always hold for $r\geq t_{n}$:\\
\indent (II.1) $A^{r-t_{n}}_{2n+1}\geq \Big\{A^{\frac r 2}_{2n+1}\Big[A^{-{\frac {t_{n}}{2}}}_{2n}\big\{A^{\frac {t_{n-1}}{2}}_{2n-1} \cdots A^{\frac {t_{2}}{2}}_{5}\big[A^{-{\frac {t_{2}}{2}}}_{4}\cdot\{A^{\frac {t_{1}}{2}}_{3}(A^{-{\frac {t_{1}}{2}}}_{2}A^{p_{1}}_{1}A^{-{\frac {t_{1}}{2}}}_{2})^{p_{2}} A^{\frac {t_{1}}{2}}_{3}\}^{p_{3}}\cdot \\
A^{-{\frac {t_{2}}{2}}}_{4}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{5}
\cdots A^{\frac {t_{n-1}}{2}}_{2n-1}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n}\Big]^{p_{2n}}A^{\frac r 2}_{2n+1}\Big\}^{\frac {r-t_{n}}{\psi [2n]-t_{n}+r}}$;\\
\indent (II.2) $A^{r-t_{n}}_{2n+1}\geq \Big\{A^{\frac r 2}_{2n+1}\Big[A^{-{\frac {t_{n}}{2}}}_{2n+1}\big\{A^{\frac {t_{n-1}}{2}}_{2n } \cdots A^{\frac {t_{2}}{2}}_{6}\big[A^{-{\frac {t_{2}}{2}}}_{5}\cdot\{A^{\frac {t_{1}}{2}}_{4}(A^{-{\frac {t_{1}}{2}}}_{3}A^{p_{1}}_{2}A^{-{\frac {t_{1}}{2}}}_{3})^{p_{2}} A^{\frac {t_{1}}{2}}_{4}\}^{p_{3}}\cdot \\
A^{-{\frac {t_{2}}{2}}}_{5}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{6}
\cdots A^{\frac {t_{n-1}}{2}}_{2n }\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n+1}\Big]^{p_{2n}}A^{\frac r 2}_{2n+1}\Big\}^{\frac {r-t_{n}}{\psi [2n]-t_{n}+r}}$;\\
\indent (II.3) $A^{r-t_{n}}_{2n+1}\geq \Big\{A^{\frac r 2}_{2n+1}\Big[A^{-{\frac {t_{n}}{2}}}_{2n+1}\big\{A^{\frac {t_{n-1}}{2}}_{2n+1 } \cdots A^{\frac {t_{2}}{2}}_{7}\big[A^{-{\frac {t_{2}}{2}}}_{6}\cdot\{A^{\frac {t_{1}}{2}}_{5}(A^{-{\frac {t_{1}}{2}}}_{4}A^{p_{1}}_{3}A^{-{\frac {t_{1}}{2}}}_{4})^{p_{2}} A^{\frac {t_{1}}{2}}_{5}\}^{p_{3}} \cdot\\
A^{-{\frac {t_{2}}{2}}}_{6}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{7}
\cdots A^{\frac {t_{n-1}}{2}}_{2n+1 }\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n+1}\Big]^{p_{2n}}A^{\frac r 2}_{2n+1}\Big\}^{\frac {r-t_{n}}{\psi [2n]-t_{n}+r}}$;\\
\indent { $\cdots \cdots \cdots \cdots$}\\
\indent (II.n) $A^{r-t_{n}}_{2n+1}\geq \Big\{A^{\frac r 2}_{2n+1}\Big[A^{-{\frac {t_{n}}{2}}}_{2n+1}\big\{A^{\frac {t_{n-1}}{2}}_{2n+1 } \cdots A^{\frac {t_{2}}{2}}_{n+4}\big[A^{-{\frac {t_{2}}{2}}}_{n+3}\{A^{\frac {t_{1}}{2}}_{n+2}(A^{-{\frac {t_{1}}{2}}}_{n+1}A^{p_{1}}_{n}A^{-{\frac {t_{1}}{2}}}_{n+1})^{p_{2}} A^{\frac {t_{1}}{2}}_{n+2}\}^{p_{3}} \\
A^{-{\frac {t_{2}}{2}}}_{n+3}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{n+4}
\cdots A^{\frac {t_{n-1}}{2}}_{2n+1 }\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n+1}\Big]^{p_{2n}}A^{\frac r 2}_{2n+1}\Big\}^{\frac {r-t_{n}}{\psi [2n]-t_{n}+r}}$;\\
\indent (II.n+1) $A^{r-t_{n}}_{ 1}\leq \Big\{A^{\frac r 2}_{ 1}\Big[A^{-{\frac {t_{n}}{2}}}_{1}\big\{A^{\frac {t_{n-1}}{2}}_{ 1 } \cdots A^{\frac {t_{2}}{2}}_{n-2}\big[A^{-{\frac {t_{2}}{2}}}_{n-1}\{A^{\frac {t_{1}}{2}}_{n }(A^{-{\frac {t_{1}}{2}}}_{n+1}A^{p_{1}}_{n+2}A^{-{\frac {t_{1}}{2}}}_{n+1})^{p_{2}} A^{\frac {t_{1}}{2}}_{n }\}^{p_{3}} \\
A^{-{\frac {t_{2}}{2}}}_{n-1}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{n-2}
\cdots A^{\frac {t_{n-1}}{2}}_{1}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{1}\Big]^{p_{2n}}A^{\frac r 2}_{1}\Big\}^{\frac {r-t_{n}}{\psi [2n]-t_{n}+r}}$;\\
\indent { $\cdots \cdots \cdots \cdots$}\\
\indent (II.2n-2) $A^{r-t_{n}}_{ 1}\leq \Big\{A^{\frac r 2}_{ 1}\Big[A^{-{\frac {t_{n}}{2}}}_{1}\big\{A^{\frac {t_{n-1}}{2}}_{1} \cdots A^{\frac {t_{2}}{2}}_{2n-5}\big[A^{-{\frac {t_{2}}{2}}}_{2n-4}\{A^{\frac {t_{1}}{2}}_{2n-3 }\cdot(A^{-{\frac {t_{1}}{2}}}_{2n-2}A^{p_{1}}_{2n-1}A^{-{\frac {t_{1}}{2}}}_{2n-2})^{p_{2}}\cdot \\ A^{\frac {t_{1}}{2}}_{2n-3}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{2n-4}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{2n-5}
\cdots A^{\frac {t_{n-1}}{2}}_{1}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{1}\Big]^{p_{2n}}A^{\frac r 2}_{1}\Big\}^{\frac {r-t_{n}}{\psi [2n]-t_{n}+r}}$;\\
\indent (II.2n-1) $A^{r-t_{n}}_{ 1}\leq \Big\{A^{\frac r 2}_{ 1}\Big[A^{-{\frac {t_{n}}{2}}}_{1}\big\{A^{\frac {t_{n-1}}{2}}_{2} \cdots A^{\frac {t_{2}}{2}}_{2n-4}\big[A^{-{\frac {t_{2}}{2}}}_{2n-3}\{A^{\frac {t_{1}}{2}}_{2n-2 }\cdot(A^{-{\frac {t_{1}}{2}}}_{2n-1}A^{p_{1}}_{2n}A^{-{\frac {t_{1}}{2}}}_{2n-1})^{p_{2}}\cdot \\ A^{\frac {t_{1}}{2}}_{2n-2}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{2n-3}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{2n-4}
\cdots A^{\frac {t_{n-1}}{2}}_{2}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{1}\Big]^{p_{2n}}A^{\frac r 2}_{1}\Big\}^{\frac {r-t_{n}}{\psi [2n]-t_{n}+r}}$;\\
\indent (II.2n) $A^{r-t_{n}}_{ 1}\leq \Big\{A^{\frac r 2}_{ 1}\Big[A^{-{\frac {t_{n}}{2}}}_{2}\big\{A^{\frac {t_{n-1}}{2}}_{3} \cdots A^{\frac {t_{2}}{2}}_{2n-3}\big[A^{-{\frac {t_{2}}{2}}}_{2n-2}\{A^{\frac {t_{1}}{2}}_{2n-1}\cdot(A^{-{\frac {t_{1}}{2}}}_{2n}A^{p_{1}}_{2n+1}A^{-{\frac {t_{1}}{2}}}_{2n})^{p_{2}}\cdot \\ A^{\frac {t_{1}}{2}}_{2n-1}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{2n-2}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{2n-3}
\cdots A^{\frac {t_{n-1}}{2}}_{3}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2}\Big]^{p_{2n}}A^{\frac r 2}_{1}\Big\}^{\frac {r-t_{n}}{\psi [2n]-t_{n}+r}}$.\\
\noindent {\bf Proof.} (I)$\Rightarrow $(II) Applying L\"{o}wner-Heinz inequality for ${\frac {r-t_{n}}{1-t_{n}+r}}$ to further extension of Furuta inequality, (II.1) is obtained; Replacing $A_{1}, A_{2}, A_{3}, \cdots,A_{2n-1}, A_{2n}$ by $A_{2}, A_{3}, A_{4}, \cdots, A_{2n}, A_{2n+1}$ in (II.1), respectively, (II.2) is obtained; Replacing $A_{2}, A_{3}, A_{4}, \cdots, A_{2n-1}, A_{2n}$ by $A_{3}, A_{4}, A_{5}, \cdots, A_{2n}, A_{2n+1}$ in (II.2), respectively, (II.3) is obtained. Similarly, we can obtain (II.4), (II.5), $\cdots$, (II.n).
If we replace $A_{1}, A_{2}, A_{3} \cdots, A_{2n-1}, A_{2n}, A_{2n+1}$ by $A_{2n+1}^{-1}, A_{2n}^{-1}, A_{2n-1}^{-1}, \cdots, A_{3}^{-1},$ $ A_{2}^{-1}, A_{1}^{-1}$ in (II.1), (II.2), (II.3), $\cdots$, (II.n), respectively, and take reverse, then (II.2n), (II.2n-1), (II.2n-2), $\cdots$, (II.n+1) hold.
(II)$\Rightarrow$(I) Because each $A_{i}$ is strictly positive and bounded, there exist $u_{i}$ and $v_{i}$ such that $+\infty>u_{i}I\geq A_{i}\geq v_{i}I>0$ $(i=1, 2, \cdots, 2n+1)$. If we take $p_{1}=p_{3}=p_{4}=\cdots =p_{2n}=1$, $t_{1}=t_{2}=\cdots =t_{n}=1$, $r=2$ in (II.1), then we have
\begin{equation}\tag{2.1}
\begin{split}
&\ A_{2n+1}\\ \geq &\ \big\{A_{2n+1}A_{2n}^{-{\frac 1 2}}A_{2n-1}^{\frac 1 2}\cdots A_{5}^{\frac 1 2} A_{4}^{-{\frac 1 2}}A_{3}^{\frac 1 2}(A^{-{\frac 1 2}}_{2}A_{1}A^{-{\frac 1 2}}_{2})^{p_{2}}A_{3}^{\frac 1 2}A_{4}^{-{\frac 1 2}} A_{5}^{\frac 1 2}\cdots A_{2n-1}^{\frac 1 2}A_{2n}^{-{\frac 1 2}}A_{2n+1}\big\}^{\frac 1 2}.
\end{split}\end{equation}
According to Theorem 6' in \cite{Fujii_1997}: $X\geq Y>0$ with $sI\geq X\geq tI>0 \Rightarrow {\frac {(s+t)^{2}}{4st}}X^{2}\geq Y^{2}$, we can obtain the following inequality by (2.1) and $u_{2n+1}I\geq A_{2n+1}\geq v_{2n+1}I>0$.
\begin{equation}\tag{2.2}
\begin{split}
&\ {\frac {(u_{2n+1}+v_{2n+1})^{2}}{4u_{2n+1}v_{2n+1}}}A_{2n+1}^{2}\\ \geq &\ A_{2n+1}A_{2n}^{-{\frac 1 2}}A_{2n-1}^{\frac 1 2}\cdots A_{5}^{\frac 1 2} A_{4}^{-{\frac 1 2}}A_{3}^{\frac 1 2}(A^{-{\frac 1 2}}_{2}A_{1}A^{-{\frac 1 2}}_{2})^{p_{2}}A_{3}^{\frac 1 2}A_{4}^{-{\frac 1 2}} A_{5}^{\frac 1 2}\cdots A_{2n-1}^{\frac 1 2}A_{2n}^{-{\frac 1 2}}A_{2n+1}.
\end{split}
\end{equation}
Then we have
\begin{equation}\tag{2.3}
\begin{split}
&\ {\frac {(u_{2n+1}+v_{2n+1})^{2}}{4u_{2n+1}v_{2n+1}}}\cdot {\frac {u_{2n}u_{2n-2}\cdots u_{6}u_{4}}{v_{2n-1}v_{2n-3}\cdots v_{5}v_{3}}}I\\
\geq &\ {\frac {(u_{2n+1}+v_{2n+1})^{2}}{4u_{2n+1}v_{2n+1}}}A_{3}^{-{\frac 1 2}}A_{4}^{\frac 1 2}A_{5}^{-{\frac 1 2}}\cdots A_{2n-1}^{-{\frac 1 2}}A_{2n}A_{2n-1}^{-{\frac 1 2}}\cdots A_{5}^{-{\frac 1 2}} A_{4}^{\frac 1 2} A_{3}^{-{\frac 1 2}}\\
\geq &\ (A_{2}^{-{\frac 1 2}}A_{1}A_{2}^{-{\frac 1 2}})^{p_{2}}.
\end{split}
\end{equation}
Thus,
\begin{equation}\tag{2.4}
\Big[{\frac {(u_{2n+1}+v_{2n+1})^{2}}{4u_{2n+1}v_{2n+1}}}\cdot {\frac {u_{2n}u_{2n-2}\cdots u_{6}u_{4}}{v_{2n-1}v_{2n-3}\cdots v_{5}v_{3}}}\Big]^{\frac {1}{p_{2}}}I
\geq A_{2}^{-{\frac 1 2}}A_{1}A_{2}^{-{\frac 1 2}}
\end{equation}
holds for any $p_{2}\geq 1$. $A_{2}\geq A_{1}$ is obtained by taking $p_{2}\rightarrow +\infty$.
Similarly, we can obtain $A_{3}\geq A_{2}, A_{4}\geq A_{3}, \cdots, A_{n+1}\geq A_{n}$ by (II.2), (II.3), $\cdots$, (II.n), respectively.
By the same setting for (2.2n), the following inequality holds according to Theorem 6 in \cite{Fujii_1997}: $X\geq Y>0 $ with $s'I\geq Y\geq t'I>0 \Rightarrow {\frac {(s'+t')^{2}}{4s't'}}X^{2}\geq Y^{2}.$
\begin{equation}\tag{2.5}
\begin{split}
&\ {\frac {(u_{1}+v_{1})^{2}}{4u_{1}v_{1}}}A_{1}A_{2}^{-{\frac 1 2}}A_{3}^{\frac 1 2}\cdots A_{2n-2}^{-{\frac 1 2}}A_{2n-1}^{\frac 1 2}(A_{2n}^{-{\frac 1 2}}A_{2n+1}A_{2n}^{-{\frac 1 2}})^{p_{2}}A_{2n-1}^{\frac 1 2}A_{2n-2}^{-{\frac 1 2}}\cdots A_{3}^{\frac 1 2}A_{2}^{-{\frac 1 2}}A_{1}\\
\geq &\ A_{1}^{2}.
\end{split}
\end{equation}
Then we have
\begin{equation}\tag{2.6}
\begin{split}
&\ A_{2n}^{-{\frac 1 2}}A_{2n+1}A_{2n}^{-{\frac 1 2}}\\ \geq &\ \Big[{\frac {4u_{1}v_{1}}{(u_{1}+v_{1})^{2}}}A_{2n-1}^{-{\frac 1 2}}A_{2n-2}^{\frac 1 2}\cdots A_{3}^{-{\frac 1 2}}A_{2}A_{3}^{-{\frac 1 2}}\cdots A_{2n-2}^{\frac 1 2}A_{2n-1}^{-{\frac 1 2}} \Big]^{\frac {1}{p_{2}}}\\ \geq &\ \Big[{\frac {4u_{1}v_{1}}{(u_{1}+v_{1})^{2}}}\cdot {\frac {v_{2}v_{4}\cdots v_{2n-2}}{u_{3}u_{5}\cdots u_{2n-1}}}\Big]^{\frac {1}{p_{2}}}I.
\end{split}
\end{equation}
$A_{2n+1}\geq A_{2n}$ is obtained by taking $p_{2}\rightarrow +\infty$ in (2.6).
Similarly, we can obtain $A_{2n}\geq A_{2n-1}, A_{2n-1}\geq A_{2n-2}, \cdots, A_{n+2}\geq A_{n+1}$ by (II.2n-1), (II.2n-2), $\cdots$, (II.n+1), respectively. $ \square$\\
\noindent{\bf Remark 2.1.} If $n=1$, Theorem 2.1 is the main result of \cite{Lin_2011}.\\
Next, we assume that $k$ is an even integer ($k=2n$).\\
\noindent{\bf Theorem 2.2.} Let $A_{1}, A_{2}, A_{3}, \cdots, A_{2n-1}, A_{2n}$ be strictly positive operators. Then the following two assertions are equivalent:\\
(I) $A_{2n}\geq A_{2n-1}\geq \cdots \geq A_{3}\geq A_{2}\geq A_{1}$.\\
(II) If $t_{1}, t_{2}, \cdots, t_{n}\in [0, 1]$, $p_{1}, p_{2}, \cdots, p_{2n-1}, p_{2n}\geq 1$, $\psi [2n]=\{\cdots[\{[(p_{1}-t_{1})p_{2}+t_{1}]p_{3}-t_{2}\}p_{4}+t_{2}]p_{5}-\cdots -t_{n}\}p_{2n}+t_{n}$, then the following inequalities always hold for $r\geq t_{n}$:\\
\indent (II.1) $A^{r-t_{n}}_{2n}\geq \Big\{A^{\frac r 2}_{2n }\Big[A^{-{\frac {t_{n}}{2}}}_{2n}\big\{A^{\frac {t_{n-1}}{2}}_{2n-1} \cdots A^{\frac {t_{2}}{2}}_{5}\big[A^{-{\frac {t_{2}}{2}}}_{4}\cdot\{A^{\frac {t_{1}}{2}}_{3}(A^{-{\frac {t_{1}}{2}}}_{2}A^{p_{1}}_{1}A^{-{\frac {t_{1}}{2}}}_{2})^{p_{2}} A^{\frac {t_{1}}{2}}_{3}\}^{p_{3}}\cdot \\
A^{-{\frac {t_{2}}{2}}}_{4}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{5}
\cdots A^{\frac {t_{n-1}}{2}}_{2n-1}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n}\Big]^{p_{2n}}A^{\frac r 2}_{2n }\Big\}^{\frac {r-t_{n}}{\psi [2n]-t_{n}+r}}$;\\
\indent (II.2) $A^{r-t_{n}}_{2n }\geq \Big\{A^{\frac r 2}_{2n }\Big[A^{-{\frac {t_{n}}{2}}}_{2n }\big\{A^{\frac {t_{n-1}}{2}}_{2n } \cdots A^{\frac {t_{2}}{2}}_{6}\big[A^{-{\frac {t_{2}}{2}}}_{5}\cdot\{A^{\frac {t_{1}}{2}}_{4}(A^{-{\frac {t_{1}}{2}}}_{3}A^{p_{1}}_{2}A^{-{\frac {t_{1}}{2}}}_{3})^{p_{2}} A^{\frac {t_{1}}{2}}_{4}\}^{p_{3}}\cdot \\
A^{-{\frac {t_{2}}{2}}}_{5}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{6}
\cdots A^{\frac {t_{n-1}}{2}}_{2n }\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n }\Big]^{p_{2n}}A^{\frac r 2}_{2n }\Big\}^{\frac {r-t_{n}}{\psi [2n]-t_{n}+r}}$;\\
\indent (II.3) $A^{r-t_{n}}_{2n }\geq \Big\{A^{\frac r 2}_{2n }\Big[A^{-{\frac {t_{n}}{2}}}_{2n }\big\{A^{\frac {t_{n-1}}{2}}_{2n } \cdots A^{\frac {t_{2}}{2}}_{7}\big[A^{-{\frac {t_{2}}{2}}}_{6}\cdot\{A^{\frac {t_{1}}{2}}_{5}(A^{-{\frac {t_{1}}{2}}}_{4}A^{p_{1}}_{3}A^{-{\frac {t_{1}}{2}}}_{4})^{p_{2}} A^{\frac {t_{1}}{2}}_{5}\}^{p_{3}} \cdot\\
A^{-{\frac {t_{2}}{2}}}_{6}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{7}
\cdots A^{\frac {t_{n-1}}{2}}_{2n }\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n }\Big]^{p_{2n}}A^{\frac r 2}_{2n }\Big\}^{\frac {r-t_{n}}{\psi [2n]-t_{n}+r}}$;\\
\indent { $\cdots \cdots \cdots \cdots$}\\
\indent (II.n) $A^{r-t_{n}}_{2n }\geq \Big\{A^{\frac r 2}_{2n }\Big[A^{-{\frac {t_{n}}{2}}}_{2n }\big\{A^{\frac {t_{n-1}}{2}}_{2n } \cdots A^{\frac {t_{2}}{2}}_{n+4}\big[A^{-{\frac {t_{2}}{2}}}_{n+3}\{A^{\frac {t_{1}}{2}}_{n+2}(A^{-{\frac {t_{1}}{2}}}_{n+1}A^{p_{1}}_{n}A^{-{\frac {t_{1}}{2}}}_{n+1})^{p_{2}} A^{\frac {t_{1}}{2}}_{n+2}\}^{p_{3}} \\
A^{-{\frac {t_{2}}{2}}}_{n+3}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{n+4}
\cdots A^{\frac {t_{n-1}}{2}}_{2n }\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n }\Big]^{p_{2n}}A^{\frac r 2}_{2n }\Big\}^{\frac {r-t_{n}}{\psi [2n]-t_{n}+r}}$;\\
\indent (II.n+1) $A^{r-t_{n}}_{ 1}\leq \Big\{A^{\frac r 2}_{ 1}\Big[A^{-{\frac {t_{n}}{2}}}_{1}\big\{A^{\frac {t_{n-1}}{2}}_{ 1 } \cdots A^{\frac {t_{2}}{2}}_{n-2}\big[A^{-{\frac {t_{2}}{2}}}_{n-1}\{A^{\frac {t_{1}}{2}}_{n }(A^{-{\frac {t_{1}}{2}}}_{n+1}A^{p_{1}}_{n+2}A^{-{\frac {t_{1}}{2}}}_{n+1})^{p_{2}} A^{\frac {t_{1}}{2}}_{n }\}^{p_{3}} \\
A^{-{\frac {t_{2}}{2}}}_{n-1}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{n-2}
\cdots A^{\frac {t_{n-1}}{2}}_{1}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{1}\Big]^{p_{2n}}A^{\frac r 2}_{1}\Big\}^{\frac {r-t_{n}}{\psi [2n]-t_{n}+r}}$;\\
\indent { $\cdots \cdots \cdots \cdots$}\\
\indent (II.2n-2) $A^{r-t_{n}}_{ 1}\leq \Big\{A^{\frac r 2}_{ 1}\Big[A^{-{\frac {t_{n}}{2}}}_{1}\big\{A^{\frac {t_{n-1}}{2}}_{1} \cdots A^{\frac {t_{2}}{2}}_{2n-5}\big[A^{-{\frac {t_{2}}{2}}}_{2n-4}\{A^{\frac {t_{1}}{2}}_{2n-3 }\cdot(A^{-{\frac {t_{1}}{2}}}_{2n-2}A^{p_{1}}_{2n-1}A^{-{\frac {t_{1}}{2}}}_{2n-2})^{p_{2}}\cdot \\ A^{\frac {t_{1}}{2}}_{2n-3}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{2n-4}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{2n-5}
\cdots A^{\frac {t_{n-1}}{2}}_{1}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{1}\Big]^{p_{2n}}A^{\frac r 2}_{1}\Big\}^{\frac {r-t_{n}}{\psi [2n]-t_{n}+r}}$;\\
\indent (II.2n-1) $A^{r-t_{n}}_{ 1}\leq \Big\{A^{\frac r 2}_{ 1}\Big[A^{-{\frac {t_{n}}{2}}}_{1}\big\{A^{\frac {t_{n-1}}{2}}_{2} \cdots A^{\frac {t_{2}}{2}}_{2n-4}\big[A^{-{\frac {t_{2}}{2}}}_{2n-3}\{A^{\frac {t_{1}}{2}}_{2n-2 }\cdot(A^{-{\frac {t_{1}}{2}}}_{2n-1}A^{p_{1}}_{2n}A^{-{\frac {t_{1}}{2}}}_{2n-1})^{p_{2}}\cdot \\ A^{\frac {t_{1}}{2}}_{2n-2}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{2n-3}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{2n-4}
\cdots A^{\frac {t_{n-1}}{2}}_{2}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{1}\Big]^{p_{2n}}A^{\frac r 2}_{1}\Big\}^{\frac {r-t_{n}}{\psi [2n]-t_{n}+r}}$.\\
\noindent{\bf Proof.} Let $A_{2n+1}=A_{2n}$ in Theorem 2.1. $\square$\\
Together with Theorem 2.1 and Theorem 2.2, we show the characterizations of operator order $A_{k}\geq A_{k-1}\geq \cdots \geq A_{2}\geq A_{1}>0$ for any positive integer $k$. For example, if $k=5$, we have the following result:
\noindent{\bf Proposition 2.1.} Let $A_{1}, A_{2}, A_{3}, A_{4}$ and $A_{5}$ be strictly positive operators. Then $A_{5}\geq A_{4}\geq A_{3}\geq A_{2}\geq A_{1}$ if and only if the following four operator inequalities
\begin{equation}\tag{2.7}
A_{5}^{r-t_{2}}\geq \Big\{A_{5}^{\frac r 2}\Big[A_{4}^{-{\frac {t_{2}} {2}}}\big(A_{3}^{ {\frac {t_{1}}{2}}}(A_{2}^{-{\frac {t_{1}}{2}}}A_{1}^{p_{1}}A_{2}^{-{\frac {t_{1}}{2}}})^{p_{2}}A_{3}^{ {\frac {t_{1}}{2}}}\big)^{p_{3}}A_{4}^{-{\frac {t_{2}} {2}}}\Big]^{p_{4}}A_{5}^{\frac r 2}\Big\}^{\frac {r-t_{2}}{\psi [4]-t_{2}+r}},
\end{equation}
\begin{equation}\tag{2.8}
A_{5}^{r-t_{2}}\geq \Big\{A_{5}^{\frac r 2}\Big[A_{5}^{-{\frac {t_{2}} {2}}}\big(A_{4}^{ {\frac {t_{1}}{2}}}(A_{3}^{-{\frac {t_{1}}{2}}}A_{2}^{p_{1}}A_{3}^{-{\frac {t_{1}}{2}}})^{p_{2}}A_{4}^{ {\frac {t_{1}}{2}}}\big)^{p_{3}}A_{5}^{-{\frac {t_{2}} {2}}}\Big]^{p_{4}}A_{5}^{\frac r 2}\Big\}^{\frac {r-t_{2}}{\psi [4]-t_{2}+r}},
\end{equation}
\begin{equation}\tag{2.9}
A_{1}^{r-t_{2}}\leq \Big\{A_{1}^{\frac r 2}\Big[A_{1}^{-{\frac {t_{2}} {2}}}\big(A_{2}^{ {\frac {t_{1}}{2}}}(A_{3}^{-{\frac {t_{1}}{2}}}A_{4}^{p_{1}}A_{3}^{-{\frac {t_{1}}{2}}})^{p_{2}}A_{2}^{ {\frac {t_{1}}{2}}}\big)^{p_{3}}A_{1}^{-{\frac {t_{2}} {2}}}\Big]^{p_{4}}A_{1}^{\frac r 2}\Big\}^{\frac {r-t_{2}}{\psi [4]-t_{2}+r}},
\end{equation}
\begin{equation}\tag{2.10}
A_{1}^{r-t_{2}}\leq \Big\{A_{1}^{\frac r 2}\Big[A_{2}^{-{\frac {t_{2}} {2}}}\big(A_{3}^{ {\frac {t_{1}}{2}}}(A_{4}^{-{\frac {t_{1}}{2}}}A_{5}^{p_{1}}A_{4}^{-{\frac {t_{1}}{2}}})^{p_{2}}A_{3}^{ {\frac {t_{1}}{2}}}\big)^{p_{3}}A_{2}^{-{\frac {t_{2}} {2}}}\Big]^{p_{4}}A_{1}^{\frac r 2}\Big\}^{\frac {r-t_{2}}{\psi [4]-t_{2}+r}}
\end{equation}
always hold for $p_{1}, p_{2}, p_{3}, p_{4}\geq 1$, $t_{1}, t_{2}\in [0, 1]$ and $r\geq t_{2}$, where $\psi [4]=\{[(p_{1}-t_{1})p_{2}+t_{1}]p_{3}-t_{2}\}p_{4}+t_{2}$.\\
\noindent{\bf Remark 2.2.} It should be mentioned that we can not obtain $A_{5}\geq A_{4}\geq A_{3}\geq A_{2}\geq A_{1}$ only by (2.7) and (2.10). If $A_{1}=\begin{pmatrix} 1 & 0 \\ 0 & {\frac 1 u} \end{pmatrix}$, $A_{2}=\begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix}$, $A_{3}=\begin{pmatrix} 1 & 0 \\ 0 & u \end{pmatrix}$, $A_{4}=\begin{pmatrix} u & 0 \\ 0 & 1 \end{pmatrix}$, $A_{5}=\begin{pmatrix} u+\varepsilon & 0 \\ 0 & 1 \end{pmatrix}$, where $u>1$ and $\varepsilon >0$, then the five strictly positive operators satisfy (2.7) and (2.10) without satisfying $A_{4}\geq A_{3}$.
\section{An application}
In what follows we give an application of the characterizations in Theorem 2.1 and Theorem 2.2 to operator equalities.
\noindent{\bf Theorem 3.1.} If $A_{1}, A_{2}, A_{3}, \cdots, A_{2n-1}, A_{2n}, A_{2n+1}$ are strictly positive operators, $t_{1}, t_{2}, \cdots, t_{n}\in [0, 1]$, $p_{1}, p_{2}, \cdots, p_{2n} \geq 1$, $\psi [2n]=\{\cdots [\{[(p_{1}-t_{1})p_{2}+t_{1}]p_{3}-t_{2}\}p_{4}+t_{2}]p_{5}-\cdots -t_{n}\}p_{2n}+t_{n}$, $r\geq t_{n}$, $m$ is a positive integer such that $(r-t_{n})m=\psi [2n]-t_{n}+r$ with $m\geq 2$, then the following assertions are mutually equivalent:\\
(I) $A_{2n+1}\geq A_{2n}\geq A_{2n-1}\geq \cdots \geq A_{3}\geq A_{2}\geq A_{1}$.\\
(II) The following operator inequalities hold: \\
\indent (II.1) $A^{r-t_{n}}_{2n+1}\geq \Big\{A^{\frac r 2}_{2n+1}\Big[A^{-{\frac {t_{n}}{2}}}_{2n}\big\{A^{\frac {t_{n-1}}{2}}_{2n-1} \cdots A^{\frac {t_{2}}{2}}_{5}\big[A^{-{\frac {t_{2}}{2}}}_{4}\cdot\{A^{\frac {t_{1}}{2}}_{3}(A^{-{\frac {t_{1}}{2}}}_{2}A^{p_{1}}_{1}A^{-{\frac {t_{1}}{2}}}_{2})^{p_{2}} A^{\frac {t_{1}}{2}}_{3}\}^{p_{3}}\cdot \\
A^{-{\frac {t_{2}}{2}}}_{4}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{5}
\cdots A^{\frac {t_{n-1}}{2}}_{2n-1}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n}\Big]^{p_{2n}}A^{\frac r 2}_{2n+1}\Big\}^{ \frac 1 m}$;\\
\indent (II.2) $A^{r-t_{n}}_{2n+1}\geq \Big\{A^{\frac r 2}_{2n+1}\Big[A^{-{\frac {t_{n}}{2}}}_{2n+1}\big\{A^{\frac {t_{n-1}}{2}}_{2n } \cdots A^{\frac {t_{2}}{2}}_{6}\big[A^{-{\frac {t_{2}}{2}}}_{5}\cdot\{A^{\frac {t_{1}}{2}}_{4}(A^{-{\frac {t_{1}}{2}}}_{3}A^{p_{1}}_{2}A^{-{\frac {t_{1}}{2}}}_{3})^{p_{2}} A^{\frac {t_{1}}{2}}_{4}\}^{p_{3}}\cdot \\
A^{-{\frac {t_{2}}{2}}}_{5}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{6}
\cdots A^{\frac {t_{n-1}}{2}}_{2n }\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n+1}\Big]^{p_{2n}}A^{\frac r 2}_{2n+1}\Big\}^{ \frac 1 m}$;\\
\indent (II.3) $A^{r-t_{n}}_{2n+1}\geq \Big\{A^{\frac r 2}_{2n+1}\Big[A^{-{\frac {t_{n}}{2}}}_{2n+1}\big\{A^{\frac {t_{n-1}}{2}}_{2n+1 } \cdots A^{\frac {t_{2}}{2}}_{7}\big[A^{-{\frac {t_{2}}{2}}}_{6}\cdot\{A^{\frac {t_{1}}{2}}_{5}(A^{-{\frac {t_{1}}{2}}}_{4}A^{p_{1}}_{3}A^{-{\frac {t_{1}}{2}}}_{4})^{p_{2}} A^{\frac {t_{1}}{2}}_{5}\}^{p_{3}} \cdot\\
A^{-{\frac {t_{2}}{2}}}_{6}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{7}
\cdots A^{\frac {t_{n-1}}{2}}_{2n+1 }\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n+1}\Big]^{p_{2n}}A^{\frac r 2}_{2n+1}\Big\}^{ \frac 1 m}$;\\
\indent { $\cdots \cdots \cdots \cdots$}\\
\indent (II.n) $A^{r-t_{n}}_{2n+1}\geq \Big\{A^{\frac r 2}_{2n+1}\Big[A^{-{\frac {t_{n}}{2}}}_{2n+1}\big\{A^{\frac {t_{n-1}}{2}}_{2n+1 } \cdots A^{\frac {t_{2}}{2}}_{n+4}\big[A^{-{\frac {t_{2}}{2}}}_{n+3}\{A^{\frac {t_{1}}{2}}_{n+2}(A^{-{\frac {t_{1}}{2}}}_{n+1}A^{p_{1}}_{n}A^{-{\frac {t_{1}}{2}}}_{n+1})^{p_{2}} A^{\frac {t_{1}}{2}}_{n+2}\}^{p_{3}} \\
A^{-{\frac {t_{2}}{2}}}_{n+3}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{n+4}
\cdots A^{\frac {t_{n-1}}{2}}_{2n+1 }\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n+1}\Big]^{p_{2n}}A^{\frac r 2}_{2n+1}\Big\}^{ \frac 1 m}$;\\
\indent (II.n+1) $A^{r-t_{n}}_{ 1}\leq \Big\{A^{\frac r 2}_{ 1}\Big[A^{-{\frac {t_{n}}{2}}}_{1}\big\{A^{\frac {t_{n-1}}{2}}_{ 1 } \cdots A^{\frac {t_{2}}{2}}_{n-2}\big[A^{-{\frac {t_{2}}{2}}}_{n-1}\{A^{\frac {t_{1}}{2}}_{n }(A^{-{\frac {t_{1}}{2}}}_{n+1}A^{p_{1}}_{n+2}A^{-{\frac {t_{1}}{2}}}_{n+1})^{p_{2}} A^{\frac {t_{1}}{2}}_{n }\}^{p_{3}} \\
A^{-{\frac {t_{2}}{2}}}_{n-1}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{n-2}
\cdots A^{\frac {t_{n-1}}{2}}_{1}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{1}\Big]^{p_{2n}}A^{\frac r 2}_{1}\Big\}^{ \frac 1 m}$;\\
\indent { $\cdots \cdots \cdots \cdots$}\\
\indent (II.2n-2) $A^{r-t_{n}}_{ 1}\leq \Big\{A^{\frac r 2}_{ 1}\Big[A^{-{\frac {t_{n}}{2}}}_{1}\big\{A^{\frac {t_{n-1}}{2}}_{1} \cdots A^{\frac {t_{2}}{2}}_{2n-5}\big[A^{-{\frac {t_{2}}{2}}}_{2n-4}\{A^{\frac {t_{1}}{2}}_{2n-3 }\cdot(A^{-{\frac {t_{1}}{2}}}_{2n-2}A^{p_{1}}_{2n-1}A^{-{\frac {t_{1}}{2}}}_{2n-2})^{p_{2}}\cdot \\ A^{\frac {t_{1}}{2}}_{2n-3}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{2n-4}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{2n-5}
\cdots A^{\frac {t_{n-1}}{2}}_{1}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{1}\Big]^{p_{2n}}A^{\frac r 2}_{1}\Big\}^{ \frac 1 m}$;\\
\indent (II.2n-1) $A^{r-t_{n}}_{ 1}\leq \Big\{A^{\frac r 2}_{ 1}\Big[A^{-{\frac {t_{n}}{2}}}_{1}\big\{A^{\frac {t_{n-1}}{2}}_{2} \cdots A^{\frac {t_{2}}{2}}_{2n-4}\big[A^{-{\frac {t_{2}}{2}}}_{2n-3}\{A^{\frac {t_{1}}{2}}_{2n-2 }\cdot(A^{-{\frac {t_{1}}{2}}}_{2n-1}A^{p_{1}}_{2n}A^{-{\frac {t_{1}}{2}}}_{2n-1})^{p_{2}}\cdot \\ A^{\frac {t_{1}}{2}}_{2n-2}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{2n-3}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{2n-4}
\cdots A^{\frac {t_{n-1}}{2}}_{2}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{1}\Big]^{p_{2n}}A^{\frac r 2}_{1}\Big\}^{ \frac 1 m}$;\\
\indent (II.2n) $A^{r-t_{n}}_{ 1}\leq \Big\{A^{\frac r 2}_{ 1}\Big[A^{-{\frac {t_{n}}{2}}}_{2}\big\{A^{\frac {t_{n-1}}{2}}_{3} \cdots A^{\frac {t_{2}}{2}}_{2n-3}\big[A^{-{\frac {t_{2}}{2}}}_{2n-2}\{A^{\frac {t_{1}}{2}}_{2n-1}\cdot(A^{-{\frac {t_{1}}{2}}}_{2n}A^{p_{1}}_{2n+1}A^{-{\frac {t_{1}}{2}}}_{2n})^{p_{2}}\cdot \\ A^{\frac {t_{1}}{2}}_{2n-1}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{2n-2}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{2n-3}
\cdots A^{\frac {t_{n-1}}{2}}_{3}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2}\Big]^{p_{2n}}A^{\frac r 2}_{1}\Big\}^{ \frac 1 m}$.\\
(III) There exists strictly positive operators $S_{1}, S_{2}, S_{3}, \cdots, S_{2n-2}, S_{2n-1}, S_{2n }$ satisfying the following operator equalities, respectively, where each $S_{i}$ $(i=1, 2, \cdots, 2n)$ is unique with $\|S_{i}\|\leq 1$.\\
\indent(III.1) $A_{2n+1}^{-{\frac {t_{n}}{2}}}S_{1}(A_{2n+1}^{r-t_{n}}S_{1})^{m-1}A_{2n+1}^{-{\frac {t_{n}}{2}}}=A_{2n+1}^{-{\frac {t_{n}}{2}}}(S_{1}A_{2n+1}^{r-t_{n}})^{m-1}S_{1}A_{2n+1}^{-{\frac {t_{n}}{2}}}=\\ \Big[A^{-{\frac {t_{n}}{2}}}_{2n}\big\{A^{\frac {t_{n-1}}{2}}_{2n-1} \cdots A^{\frac {t_{2}}{2}}_{5}\big[A^{-{\frac {t_{2}}{2}}}_{4} \{A^{\frac {t_{1}}{2}}_{3}(A^{-{\frac {t_{1}}{2}}}_{2}A^{p_{1}}_{1}A^{-{\frac {t_{1}}{2}}}_{2})^{p_{2}} A^{\frac {t_{1}}{2}}_{3}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{4}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{5}
\cdots A^{\frac {t_{n-1}}{2}}_{2n-1}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n}\Big]^{p_{2n}}$;\\
\indent(III.2) $A_{2n+1}^{-{\frac {t_{n}}{2}}}S_{2}(A_{2n+1}^{r-t_{n}}S_{2})^{m-1}A_{2n+1}^{-{\frac {t_{n}}{2}}}=A_{2n+1}^{-{\frac {t_{n}}{2}}}(S_{2}A_{2n+1}^{r-t_{n}})^{m-1}S_{2}A_{2n+1}^{-{\frac {t_{n}}{2}}}=\\ \Big[A^{-{\frac {t_{n}}{2}}}_{2n+1}\big\{A^{\frac {t_{n-1}}{2}}_{2n } \cdots A^{\frac {t_{2}}{2}}_{6}\big[A^{-{\frac {t_{2}}{2}}}_{5} \{A^{\frac {t_{1}}{2}}_{4}(A^{-{\frac {t_{1}}{2}}}_{3}A^{p_{1}}_{2}A^{-{\frac {t_{1}}{2}}}_{3})^{p_{2}} A^{\frac {t_{1}}{2}}_{4}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{5}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{6}
\cdots A^{\frac {t_{n-1}}{2}}_{2n }\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n+1}\Big]^{p_{2n}}$;\\
\indent(III.3) $A_{2n+1}^{-{\frac {t_{n}}{2}}}S_{3}(A_{2n+1}^{r-t_{n}}S_{3})^{m-1}A_{2n+1}^{-{\frac {t_{n}}{2}}}=A_{2n+1}^{-{\frac {t_{n}}{2}}}(S_{3}A_{2n+1}^{r-t_{n}})^{m-1}S_{3}A_{2n+1}^{-{\frac {t_{n}}{2}}}=\\ \Big[A^{-{\frac {t_{n}}{2}}}_{2n+1}\big\{A^{\frac {t_{n-1}}{2}}_{2n+1 } \cdots A^{\frac {t_{2}}{2}}_{7}\big[A^{-{\frac {t_{2}}{2}}}_{6} \{A^{\frac {t_{1}}{2}}_{5}(A^{-{\frac {t_{1}}{2}}}_{4}A^{p_{1}}_{3}A^{-{\frac {t_{1}}{2}}}_{4})^{p_{2}} A^{\frac {t_{1}}{2}}_{5}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{6}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{7}
\cdots A^{\frac {t_{n-1}}{2}}_{2n+1 }\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n+1}\Big]^{p_{2n}}$;\\
\indent { $\cdots \cdots \cdots \cdots$}\\
\indent(III.n) $A_{2n+1}^{-{\frac {t_{n}}{2}}}S_{n}(A_{2n+1}^{r-t_{n}}S_{n})^{m-1}A_{2n+1}^{-{\frac {t_{n}}{2}}}=A_{2n+1}^{-{\frac {t_{n}}{2}}}(S_{n}A_{2n+1}^{r-t_{n}})^{m-1}S_{n}A_{2n+1}^{-{\frac {t_{n}}{2}}}=\\ \Big[A^{-{\frac {t_{n}}{2}}}_{2n+1}\big\{A^{\frac {t_{n-1}}{2}}_{2n+1 } \cdots A^{\frac {t_{2}}{2}}_{n+4}\big[A^{-{\frac {t_{2}}{2}}}_{n+3}\{A^{\frac {t_{1}}{2}}_{n+2}(A^{-{\frac {t_{1}}{2}}}_{n+1}A^{p_{1}}_{n}A^{-{\frac {t_{1}}{2}}}_{n+1})^{p_{2}} A^{\frac {t_{1}}{2}}_{n+2}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{n+3}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{n+4}
\cdots A^{\frac {t_{n-1}}{2}}_{2n+1 }\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n+1}\Big]^{p_{2n}}$;\\
\indent(III.n+1) $A_{1}^{-{\frac {t_{n}}{2}}}S_{n+1}^{-1}(A_{1}^{r-t_{n}}S_{n+1}^{-1})^{m-1}A_{1}^{-{\frac {t_{n}}{2}}}=A_{1}^{-{\frac {t_{n}}{2}}}(S_{n+1}^{-1}A_{1}^{r-t_{n}})^{m-1}S_{n+1}^{-1}A_{1}^{-{\frac {t_{n}}{2}}}=\\\Big[A^{-{\frac {t_{n}}{2}}}_{1}\big\{A^{\frac {t_{n-1}}{2}}_{ 1 } \cdots A^{\frac {t_{2}}{2}}_{n-2}\big[A^{-{\frac {t_{2}}{2}}}_{n-1}\{A^{\frac {t_{1}}{2}}_{n }(A^{-{\frac {t_{1}}{2}}}_{n+1}A^{p_{1}}_{n+2}A^{-{\frac {t_{1}}{2}}}_{n+1})^{p_{2}} A^{\frac {t_{1}}{2}}_{n }\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{n-1}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{n-2}
\cdots A^{\frac {t_{n-1}}{2}}_{1}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{1}\Big]^{p_{2n}}$;\\
\indent { $\cdots \cdots \cdots \cdots$}\\
\indent(III.2n-2) $ A_{1}^{-{\frac {t_{n}}{2}}}S_{2n-2}^{-1}(A_{1}^{r-t_{n}}S_{2n-2}^{-1})^{m-1}A_{1}^{-{\frac {t_{n}}{2}}}=A_{1}^{-{\frac {t_{n}}{2}}}(S_{2n-2}^{-1}A_{1}^{r-t_{n}})^{m-1}S_{2n-2}^{-1}A_{1}^{-{\frac {t_{n}}{2}}}=\\ \Big[A^{-{\frac {t_{n}}{2}}}_{1} \big\{A^{\frac {t_{n-1}}{2}}_{1} \cdots \big[A^{-{\frac {t_{2}}{2}}}_{2n-4}\{A^{\frac {t_{1}}{2}}_{2n-3 } (A^{-{\frac {t_{1}}{2}}}_{2n-2}A^{p_{1}}_{2n-1}A^{-{\frac {t_{1}}{2}}}_{2n-2})^{p_{2}} A^{\frac {t_{1}}{2}}_{2n-3}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{2n-4}\big]^{p_{4}}
\cdots A^{\frac {t_{n-1}}{2}}_{1}\big\}^{p_{2n-1}} A^{-{\frac {t_{n}}{2}}}_{1}\Big]^{p_{2n}}$;\\
\indent(III.2n-1) $ A_{1}^{-{\frac {t_{n}}{2}}}S_{2n-1}^{-1}(A_{1}^{r-t_{n}}S_{2n-1}^{-1})^{m-1}A_{1}^{-{\frac {t_{n}}{2}}}=A_{1}^{-{\frac {t_{n}}{2}}}(S_{2n-1}^{-1}A_{1}^{r-t_{n}})^{m-1}S_{2n-1}^{-1}A_{1}^{-{\frac {t_{n}}{2}}}=\\ \Big[A^{-{\frac {t_{n}}{2}}}_{1}\big\{A^{\frac {t_{n-1}}{2}}_{2} \cdots \big[A^{-{\frac {t_{2}}{2}}}_{2n-3}\{A^{\frac {t_{1}}{2}}_{2n-2 } (A^{-{\frac {t_{1}}{2}}}_{2n-1}A^{p_{1}}_{2n}A^{-{\frac {t_{1}}{2}}}_{2n-1})^{p_{2}} A^{\frac {t_{1}}{2}}_{2n-2}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{2n-3}\big]^{p_{4}}
\cdots A^{\frac {t_{n-1}}{2}}_{2}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{1}\Big]^{p_{2n}}$;\\
\indent(III.2n) $A_{1}^{-{\frac {t_{n}}{2}}}S_{2n }^{-1}(A_{1}^{r-t_{n}}S_{2n }^{-1})^{m-1}A_{1}^{-{\frac {t_{n}}{2}}}=A_{1}^{-{\frac {t_{n}}{2}}}(S_{2n }^{-1}A_{1}^{r-t_{n}})^{m-1}S_{2n }^{-1}A_{1}^{-{\frac {t_{n}}{2}}}=\\ \Big[A^{-{\frac {t_{n}}{2}}}_{2}\big\{A^{\frac {t_{n-1}}{2}}_{3} \cdots \big[A^{-{\frac {t_{2}}{2}}}_{2n-2}\{A^{\frac {t_{1}}{2}}_{2n-1} (A^{-{\frac {t_{1}}{2}}}_{2n}A^{p_{1}}_{2n+1}A^{-{\frac {t_{1}}{2}}}_{2n})^{p_{2}} A^{\frac {t_{1}}{2}}_{2n-1}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{2n-2}\big]^{p_{4}}
\cdots A^{\frac {t_{n-1}}{2}}_{3}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2}\Big]^{p_{2n}}$.\\
\noindent{\bf Proof.} Because (II)$\Leftrightarrow$(III) holds obviously by Theorem 2.1, we only need to prove that (II)$\Leftrightarrow$(III).
Firstly, let us prove that (II.1)$\Rightarrow$(III.1). We recall Douglas's majorization and factorization theorem in \cite{Douglas_1966}: $SS^{\ast}\leq \lambda^{2}TT^{\ast}\Leftrightarrow $ there exists an operator $Q$ s.t. $TQ=S$, where $\|Q\|^{2}=inf\{\mu: SS^{\ast}\leq \mu TT^{\ast}\}$.\\
By (II.1), there exists an operator $E_{1}$ with $\|E_{1}\|\leq 1$ such that
\begin{equation}
\tag{3.1}
\begin{split}
&\ A_{2n+1}^{\frac {r-t_{n}}{2}}E_{1}=E_{1}^{\ast}A_{2n+1}^{\frac {r-t_{n}}{2}}\\
=&\ \Big\{A^{\frac r 2}_{2n+1}\Big[A^{-{\frac {t_{n}}{2}}}_{2n}\big\{A^{\frac {t_{n-1}}{2}}_{2n-1} \cdots \{A^{\frac {t_{1}}{2}}_{3}(A^{-{\frac {t_{1}}{2}}}_{2}A^{p_{1}}_{1}A^{-{\frac {t_{1}}{2}}}_{2})^{p_{2}} A^{\frac {t_{1}}{2}}_{3}\}^{p_{3}}
\cdots A^{\frac {t_{n-1}}{2}}_{2n-1}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n}\Big]^{p_{2n}}A^{\frac r 2}_{2n+1}\Big\}^{ \frac {1} {2m}}.
\end{split}
\end{equation}
Taking $S_{1}=E_{1}E_{1}^{\ast}$, we have
\begin{equation}
\tag{3.2}
\begin{split}
&\ A_{2n+1}^{\frac {r-t_{n}}{2}}S_{1}A_{2n+1}^{\frac {r-t_{n}}{2}}\\
=&\ \Big\{A^{\frac r 2}_{2n+1}\Big[A^{-{\frac {t_{n}}{2}}}_{2n}\big\{A^{\frac {t_{n-1}}{2}}_{2n-1} \cdots \{A^{\frac {t_{1}}{2}}_{3}(A^{-{\frac {t_{1}}{2}}}_{2}A^{p_{1}}_{1}A^{-{\frac {t_{1}}{2}}}_{2})^{p_{2}} A^{\frac {t_{1}}{2}}_{3}\}^{p_{3}}
\cdots A^{\frac {t_{n-1}}{2}}_{2n-1}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n}\Big]^{p_{2n}}A^{\frac r 2}_{2n+1}\Big\}^{ \frac {1} { m}}.
\end{split}
\end{equation}
According to (3.2) and $S_{1}=E_{1}E_{1}^{\ast}$, $S_{1}$ is unique and strictly positive with $\|S_{1}\|\leq 1$. (3.2) also implies that
\begin{equation}
\tag{3.3}
\begin{split}
&\ (A_{2n+1}^{\frac {r-t_{n}}{2}}S_{1}A_{2n+1}^{\frac {r-t_{n}}{2}})^{m}
= A_{2n+1}^{ {\frac {r-t_{n}}{2}}}S_{1}(A_{2n+1}^{r-t_{n}}S_{1})^{m-1}A_{2n+1}^{ {\frac {r-t_{n}}{2}}}
= A_{2n+1}^{ {\frac {r-t_{n}}{2}}}(S_{1}A_{2n+1}^{r-t_{n}})^{m-1}S_{1}A_{2n+1}^{ {\frac {r-t_{n}}{2}}}\\
=&\ A^{\frac r 2}_{2n+1}\Big[A^{-{\frac {t_{n}}{2}}}_{2n}\big\{A^{\frac {t_{n-1}}{2}}_{2n-1} \cdots \{A^{\frac {t_{1}}{2}}_{3}(A^{-{\frac {t_{1}}{2}}}_{2}A^{p_{1}}_{1}A^{-{\frac {t_{1}}{2}}}_{2})^{p_{2}} A^{\frac {t_{1}}{2}}_{3}\}^{p_{3}}
\cdots A^{\frac {t_{n-1}}{2}}_{2n-1}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n}\Big]^{p_{2n}}A^{\frac r 2}_{2n+1}.
\end{split}
\end{equation}
Then (III.1) holds by (3.3).
Secondly, we prove that (III.1)$\Rightarrow$ (II.1). By (III.1),
\begin{equation*}
\begin{split}
&\ \Big\{A^{\frac r 2}_{2n+1}\Big[A^{-{\frac {t_{n}}{2}}}_{2n}\big\{A^{\frac {t_{n-1}}{2}}_{2n-1} \cdots \{A^{\frac {t_{1}}{2}}_{3}(A^{-{\frac {t_{1}}{2}}}_{2}A^{p_{1}}_{1}A^{-{\frac {t_{1}}{2}}}_{2})^{p_{2}} A^{\frac {t_{1}}{2}}_{3}\}^{p_{3}}
\cdots A^{\frac {t_{n-1}}{2}}_{2n-1}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n}\Big]^{p_{2n}}A^{\frac r 2}_{2n+1}\Big\}^{ \frac 1 m}\\
= &\ \big\{A_{2n+1}^{ {\frac {r-t_{n}}{2}}}S_{1}(A_{2n+1}^{r-t_{n}}S_{1})^{m-1}A_{2n+1}^{ {\frac {r-t_{n}}{2}}}\big\}^{\frac 1 m}\\
=&\ \big\{(A_{2n+1}^{ {\frac {r-t_{n}}{2}}}S_{1}A_{2n+1}^{ {\frac {r-t_{n}}{2}}})\cdot(A_{2n+1}^{ {\frac {r-t_{n}}{2}}}S_{1}A_{2n+1}^{ {\frac {r-t_{n}}{2}}})\cdots (A_{2n+1}^{ {\frac {r-t_{n}}{2}}}S_{1}A_{2n+1}^{ {\frac {r-t_{n}}{2}}})\big\}^{\frac 1 m}\\
=&\ A_{2n+1}^{ {\frac {r-t_{n}}{2}}}S_{1}A_{2n+1}^{ {\frac {r-t_{n}}{2}}}\\
\leq &\ A_{2n+1}^{r-t_{n}}.
\end{split}
\end{equation*}
The inequality follows form the fact that $S_{1}\leq \|S_{1}\|I\leq I$, and then (II.1) holds.
By using the same method above, we can prove that (II.2)$\Leftrightarrow$ (III.2), (II.3)$\Leftrightarrow$ (III.3), $\cdots$, (II.n)$\Leftrightarrow$ (III.n), respectively.
Next, we show that (II.2n)$\Leftrightarrow$(III.2n). Notice that for two strictly positive operators $S$ and $T$, $S\geq T$ if and only if $T^{-1}\geq S^{-1}$. Then (II.2n) is equivalent to
\begin{equation}
\tag{3.4}
\begin{split}
A^{-(r-t_{n})}_{ 1}\geq &\ \Big\{A^{-{\frac r 2}}_{ 1}\Big[A^{-{\frac {t_{n}}{2}}}_{2}\big\{A^{\frac {t_{n-1}}{2}}_{3} \cdots A^{\frac {t_{2}}{2}}_{2n-3}\big[A^{-{\frac {t_{2}}{2}}}_{2n-2}\{A^{\frac {t_{1}}{2}}_{2n-1} (A^{-{\frac {t_{1}}{2}}}_{2n}A^{p_{1}}_{2n+1}A^{-{\frac {t_{1}}{2}}}_{2n})^{p_{2}} \\&\ A^{\frac {t_{1}}{2}}_{2n-1}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{2n-2}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{2n-3}
\cdots A^{\frac {t_{n-1}}{2}}_{3}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2}\Big]^{-{p_{2n}}}A^{-{\frac r 2}}_{1}\Big\}^{ \frac 1 m}.
\end{split}
\end{equation}
The proof that (3.4)$\Leftrightarrow$(III.2n) is similar to the proof of that (II.1)$\Leftrightarrow$(III.1), so we omit it here.
Repeat the method above, we can prove that (II.2n-1)$\Leftrightarrow$(III.2n-1), (II.2n-2)$\Leftrightarrow$(III.2n-2), $\cdots $, (II.n+1)$\Leftrightarrow$(III.n+1). $\square$\\
\noindent{\bf Theorem 3.2.} If $A_{1}, A_{2}, A_{3}, \cdots, A_{2n-1}, A_{2n}$ are strictly positive operators, $t_{1}, t_{2}, \cdots$, $t_{n}\in [0, 1]$, $p_{1}, p_{2}, \cdots, p_{2n} \geq 1$, $\psi [2n]=\{\cdots [\{[(p_{1}-t_{1})p_{2}+t_{1}]p_{3}-t_{2}\}p_{4}+t_{2}]p_{5}-\cdots -t_{n}\}p_{2n}+t_{n}$, $r\geq t_{n}$, $m$ is a positive integer such that $(r-t_{n})m=\psi [2n]-t_{n}+r$ with $m\geq 2$, then the following assertions are mutually equivalent:\\
(I) $ A_{2n}\geq A_{2n-1}\geq \cdots \geq A_{3}\geq A_{2}\geq A_{1}$.\\
(II) The following operator inequalities hold: \\
\indent (II.1) $A^{r-t_{n}}_{2n}\geq \Big\{A^{\frac r 2}_{2n }\Big[A^{-{\frac {t_{n}}{2}}}_{2n}\big\{A^{\frac {t_{n-1}}{2}}_{2n-1} \cdots A^{\frac {t_{2}}{2}}_{5}\big[A^{-{\frac {t_{2}}{2}}}_{4}\cdot\{A^{\frac {t_{1}}{2}}_{3}(A^{-{\frac {t_{1}}{2}}}_{2}A^{p_{1}}_{1}A^{-{\frac {t_{1}}{2}}}_{2})^{p_{2}} A^{\frac {t_{1}}{2}}_{3}\}^{p_{3}}\cdot \\
A^{-{\frac {t_{2}}{2}}}_{4}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{5}
\cdots A^{\frac {t_{n-1}}{2}}_{2n-1}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n}\Big]^{p_{2n}}A^{\frac r 2}_{2n }\Big\}^{\frac 1 m }$;\\
\indent (II.2) $A^{r-t_{n}}_{2n }\geq \Big\{A^{\frac r 2}_{2n }\Big[A^{-{\frac {t_{n}}{2}}}_{2n }\big\{A^{\frac {t_{n-1}}{2}}_{2n } \cdots A^{\frac {t_{2}}{2}}_{6}\big[A^{-{\frac {t_{2}}{2}}}_{5}\cdot\{A^{\frac {t_{1}}{2}}_{4}(A^{-{\frac {t_{1}}{2}}}_{3}A^{p_{1}}_{2}A^{-{\frac {t_{1}}{2}}}_{3})^{p_{2}} A^{\frac {t_{1}}{2}}_{4}\}^{p_{3}}\cdot \\
A^{-{\frac {t_{2}}{2}}}_{5}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{6}
\cdots A^{\frac {t_{n-1}}{2}}_{2n }\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n }\Big]^{p_{2n}}A^{\frac r 2}_{2n }\Big\}^{\frac 1 m }$;\\
\indent (II.3) $A^{r-t_{n}}_{2n }\geq \Big\{A^{\frac r 2}_{2n }\Big[A^{-{\frac {t_{n}}{2}}}_{2n }\big\{A^{\frac {t_{n-1}}{2}}_{2n } \cdots A^{\frac {t_{2}}{2}}_{7}\big[A^{-{\frac {t_{2}}{2}}}_{6}\cdot\{A^{\frac {t_{1}}{2}}_{5}(A^{-{\frac {t_{1}}{2}}}_{4}A^{p_{1}}_{3}A^{-{\frac {t_{1}}{2}}}_{4})^{p_{2}} A^{\frac {t_{1}}{2}}_{5}\}^{p_{3}} \cdot\\
A^{-{\frac {t_{2}}{2}}}_{6}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{7}
\cdots A^{\frac {t_{n-1}}{2}}_{2n }\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n }\Big]^{p_{2n}}A^{\frac r 2}_{2n }\Big\}^{\frac 1 m }$;\\
\indent { $\cdots \cdots \cdots \cdots$}\\
\indent (II.n) $A^{r-t_{n}}_{2n }\geq \Big\{A^{\frac r 2}_{2n }\Big[A^{-{\frac {t_{n}}{2}}}_{2n }\big\{A^{\frac {t_{n-1}}{2}}_{2n } \cdots A^{\frac {t_{2}}{2}}_{n+4}\big[A^{-{\frac {t_{2}}{2}}}_{n+3}\{A^{\frac {t_{1}}{2}}_{n+2}(A^{-{\frac {t_{1}}{2}}}_{n+1}A^{p_{1}}_{n}A^{-{\frac {t_{1}}{2}}}_{n+1})^{p_{2}} A^{\frac {t_{1}}{2}}_{n+2}\}^{p_{3}} \\
A^{-{\frac {t_{2}}{2}}}_{n+3}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{n+4}
\cdots A^{\frac {t_{n-1}}{2}}_{2n }\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n }\Big]^{p_{2n}}A^{\frac r 2}_{2n }\Big\}^{\frac 1 m }$;\\
\indent (II.n+1) $A^{r-t_{n}}_{ 1}\leq \Big\{A^{\frac r 2}_{ 1}\Big[A^{-{\frac {t_{n}}{2}}}_{1}\big\{A^{\frac {t_{n-1}}{2}}_{ 1 } \cdots A^{\frac {t_{2}}{2}}_{n-2}\big[A^{-{\frac {t_{2}}{2}}}_{n-1}\{A^{\frac {t_{1}}{2}}_{n }(A^{-{\frac {t_{1}}{2}}}_{n+1}A^{p_{1}}_{n+2}A^{-{\frac {t_{1}}{2}}}_{n+1})^{p_{2}} A^{\frac {t_{1}}{2}}_{n }\}^{p_{3}} \\
A^{-{\frac {t_{2}}{2}}}_{n-1}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{n-2}
\cdots A^{\frac {t_{n-1}}{2}}_{1}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{1}\Big]^{p_{2n}}A^{\frac r 2}_{1}\Big\}^{\frac 1 m }$;\\
\indent { $\cdots \cdots \cdots \cdots$}\\
\indent (II.2n-2) $A^{r-t_{n}}_{ 1}\leq \Big\{A^{\frac r 2}_{ 1}\Big[A^{-{\frac {t_{n}}{2}}}_{1}\big\{A^{\frac {t_{n-1}}{2}}_{1} \cdots A^{\frac {t_{2}}{2}}_{2n-5}\big[A^{-{\frac {t_{2}}{2}}}_{2n-4}\{A^{\frac {t_{1}}{2}}_{2n-3 }\cdot(A^{-{\frac {t_{1}}{2}}}_{2n-2}A^{p_{1}}_{2n-1}A^{-{\frac {t_{1}}{2}}}_{2n-2})^{p_{2}}\cdot \\ A^{\frac {t_{1}}{2}}_{2n-3}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{2n-4}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{2n-5}
\cdots A^{\frac {t_{n-1}}{2}}_{1}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{1}\Big]^{p_{2n}}A^{\frac r 2}_{1}\Big\}^{\frac 1 m }$;\\
\indent (II.2n-1) $A^{r-t_{n}}_{ 1}\leq \Big\{A^{\frac r 2}_{ 1}\Big[A^{-{\frac {t_{n}}{2}}}_{1}\big\{A^{\frac {t_{n-1}}{2}}_{2} \cdots A^{\frac {t_{2}}{2}}_{2n-4}\big[A^{-{\frac {t_{2}}{2}}}_{2n-3}\{A^{\frac {t_{1}}{2}}_{2n-2 }\cdot(A^{-{\frac {t_{1}}{2}}}_{2n-1}A^{p_{1}}_{2n}A^{-{\frac {t_{1}}{2}}}_{2n-1})^{p_{2}}\cdot \\ A^{\frac {t_{1}}{2}}_{2n-2}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{2n-3}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{2n-4}
\cdots A^{\frac {t_{n-1}}{2}}_{2}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{1}\Big]^{p_{2n}}A^{\frac r 2}_{1}\Big\}^{\frac 1 m }$.\\
(III) There exists strictly positive operators $S_{1}, S_{2}, S_{3}, \cdots, S_{2n-2}, S_{2n-1}$ satisfying the following operator equalities, respectively, where each $S_{i}$ $(i=1, 2, \cdots, 2n-1)$ is unique with $\|S_{i}\|\leq 1$.\\
\indent(III.1) $A_{2n}^{-{\frac {t_{n}}{2}}}S_{1}(A_{2n}^{r-t_{n}}S_{1})^{m-1}A_{2n}^{-{\frac {t_{n}}{2}}}=A_{2n}^{-{\frac {t_{n}}{2}}}(S_{1}A_{2n}^{r-t_{n}})^{m-1}S_{1}A_{2n}^{-{\frac {t_{n}}{2}}}=\\ \Big[A^{-{\frac {t_{n}}{2}}}_{2n}\big\{A^{\frac {t_{n-1}}{2}}_{2n-1} \cdots A^{\frac {t_{2}}{2}}_{5}\big[A^{-{\frac {t_{2}}{2}}}_{4} \{A^{\frac {t_{1}}{2}}_{3}(A^{-{\frac {t_{1}}{2}}}_{2}A^{p_{1}}_{1}A^{-{\frac {t_{1}}{2}}}_{2})^{p_{2}} A^{\frac {t_{1}}{2}}_{3}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{4}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{5}
\cdots A^{\frac {t_{n-1}}{2}}_{2n-1}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n}\Big]^{p_{2n}}$;\\
\indent(III.2) $A_{2n}^{-{\frac {t_{n}}{2}}}S_{2}(A_{2n}^{r-t_{n}}S_{2})^{m-1}A_{2n}^{-{\frac {t_{n}}{2}}}=A_{2n}^{-{\frac {t_{n}}{2}}}(S_{2}A_{2n}^{r-t_{n}})^{m-1}S_{2}A_{2n}^{-{\frac {t_{n}}{2}}}=\\ \Big[A^{-{\frac {t_{n}}{2}}}_{2n}\big\{A^{\frac {t_{n-1}}{2}}_{2n } \cdots A^{\frac {t_{2}}{2}}_{6}\big[A^{-{\frac {t_{2}}{2}}}_{5} \{A^{\frac {t_{1}}{2}}_{4}(A^{-{\frac {t_{1}}{2}}}_{3}A^{p_{1}}_{2}A^{-{\frac {t_{1}}{2}}}_{3})^{p_{2}} A^{\frac {t_{1}}{2}}_{4}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{5}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{6}
\cdots A^{\frac {t_{n-1}}{2}}_{2n }\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n}\Big]^{p_{2n}}$;\\
\indent(III.3) $A_{2n}^{-{\frac {t_{n}}{2}}}S_{3}(A_{2n}^{r-t_{n}}S_{3})^{m-1}A_{2n}^{-{\frac {t_{n}}{2}}}=A_{2n}^{-{\frac {t_{n}}{2}}}(S_{3}A_{2n}^{r-t_{n}})^{m-1}S_{3}A_{2n}^{-{\frac {t_{n}}{2}}}=\\ \Big[A^{-{\frac {t_{n}}{2}}}_{2n }\big\{A^{\frac {t_{n-1}}{2}}_{2n } \cdots A^{\frac {t_{2}}{2}}_{7}\big[A^{-{\frac {t_{2}}{2}}}_{6} \{A^{\frac {t_{1}}{2}}_{5}(A^{-{\frac {t_{1}}{2}}}_{4}A^{p_{1}}_{3}A^{-{\frac {t_{1}}{2}}}_{4})^{p_{2}} A^{\frac {t_{1}}{2}}_{5}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{6}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{7}
\cdots A^{\frac {t_{n-1}}{2}}_{2n }\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n }\Big]^{p_{2n}}$;\\
\indent { $\cdots \cdots \cdots \cdots$}\\
\indent(III.n) $A_{2n}^{-{\frac {t_{n}}{2}}}S_{n}(A_{2n}^{r-t_{n}}S_{n})^{m-1}A_{2n}^{-{\frac {t_{n}}{2}}}=A_{2n}^{-{\frac {t_{n}}{2}}}(S_{n}A_{2n}^{r-t_{n}})^{m-1}S_{n}A_{2n}^{-{\frac {t_{n}}{2}}}=\\ \Big[A^{-{\frac {t_{n}}{2}}}_{2n }\big\{A^{\frac {t_{n-1}}{2}}_{2n } \cdots A^{\frac {t_{2}}{2}}_{n+4}\big[A^{-{\frac {t_{2}}{2}}}_{n+3}\{A^{\frac {t_{1}}{2}}_{n+2}(A^{-{\frac {t_{1}}{2}}}_{n+1}A^{p_{1}}_{n}A^{-{\frac {t_{1}}{2}}}_{n+1})^{p_{2}} A^{\frac {t_{1}}{2}}_{n+2}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{n+3}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{n+4}
\cdots A^{\frac {t_{n-1}}{2}}_{2n }\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{2n }\Big]^{p_{2n}}$;\\
\indent(III.n+1) $A_{1}^{-{\frac {t_{n}}{2}}}S_{n+1}^{-1}(A_{1}^{r-t_{n}}S_{n+1}^{-1})^{m-1}A_{1}^{-{\frac {t_{n}}{2}}}=A_{1}^{-{\frac {t_{n}}{2}}}(S_{n+1}^{-1}A_{1}^{r-t_{n}})^{m-1}S_{n+1}^{-1}A_{1}^{-{\frac {t_{n}}{2}}}=\\\Big[A^{-{\frac {t_{n}}{2}}}_{1}\big\{A^{\frac {t_{n-1}}{2}}_{ 1 } \cdots A^{\frac {t_{2}}{2}}_{n-2}\big[A^{-{\frac {t_{2}}{2}}}_{n-1}\{A^{\frac {t_{1}}{2}}_{n }(A^{-{\frac {t_{1}}{2}}}_{n+1}A^{p_{1}}_{n+2}A^{-{\frac {t_{1}}{2}}}_{n+1})^{p_{2}} A^{\frac {t_{1}}{2}}_{n }\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{n-1}\big]^{p_{4}}A^{\frac {t_{2}}{2}}_{n-2}
\cdots A^{\frac {t_{n-1}}{2}}_{1}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{1}\Big]^{p_{2n}}$;\\
\indent { $\cdots \cdots \cdots \cdots$}\\
\indent(III.2n-2) $ A_{1}^{-{\frac {t_{n}}{2}}}S_{2n-2}^{-1}(A_{1}^{r-t_{n}}S_{2n-2}^{-1})^{m-1}A_{1}^{-{\frac {t_{n}}{2}}}=A_{1}^{-{\frac {t_{n}}{2}}}(S_{2n-2}^{-1}A_{1}^{r-t_{n}})^{m-1}S_{2n-2}^{-1}A_{1}^{-{\frac {t_{n}}{2}}}=\\ \Big[A^{-{\frac {t_{n}}{2}}}_{1} \big\{A^{\frac {t_{n-1}}{2}}_{1} \cdots \big[A^{-{\frac {t_{2}}{2}}}_{2n-4}\{A^{\frac {t_{1}}{2}}_{2n-3 } (A^{-{\frac {t_{1}}{2}}}_{2n-2}A^{p_{1}}_{2n-1}A^{-{\frac {t_{1}}{2}}}_{2n-2})^{p_{2}} A^{\frac {t_{1}}{2}}_{2n-3}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{2n-4}\big]^{p_{4}}
\cdots A^{\frac {t_{n-1}}{2}}_{1}\big\}^{p_{2n-1}} A^{-{\frac {t_{n}}{2}}}_{1}\Big]^{p_{2n}}$;\\
\indent(III.2n-1) $ A_{1}^{-{\frac {t_{n}}{2}}}S_{2n-1}^{-1}(A_{1}^{r-t_{n}}S_{2n-1}^{-1})^{m-1}A_{1}^{-{\frac {t_{n}}{2}}}=A_{1}^{-{\frac {t_{n}}{2}}}(S_{2n-1}^{-1}A_{1}^{r-t_{n}})^{m-1}S_{2n-1}^{-1}A_{1}^{-{\frac {t_{n}}{2}}}=\\ \Big[A^{-{\frac {t_{n}}{2}}}_{1}\big\{A^{\frac {t_{n-1}}{2}}_{2} \cdots \big[A^{-{\frac {t_{2}}{2}}}_{2n-3}\{A^{\frac {t_{1}}{2}}_{2n-2 } (A^{-{\frac {t_{1}}{2}}}_{2n-1}A^{p_{1}}_{2n}A^{-{\frac {t_{1}}{2}}}_{2n-1})^{p_{2}} A^{\frac {t_{1}}{2}}_{2n-2}\}^{p_{3}}
A^{-{\frac {t_{2}}{2}}}_{2n-3}\big]^{p_{4}}
\cdots A^{\frac {t_{n-1}}{2}}_{2}\big\}^{p_{2n-1}}A^{-{\frac {t_{n}}{2}}}_{1}\Big]^{p_{2n}}$;\\
\noindent{\bf Proof.} Let $A_{2n+1}=A_{2n}$ in Theorem 3.1. $\square$\\
Together with Theorem 3.1 and Theorem 3.2, we give an application of the characterizations of $A_{k}\geq A_{k-1}\geq \cdots \geq A_{2}\geq A_{1}>0$ to operator equalities for any positive integer $k$.
\begin{center}
|
\section{Introduction}
High-order harmonic generation (HHG) is a reliable source of coherent soft x-ray radiation in the nonrelativistic regime. With the state-of-the-art technology, coherent x-rays of up to $\sim3$~keV~\cite{Seres:XA-06} of photon energy can be generated.
The most favorable conversion efficiency for nonrelativistic multi-keV harmonics is anticipated with mid-infrared driving laser fields at high gas pressures~\cite{Popmintchev,Kapteyn,Murnane}.
A further increase of the photon energy can in principle be achieved by
increasing the laser intensity. However, the applicable laser intensity is
limited in two ways~\cite{Piazza:Re-12}. First, the relativistic electron drift prevents recollision
and results in a dramatic suppression of the HHG efficiency. And second, the
strong field causes rapid ionization of the medium leading to a large free
electron dispersion and along with a significant phase-mismatch. The electron
recollision is suppressed when the drift distance becomes larger than the
electron wave packet size at the moment of recollision~\cite{Walker}. This
happens when the laser intensity exceeds $10^{16}-10^{17}$ W/cm$^2$ for infrared
(IR) wavelengths. The ponderomotive potential of the laser field in this case
amounts to $U_p\approx 3$ keV and the achievable cutoff frequency for HHG
to $\omega_{c}\approx 10$ keV. This indicates the limit of nonrelativistic HHG.
Various methods to counteract the relativistic drift have been proposed. To suppress the drift, different laser field geometries, in some cases with an additional field, can be applied~\cite{Lin:HH-06,KYLSTRA:BS-00,TARANUKHIN:RH-00,VERSCHL:SW-07,MILOSEVIC:DA-04,LIU:LG-09,CHIRILA:ND-02,VERSCHL:RC-07,VERSCHL:RF-07,KLAIBER:TA-06,KLAIBER:LO-07,KLAIBER:CH-08,HATSAGORTSYAN:UFO-08,VERSCHL:LR-08}. Highly charged ions moving relativistically~\cite{MOCKEN:BA-04,CHIRILA:IS-04} or a gas of positronium atoms~\cite{HENRICH:PI-04,HATSAGORTSYAN:ML-06} can also be employed for this purpose. However, all these efforts have only addressed the drift suppression problem for the emission from a single atom rather than coherent emission from a macroscopic gas target where phase-matching becomes crucial. For the first time both
problems of relativistic HHG, namely, the relativistic drift and the phase-matching, have been solved at the same time in~\cite{KOHLER:HX-11} where the macroscopic yield of HHG has been calculated in the setup consisting of two counterpropagating attosecond pulse trains~\cite{HATSAGORTSYAN:UFO-08}. In this setup, the relativistic drift caused by the ionizing laser pulse is reverted by the counter-propagating pulse inducing recombination. It appears that, specific to this setup, an additional harmonic phase exists which depends on the time delay between the driving pulse trains. We have shown in~\cite{KOHLER:HX-11} that this additional phase of the emitted harmonics can be tuned to compensate the phase mismatch caused by the free electron background. However, the setup of counter-propagating pulses is rather challenging, for instance, the requirement for a small pulse distortion imposes a rather strong restriction on the medium length. Additionally, a precise modulation of the laser intensity along the propagation direction is required.
Another appealing scheme for relativistic HHG exists based on XUV assistance~\cite{KLAIBER:CH-08} which seems experimentally less demanding than the scheme with counter-propagating attosecond pulses.
The usefulness of XUV light assisting a strong laser field has been demonstrated in the nonrelativistic regime for various purposes.
It has been used to enhance HHG by many orders of magnitude compared with the case via a fundamental laser pulse alone~\cite{ISHIKAWA:PI-03,TAKAHASHI:DE-07}.
When the XUV field has the form of an attosecond pulse train a single quantum path can be selected to contribute to HHG and in this way allowing to manipulate the time-frequency properties of harmonics as well as to enhance a selected bandwidth of harmonics~\cite{SCHAFER:PC-04,GAARDE:DE-05,FARIA:HO-07}. Tuning the XUV field to a resonance between a core and valence state can lead to the emergence of a second plateau that is shifted to higher energies by the former resonance energy with respect to the first plateau~\cite{Buth:HO-up}.
In the relativistic regime the XUV/x-ray assistance can be employed to overcome the relativistic drift motion~\cite{KLAIBER:CH-08}. Thereby, the XUV frequency requires to exceed the ionization energy to
liberate the electron with a single photon and to deliver a significant initial momentum to the freed electron. This way, the electron can obtain sufficient momentum in the direction opposite to the laser propagation direction to compensate for subsequent drift motion and return to the atomic core, recombine and emit harmonics after the excursion in the relativistically strong laser field.
The medium is a gas of multiply charged ions with an ionization energy large enough to withstand the strong optical laser field.
How much the XUV assisted setup for relativistic HHG favors phase-matching needs investigation.
Formalisms describing macroscopic effects due to ionization and phase matching in HHG
are restricted to the non-relativistic regime, for a recent review see, e.g.,~\cite{GAARDE:MA-08}.
In this paper, we investigate the feasibility of phase-matched emission and the macroscopic yield of harmonics in the relativistic regime of the x-ray assisted HHG setup in a strong IR laser field.
Generally, the efficiency of HHG is rather small even in the nonrelativistic regime due to the wave packet spreading. In the relativistic regime, the single-atom HHG emission rate continues to decrease even when the relativistic drift is compensated~\cite{KOHLER:HX-11}. Thus, a large phase-matching volume is crucial in order to achieve a significant HHG yield. Furthermore, the large ponderomotive potential is likely to result in rapid phase changes if ions emit under different conditions. For generating relativistic harmonics both challenges have to be met: circumventing the drift and having the setup stable against phase changes. Our presented setup overcomes both issues and renders a measurable HHG yield in the relativistic regime possible.
The structure of the paper is the following. In Sec.~\ref{sec:relHHG_macr}, the theory of macroscopic HHG is presented applicable for any field geometry in the relativistic regime. In Sec.~\ref{sec:relxray}, the developed theory is applied to calculate the macroscopic HHG yield for the setup of x-ray assisted relativistic HHG. Our conclusion is presented in Sec.~\ref{sec:conclusion}.
\section{Macroscopic model for relativistic HHG}\label{sec:relHHG_macr}
\subsection{Macroscopic HHG yield}\label{sec:relHHG_macr1}
In this section, our model is presented for the calculation of
the harmonic spectrum from a macroscopic gas target suitable for relativistic laser intensities. In the non-relativistic regime, the standard approach for the calculation of the macroscopic HHG response incorporates the single-atom contribution via the time-dependent dipole moment~\cite{HUILLIER:TA-91,PRIORI:PM-00,GAARDE:MA-08}. However, employing the dipole moment for the radiation response assumes that an emitted harmonic wavelength is much longer than the spatial extensions of the emitter. This approach fails for sufficiently small wavelengths because then the retardation between different points of the emitting wave packet becomes important~\cite{HERNANDEZ:HP-91,MOCKEN:RS-05}. Our approach uses the complete current density distribution of each atom rather than the dipole moment. Retardation between different emission points within the distribution is taken into account by a phase factor.
The link between the microscopic (atomic) current density $\mathbf{j}$ and the macroscopically emitted harmonic electric field $\mathbf{E}\I{H}$ is obtained from Maxwell's equations similar to the non-relativistic approaches~\cite{HUILLIER:TA-91,PRIORI:PM-00,GAARDE:MA-08}. The Fourier component of the emitted harmonic electric field from a gas target is given by~\cite{JACKSON:CE-98}
\begin{equation}\label{eq:EH_general}
\tilde{\mathbf{E}}\I{H}(\mathbf{x}',\omega\I{H})={\rm i} \frac{\omega\I{H}}{c^2} \int\mathrm{d}^3 \textbf{x}
\frac{\mathbf{n}'\times(\tilde{\mathbf{j}}(\mathbf{x},\omega\I{H})\times \mathbf{n}')}{R}
\textrm{e}^{{\rm i} k\I{H} R}\, ,
\end{equation}
where $\tilde{\mathbf{j}}(\mathbf{x},\omega\I{H})$ is the Fourier component of the current density,
$\omega\I{H}$ is the frequency of the emitted harmonic light, $\mathbf{k}\I{H}=k\I{H}\mathbf{n}'$ the wave vector, $R=\vert \mathbf{x}-\mathbf{x'}\vert$ the distance between the emission and observation point, $\mathbf{x'}$ the coordinate of the observation point and
$\mathbf{n}'=\mathbf{x}'/|\mathbf{x}'|$ the unit vector in the observation direction (see Fig.~\ref{HHG_geometry}).
Absorption of the harmonic photons is neglected~\cite{PRIORI:PM-00} because their energy is much higher than the largest atomic transition energy. The current density $\mathbf j$ is exclusively determined by the HHG process. For the evaluation of Eq.~\eqref{eq:EH_general}, we restrict ourselves
to the far-field zone which is sufficient for calculating the overall HHG photon yield. The far-field zone is determined by the conditions that the distances from the emitters to the observation point are larger than the wavelength of the emitted radiation as well as the size $a$ of the emitting region ($k\I{H} R\gg1$ and $R\gg a$). We thus can expand Eq.~\eqref{eq:EH_general} over a small parameter $a/R$ using
$R=\vert \mathbf{x}-\mathbf{x'} \vert \simeq |\mathbf{x}'| - \mathbf{x}\cdot \mathbf{n}'$~\cite{JACKSON:CE-98}. When inserting this expression into Eq.~\eqref{eq:EH_general}, the exponential function splits up into two parts. One term contains $\textbf{x}'$ and is thus a general phase factor depending on the constant observation point that can be separated: $\tilde{\mathbf{E}}\I{H}(\mathbf{x}',\omega\I{H})=\textrm{e}^{{\rm i} k\I{H} |\textbf{x}'|}\tilde{\mathbf{E}}\I{H,0}(\mathbf{n}',\omega\I{H})$. In the following we consider $\tilde{\mathbf{E}}\I{H,0}(\mathbf{n}',\omega\I{H})$ only and find
\begin{equation}
\tilde{\mathbf{E}}\I{H,0}(\mathbf{n}',\omega\I{H})={\rm i} \frac{\omega\I{H}}{c^2R}\int\mathrm{d}^3 \mathbf{x}\, \mathbf{n}'\times(\tilde{\mathbf{j}}(\mathbf{x},\omega\I{H})\times \mathbf{n}') \textrm{e}^{-{\rm i} \mathbf{k}\I{H}\cdot \mathbf{x}} \label{E01}\, .
\end{equation}
The total current density distribution consists of a sum of the current densities of the single atoms $\mathbf{j}\I{a}(\mathbf{x}\I{a},\mathbf{x},t)$ with positions $\mathbf{x}\I{a}$:
\begin{equation}
\tilde{\mathbf{j}}(\mathbf{x},\omega\I{H})=\int \mathrm{d}^3\textbf{x}\I{a} \rho(\textbf{x}\I{a})\int \mathrm{d} t\, \mathbf{j}\I{a}(\mathbf{x}\I{a},\mathbf{x},t) \textrm{e}^{{\rm i} \omega\I{H} t} \label{curint}
\end{equation}
where $\rho(\textbf{x}\I{a})$ is the atomic number density.
\begin{figure}[b]
\begin{center}
\includegraphics[width=0.48\textwidth]{kohlerfig1}
\caption{(color online). Geometry of the medium and detector including the definitions of the coordinates. The dashed lines denote the divergence angle of the harmonic radiation. The box on the right schematically shows the measured angular distribution.} \label{HHG_geometry}
\end{center}
\end{figure}
Inserting Eq.~\eqref{curint} into~Eq.~\eqref{E01} yields the final expression for the macroscopically emitted harmonic field
\begin{eqnarray}
\tilde{\mathbf{E}}\I{H,0}&&\!\!\!\!\!\!(\mathbf{n}', \omega\I{H})\nonumber\\
&&\!\!\!\!\!\!={\rm i} \frac{\omega\I{H}}{Rc^2} \int \mathrm{d}^3\textbf{x}\I{a} \rho(\textbf{x}\I{a}) \int \mathrm{d}^3 \textbf{x} \nonumber\\ &&\int \mathrm{d} t\,\mathbf{n}'\times(\mathbf{j}\I{a}(\mathbf{x}\I{a},\mathbf{x},t)\times \mathbf{n}') \textrm{e}^{-{\rm i} k\I{H} \mathbf{x}\cdot{\mathbf{n}'}+{\rm i} \omega\I{H} t} \label{E02}\nonumber\\
&&\!\!\!\!\!\!= {\rm i} \frac{\omega\I{H}}{Rc^2} \int \mathrm{d}^3\mathbf{x}\I{a}\, \rho(\textbf{x}\I{a})\, \mathbf{n}'\times(\tilde{\mathbf{j}}\I{a}(\mathbf{x}\I{a},\mathbf{n}',\omega\I{H})\times \mathbf{n}')\label{phase_match_integral}\, ,
\end{eqnarray}
where $\tilde{\mathbf{j}}\I{a}(\mathbf{x}\I{a},\mathbf{n}',\omega\I{H})\equiv \int \mathrm{d}^3 \textbf{x}\int\mathrm{d} t\mathbf{j}\I{a}(\mathbf{x}\I{a},\mathbf{x},t)\exp\{-{\rm i} k\I{H} \mathbf{x}\cdot{\mathbf{n}'}+{\rm i} \omega\I{H} t\}$.
Note that the combination of outer products in Eq.~\eqref{phase_match_integral} prevents mathematically from emission in the direction of the current density vector.
However, due to phase-matching, the macroscopic emission is in many cases mainly along the propagation direction of the laser $\mathbf{n}'\approx\hat{\mathbf z}$, (see Fig.~\ref{HHG_geometry}). When the laser is linearly polarized in $\hat{\mathbf x}$ direction, the current density vector is parallel to the $\hat{\mathbf x}-\hat{\mathbf z}$ plane. Thus, we can approximate
$\mathbf{n}'\times(\tilde{\mathbf{j}}\times\mathbf{n}')=\tilde{\mathbf{j}}-(\mathbf{n}'\cdot\tilde{\mathbf{j}})\mathbf{n}'\approx \tilde{j\I{x}} \hat{\mathbf x}$
and can restrict ourselves to the $\hat{\mathbf x}$ component
\begin{eqnarray}
\tilde{E}\I{H,0,x}(\mathbf{n}', \omega\I{H})
= {\rm i} \frac{\omega\I{H}}{Rc^2} \int \mathrm{d}^3\textbf{x}\I{a}\,\rho(\textbf{x}\I{a}) \tilde{j}\I{a,x}(\mathbf{x}\I{a},\mathbf{n}',\omega\I{H})\label{phase_match_integralEx}
\end{eqnarray}
to describe emission in this case.
The overall emitted energy can be obtained via integrating the Poynting vector $\mathbf{S}(\mathbf{r},t)=\frac{c}{4\pi}\mathbf{E}^2\I{H}(\mathbf x', t)\mathbf{n}'$ overall emission directions in the far field
\begin{equation}
W=\frac{c R^2}{4\pi}\int \mathrm{d} t \int \mathrm{d} \Omega\, \mathbf{E}\I{H}^2(\mathbf{x},t) \label{W_integral0}\, .
\end{equation}
By inserting
\begin{equation}
\mathbf{E}\I{H}(\mathbf{x}',t)=\frac{1}{2\pi}\int_{-\infty}^{\infty} \mathrm{d} \omega\I{H} \textrm{e}^{-{\rm i} \omega\I{H} t}\tilde{\mathbf{E}}\I{H}(\mathbf{n}',\omega\I{H})\,
\end{equation}
into ~Eq.~\eqref{W_integral0}, the energy can be calculated via an integration over the spectrum~\cite{LANDAU:CF-62}:
\begin{equation}
W=\frac{c R^2}{(2\pi)^2} \int \mathrm{d} \Omega \int_{0}^{\infty} \mathrm{d} \omega\I{H} \vert\tilde{\mathbf{E}}\I{H,0}(\mathbf{n}',\omega\I{H}) \vert^2.\label{W_integral}
\end{equation}
The emitted spectral photon number per solid angle from Eq.~\eqref{W_integral} is given by:
\begin{equation}
\frac{\mathrm{d} N}{\mathrm{d} \omega\I{H} \mathrm{d}\Omega}=\frac{c R^2}{(2\pi)^2\omega\I{H}} \vert\tilde{\mathbf{E}}\I{H,0}(\mathbf{n}',\omega\I{H}) \vert^2. \label{eq:spectr-phot-num}
\end{equation}
\subsection{Single-atom current density} \label{quantum_current}
In Sec.~\ref{sec:relHHG_macr1}, the macroscopic HHG yield has been calculated via classical electrodynamics. Since the expression Eq.~\eqref{phase_match_integralEx} for the emission field relies on the current densities of a single atom in the gas, we continue to derive the single-atom current density quantum mechanically in the relativistic regime via the Klein-Gordon equation. The Klein-Gordon current density of a particular atom at position $\mathbf{x}\I{a}$ in the laser field is given by~\cite{BJORKEN:RQ-64}
\begin{equation}\label{eq:KG-current}
\mathbf{j}\I{a}(\mathbf{x}\I{a},\mathbf{x},t)=\Bigl(\Psi^*_{\mathbf{x}\I{a}}(x)\hat{\mathbf{j}}\Psi_{\mathbf{x}\I{a}}(x)+\Psi_{\mathbf{x}\I{a}}(x)\hat{\mathbf{j}}^* \Psi^*_{\mathbf{x}\I{a}}(x)\Bigr)
\end{equation}
where $ \hat{\mathbf{j}}=\hat{\mathbf{p}}+\mathbf{A\I{L}}(x)/c$ and $\Psi_{\mathbf{x}\I{a}}(x)$ is the solution of the Klein-Gordon equation when the binding potential is centered around $\mathbf{x}\I{a}$. In the following, the time-space coordinate is $x=(ct,\mathbf{x})$, the wave four-vector of the laser field $k_L=(\omega\I{L}/c,\mathbf k\I{L})$, and the metric tensor $g_{\mu \nu}=\textrm{diag}(1,-1,-1,-1)$.
By a Fourier transformation of Eq.~\eqref{eq:KG-current} and partial integration, the spectral current is obtained:
\begin{eqnarray}
&&\tilde{\mathbf{j}}\I{a}(\mathbf{x}\I{a},\omega\I{H},\mathbf{n}') \nonumber\\
&=&\frac{1}{c}\int \mathrm{d}^4 x \, \textrm{e}^{{\rm i} k\I{H} x} \bigl(\Psi^*_{\mathbf{x}\I{a}}(x)\hat{\mathbf{j}}\Psi_{\mathbf{x}\I{a}}(x) +\Psi_{\mathbf{x}\I{a}}(x)\hat{\mathbf{j}}^* \Psi^*_{\mathbf{x}\I{a}}(x)\bigr) \nonumber \\
&=& \frac{2}{c}\int \mathrm{d}^4 x \, \textrm{e}^{{\rm i} k\I{H} x} \Psi_{\mathbf{x}\I{a}}^*(x)(\hat{\mathbf{j}}-\tfrac{1}{2} \mathbf{k}\I{H})\Psi_{\mathbf{x}\I{a}}(x) \label{eq:KG-jw}
\end{eqnarray}
We calculate the electron wave function in the field of the laser and the
ionic core by means of the strong-field approximation (SFA)~\cite{REISS:CK-90,REISS:RS-90}
\begin{equation}\label{eq:SFA_WF}
\Psi_{\mathbf{x}\I{a}}(x)=\phi(\mathbf{x}-\mathbf{x}\I{a},t)+\int d^4 x' G\I{L}(x,x')V\I{I}(x')\phi(\mathbf{x'}-\mathbf{x}\I{a},t')
\end{equation}
where $V\I{I}=2{\rm i}(\textbf{A}\I{L}(\eta)/c)\nabla -\textbf{A}\I{L}^2(\eta)/c^2)$
is the term in the Hamiltonian describing the electron interaction with the
laser field with the vector-potential $\textbf{A}_L(\eta)$ and phase $\eta=
k\I{L}\cdot x$.
The Volkov propagator $G\I{L}(x,x')$ in a plane wave laser field is given by~\cite{MILOSEVIC:UH-00,MILOSEVIC:UH-02}
\begin{equation}\label{eq:relHH_Volk_prop}
G\I{L}(x,x')=-{\rm i}\,\theta(t-t')\int
\frac{c \,\mathrm{d}^3\mathbf{q} }{2\varepsilon_{\mathbf{q}} (2\pi)^3}
\exp\big[-{\rm i}\,S\I{L}(x,x')\big]\\
\end{equation}
with the classical action of an electron in the laser field
\begin{equation}\label{eq:SL}
S\I{L}(x,x')=q\cdot\left(x-x'\right)
+\int^{\eta}_{\eta'}\mathrm{d}\tilde{\eta}\left[\frac{\left(\mathbf{q}+
\mathbf{A}_L(\tilde{\eta})/2c\right)\cdot \mathbf{A}\I{L}(\tilde{\eta})/c}
{k\I{L}\cdot q}\right]\, ,
\end{equation}
the energy-momentum four-vector $q=(\varepsilon_{\mathbf{q}}/c,\mathbf{q})$, the
energy $\varepsilon_{\mathbf{q}}=\sqrt{c^2\mathbf{q}^2+c^4}$.
Here $\phi(x)=\frac{\phi_0(\mathbf{x})c}{\sqrt{2(c^2-I\I{p})}}\exp\{-{\rm i}[(c^2-I\I{p})t+\mathbf{x}\cdot\mathbf{A}\I{L}/c]\}$, where $\phi_0(\mathbf{x})$ is the nonrelativistic ground-state wave function.
Inserting~Eq.~\eqref{eq:SFA_WF} into~Eq.~\eqref{eq:KG-jw} and applying the
usual assumptions~\cite{Lewenstein:HH-94,MILOSEVIC:UH-02} of neglecting bound--bound and continuum--continuum transitions and the time--inverted process, we obtain
\begin{eqnarray}
&&\!\!\!\!\!\!\!\!\!\!\!\!\tilde{\mathbf{j}}\I{a}(\mathbf{x}\I{a},\omega\I{H},\mathbf{n}')\nonumber\\
&&\approx\frac{2}{c} \int \mathrm{d}^4 x\int \mathrm{d}^4 x'\textrm{e}^{{\rm i} k\I{H}\cdot x}\phi^*(\mathbf{x}-\mathbf{x}\I{a},t)(\hat{\mathbf{j}}-\tfrac{1}{2} \mathbf{k}\I{H}) \nonumber\\
&&\,\,\,\,\times G\I{L}(x,x') V\I{I}(x')\phi(\mathbf{x'}-\mathbf{x}\I{a},t')\, .\label{j_expr_L}
\end{eqnarray}
\subsection{Electron wave function in a distorted plane wave laser field} \label{sec:eikonal}
The Volkov propagator Eq.~\eqref{eq:relHH_Volk_prop} describes the evolution of the wave function of an electron in a plane wave laser field with vector potential $\mathbf{A}\I{L}(\eta)$. However, in many practical situations this assumption on the laser field is not met,
in particular, when a focused laser field or multiple laser beams are applied~\cite{PEATROSS:SZ-97,COHEN:GA-99,SERRAT:SE-10}, or when the dispersion distorts the laser pulse. For this reason, we need to find the electron wave function in an external laser field where the vector potential depends not only on phase $\eta$ but also on the position $z$ along propagation.
The deviation of the laser field from the plane wave form is assumed to be a perturbation, so that the total vector potential reads
\begin{equation}
\mathbf{A}(\eta,\mathbf{x})=\mathbf{A}\I{L}(\eta)+\mathbf{A}\I{P}(\eta,\mathbf{x}),
\label{A}
\end{equation}
with $\vert\mathbf{A}\I{P}(\eta,\mathbf{x})\vert\ll \vert\mathbf{A}\I{L}(\eta)\vert$. We
find the solution of the Klein-Gordon equation in the field of Eq.~\eqref{A}
using the eikonal approximation, see e.g.~\cite{AVETISSIAN:GE-99,SMIRNOVA:EI-08},
in which the impact of the perturbation onto the wave function is taken into account by an expansion of the wave function phase.
The electron wave function is described by the Klein-Gordon equation
\begin{equation}\label{eq:rel_KGeik}
(\partial^\mu\partial_\mu +c^2)\Psi(x)=V\Psi(x)\, ,
\end{equation}
with $V=2{\rm i}(\mathbf{A}(\eta)/c)\nabla-A^2(\eta)/c^2$. In order to solve Eq.~\eqref{eq:rel_KGeik}, the ansatz
\begin{equation}\label{eq:rel_ansatzeik}
\Psi(x)=\frac{\sqrt{c}}{\sqrt{2(2\pi)^3\varepsilon_{\mathbf{p}}}}\exp\{-{\rm i}[S\I{L}
(x,x\I{0})+S\I{P}(x,x\I{0})]\} \, .
\end{equation}
is employed. Here, $S\I{L}(x,x\I{0})$ is defined in Eq.~\eqref{eq:SL}, with an
arbitrary constant $x_0$, and, consequently, $\exp(-{\rm i} S\I{L}(x,x\I{0}))$
solves the unperturbed equation \eqref{eq:rel_KGeik} with
$\mathbf{A}(\eta,z)=\mathbf{A}\I{L}(\eta)$.
Inserting the ansatz Eq.~\eqref{eq:rel_ansatzeik} into Eq.~\eqref{eq:rel_KGeik},
one finds
\begin{eqnarray}
&&\frac{1}{c^2}[-{\rm i}\partial_t^2 S\I{P} -2\partial_t S\I{L}\partial_t S\I{P} -(\partial_t S\I{P})^2 ]\nonumber\\
&&\;\;\;\;\;-[-{\rm i}\nabla^2 S\I{P}-2\nabla S\I{L}\nabla S\I{P}-(\nabla S\I{P})^2]\nonumber\\
&&\;\;\;\;\;\;\;\;\;\;=2\frac{\mathbf{A}\I{L}}{c}\nabla S\I{P}+2\frac{\mathbf{A}\I{P}}{c}\nabla(S\I{L}+S\I{P})
\nonumber\\
&&\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;-\frac{A\I{P}^2}{c^2}-2\frac{A\I{L} A\I{P}}{c^2}.
\end{eqnarray}
In the applied approximation, $|\nabla S\I{p}|\ll |\textbf{A}\I{P}/c|$ and
$|\partial_t S\I{P}|\ll |\textbf{A}_P|$, which allows to neglect the $(\nabla
S\I{P})^2$ and $ (\partial_t S\I{P})^2$ terms. When additionally $\xi\equiv
E\I{L}/c\omega\I{L}\gg \omega\I{L}/c^2$, one has $|\nabla^2 S\I{p}|\ll
|(\textbf{A}_P/c)\nabla S\I{p}|$ and $|\partial_t^2 S\I{p}|\ll |\partial_t
S\I{L} \partial_t S\I{p}|$. Then, $\nabla^2 S\I{P}$, $\partial_t^2S\I{P}$ terms
also can be neglected, yielding
\begin{eqnarray}
&& 2\frac{A\I{L} A\I{P}}{c^2}+\frac{A\I{P}^2}{c^2}+2\frac{\mathbf{A}\I{P}}{c}\nabla S\I{L} \label{Sp}\\
&&=\frac{2}{c^2}\partial_t S\I{L}\partial_t S\I{P}+\Bigl[-2\nabla S\I{L}+\frac{2}{c}(\mathbf{A}\I{L}+\mathbf{A}\I{P})\Bigr]\nabla S\I{P}\, .\nonumber
\end{eqnarray}
The equations for the characteristics of the first order partial differential equation \eqref{Sp} are
\begin{eqnarray}
\frac{\partial t(u)}{\partial u}&=&\frac{2}{c^2}\partial_t S\I{L}\label{eq:eik2t}\\
\frac{\partial {\bf r}(u)}{\partial u}&=&-2\nabla S\I{L}+\frac{2}{c}(\mathbf{A}\I{L}+\mathbf{A}\I{P}).\label{eq:eik3r}
\end{eqnarray}
It follows from~Eq.~\eqref{eq:eik2t} and~Eq.~\eqref{eq:eik3r} that
\begin{equation}
-2k\I{L}\cdot p=\mathbf{k}\I{L}\frac{\partial \mathbf{r}}{\partial u}-\omega\I{L} \frac{\partial t}{\partial u}
\end{equation}
and
\begin{equation}
u=-(\mathbf{k}\I{L}\mathbf{r}-\omega\I{L} t)/(2k\I{L}\cdot p)=\eta/(2k\I{L}\cdot p)\, .\label{eq:defu}
\end{equation}
Integrating~Eq.~\eqref{Sp} and employing~Eq.~\eqref{eq:defu} we derive
\begin{eqnarray}
S\I{P} (x,x\I{0})\!\!\!&=&\!\!\!\int_{\eta_0}^\eta \!\!\!\!\mathrm{d}\tilde{\eta} \frac{1}{k\I{L}\cdot p}\Bigl[\nabla S\I{L}+\frac{\mathbf{A}\I{L}(\tilde{\eta})}{c}+\frac{\mathbf{A}\I{P}(\tilde{\eta},\tilde{\mathbf{x}}\I{L}(\mathbf{x},\tilde{\eta},\eta))}{2c}\Bigr]\nonumber\\
&&\;\;\;\;\times\mathbf{A}\I{P}(\tilde{\eta},\tilde{\mathbf{x}}\I{L}(\mathbf{x},\tilde{\eta},\eta))/c\,
\end{eqnarray}
with
\begin{eqnarray}
\tilde{\mathbf{x}}\I{L}(\mathbf{x},\tilde{\eta},\eta)&=& \mathbf{x}-\int_{\tilde{\eta}}^\eta \mathrm{d}\eta'\frac{1}{k\I{L}\cdot p}\Bigg\{\mathbf{p}+\frac{\mathbf{A}\I{L}(\eta')}{c}+\frac{\mathbf{k}\I{L}}{k\I{L}\cdot p} \nonumber\\
&&\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\times\Bigl[p_x+\frac{A\I{L}(\eta')}{2c}\Bigr]\frac{A\I{L}(\eta')}{c}\Bigg\} \label{ztraj}\, .
\end{eqnarray}
The integral in Eq.~\eqref{ztraj} can be omitted in the case if the z-dependence of $A\I{P}$ along the classical trajectory of the particle in the laser field is negligible.
For the total phase of the wave function, we therefore find
\begin{eqnarray}
S\I{T}(x,x\I{0})&=&S\I{L}(x,x\I{0})+S\I{P}(x,x\I{0})\nonumber\\
&=&\,p\cdot (x-x_0) \label{eq:rel_pertS}\\ \nonumber
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&+&\int_{\eta_0}^\eta \mathrm{d}\tilde{\eta}\frac{\left[\mathbf{p}+\frac{\mathbf{A}(\tilde{\eta},\tilde{\mathbf{x}}\I{L}(\mathbf{x},\tilde{\eta},\eta))}{2c}\right]\frac{\mathbf{A}(\tilde{\eta},\tilde{\mathbf{x}}\I{L}(\mathbf{x},\tilde{\eta},\eta))}{c}}{k\I{L}\cdot p}\, ,
\end{eqnarray}
where $\eta_0$ and $x_0$ are arbitrary constants.
We derive the propagator as follows
\begin{equation}\label{eq:relHH_prop}
G\I{T}(x,x')\approx-{\rm i}\,\theta(t-t')\int
\frac{c \,\mathrm{d}^3\mathbf{q} }{2\varepsilon_{\mathbf{q}} (2\pi)^3}
\exp\big[-{\rm i}\,\Phi(x,x')\big]\\ \,,
\end{equation}
where $\Phi(x,x')\equiv S\I{T}(x,x\I{0})- S\I{T}(x',x\I{0})$. The latter can be represented as
\begin{eqnarray}
\Phi(x,x')=S\I{T}(x,x')+ \Delta\I{T}(x,x')\,
\end{eqnarray}
with
\begin{eqnarray}
&& \Delta\I{T}(x,x')\nonumber\\
&=&\frac{1}{k\I{L}\cdot p}\int_{\eta_0}^\eta \mathrm{d}\tilde{\eta}\bigg\{\left[\mathbf{p}+\frac{\mathbf{A}(\tilde{\eta},\tilde{\mathbf{x}}\I{L}(\mathbf{x},\tilde{\eta},\eta))}{2c}\right]\frac{\mathbf{A}(\tilde{\eta},\tilde{\mathbf{x}}\I{L}(\mathbf{x},\tilde{\eta},\eta))}{c} \nonumber\\
&&- \left[\mathbf{p}+\frac{\mathbf{A}(\tilde{\eta},\tilde{\mathbf{x}}\I{L}(\mathbf{x},\tilde{\eta},\eta'))}{2c}\right]\frac{\mathbf{A}(\tilde{\eta},\tilde{\mathbf{x}}\I{L}(\mathbf{x},\tilde{\eta},\eta'))}{c} \bigg\}\, .\label{eq:Delta-eikonal}
\end{eqnarray}
The two terms in the integrand of Eq.~\eqref{eq:Delta-eikonal} deviate only in $\tilde{\mathbf{x}}\I{L}(\mathbf{x},\tilde{\eta},\eta)$ and $\tilde{\mathbf{x}}\I{L}(\mathbf{x},\tilde{\eta},\eta')$. Within the saddle-point approximation applied later, $\eta$ and $\eta'$ are the phase of recollision and ionization, respectively. Therefore, $\tilde{\mathbf{x}}\I{L}(\mathbf{x},\tilde{\eta},\eta)$ and $\tilde{\mathbf{x}}\I{L}(\mathbf{x},\tilde{\eta},\eta')$ differ only by the distance in space between ionization and recollision which is zero. Thus, we can omit $\Delta\I{T}(x,x')$.
\section{Relativistic phase-matched x-ray assisted HHG}\label{sec:relxray}
In this section, the theory developed in Sec.~\ref{sec:relHHG_macr} is employed to investigate macroscopic harmonic emission in the relativistic regime for a
HHG setup where an IR laser field of relativistic intensity is assisted by an x-ray field.
In this scheme, the x-ray frequency $\omega\I{X}$ exceeds the binding energy $I\I{p,x}$ of the electron and thus delivers an initial momentum to the freed electron which can balance the subsequent drift motion.
The laser alignment of the setup and an example of a classical trajectory that recollides are illustrated in Fig.~\ref{xray_setup2}. The weak counterpropagating IR field (brown) is important for a phase-matched macroscopic response and can be ignored when discussing the process for a single atom.
The HHG medium is a macroscopic gas of multiply-charged ions.
\begin{figure}[b]
\begin{center}
\includegraphics[width=0.45\textwidth]{kohlerfig2}
\caption{(color online). Geometry of the HHG process for a collinear alignment of the x-ray and laser field. The copropagating x-ray field (orange) has a frequency above the ionization energy to achieve drift compensation. The weak IR field (brown) is employed to accomplish phase-matching.} \label{xray_setup2}
\end{center}
\end{figure}
\subsection{Current density}\label{sec:current density}
We begin with adapting the general equation for the current density Eq.~\eqref{j_expr_L} to the present setup.
Presuming that for the realization of phase-matching additional weak fields $\mathbf{A}\I{P}(\eta,\mathbf{x})$
will be required, we employ the Green function $G\I{T}(x,x^{\prime})$ of Eq.~\eqref{eq:relHH_prop} instead of the Volkov propagator
$G\I{L}(x,x^{\prime})$ in the general expression for the current density of Eq.~\eqref{j_expr_L}:
\begin{eqnarray}
\tilde{j\I{a}}(\mathbf{x\I{a}},\omega\I{H},\mathbf{n}')\nonumber\\
&&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!=\frac{2}{c} \int \mathrm{d}^4 x\int \mathrm{d}^4 x'\textrm{e}^{{\rm i}\omega\I{H} t-{\rm i}\mathbf{k}_H\mathbf{r}}\phi^*(\mathbf{x}-\mathbf{x}\I{a})\nonumber\\
&&\!\!\!\!\!\!\!\!\!\!\!\!\times(\hat{\mathbf{j}}-\tfrac{1}{2} \mathbf{k}\I{H})G\I{T}(x,x')V\I{I}(x')\phi(\mathbf{x}'-\mathbf{x}\I{a}) \label{eq:current-int}\, .
\end{eqnarray}
In this expression, the x-ray field enters only into the potential
$V\I{I}(x)=2\,\mathbf{x}\cdot (\mathbf{E\I{X}}+\mathbf{E\I{L}})$ in the Klein-Gordon formalism.
Ignoring tunnel ionization by the IR laser field which is justified in the
considered setup, we drop the laser field in this term and approximate $V\I{I}(x) \approx 2\,\mathbf{x}\cdot
\mathbf{E\I{X}}\approx \mathbf{x}\cdot \mathbf{E}_{X0} \textrm{e}^{-{\rm i}(\omega\I{X} t-
\mathbf{k}\I{X}\cdot \mathbf{x})}$. In the last step, the exponential function with the positive argument is dropped because
it leads to an unphysical solution of the saddle point equations~\cite{FARIA:HO-07}.
Due to its negligible ponderomotive potential, $\mathbf{E\I{X}}$ can be neglected for the continuum propagation of the electron and, thus, it is neglected in the propagator.
The dependence of the current density Eq.~\eqref{eq:current-int} on the position of the atom $\mathbf{x}\I{a}$ is given by the bound wave functions $\phi(\mathbf{x}-\mathbf{x}\I{a},t)$. To separate out phase factors that highly oscillate with $\mathbf{x}\I{a}$, a coordinate transformation is applied: $\mathbf{\tilde{x}}=\mathbf{x}-\mathbf{x}\I{a}$. Thereafter, time integration is transformed to an integration over the laser phase: $\eta=\omega\I{L} t- \mathbf{k}\I{L}\cdot \mathbf{x}=\omega\I{L} t- \mathbf{k}\I{L}\cdot \mathbf{x}\I{a}- \mathbf{k}\I{L}\cdot \mathbf{\tilde{x}}$.
Finally, we obtain an expression for the current density that can be evaluated within the saddle-point approximation:
\begin{eqnarray}
&&\tilde{\mathbf{j}}\I{a}(\mathbf{x}\I{a},\omega\I{H},\mathbf{n}') =\int^{\omega\I{L}T\I{P}}_{0}\mathrm{d}\eta\int^{\eta}_{-\infty}\mathrm{d}\eta'\int \mathrm{d}^3\mathbf{q}\,
m^j(\mathbf{q},\eta,\eta',\mathbf{x}\I{a}) \nonumber\\
&&\times \exp\left[-{\rm i}( \tilde{S}\I{P}(\mathbf{q},\eta,\eta',\mathbf{x}\I{a})+\frac{\omega\I{X}}{\omega\I{L}} \eta'-\frac{\omega\I{H}}{\omega\I{L}}\eta)\right]\nonumber \\
&&\times \exp\left[{\rm i}(\frac{\omega\I{H}}{\omega\I{L}}\mathbf{k}\I{L}-\mathbf{k}\I{H}+\mathbf{k}\I{X}-\frac{\omega\I{X}}{\omega\I{L}}\mathbf{k}\I{L})\mathbf{x}\I{a}\right]
\label{current_dens_ass}
\end{eqnarray}
where $T\I{P}$ is the laser pulse duration and
\begin{eqnarray}
&&m^j(\mathbf{q},\eta,\eta',\mathbf{x}\I{a})=-{\rm i}\frac{c^2 (\mathbf{q}+\frac{\mathbf{A}(\eta,\mathbf{x}\I{a})}{c}-\frac{{\bf
k}\I{L}}{\omega\I{L}}\varepsilon_{{\bf q}} -\tfrac{1}{2} \mathbf{k}\I{H} )}
{\varepsilon_\mathbf{q}\omega\I{L}^2}
\nonumber\\&&\times \left<0|\mathbf{q}+\frac{\mathbf{A}(\eta,\mathbf{x}\I{a})}{c}-\frac{{\bf k}\I{L}}{\omega\I{L}}(\varepsilon_{{\bf q}}+I\I{p,x}-c^2)+\frac{\omega\I{H}}{\omega\I{L}}\mathbf{k}\I{L}-\mathbf{k}\I{H}\right> \nonumber \\
&&\times \left<\mathbf{q}+\frac{\mathbf{A}(\eta^{\prime },\mathbf{x}\I{a})}{c}-\frac{{\bf k}\I{L}}{\omega\I{L}}(\varepsilon_{{\bf q}}+I\I{p,x}-c^2)
\right.\nonumber\\ && \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;+ \left.
\frac{\omega\I{X}}{\omega\I{L}}\mathbf{k}\I{L}-\mathbf{k}\I{X}\left|\mathbf{x}\cdot \mathbf{E}_{X0}\right|0\right>\, ; \label{mh5}
\end{eqnarray}
Further,
\begin{equation}\label{eq:tildeSpert}
\tilde{S}_{\mathrm P}(\mathbf{p},\eta,\eta',\mathbf{x}\I{a})=
\int^{\eta}_{\eta'}
d\tilde{\eta}\left(\tilde{\varepsilon}_{\mathbf{q}}^{\mathrm P}
(\tilde{\eta},\tilde{\mathbf{x}}(\mathbf{x}\I{a},\tilde{\eta},\eta))-c^2+I\I{p,x}\right)/\omega\I{L}
\end{equation}
with the relativistic energy of the electron in the position dependent laser field given by
\begin{eqnarray}
\tilde{\varepsilon}_{\mathbf{q}}^{\mathrm P}(\eta,\mathbf{x})=\varepsilon_{\mathbf{q}}
+\frac{\omega\I{L}}{k\I{L}\cdot q}
\left(\mathbf{q}+\frac{\mathbf{A}(\eta,\mathbf{x})}{2c}\right)\cdot\frac{\mathbf{A}
(\eta,\mathbf{x})}{c}\, .
\end{eqnarray}
The wave vector of the laser is $\mathbf{k}\I{L}=n\I{ref}\frac{\omega\I{L}}{c} \hat{\mathbf{e}}\I{z}$ and
\begin{equation}
n\I{ref}=\sqrt{1-\frac{\omega\I{p}^2}{\omega\I{L}^2}}\label{eq:plasma_n}
\end{equation}
is the refractive index of the plasma, with the plasma frequency $\omega\I{p}=\sqrt{4\pi Z\rho}$ and the ion charge number $Z$.
This way, we take into account the change of the phase velocity caused by the free electrons but ignore pulse deformation.
Further, we restrict ourselves to a weakly focused laser field which can be treated as a plane wave and ignore the transversal variation of the vector potential with respect to the propagation direction.
The propagation direction and the frequency of the assisting x-ray field
$\mathbf{E\I{X}}$ have to be chosen in such a way to facilitate the phase-matching
and to counteract the relativistic drift effectively. Let us first consider the
feasibility of phase-matching by analyzing the harmonic emission phase
of different ions in the medium via Eq.~\eqref{current_dens_ass}. Generally,
the vector potential consists of two terms as in section~\ref{sec:eikonal}
\begin{equation}\label{eq:totalA}
\mathbf{A}(\eta,\mathbf{x})=\mathbf{A}\I{L}(\eta)+\mathbf{A}\I{P}(\eta,\mathbf{x}).
\end{equation}
To understand the most favorable condition for phase-matching, let us assume for a moment that there are no additional fields for facilitating quasi-phase matching and ignore deformation of the laser pulse. Then, we can omit $\mathbf{A}\I{P}(\eta,\mathbf{x})$ in~Eq.~\eqref{eq:totalA}
[a modified setup where the additional $\mathbf{A}\I{P}(\eta,\mathbf{x})$ field is taken into account will be considered in Sec. \ref{sec:macro}].
In this case, we have no $\mathbf{x}_A$ dependence of
$m^j(\mathbf{q},\eta',\eta'',\mathbf{x}_A)$ and $\tilde{S}\I{P}(\mathbf{q},\eta',\eta'',\mathbf{x}_A)$.
The only contribution to the variation of the harmonic emission phase along the medium comes from the exponential function $\exp({\rm i} \Delta \mathbf{k}\, \textbf{r}_a)$ in~Eq.~\eqref{current_dens_ass} determining the coherence length $l_{coh}=\pi/|\Delta \mathbf{k}|$, with the phase-mismatching wave-vector
\begin{equation}
\Delta\mathbf{ k}=\frac{\omega\I{H}}{\omega\I{L}}\mathbf{k}\I{L}-\mathbf{k}\I{H}+\mathbf{k}\I{X}-\frac{\omega\I{X}}{\omega\I{L}}\mathbf{k}\I{L}\, ,\label{eq:xrayass_mismatch}
\end{equation}
caused by the refractive index difference between the fields. The phase-mismatch Eq.~\eqref{eq:xrayass_mismatch} can be split up in two parts. One is due to the phase-mismatch between the harmonics and the laser and the other one due to the phase-mismatch between the ionizing x rays and the laser. Interestingly, due to the different signs both can partially cancel out. This fact also holds true for this setup in the non-relativistic regime. Note that $\frac{\omega\I{H}}{\omega\I{L}}\mathbf{k}\I{L}$ and $\mathbf{k}\I{H}$ are much larger than the other two terms left in~Eq.~\eqref{eq:xrayass_mismatch} for the case $\omega\I{H}\gg\omega\I{X}$. The phase mismatch is smallest in the case of a collinear alignment of the x-ray and laser fields $\mathbf{k}\I{H}\parallel \mathbf{k}\I{L}$ (see Fig.~\ref{xray_setup2}) with propagation in the z direction.
Accordingly, only the x component of the spectral current density of Eq.~\eqref{current_dens_ass} will contribute to the harmonic emission in the z direction.
The drift compensation can be achieved with an appropriate choice of the x-ray frequency which is considered in Sec.~\ref{sec:relHHGass_saddlepoints}.
\subsection{Single-atom response}\label{sec:single-atom}
\subsubsection{Singe-atom HHG rate}
The single-atom photon emission rate per solid angle of the $n^{{\rm
th}}$ harmonic can be calculated from
\begin{eqnarray}
\frac{dw_n}{d\Omega}&=& \frac{1}{T\I{P}} \int_{n \omega\I{L}-\omega\I{L}}^{n \omega\I{L}+\omega\I{L}}\mathrm{d} \omega\I{H} \frac{\mathrm{d} N}{\mathrm{d} \omega\I{H} \mathrm{d}\Omega} \\ \nonumber
&=& \frac{c R^2}{(2\pi)^2\omega\I{H} T\I{P}} \int_{n \omega\I{L}-\omega\I{L}}^{n \omega\I{L}+\omega\I{L}}\mathrm{d} \omega\I{H} \vert\tilde{\mathbf{E}}\I{H,0}(\mathbf{n}',\omega\I{H}) \vert^2\, ,
\label{mhhg3a}
\end{eqnarray}
using Eq.~\eqref{phase_match_integralEx} for
$\tilde{\mathbf{E}}\I{H,0}(\mathbf{n}',\omega\I{H})$ and the density
$\rho(\textbf{x}\I{a})=\delta(\textbf{x}\I{a})$ which yields
\begin{equation}
\frac{dw_n}{d\Omega}= \frac{\omega\I{L}^2 \omega\I{H}}{(2 \pi c)^3 } \vert\tilde{\mathbf{j}}\I{a}(0,n \omega\I{L},\mathbf{n}') \vert^2\, ,
\label{mhhg4a}
\end{equation}
where in the expression for $\tilde{\mathbf{j}}\I{a}(0,n \omega\I{L},\mathbf{n}')$ given by Eq.~(\ref{current_dens_ass}), $T\I{P}$ is replaced by $\frac{2\pi}{\omega\I{L}}$ to confine the emission to one laser period [the rate of Eq.~(\ref{mhhg4a}) is identical to ones in Refs.~\cite{HATSAGORTSYAN:UFO-08,KLAIBER:CH-08}].
A typical HHG spectrum for the considered setup calculated within the
saddle-point approximation is displayed in Fig.~\ref{fig:single_atom_xray}~(a)
in blue for the set of parameters denoted in the caption of the figure. In this
paper we employ a zero-range potential~\cite{BECKER:ZP-94} to model the
binding potential.
Analytical expressions of the matrix elements which appear
in Eq.~\eqref{mh5} can be found in, e.g.,~\cite{MILOSEVIC:RL-02,KLAIBER:LO-07}.
The ionization potential $I\I{p,x}=8\U{a.u.}$ is chosen large enough such that tunnel ionization by the strong optical laser field does not lead to depletion of the bound wave function. We compare the spectrum obtained from the x-ray assisted setup (blue line) with the spectrum of a
conventional HHG setup where no x-ray field is present, calculated either fully
relativistically~\cite{KLAIBER:LO-07} (dashed black) or within the
dipole approximation (DA)~\cite{MILOSEVIC:RL-02} (gray).
The ionization matrix elements for the conventional HHG setups was multiplied by the factor
$\frac{2\sqrt{2}\kappa\I{t}^3}{\vert E\I{L}(t') \vert}$~\cite{MILOSEVIC:RT-02} to account for the underestimation of the tunneling rate when employing the zero-range potential. Here, $t'$ is the ionization time and $\kappa\I{t}=\sqrt{2I\I{p,t}}$, with $I\I{p,t}$ being the ionization potential in the case of the conventional setup.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.4\textwidth]{kohlerfig3a}
\vskip0.2cm
\includegraphics[width=0.4\textwidth]{kohlerfig3b}
\caption{(color online). (a) Single-atom emission probability for $E_0=2.5\U{a.u.}$, $I\I{p,x}=8\U{a.u.}$, $\omega\I{X}=14\U{a.u.}$ and $E\I{X}=0.65\U{a.u.}$ (blue, dark gray) and a conventional laser field ($E\I{X}=0$) with $E_0=2.5\U{a.u.}$ and $I\I{p,t}=4.8\U{a.u.}$ (dashed black) and the same configuration in the DA (gray). For the second configuration, $I\I{p,t}$ is chosen such that the average tunnel-ionization rate is the same as the single-photon ionization rate in the case before. (b) Separate HHG yields of the three contributing quasi-classical trajectories for the discussed setup [blue line in (a)]. The dotted red contribution (long trajectory) is suppressed because the drift for the trajectory is not completely compensated (as discussed in Sec.~\ref{sec:relHHGass_saddlepoints}). The short trajectory contributions (solid lines in dark blue, dark gray and orange, gray) are nearly identical.
} \label{fig:single_atom_xray}
\end{center}
\end{figure}
In order to have a fair comparison of the x-ray assisted setup with the conventional one, we have to choose the ionization potential of the conventional setup ($I\I{p,t}$) different from that of the x-ray assisted one ($I\I{p,x}$) such that the ionization rates of both setups were the same.
For this purpose, the ionization rate for the conventional setup is calculated from the Perelomov, Popov, Terent'ev (PPT) tunneling rate~\cite{PERELOMOV:IA-67,AMMOSOV:TR-86}, while the single-photon ionization rate from the zero-range potential is derived using the differential photoionization cross section in nonrelativisitic, dipole approximation and approximating the ionized electron wave function by a plane wave (in analogy to Ref.~\cite{bethe1968intermediate}):
\begin{eqnarray}
\frac{\mathrm{d} \sigma\I{x,1ph}}{\mathrm{d} \Omega}=\frac{p}{2\pi c \omega\I{X}}\vert\langle
\mathbf{p}\vert z\vert 0\rangle \vert^2=\frac{2^4 p^3\omega\I{X} \kappa\I{z}}{c}\frac{\cos^2 \theta}{(\kappa\I{x}^2+p^2)^4}\label{eq:diff_1-photon}
\end{eqnarray}
where $\theta$ is the angle between the electron momentum $\mathbf{p}$ and the polarization direction $\hat{\mathbf{z}}$ of the x-ray field. An integration over all emission angles yields the total cross section
\begin{equation}\label{eq:io-rate-zr}
\sigma_{\delta,1ph}=\frac{4 \kappa\I{x}\pi}{3c}\frac{(2\omega\I{X}-\kappa\I{x}^2)^{3/2}}{\omega\I{X}^3}\, ,
\end{equation}
where we used $p=\sqrt{2\omega\I{x}-\kappa\I{x}^2}$.
The main message of Fig.~\ref{fig:single_atom_xray}~(a) is that the relativistic
drift can be fully compensated in the x-ray assisted HHG setup
(in the case of the chosen x-ray frequency $\omega\I{X}-I\I{p,x}= 6$ a.u.,
the gray and blue curve are of comparable order, small suppression arises from the different spreading behavior). The yield of the considered setup is much higher than that for the conventional setup (dashed black), the latter being suppressed by the drift.
In the next section, we explain how the single-x-ray-photon ionization provides the necessary initial momentum for the electron (opposite to the IR laser propagation direction) to counteract the relativistic drift in the case when the x-rays propagates along the strong IR laser field. We discuss also the optimization of the applied x-ray frequency for the HHG process.
\subsubsection{Drift compensation and influence of x-ray frequency}\label{sec:relHHGass_saddlepoints}
The integration in~Eq.~\eqref{current_dens_ass} is carried out via the
saddle-point integration method~\cite{Arfken:MM-05,Lewenstein:HH-94}. This
means that instead of the integration we only need to sum the integrand over a
small number of saddle points for each energy $\omega\I{H}$. A saddle point
$(\eta,\eta',\mathbf{q})$ determines the ionization and recollision times and
the canonical momentum for the electron classical trajectory leading to the
harmonic energy under consideration. In general, they are complex expressing
non-classical dynamics during tunneling ionization.
For the parameters chosen above, 3 quantum paths (saddle points) contribute to the spectrum for each energy being equivalent to three classical trajectories that recollide with that energy. The separate contributions of each quantum path to the spectrum are shown in Fig.~\ref{fig:single_atom_xray}~(b). The two paths marked in blue and orange (solid lines) have nearly the same yield whereas the dotted red line is suppressed by several orders of magnitude. In the following we explain the reason for the difference and discuss the influence of the x-ray frequency on the dynamics.
In order to understand the number of contributing trajectories in Fig.~\ref{fig:single_atom_xray}~(b), we calculate the saddle-point solutions for different x-ray frequencies $\omega\I{X}$ for the harmonic emission at $50\U{keV}$ and show the ionization phase saddle point in Fig.~\ref{fig:xray_ass_wx-dep}~(a). For small initial energies $\omega\I{X}-I\I{p,x}$, two saddle points contribute to harmonic emission as in the usual case of HHG in a laser field only. Both saddle points, the long ($\mathrm{Re}\,\eta_2\approx -1.345$) and short ($\mathrm{Re}\,\eta_2\approx -1.115$) trajectory, are complex [their real part is shown in the graph] and their HHG amplitude is very tiny due to the missing drift compensation which is indicated by the complex value. When increasing $\omega\I{X}$, first the short trajectory and then the long trajectory split up into two parts. These branches are called uphill and downhill trajectories, respectively, because their initial momentum component along the laser polarization is either positive or negative~\cite{FARIA:HO-07}. After the splitting at about $\omega\I{X}-I\I{p,x}\approx4$ a.u. and $\omega\I{X}-I\I{p,x}\approx 8.5$ a.u., the respective ionization phase is purely real which indicates that the initial momentum is sufficient to compensate the subsequent relativistic drift. The short trajectory reaches drift compensation earlier because it spends less time in the continuum and, therefore, undergoes a smaller drift that requires compensation.
\begin{figure}[h!] \begin{center}
\includegraphics[width=0.45\textwidth]{kohlerfig4a}
\vskip0.2cm \includegraphics[width=0.45\textwidth]{kohlerfig4b}
\vskip0.2cm\includegraphics[width=0.45\textwidth]{kohlerfig4c}
\caption{(color online). (a) We show the ionization phase saddle points of the $50\U{keV}$ trajectory for different x-ray frequencies which is the same as ionization time in units of radian. The dashed line indicates the value chosen in Fig.~\ref{fig:single_atom_xray}. The dotted and solid lines belong to the long and short trajectories, respectivly. (b) displays the initial momentum direction for different initial energies $\omega\I{X}-I\I{p,x}$ [as indicated in a.u. next to the respective arrow] needed for the emission of a $50\U{keV}$ photon. The uper and lower branch correspond to the short up- and downhill trajectories, respectivly, where the color indication coincides with the one in (a). Note that the ionization phases $\eta_2$ are different for the up- and downhill trajectories. $\hat{\mathbf{x}}$ and $\hat{\mathbf{z}}$ are the propagation and polarization directions of the laser, respectively. (c) Differential ionization rate depending on the initial energy $\omega\I{X}-I\I{p,x}$ from Eq.~\eqref{eq:diff_1-photon}. The considered direction is in the initial momentum direction determined by the saddle point equations. From (b) we see that the z-component is approximately $p\I{z,c}=2.9\U{a.u.}$ and the x-component depends on $\omega\I{X}-I\I{p,x}$.} \label{fig:xray_ass_wx-dep}
\end{center}
\end{figure}
The dashed line in Fig.~\ref{fig:xray_ass_wx-dep}~(a) denotes the x-ray frequency $\omega\I{X}$ that was chosen in Fig.~\ref{fig:single_atom_xray}~(b). The short trajectory has two contributions (blue and orange) whereas the long trajectory (dotted) has only one contribution (red). The contribution of the long trajectory is suppressed by about 3 orders of magnitude compared to the short contributions which is visible from Fig.~\ref{fig:single_atom_xray}~(b). This is because the long trajectory spends more time in the continuum and experiences a larger relativistic drift which cannot be fully compensated for. In this case, the ionization saddle point is complex leading to a damping in the exponential function in the respective amplitude~Eq.~\eqref{current_dens_ass}. By increasing $\omega\I{X}$ above $\omega\I{X}-I\I{p}\approx8.5$ a.u., the drift compensation could also be achieved for the long trajectory and the dashed red contribution in the spectrum could be enhanced leading to a larger single-atom yield. However, only one of the trajectories can be phase-matched in many cases and the enhancement of the other trajectories would not be useful.
In Fig.~\ref{fig:xray_ass_wx-dep}~(b), the initial momentum vectors of the
ionized electron $\mathbf p(\eta^{\prime},\mathbf{q})=\mathbf{q}+\frac{\mathbf{A}(\eta^{\prime})}{c
}-\frac{{\bf k}}{\omega\I{L}}(\varepsilon_{{\bf q}}+I\I{p,x}-c^2)$, which
correspond to solutions of the saddle-point equations, are displayed for
different $\omega\I{X}$. When $\omega\I{X}-I\I{p,x}\approx4.2\U{a.u.}$, the momentum
required for drift compensation of the short trajectories is just reached [see
Fig.~\ref{fig:xray_ass_wx-dep}~(a)]. In this case, the initial momentum is
directed mainly along the z-direction [arrows marked with $4.2$ in
Fig.~\ref{fig:xray_ass_wx-dep}~(b)]. When $\omega\I{X}$ is increased, only the
$p\I{x}$ component changes; the $p\I{z}$ component approximately remains
constant because it is determined by the drift compensation condition. The
electrons with an appropriate initial momentum vector can be provided by the
x-ray single-photon photoionization because the latter happens with a large
angular distribution with a maximum around x-ray polarization direction as can
be seen from the differential ionization cross section
of~Eq.~\eqref{eq:diff_1-photon}. Because the HHG amplitude for each trajectory
contains the differential ionization cross section~Eq.~\eqref{eq:diff_1-photon},
the efficiency in each case depends on the scalar product between required
ionization direction $\mathbf p$ and x-ray field polarization direction. This
results in some freedom in choosing the direction of $\mathbf{E}\I{X}$. Only if $\mathbf
p$ and $\mathbf{E}\I{X}$ were close to perpendicular [$\theta\approx \pi/2$ in
Eq.~\eqref{eq:diff_1-photon}], the differential ionization probability would be
close to zero. For realization of phase-matching,
as it is shown above in Sec. \ref{sec:current density},
the collinear propagation of the laser and x-ray field is advantageous. For this case of a collinear alignment, we show the differential ionization probability $\mathrm{d} \sigma\I{x,1ph}/\mathrm{d} \Omega$ from Eq.~\eqref{eq:diff_1-photon} for different x-ray frequencies $\omega\I{X}$ in Fig.~\ref{fig:xray_ass_wx-dep}~(c). The emission angle of interest is estimated by the initial momentum $p=\sqrt{2 (\omega\I{X}-I\I{p,x})}$ and its z-component $p\I{z,c}=2.9\U{a.u.}$ taken from Fig.~\ref{fig:xray_ass_wx-dep}~(b) via $\sin \theta=p\I{z,c}/p$. For initial energies just above $4$ [e.g., 4.2 corresponding to the nearly horizontal vectors in Fig.~\ref{fig:xray_ass_wx-dep}~(b)], $\mathrm{d} \sigma\I{x,1ph}/\mathrm{d} \Omega$ is vanishing because the momentum direction and the direction of $\mathbf{E}\I{X}$ are perpendicular. The angle of emission $\theta$ will increase with rising $\omega_X$, increasing the ionization probability. On the other side, large values for $\omega\I{X}$ decrease the overall ionization probability due to the denominator of Eq. (\ref{eq:diff_1-photon}). These two competing tendencies creates the maximum in the ionization probability in Fig.~\ref{fig:xray_ass_wx-dep}~(c). We see that the chosen value $\omega\I{X}-I\I{p,x}=6\U{a.u.}$ ($\omega\I{X}=14$) is close to the optimal conditions.
The former trajectory-based discussion can also be seen from a wave packet perspective. The single-photon ionization mechanism with a large initial kinetic energy obeys a dipole angular distribution of the ejected wave packet, i.e. the wave packet has an increased spreading velocity compared to tunnel ionization. The increased spatial dimension of the recolliding wave packet is exploited to overcome the drift.
\subsection{Macroscopic HHG emission}\label{sec:macro}
After discussing the single-atom yield of the x-ray assisted setup, we continue to elaborate on the macroscopic aspect of the emission from a gas target.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.4\textwidth]{kohlerfig5a}
\includegraphics[width=0.4\textwidth]{kohlerfig5b}
\caption{(color online). (a) Real part of spectral component of the locally emitted HHG field at $48.6\U{keV}$ at different positions along the propagation direction. The blue dashed line is for HHG without the quasi-phase-matching scheme. The red line is for the case of adding the weak counterpropagating field to achieve QPM. (b) The macroscopically emitted spectral photon number via Eq.~\eqref{eq:spectr-phot-num} is displayed for the QPM scenario.} \label{fig:hhgass-phase-mod-spectr}
\end{center}
\end{figure}
We inspect the emission from a Be$^{3+}$ gas of homogeneous density $\rho=5\times10^{16}/\text{cm}^3$ with the same parameters as in Fig.~\ref{fig:single_atom_xray}. The plasma refractive index at the laser frequency is $n\I{L}=5\times10^{-5}$.
The phase mismatching wave-vector of Eq.~\eqref{eq:xrayass_mismatch} at the harmonic emission energy of $50\U{keV}$ is then $\Delta k=6\times10^{-4}\U{a.u.}$ and the coherence length $l\I{coh}=\pi/\Delta k=0.25\,\mu \text{m}$.
In order to increase the coherence length, a quasi-phase matching (QPM) scheme can be employed~\cite{PEATROSS:SZ-97,COHEN:GA-99,SERRAT:SE-10}.
We propose to use a weak counterpropagating IR field with the parameters $E_2=5\times10^{-5}\U{a.u.}$ and $\omega_2=0.0418\U{a.u.}$ to achieve quasi-phase matching (QPM). The additional field is denoted by a brown line in Fig.~\ref{xray_setup2}. It is included into our mathematical formalism by $\mathbf{A}\I{P}(\eta,\mathbf{x})$ in~Eq.~\eqref{eq:totalA}. In that way a dependence on $\mathbf{x}\I{a}$ is introduced into $\tilde{S}\I{P}(\mathbf{q},\eta,\eta',\mathbf{x}\I{a})$ and the saddle points for the integration in~Eq.~\eqref{current_dens_ass} depend on the position within the medium. Thereby, $\tilde{\mathbf{x}}(\mathbf{x}\I{a},\tilde{\eta},\eta)$ in~Eq.~\eqref{eq:tildeSpert} contains the variation of the weak field seen by the electron along the z-direction. For the chosen set of parameters, the approximation of $\tilde{\mathbf{x}}(\mathbf{x}\I{a},\tilde{\eta},\eta)$
by $\mathbf{x}\I{a}$ does not lead to a significant change of the final results and, thus, can be done to save computation time. The impact of the additional field can be observed in Fig.~\ref{fig:hhgass-phase-mod-spectr}~(a). The real part of the spectral current density Eq.~\eqref{current_dens_ass} at the respective position is shown. The emitted total field is given by a spatial integral over all contributions of the current density [see Eq.~\eqref{phase_match_integral}].
Without QPM (blue dashed line), the single-atom contributions oscillate on the scale of
the coherence length estimated previously. An integration over all contributions
results in extensive cancellation. However, when applying the additional field,
the symmetry between the positive and negative contributions is broken (see the
red line)
and both parts only partially cancel thus achieving quasi-phase-matching and a nonzero value of the integral. The parameters of the additional field were chosen to optimize the photon energy at $48.6\U{keV}$. A medium length of 100~$\mu$m was chosen whereas the diameter is 500~$\mu$m. The assumed laser and x-ray pulse duration is 10 cycles. The length is limited due to the assumed bandwidth of the weak QPM field $\Delta \omega_2 \sim 0.1\%$.
The spectrum is shown in Fig.~\ref{fig:hhgass-phase-mod-spectr}~(b) and an integral over the spectrum yields the final result of $5\times10^{-7}$ emitted photons per shot. The number is of similar order of magnitude as in the other relativistic HHG setup based on the counterpropagating attosecond pulse trains for driving harmonics~\cite{KOHLER:HX-11}.
\subsection{Efficiency analysis}\label{sec:relHHGefficiency}
We continue with a discussion about the small HHG yield in the relativistic regime and identify several reasons for it that are either general to the relativistic regime or specific to this setup. First, we specify an estimate expression for the emitted photon number:
\begin{equation}
N=\frac{\mathrm{d} w_n}{\mathrm{d} \Omega} \times \Delta n \times \Delta\Omega \times \Delta t \times V^2 \rho^2 .\label{estimation}
\end{equation}
It allows in a simple way to estimate the HHG yield by an order of magnitude
and to single out the different issues influencing the HHG yield.
In~Eq.~\eqref{estimation}, $\mathrm{d} w_n/\mathrm{d} \Omega$ is the single-atom emission rate, $\Delta n$ is the number of harmonics within the phase-matched frequency bandwidth, $\Delta t$ the interaction time that is approximately the delay between both pulses, $\Delta \Omega$ the solid angle of emitted harmonics,
$V$ the volume of coherently emitting atoms [perfect phase-matching is assumed in this volume], $\rho$ the atomic density.
First, we demonstrate the usefulness of the expression by estimating the photon number for the proposed setup and show that the result of the former exact calculation can be reproduced. We estimate the terms in Eq.~\eqref{estimation} as follows: the single atom emission rate is $\mathrm{d} w_n/\mathrm{d} \Omega\approx 10^{-21}$ [see Fig.~\ref{fig:single_atom_xray}~(b)]; the phase-matched frequency bandwidth can be deduced from Fig.~\ref{fig:hhgass-phase-mod-spectr}~(b): $\Delta \omega\I{H}\approx 0.2\U{keV}$ which gives $\Delta n\sim 10^2$; the solid angle of phase-matched emission in the far field is determined by the interference pattern of a circular aperture $\Delta \Omega\sim\pi(2\pi c/D\omega\I{H})^2\approx10^{-14}$, where $\omega\I{H}=50\U{keV}$ and medium radius $r\I{a}=5\times10^6$a.u. are assumed; the interaction time $\Delta t\approx10^3\U{a.u.}$ is the laser pulse duration; the volume is cylindrical $V=\pi r\I{a}^2 \Delta z\I{a}=10^{20}$; the plasma density restricted by dispersion is $\rho=5\times10^{16}/$cm$^3$.
Taking all pieces together, the emitted photon number under phase-matched conditions is
\begin{eqnarray}\label{eq:estcprop}
N\I{rel}^{50\U{keV}}&=&\frac{\mathrm{d} w_n}{\mathrm{d} \Omega}{\Big|}\I{x,s} \times \Delta n \times \Delta\Omega \times \Delta t \times V^2 \rho^2 \\&=& 10^{-21}\times10^{2}\times10^{-14}\times10^{3}\times10^{40}\times10^{-16}\nonumber \\&=&10^{-6}\nonumber
\end{eqnarray}
in agreement with the previous accurate calculation. The subindex x stands for the x-ray assisted setup whereas s denotes that the short trajectory contribution is taken into account only.
To explain the low yield in the relativistic regime we perform the same kind of estimation for a state-of-the-art HHG experiment~\cite{Goulielmakis:SC-08} in the non-relativistic regime where $80\U{as}$ pulses were generated from harmonics below $100\U{eV}$. We can estimate the emitted photon number in this case:
\begin{eqnarray}\label{eq:estmpy}
N\I{non-rel}^{50\U{eV}}&=&\frac{\mathrm{d} w_n}{\mathrm{d} \Omega}{\Big|}\I{t,s} \times \Delta n \times \Delta\Omega \times \Delta t \times V^2 \rho^2\\
&=& 10^{-16}\times10^{1}\times10^{-7}\times10^{2}\times10^{41}\times10^{-12}\nonumber \\
&=&10^{9}\nonumber
\end{eqnarray}
The single-atom contribution $\mathrm{d} w_n/\mathrm{d} \Omega|\I{t,s}$ was calculated from~\cite{MILOSEVIC:RL-02} where we additionally inserted a correction factor accounting for the underestimation of the tunneling rate when using the zero-range potential as described in~\cite{KOHLER:HX-11}.
By comparing~Eq.~\eqref{eq:estcprop} and~Eq.~\eqref{eq:estmpy}, one observes a dramatic suppression of 15 orders of magnitude when rising the HHG energy by about 3 orders of magnitude. It arises mainly due to the single-atom yield $\mathrm{d} w_n/\mathrm{d} \Omega$, the phase-matched emission angle $\Delta\Omega$ and the gas density.
The single-atom contribution will be investigated separately below.
The estimated solid angle emission angle decreases quadratically with the harmonic energy. This is because a smaller harmonic wavelength leads to a smaller angle of the first interference minimum. The gas density depends on the phase-matching conditions which are much more difficult to fulfill in the relativistic regime and thus the gas density is lower in this case.
In the following, we inspect the ratio between the single-atom yields $\mathrm{d}
w_n/\mathrm{d} \Omega$ of Eq.~\eqref{eq:estcprop} and Eq.~\eqref{eq:estmpy} closer.
In each case, we concentrate on a single (short) trajectory at $50\U{eV}$ and
$50\U{keV}$, respectively.
Then each single-atom rate can be estimated~\cite{HATSAGORTSYAN:UFO-08,KLAIBER:CH-08},
\begin{eqnarray}\label{eq:1atom_rel}
\frac{\mathrm{d} w_n}{\mathrm{d} \Omega}{\Big|}\I{x,s}&=&\frac{1}{(2\pi c)^3\omega\I{L}^2}\omega\I{H}^2 \Big\vert\sqrt{\frac{(-2\pi i)^5}{D(\mathbf{q},\eta,\eta')}}\Big\vert^2\\
&&\!\!\!\!\!\!\times
\Big\vert\frac{c^2 p_x(\eta,\mathbf{q})}{\varepsilon_\mathbf{q}\sqrt{\omega\I{H}}}\langle0\vert\mathbf{p}(\eta,\mathbf{q})\rangle\Big\vert^2 \Big\vert \langle \mathbf{p}(\eta',\mathbf{q})\vert E\I{X} x\vert 0\rangle \Big\vert^2 \nonumber
\end{eqnarray}
for the considered relativistic setup [see also Eq.~\eqref{mhhg4a}], and by~\cite{MILOSEVIC:RL-02}
\begin{eqnarray}\label{eq:1atom_nonrel}
\frac{\mathrm{d} w_n}{\mathrm{d} \Omega}{\Big|}\I{t,s}&=&\frac{\omega\I{L}^2}{(2\pi c)^3} \omega\I{H}^2 \Bigg|\frac{1}{\omega\I{L}^2} \sqrt{\frac{(-2\pi i)^5}{D(\mathbf{p},\eta,\eta')}}\Bigg|^2
\nonumber\\
&&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\times\Big\vert\sqrt{\omega\I{H}}\langle0\vert x\vert p_x+A(\eta)/c\rangle\Big\vert^2 \\
&&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\times\Big\vert \frac{2\sqrt{2}\kappa^3}{\vert E(\eta')\vert} \langle p_x+A(\eta')/c\vert V\vert 0\rangle e^{-{\rm i}(\tilde{S}(\mathbf{q},\eta,\eta')+\frac{\omega\I{X}}{\omega\I{L}} \eta'-\frac{\omega\I{H}}{\omega\I{L}}\eta)}\Big\vert^2 \nonumber
\end{eqnarray}
for a conventional nonrelativistic setup where we inserted the tunneling correction factor of Ref.~\cite{MILOSEVIC:RT-02} and where $D(\mathbf{p},\eta,\eta')=\mathrm{det}\frac{\partial \tilde{S}(\mathbf{p},\eta,\eta')}{\partial (\mathbf{p},\eta,\eta')}$ is the functional determinant of the respective action.
Both,~Eq.~\eqref{eq:1atom_rel} and~Eq.~\eqref{eq:1atom_nonrel} are evaluated at the saddle point belonging to the short trajectory of the respective energy.
All factors in Eq.~\eqref{eq:1atom_rel} and Eq.~\eqref{eq:1atom_nonrel} were ordered in the same way and a distinct physical meaning can be assigned to them~\cite{IVANOV:CO-96,KOHLER:HX-11}:
\begin{equation}
\frac{\mathrm{d} w_n}{\mathrm{d} \Omega}
\propto \omega\I{H}^2 \left(\vert \mathbf{p}_f\vert \frac{\partial\omega\I{H}}{\partial t\I{i}}\right)^{-1}
\left| a\I{rec}(\mathbf{p}\I{f})\right|^2\frac{\mathrm{d}^3 w\I{i}(t\I{i},\mathbf{p\I{i}})}{\mathrm{d} \mathbf{p\I{i}}^3} \, .
\label{wn}
\end{equation}
The factor $\omega\I{H}^2$ accounts for the phase space and converts the matrix element into the probability, $\mathrm{d}^3 w\I{i}(t\I{i},\mathbf{p\I{i}})/\mathrm{d} \mathbf{p\I{i}}^3$ is the differential ionization rate with the ionization time $t\I{i}$ in momentum direction $p\I{i}$, $a\I{rec}(\mathbf{p}\I{f})$ is the recombination amplitude and the last factor accounts for the dynamical properties of the wave, $\mathbf{p}\I{f}$ is the final momentum at recollision, $ \partial \omega\I{H}/\partial t\I{i}$ is the electron wave packet chirping factor discussed in~\cite{KOHLER:HX-11}.
We compare all factors in~Eq.~\eqref{wn} between the two cases, to identify the reasons for the five orders of magnitude suppression of the single-atom yield in the relativistic regime. Since the harmonic energy increases by a factor of $10^3$ by going to relativistic case, the factor $\omega\I{H}^2$ yields an increase of $6$ orders of magnitude. The differential ionization rate of the particular trajectory is reduced by a factor of $10^{-2}$. Three properties contribute to estimate this ratio: The electron angular distribution of ionization is much broader for the one-photon ionization $p^2\sim(\omega\I{X}-I\I{p,x})\sim6$ than for tunneling ionization $p^2\sim\frac{3E}{\sqrt{2 I\I{p,t}}}\sim0.23$~\cite{HATSAGORTSYAN:UFO-08} yielding a factor of $4\times 10^{-2}$. On the other hand, the total (constant) ionization rate of the relativistic example [see Eq.~\eqref{eq:io-rate-zr}] is by a factor 2 higher than the instantaneous rate of non-relativistic example~\cite{PERELOMOV:IA-67,AMMOSOV:TR-86}. Third, in the relativistic setup, the relevant electron trajectory starts with a certain angle $\theta$ off the x-ray field direction resulting in a factor $\cos^2\theta\approx0.3$ in Eq.~\eqref{eq:io-rate-zr}. Together, we find the ratio of the differential rate $0.3\times2\times4\times 10^{-2}\approx 10^{-2}$. Note that the total time-average ionization rate is a factor of 10 lower for the non-relativistic example than for the relativistic one [see Eq.~\eqref{eq:io-rate-zr}]. The recombination amplitude is reduced by a factor of $10^{-5}$ as discussed in~\cite{KOHLER:HX-11}. The factors $p\I{f}$ and $\partial\omega\I{H}/\partial t\I{i}$ contained in the functional determinant reduce the relativistic yield by $10^{-4}$.
Collecting all factors, we find the suppression of $10^6\times10^{-2}\times10^{-5}\times10^{-4}=10^{-5}$ of the single-atom yield according to ratio between the respective terms in Eq.~\eqref{eq:estcprop} and~Eq.~\eqref{eq:estmpy}.
\section{Conclusion }
\label{sec:conclusion}
Extending table-top HHG to the hard x-ray domain is an exciting prospect, especially because many research labs already use HHG as a XUV sources and other approaches to generate hard x rays, like free electron lasers, require large scale facilities.
The present study discussed several difficulties that need to be overcome in order to realize the idea. The relativistic drift has been extensively discussed in the literature. Each proposed geometry has its own advantages and disadvantages regarding phase-matching.
Generally, increasing the harmonic energy renders phase-matching more difficult for many reasons: the emission phase of the harmonics depends approximately linearly on the intensity ($\phi\sim U\I{p}\tau$). Small intensity variation, e.g., in a Gaussian focus, immediately results in phase difference much larger than $\pi$. On the other hand, differences in the phase velocities between the harmonics and the laser wave lead to a slip in space between both waves. This results in phase-mismatch as soon as the slip is comparable to the harmonic wavelength which happens earlier for shorter harmonic wavelengths. Additionally, relativistic HHG is always accompanied by a large ionization
leading to an enormous plasma dispersion.
For these reasons, in the best case, we obtain realizable medium lengths of only a few tens of~$\mu$m reducing the expectable macroscopic yield.
Apart from the relativistic drift and phase-matching, we identified further issues decreasing the harmonic emission in relativistic HHG connected with the single-atom yield. First, recombination of the recolliding electron becomes less likely for high momenta: scattering is favored instead. Secondly, the electronic wave function is spread over a larger energy bandwidth. If phase-matching cannot be achieved for the whole bandwidth, however, a large part of the harmonic radiation is lost. This was expressed by the chirping factor. Third, the solid emission angle decreases quadratically with the harmonic energy increase.
Regarding the harmonic yield, the present setup for relativistic HHG as well as the one in~\cite{KOHLER:HX-11} yield a small photon number for emitted harmonics that are both of similar order. On the bottom line, we think that the setup considered in this paper is more promising than that of~\cite{KOHLER:HX-11} because the required laser intensities are lower and the phase-matching scheme is more practical.
One important conclusion of the paper is that phase-matching favors the collinear alignment of the laser and x-ray beams for the x-ray assisted relativistic HHG setup.
This co-propagation is sufficient to induce drift compensation and no perpendicular alignment of both beams is required. Note that the collinear geometry has already been used in various experiments~\cite{ISHIKAWA:PI-03,TAKAHASHI:DE-07,GAARDE:DE-05}.
\section{Acknowledgements}
We would like to thank Christoph H. Keitel and Michael Klaiber for fruitful discussions.
|
\section{Introduction}
Let \(p_m\) denote the \(m\)-th prime number (\(p_1=2, p_2 = 3, p_3=5, ...\)). We call \(m\) the \emph{order} of \(p_m\). For a positive integer \(n\), we denote by \(\Omega(n)\) the number of prime divisors of n, counted with multiplicities.
\\
For a rooted tree \(T\), its \emph{Matula number} \(\mu(T)\) is defined recursively in the following manner \cite{Matula}. For the 1-vertex tree \(T\) we define \(\mu(T)=1\). Otherwise, let \(T_1, T_2, ..., T_d\) be the subtrees of \(T\) rooted at the vertices \(v_1, v_2, ...,v_d\), adjacent to the root of \(T\) (see Fig. 1). Then, we define
\[
\mu(T)=p_{\mu(T_1)}p_{\mu(T_2)}...p_{\mu(T_d)}.
\]
\begin{figure}
\begin{center}
\input{fig1.tex}
\end{center}
\caption{}
\end{figure}
Conversely, given a natural number \(n\), the above procedure can be done in reverse, leading to a unique rooted tree \(\tau(n)\) with Matula number \(n\). We illustrate this on the arbitrarily chosen \(n=987654321\). We label the root of the desired \(\tau(n)\) by \(987654321=3*3*17*17*379721\). Since we have 5 prime factors, the degree of the root is 5. We draw 5 edges from the root and we label their endpoints, for example from left to right, by the orders of these 5 primes, i.e. by 2, 2, 7, 7, 32277 since \(p_2=3, p_7=17\), and \(p_{32277}=379721\). More conveniently, instead of 32277 we use the label \(32277=3*7*29*53\). Now we repeat this procedure for each of the 5 newly labeled vertices. For example, considering one of the vertices labeled 2, we draw one edge whose endpoint will have the label 1 (the order of the prime 2). This is a leaf of the required tree \(\tau(n)\) and the procedure stops along this branch. The reader is asked to follow in detail the way the tree \(\tau(987654321)\) has been obtained (Fig. 2).
\\
\begin{figure}
\begin{center}
\input{fig2.tex}
\end{center}
\caption{}
\end{figure}
We will use the following slightly modified description of the above defined correspondence. We give the mapping \(\tau\) from natural numbers to rooted trees, i.e. the inverse of the \(\mu\) function. We define \(\tau(1)\) to be the 1-vertex tree. For a prime number \(p_t\) (the \(t\)-th prime), we define \(\tau(p_t)\) recursively as shown in Fig. 3a. If \(n\) is composite, \(n=rs\) (\(r,s \geq 2)\), then \(\tau(rs)\) is defined recursively to be the tree shown in Fig. 3b, i.e. the trees \(\tau(r)\) and \(\tau(s)\) joined at their roots. Clearly, the obtained rooted tree does not depend on the used factorization \(rs\) of \(n\).
\\
{\bf Remark.} The prime/composite dichotomy has led us to the same recursive construction of the rooted trees that is mentioned by Czabarka, Sz\'ekely, and Wagner \cite{Czabarka}. Although that paper does not consider the Matula bijection, there may be some overlap when the same statistic is investigated.
\\
As pointed out by Ivan Gutman and Yeong-Nan Yeh \cite{GutmanYeong}, it is of interest to find statistics on rooted trees \(T\) directly from their Matula numbers \(\mu(T)\). This has been done for several statistics in \cite{ElkGutman}, \cite{GutmanMath}, \cite{GutmanIvic}.
\\
In this paper we will do the same in a slightly different way and also for several other properties like, for example, various topological indices (Wiener, Zagreb1, Zagreb2, Randi\'c, etc.). We will also consider polynomial-valued statistics (i.e. a finite sequence of statistics), namely the partial Wiener polynomial, the Wiener polynomial, the degree sequence polynomial, and the exit-distance polynomial (the last two terms will be defined below).
\\
\section{Terminology}
A \emph{rooted tree} is a tree having a distinguished vertex, called the \emph{root}.
In a rooted tree the \emph{level} of a vertex \(v\) is its distance from the root, i.e. the length of the unique path from the root to \(v\).
The \emph{height} (called also \emph{depth}) of a rooted tree is the length of the longest path from the root.
In a rooted tree if vertex \(v\) immediately precedes vertex \(w\) on the path from the root to \(w\), then \(v\) is the \emph{parent} of \(w\) and \(w\) is the \emph{child} of \(v\).
In a rooted tree vertices having the same parent are called \emph{siblings}.
A \emph{leaf} in a rooted tree is any vertex having no children. However, in the 1-vertex tree, the root is not considered to be a leaf. Given a rooted tree, the \emph{path length} is the sum of the levels of each of the nodes. The \emph{external path length} is the sum of the levels of each of the leaves (see, for example, \cite{SedgewickFlajolet}). The \emph{distance} between vertices \(i\) and \(j\) of a tree \(T\) is the number of edges on the unique path from \(i\) to \(j\); it is denoted by \(d_T(i,j)\) (or \(d(i,j)\) when no ambiguity is possible). The \emph{diameter} of a tree is the greatest distance between pairs of vertices.
The \emph{degree} of a vertex \(i\) of a tree \(T\) is the number of edges emanating from \(i\); it is denoted by \(deg_T(i)\) (or \(deg(i)\) when no ambiguity is possible).
A vertex of degree 1 is called a \emph{pendant vertex} while a vertex of degree \(\geq\) 3 is called a \emph{branching vertex}. A \emph{root subtree} of a rooted tree is any subtree that contains the root.
\\
The \emph{degree sequence} of a graph is the list of vertex degrees, written in nonincreasing order. This is basically the \emph{comparability index} of a tree introduced by Gutman and Randi\'c \cite{GutmanRandic} (see also \cite{Trinajstic}). For example, for the path on 5 vertices, the degree sequence is \((2,2,2,1,1)\). We define the \emph{degree sequence polynomial} of a graph with vertex set \(\{1,2,...,n\}\) as \(\sum_{i=1}^n x^{deg(i)}\) (it is the generating polynomial of the vertices of the graph with respect to vertex degree). For example, for the path on 5 vertices the degree sequence polynomial is \(2x+3x^2\).
The \emph{visitation length} of a rooted tree \(T\) is defined as the sum of the number of nodes of \(T\) and of its path length \cite{KeijzerFoster}.
\\
Given a vertex \(v\) in a rooted tree \(T\), we define the \emph{exit distance} of \(v\), denoted \(\lambda(v)\), to be the distance from \(v\) to the nearest leaf of \(T\) that is a descendant of \(v\). We are interested in three statistics on a rooted tree \(T\), connected with this new concept: (i) the sum of the exit distances of all vertices of \(T\), (ii) the largest exit distance, and (iii) the number of vertices for which the largest exit distance is attained. To obtain all three statistics we define the \emph{exit-distance polynomial} of \(T\) by \(\sum x^{\lambda(v)}\), where the summation extends over all vertices \(v\) of \(T\); it is the generating polynomial of the vertices of the tree with respect to exit distance. To label the vertices of a rooted tree by their exit distances, label the leaves with 0, label their parents with 1, label the so far unlabelled parents of the 1's with 2, label the so far unlabelled parents of the 2's with 3, and so on. Note the monotonicity of the coefficients of this polynomial, due to the fact that each vertex with exit distance \(k\) (\(k\geq 1\)) is the parent of some vertex with exit distance \(k-1\). The reader is asked to verify that the exit-distance polynomial of the rooted tree in Fig. 2 is \(15+9x+5x^2\).
\\
\begin{figure}
\begin{center}
\input{fig3.tex}
\end{center}
\caption{}
\end{figure}
The \emph{Wiener index} of a tree \(T\) is the sum of the distances between all unordered pairs of vertices of T \cite{Wiener}, \cite{DobryninEntringerGutman}.
The \emph{terminal Wiener index} of a tree \(T\) is the sum of the distances between all unordered pairs of pendant vertices of T \cite{GutmanFurtulaPetrovic}, \cite{SzekelyWangWu}.
The \emph{hyper-Wiener index} of a tree \(T\) has been defined by M. Randi\'c \cite{Randic}, \cite{GutmanChemical}. Later, Klein, Lukovits, and Gutman \cite{KleinLukovitsGutman} have derived the formula \(\frac{1}{2}(\sum d(i,j)^2 +\sum d(i,j))\), where the summations extend over all unordered pairs of vertices. This is now accepted as the definition of the hyper-Wiener index of a connected graph.
The \emph{multiplicative Wiener index} of a tree is the product of the distances between all unordered pairs of vertices of T \cite{GutmanLinertLukovitsTomovic}.
The \emph{Wiener polarity index} of a tree \(T\) is defined as the number of unordered pairs of vertices \({i,j}\) of \(T\) such that \(d(i,j)=3\) \cite{Wiener}.
The \emph{first Zagreb index} of a tree \(T\) is defined as \(\sum deg(i)^2\), where the
summation is over all the vertices \(i\) of \(T\) \cite{GutmanTrinajstic}, \cite{GutmanDasKinkar}, \cite{NikolicKovacevicMilicevicTrinajstic}.
The \emph{second Zagreb index} of a tree \(T\)
is defined as \(\sum deg(i)deg(j)\), where the summation is over all the edges \(ij\) of \(T\)
\cite{GutmanTrinajstic}, \cite{NikolicKovacevicMilicevicTrinajstic}, \cite{DasKinkarGutman}.
The \emph{Narumi-Katayama index} of a tree \(T\) is defined as \(\prod deg(i)\), where the product is taken over all the vertices \(i\) of \(T\)
\cite{NarumiKatayama}.
The \emph{first multiplicative Zagreb index} of a tree \(T\) is defined as \(\prod deg(i)^2 \), where the product is taken over all the vertices \(i\) of \(T\) \cite{GutmanMath}. It is the square of the Narumi-Katayama index.
The \emph{second multiplicative Zagreb index} of a tree \(T\) is defined as \(\prod deg(i)deg(j) \), where the product is taken over all the edges \(ij\) of \(T\) \cite{GutmanMath}. An other equivalent expression is \(\prod deg(i)^{deg(i)}\), where the product is taken over all vertices \(i\) of \(T\) (see Lemma 3.1 in \cite{GutmanMath}).
\\
The \emph{Randi\'c (connectivity) index} of a tree \(T\) is defined as \(\sum (deg(i)deg(j))^{-\frac{1}{2}}\), where the summation is over all the edges \(ij\) of \(T\). It was defined by Randi'c \cite{Randic} under the name branching index. Bollob\'as and Erd\H{o}s \cite{BollobasBellaErdos} generalized these indices by defining the \emph{general Randi\'c index} as \(\sum (deg(i)deg(j))^{\alpha}\), where \(\alpha\) is any real number and the summation is again over all the edges \(ij\) of \(T\). Note that for \(\alpha = 1\) this becomes the second Zagreb index, defined above.
\\
The \emph{Wiener polynomial} (called sometimes Hosoya polynomial or Wiener-Hosoya polynomial) of a tree \(T\) is defined as \(\sum x^{d(i,j)}\), where the sum is taken over all unordered pairs of distinct vertices \((i,j)\) of \(T\)
\cite{Hosoya}, \cite{SaganYehZhang}.
For a tree \(T\), the \emph{partial Wiener polynomial} with respect to a vertex \(w\) of \(T\) is defined as \(\sum x^{d(i,w)}\), where the sum is taken over all vertices \(i\) of \(T\), distinct from \(w\) \cite{DoslicTomsilav}.
\section{Statements and Proofs}
The proofs of our statements are based on examining how a particular statistic S on each of the two trees of Fig. 3 is obtained from the values of S and possibly of some auxiliary statistics on the branches \(\tau(t), \tau(r), \tau(s)\).
We will use frequently the fact that the degree of the root of \(\tau(n)\) is equal to the number of prime divisors of \(n\), counted with multiplicities.
The sequences obtained by all of the following propositions can be found in OEIS \cite{Encyclopedia} under the indicated number Axxxxxx, where also Maple programs are provided.
Abusing notation, given a statistic \(S\) on rooted trees, we shall denote \(S(\tau(n))\) by \(S(n)\). In other words, \(S(n)\) is the value of the statistic \(S\) on the rooted tree having Matula number \(n\).
We would like to point out that some of the statistics we consider have been considered previously in \cite{ElkGutman}, \cite{GutmanMath} and \cite{GutmanIvic}. We include them here in order that the paper be self-contained (some of them play an auxiliary role at other statistics) and because our approach is slightly different.
The symbols for the various statistics have been selected in the hope that they have at least some limited mnemonic value.
\begin{prop}
Let \(V\) denote "number of vertices". Then
\begin{equation}
V(n)=
\begin{cases}
1, &\text{if $n=1$;} \\
1+V(t), &\text{if $n=p_t$;} \\
V(r)+V(s)-1, &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}
\end{prop}
\begin{proof}
Follows at once by examining Fig. 3.(A061775)
\end{proof}
\begin{prop}
Let \(E\) denote "number of edges". Then
\begin{equation}
E(n)=
\begin{cases}
0, &\text{if $n=1$;} \\
1+E(t), &\text{if $n=p_t$;} \\
E(r)+E(s), &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}
\end{prop}
\begin{proof}
Follows at once by examining Fig. 3. Obviously, \(E(n)=V(n)-1\). (A196050)
\end{proof}
\begin{prop}
Let \(H\) denote "height". Then
\begin{equation}
H(n)=
\begin{cases}
0, &\text{if $n=1$;} \\
1+H(t), &\text{if $n=p_t$;} \\
max(H(r),H(s)), &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}
\end{prop}
\begin{proof}
Follows at once by examining Fig. 3. (A109082)
\end{proof}
\begin{prop}
Let \(LLL\) denote "level of lowest leaf". Then
\begin{equation}
LLL(n)=
\begin{cases}
1+LLL(t), &\text{if $n=p_t$;} \\
min(LLL(r),LLL(s)), &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}\end{prop}
\begin{proof}
Follows at once by examining Fig. 3. The 1-vertex tree, corresponding to \(n=1\), has no leaves. (A184166)
\end{proof}
\begin{prop}
Let \(LV\) denote "number of leaves". Then
\begin{equation}
LV(n)=
\begin{cases}
0, &\text{if $n=1$;} \\
1, &\text{if $n=2$;} \\
LV(t), &\text{if $n=p_t$, $t\geq 2$;} \\
LV(r)+LV(s), &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}\end{prop}
\begin{proof}
Follows at once by examining Fig. 3.(A109129)
\end{proof}
\begin{prop}Let \(MD\) denote "maximum vertex degree". Then
\begin{equation}
MD(n)=
\begin{cases}
0, &\text{if $n=1$;} \\
max(MD(t), 1+\Omega(t)), &\text{if $n=p_t$;} \\
max(MD(r),MD(s), \Omega(r)+\Omega(s)), &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}
\end{prop}
\begin{proof}
Follows at once by examining Fig. 3. (A196046)
\end{proof}
\begin{prop}
Let \(DM\) denote "diameter". Then
\begin{equation}
DM(n)=
\begin{cases}
0, &\text{if $n=1$;} \\
max(DM(t),1+H(t)), &\text{if $n=p_t$;} \\
max(DM(r),DM(s),H(r)+H(s)), &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}\end{prop}
\begin{proof}
(i) If the diameter of \(\tau(t)\) is not the diameter of \(\tau(p_t)\), then clearly the latter is given by the path going from the root of \(\tau(p_t)\) to a leaf of \(\tau(t)\) of maximum height.
(ii) Similarly, if neither \(DM(r)\) nor \(DM(s)\) is the diameter of the entire tree \(\tau(rs)\), then clearly this diameter is given by a path passing through the root. (A196058)
\end{proof}
\begin{prop} Let \(PL\) denote "path length". Then
\begin{equation}
PL(n)=
\begin{cases}
0, &\text{if $n=1$;} \\
PL(t)+V(t), &\text{if $n=p_t$;} \\
PL(r)+PL(s), &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}\end{prop}
\begin{proof} Each of the \(V(t)\) paths that contribute to \(PL(p_t)\) is 1 unit longer than those that make up \(PL(t)\). The case \(n=rs\) follows at once from Fig. 3b. (A196047)
\end{proof}
\begin{prop} Let \(EPL\) denote "external path length". Then
\begin{equation}
EPL(n)=
\begin{cases}
0, &\text{if $n=1$;} \\
1, &\text{if $n=2$;} \\
EPL(t)+LV(t), &\text{if $n=p_t$, $t\geq 2$;} \\
EPL(r)+EPL(s), &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}\end{prop}
\begin{proof} Each of the \(LV(t)\) paths that contribute to \(EPL(p_t)\) is 1 unit longer than those that make up \(EPL(t)\). The case \(n=rs\) follows at once from Fig. 3b. (A196048)
\end{proof}
\begin{prop} Let \(BV\) denote "number of branching vertices". Then, assuming, without loss of generality, that \(r\) is prime, we have
\begin{equation}
BV(n)=
\begin{cases}
0, &\text{if $n=1$;} \\
BV(t), &\text{if $n=p_t$, $\Omega(t) \ne 2$;} \\
1+BV(t), &\text{if $n=p_t$, $\Omega(t)=2$;} \\
BV(r)+BV(s), &\text{if $n=rs$, $r,s \geq 2$, $\Omega(s)\ne 2$;} \\
BV(r)+BV(s)+1, &\text{if $n=rs$, $r,s \geq 2$, $\Omega(s)=2$.}
\end{cases}
\end{equation}\end{prop}
\begin{proof} (i) In the case \(n=p_t\) we gain a branching vertex only if the root of \(\tau(t)\) has degree 2 in \(\tau(t)\). (ii) In the case \(n=rs\), by joining the trees \(\tau(r)\) and \(\tau(s)\) at their roots, we create a new branching vertex only if the degree of the root of \(\tau(s)\) in the tree \(\tau(s)\) is 2, i.e. \(\Omega(s)=2\). (A196049)
\end{proof}
\begin{prop} Let \(PV\) denote "number of pendant vertices". Then
\begin{equation}
PV(n)=
\begin{cases}
0, &\text{if $n=1$;} \\
2, &\text{if $n=2$;} \\
1+LV(t), &\text{if $n=p_t$, $n > 2$;} \\
LV(r)+ LV(s), &\text{if $n=rs$, $r,s \geq 2$. }
\end{cases}
\end{equation}\end{prop}
\begin{proof} (i) If \(n=p_t\), \(n>2\), then the pendant vertices of \(\tau(n)\) consist of the root and the leaves of \(\tau(t)\). (ii) If \(n\) is composite, \(n=rs\), then the pendant vertices of \(\tau(n)\) consist of the leaves of \(\tau(r)\) and those of \(\tau(s)\). (A196067)
\end{proof}
\begin{prop} Let \(SP\) denote "number of sibling pairs". Then
\begin{equation}
SP(n)=
\begin{cases}
0, &\text{if $n=1$;} \\
SP(t), &\text{if $n=p_t$;} \\
SP(r)+SP(s)+\Omega(r)\Omega(s), &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}\end{prop}
\begin{proof} (i) The trees \(\tau(p_t)\) and \(\tau(t)\) have the same number of sibling pairs. (ii) In addition to the sibling pairs of \(\tau(r)\) and those of \(\tau(s)\) we have to count the pairs \((u,v)\), where \(u\) (\(v\)) is a child of the root of \(\tau(rs)\) in \(\tau(r)\) (\(\tau(s)\)). (A196057)
\end{proof}
\begin{prop} Let \(VL\) denote "visitation length". Then
\begin{equation}
VL(n)=
\begin{cases}
1, &\text{if $n=1$;} \\
VL(t)+V(t)+1, &\text{if $n=p_t$;} \\
VL(r)+VL(s)-1, &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}\end{prop}
\begin{proof} (i) When going from \(\tau(t)\) to \(\tau(p_t)\) the path length increases by \(V(t)\) and the number of nodes increases by 1.
(ii) The term \(-1\) is due to the fact that in \(VL(r)+VL(s)\) the root is counted twice. (A196068)
\end{proof}
\begin{prop} Let \(RST\) denote "number of root subtrees". Then
\begin{equation}
RST(n)=
\begin{cases}
1, &\text{if $n=1$;} \\
1+RST(t), &\text{if $n=p_t$;} \\
RST(r)RST(s), &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}\end{prop}
\begin{proof} (i) The root subtrees of \(\tau(p_t)\) are the "extended" root subtrees of \(\tau(t)\) and the root of \(\tau(p_t)\) (a 1-vertex tree).
(ii) Each root subtree of \(\tau(rs)\) is obtained by joining a root subtree of \(\tau(r)\) and a root subtree of \(\tau(s)\) at their roots. (A184160)
\end{proof}
\begin{prop} Let \(ST\) denote "number of subtrees". Then
\begin{equation}
ST(n)=
\begin{cases}
1, &\text{if $n=1$;} \\
1+ST(t)+RST(t), &\text{if $n=p_t$;} \\
ST(r)+ST(s)+(RST(r)-1)(RST(s)-1)-1, &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}\end{prop}
\begin{proof} (i) As shown in the proof of the previous proposition, \(1+RST(t)\) is the number of subtrees of \(\tau(p_t)\) that do contain the root; those that do not contain the root are clearly the subtrees of \(\tau(t)\), counted by \(ST(t)\). (ii) The subtrees that belong entirely to \(\tau(r)\) or entirely to \(\tau(s)\) are counted by \(ST(r)+ST(s)-1\); the remaining ones are obtained by joining at their roots a non-1-vertex root subtree of \(\tau(r)\) with a non-1-vertex root subtree of \(\tau(s)\). (A84161)
\end{proof}
The statistic, depending on a real parameter \(\alpha\), considered in the following proposition is needed at the statistics "2nd Zagreb index" and "general Randi\'c index".
\begin{prop} Let \(A_{\alpha} = \sum deg(i)^{\alpha}\), where the summation extends over all vertices at level 1. Then
\begin{equation}
A_{\alpha}(n)=
\begin{cases}
0, &\text{if $n=1$;} \\
(1+\Omega(t))^{\alpha}, &\text{if $n=p_t$;} \\
A_{\alpha}(r)+A_{\alpha}(s), &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}\end{prop}
\begin{proof} In \(\tau(p_t)\), there is only one vertex at level 1; its degree is \(1+\Omega(t)\).
The other cases are obviously true. (A196052)
\end{proof}
\begin{prop} Let \(W\) denote "Wiener index". Then
\begin{equation}
W(n)=
\begin{cases}
0, &\text{if $n=1$;} \\
W(t)+PL(t)+E(t)+1, &\text{if $n=p_t$;} \\
W(r)+W(s)+PL(r)E(s)+PL(s)E(r), &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}\end{prop}
\begin{proof}
(i) In the case \(n=p_t\), the first term takes care of the distances within \(\tau(t)\) while the last 3 terms give the sum of the distances between the vertices of \(\tau(t)\) and the root. (ii) In the case \(n=rs\) the first two terms take care of the distances within \(\tau(r)\) and those within \(\tau(s)\). In the sum of the distances between the vertices of \(\tau(r)\) and those of \(\tau(s)\), the sum \(PL(r)\) (\(PL(s)\)) occurs for each nonroot vertex of \(\tau(s)\) (\(\tau(r)\)) i.e. \(E(s)\) (\(E(r)\)) times. (A196051)
\end{proof}
\begin{prop} Let \(TW\) denote "terminal Wiener index". Then, assuming, without loss of generality, that \(r\) is prime, we have
\begin{equation}
TW(n)=
\begin{cases}
0, &\text{if $n=1$;} \\
1, &\text{if $n=2$;} \\
TW(t)+LV(t), &\text{if $n=p_t$ and } \\
&\text{$\Omega(t)=1$ i.e. $t$ is prime}\\
TW(t)+EPL(t)+LV(t), &\text{if $n=p_t$} \\
&\text{and $\Omega(t)>1$;}\\
TW(r)-EPL(r)+TW(s)-& \\
EPL(s)+EPL(r)LV(s)+EPL(s)LV(r), &\text{if $n=rs$, $r,s \geq 2$, $s$ is prime} \\
TW(r)-EPL(r)+TW(s)+& \\
EPL(r)LV(s)+EPL(s)LV(r), &\text{if $n=rs$, $r,s \geq 2$, $s$ is not prime.}
\end{cases}
\end{equation}\end{prop}
\begin{proof}
(i) If both \(n\) and \(t\) are prime, then the distance between each leaf of \(\tau(t)\) and the root of \(\tau(t)\) has to be increased by 1 unit, explaining the term \(LV(t)\). (ii) If \(n=p_t\) and \(\Omega(t)>1\), then the pendant vertices of \(\tau(p_t)\) consist of those of \(\tau(t)\) and the root of \(\tau(p_t)\). Consequently, to \(TW(t)\) we have to add the sum of the additional distances, namely \(EPL(t)+LV(t)\). (iii) In the case \(n=rs\), \(\Omega(s)=1\), the root of \(\tau(rs)\) is definitely not a pendant vertex. Consequently, the contribution to \(TW(rs)\) of the distances between the pendant vertices within \(\tau(r)\) is \(TW(r)-EPL(r)\); same explanation for \(TW(s)-EPL(s)\). In the sum of the distances between the leaves of \(\tau(r)\) and those of \(\tau(s)\), the sum \(EPL(r)\) occurs \(LV(s)\) times and the sum \(EPL(s)\) occurs \(LV(r)\) times. (iv) The case \(n=rs\), \(\Omega(s)>1\) is similar to the previous one, except that now the root of \(\tau(s)\) is not a pendant vertex in \(\tau(s)\) and, therefore, there is no need to correct \(TW(s)\) by a subtraction. (A196055)
\end{proof}
\begin{prop} Let \(Z1\) denote "first Zagreb index". Then
\begin{equation}
Z1(n)=
\begin{cases}
0, &\text{if $n=1$;} \\
Z1(t)+2+2\Omega(t), &\text{if $n=p_t$;} \\
Z1(r)+Z1(s)-\Omega(r)^2-\Omega(s)^2+\Omega(n)^2, &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}\end{prop}
\begin{proof}
(i) In the case \(n=p_t\), to \(Z1(t)\) we add 1 (the squared degree of the root of \(\tau(p_t)\) and we take into account that the degree of the root of \(\tau(t)\) has increased by 1 unit (i.e. we add \((1+\Omega(t))^2-\Omega(t)^2\)). (ii) In the case \(n=rs\) we take into account that the degree of the root is \(\Omega(n)\). (A196053)
\end{proof}
\begin{prop} Let \(Z2\) denote "second Zagreb index". Then
\begin{equation}
Z2(n)=
\begin{cases}
0, &\text{if $n=1$;} \\
Z2(t)+A(t)+\Omega(t)+1, &\text{if $n=p_t$;} \\
Z2(r)+Z2(s)+A(r)\Omega(s)+A(s)\Omega(r), &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}\end{prop}
\begin{proof} Here we need the auxiliary statistic "sum of degrees of vertices at level 1", denoted by \(A_1\) and considered in proposition 3.16.
(i) Assume that \(n=p_t\). Let \(u\) be the root of \(\tau(t)\) and let \(ux\) be an edge in
\(\tau(t)\). This edge contributes \(deg(u)deg(x)\) to \(Z2(t)\) but to \(Z2(p_t)\) it has to
contribute \(1+deg(u))deg(x)\). The difference is \(deg(x)\), which, summed over all the edges \(ux\) in \(\tau(t)\), yields \(A_1(t)\). The sum \(1+\Omega(t)\) is the contribution of the edge emanating from the root of \(\tau(p_t)\).
(ii) Assume that \(n=rs\). Let \(v\) be the root of \(\tau(rs)\) and let \(vy\) be an edge in \(\tau(r)\). This edge contributes \(deg_{\tau(r)}(y)deg_{\tau(r)}(v)\) to \(Z2(r)\) but to \(Z2(n)\) it has to contribute \(deg_{\tau(r)}(y)deg_{\tau(n)}(v)\). The difference is \(deg_{\tau(r)}(y)(deg_{\tau(n)}(v)-deg_{\tau(r)}(v))=deg_{\tau(r)}(y)deg_{\tau(s)}(v)=deg_{\tau(r)}(y)\Omega(s)\). Summing over all the edges in \(\tau(r)\) that emanate from \(v\), we obtain \(A_1(r)\Omega(s)\). Same explanation for the edges within \(\tau(s)\). (A196054)
\end{proof}
\begin{prop} Let \(NK\) denote "Narumi-Katayama index". Then
\begin{equation}
NK(n)=
\begin{cases}
0, &\text{if $n=1$;} \\
1, &\text{if $n=2$;} \\
NK(t)(1+\frac{1}{\Omega(t)}), &\text{if $n=p_t$, $t\geq 2$;} \\
NK(r)NK(s)(\frac{1}{\Omega(r)}+\frac{1}{\Omega(s)}), &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}\end{prop}
\begin{proof}
(i) If \(n=p_t\), then in the Narumi-Katayama index \(NK(t)\) of \(\tau(t)\) we replace the degree of the root of \(\tau(t)\) in \(\tau(t)\) by its degree in \(\tau(p_t)\), i.e. \(\Omega(t)\) by \(1+\Omega(t)\).
(ii) If \(n=rs\), then in the product \(NK(r)NK(s)\) we replace the product \(\Omega(r)\Omega(s)\) of the root degrees in \(\tau(r)\) and \(\tau(s)\) by the root degree \(\Omega(n)=\Omega(r)+\Omega(s)\) in \(\tau(n)\). (A196063)
\end{proof}
\begin{prop} Let \(MZ1\) denote "first multiplicative Zagreb index". Then
\begin{equation}
MZ1(n)=
\begin{cases}
0, &\text{if $n=1$;} \\
1, &\text{if $n=2$;} \\
MZ1(t)(1+\frac{1}{\Omega(t)})^2, &\text{if $n=p_t$, $t\geq 2$;} \\
MZ1(r)MZ1(s)(\frac{1}{\Omega(r)}+\frac{1}{\Omega(s)})^2, &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}\end{prop}
\begin{proof} Since \(MZ1(n)=NK(n)^2\), the proof is entirely similar to that of proposition 3.21. (A196065)
\end{proof}
\begin{prop} Let \(MZ2\) denote "second multiplicative Zagreb index". Then
\begin{equation}
MZ2(n)=
\begin{cases}
0, &\text{if $n=1$;} \\
1, &\text{if $n=2$;} \\
\frac{MZ2(t)}{\Omega(t)^{\Omega(t)}}(1+\Omega(t))^{1+\Omega(t)}, &\text{if $n=p_t$, $t\geq 2$;} \\
\frac{MZ2(r)MZ2(s)\Omega(n)^{\Omega(n)}}{\Omega(r)^{\Omega(r)}\Omega(s)^{\Omega(s)}}, &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}\end{prop}
\begin{proof}
As mentioned in the introduction, the 2nd multiplicative Zagreb index can be expressed also as
\(\prod deg(i)^{deg(i)}\), where the product is taken over all vertices \(i\) of \(T\). Making use of this, the proof is similar to the proof of proposition 3.21. (A196064)
\end{proof}
\begin{prop} Let \(R_{\alpha}\) denote "general Randi\'c index". Then
\begin{equation}
R_{\alpha}(n)=
\begin{cases}
0, &\text{if $n=1$;} \\
R_{\alpha}(t)+A_{\alpha}(t)[(1+\Omega(t))^{\alpha}-\Omega(t)^{\alpha}]+ (1+ \Omega(t))^{\alpha}, &\text{if $n=p_t$;} \\
R_{\alpha}(r)+R_{\alpha}(s)+ A_{\alpha}(r)[\Omega(n)^{\alpha}-\Omega(r)^{\alpha}]+ A_{\alpha}(s)[\Omega(n)^{\alpha}-\Omega(s)^{\alpha}] &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}\end{prop}
\begin{proof}
The reasoning is basically the same as that proposition 3.20 which is a special case of this one (\(\alpha = 1\)). (i) In the case \(n=p_t\), the single edge emanating from the root brings in the term \((1+ \Omega(t))^{\alpha}\), while the edges in \(\tau(t)\) have the same contribution to \(R_{\alpha}(n)\) as to \(R_{\alpha}(t)\), except for those emanating from the root of \(\tau(t)\) because for these the degree of the lower endpoints has increased from \(\Omega(t)\) to \(1+\Omega(t)\). (ii) In the case \(n=rs\), the only edges in \(\tau(n)\) which do not have the same contribution to \(R_{\alpha}(n)\) as to \(R_{\alpha}(r) + R_{\alpha}(s)\) are those emanating from the root. The difference in these contributions is taken care of by the last two terms in the right-hand side.
\end{proof}
\begin{prop}5. Let \(PWP\) denote "partial Wiener polynomial with respect to the root". Then
\begin{equation}
PWP(n)=
\begin{cases}
0, &\text{if $n=1$;} \\
x+xPWP(t), &\text{if $n=p_t$;} \\
PWP(r)+PWP(s), &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}\end{prop}
\begin{proof}
If \(n=p_t\), then, in going from \(\tau(t)\) to \(\tau(p_t)\), the distances from the vertices of \(\tau(t)\) to the root of \(\tau(t)\) are increased by 1 unit; this explains the factor \(x\) in \(xPWP(t)\). The additional distance between the root of \(\tau(t)\) and that of \(\tau(p_t)\) explains the term \(x\). The case \(n=rs\) (\(r,s\geq2\)) follows at once from Fig. 3b. (A196056)
\end{proof}
\begin{prop} Let \(WP\) denote "Wiener polynomial". Then
\begin{equation}
WP(n)=
\begin{cases}
0, &\text{if $n=1$;} \\
WP(t)+xPWP(t)+x, &\text{if $n=p_t$;} \\
WP(r)+WP(s)+PWP(r)PWP(s), &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}\end{prop}
\begin{proof}
(i) \(WP(t)\) takes care of the distances within \(\tau(t)\), \(xPWP(t)\) takes care of the distances between the vertices of \(\tau(t)\) and the root of \(\tau(p_t)\), and the term \(x\) is due to the distance between the root of \(\tau(t)\) and the root of \(\tau(p_t)\).
(ii) In the case \(n=rs\) (\(r,s \geq 2\), the three terms give the contributions of the distances within \(\tau(r)\), within \(\tau(s)\), and between the vertices in \(\tau(r)\) and those in \(\tau(s)\), respectively. (A196059)
\end{proof}
\begin{prop} Let \(DSP\) denote "degree sequence polynomial". Then
\begin{equation}
DSP(n)=
\begin{cases}
1, &\text{if $n=1$;} \\
DSP(t)+x^{\Omega(t)}(x-1)+x, &\text{if $n=p_t$;} \\
DSP(r)+DSP(s)-x^{\Omega(r)}-x^{\Omega(s)}+x^{\Omega(n)}, &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}\end{prop}
\begin{proof}
(i) When we go from \(\tau(t)\) to \(\tau(p_t)\), the degree of the root of \(\tau(t)\) increases by 1 unit, explaining the term \(x^{\Omega(t)}(x-1)=-x^{\Omega(t)}+x^{1+\Omega(t)}\). We also have an additional vertex of degree 1 (the root of \(\tau(p_t)\)), explaining the term \(x\).
(ii) We are replacing the terms corresponding to the root of \(\tau(r)\) and to the root \(\tau(s)\) by the term corresponding to the root of \(\tau(n)\).(A182907)
\end{proof}
\begin{prop} Let \(EDP\) denote "exit-distance polynomial". Then
\begin{equation}
EDP(n)=
\begin{cases}
1, &\text{if $n=1$;} \\
EDP(t)+x^{1+LLL(t)}, &\text{if $n=p_t$;} \\
EDP(r)+EDP(s)-x^{max(LLL(r),LLL(s))} &\text{if $n=rs$, $r,s \geq 2$.}
\end{cases}
\end{equation}\end{prop}
\begin{proof}
(i) The vertices in \(\tau(t)\) have the same exit distances in \(\tau(p_t)\) as in \(\tau(t)\) while the root of \(\tau(p_t)\) has exit distance \(1+LLL(t)\); (ii) from the two terms corresponding to the roots of \(\tau(r)\) and \(\tau(s)\) we remove the term corresponding to the larger exit distance. (A184167)
\end{proof}
{\bf Remark}
From the propositions that consider polynomial-valued statistics we can re-obtain the results regarding several of the above considered statistics. Without going into details,
(i) denoting \(f(x)=PWP(n)\), we have \(H(n)=\) degree of \(f(x)\), \(E(n)=f(1)\), and \(PL(n)=\frac{df}{dx}|_{x=1}\); also, the number of vertices at level \(k\) \((k\geq 1)\) in the tree \(\tau(n)\) is equal to \([x^k]f(x)\);
(ii) denoting \(g(x)=WP(n)\), we have \(DM(n)=\) degree of \(g(x)\) and \(W(n)= \frac{dg}{dx}|_{x=1}\).
(iii) denoting \(h(x)=DSP(n)\), we have \(V(n)=h(1)\), \(MD(n)=\) degree of \(h(x)\), \(PV(n)=[x]h(x)\), and \(BV(n)=h(1)-[x]h(x)-[x^2]h(x)\).
Moreover, making use of some of these polynomial-valued statistics, we can consider also some statistics that are not included in the propositions given above.
(a) \emph{The hyper-Wiener index}. It has been proved in \cite{CashGG} that the hyper-Wiener index of a connected graph is equal to \(g'(1)+\frac{1}{2}g"(1)\), where \(g(x)\) is the Wiener polynomial of \(G\) (A196060).
(b) \emph{The multiplicative Wiener index}. From the definition there follows easily that the multiplicative Wiener index of a graph \(G\) is equal to \(\prod_{k>1} k^{\delta(G,k)}\), where
\(\delta(G,k)\) is the number of vertex pairs in \(G\) that are at distance \(k\) (see Eq. (5a) in \cite{GutmanLinertLukovitsTomovic} (A196061).
(c) \emph{The Wiener polarity index}. This is the coefficient of \(x^3\) in the Wiener polynomial of the graph (A184156). Obviously, taking other coefficients in the Wiener polynomial, one obtains generalized Wiener polarity indices (see \cite{IlicIlic}).
(d) In \cite{IvanciucIvanciucKleinSeitzBalaban} one introduces a family of statistics on a connected graph \(G\), based on the graph distances. In particular, Sum\(E(1,G)\) (Sum\(O(1,G)\)) denotes the sum of the even (odd) distances between unordered pairs of vertices of \(G\). Clearly, it can be obtained by evaluating at \(x=1\) the derivative of the even (odd) part of the Wiener polynomial of \(G\) (A184157, A184158).
(e) Denoting \(m(x)=EDP(n)\), it is easy to see that (i) the sum of the exit distances of all vertices of \(\tau(n)\) is equal to \(\frac{dm}{dx}|_{x=1}\) (A184168) ,(ii) the maximum exit distance over the vertices of \(\tau(n)\) is the degree of the polynomial \(m(x)\) (A184169), and (iii) the number of vertices of \(\tau(n)\) having maximum exit distance is the coefficient of the highest power of \(x\) in \(m(x)\) (A184170).
\newpage
|
\section{Introduction}
In his seminal dissertation about differentiable dynamical systems~\cite{smale-dynamics},
Smale described the recurrence of hyperbolic diffeomorphisms:
\begin{spectral}
Consider a diffeomorphism $f$ of a compact manifold.
If the non-wandering set $\Omega(f)$ is hyperbolic and contains a dense set of periodic points then it decomposes uniquely
as the finite union $\Omega(f)=\Omega_1\cup\dots\cup\Omega_s$ of disjoint, closed, invariant subsets on each of which $f$
is topologically transitive.
\end{spectral}
Recall that the restriction of $f$ to an invariant compact set $\Lambda$
is \emph{topologically transitive} if there exists a dense forward orbit, or equivalently,
if for any non-empty open sets
$U,V$ of $\Lambda$, there exists $n\geq 1$ such that $f^n(U)\cap V\neq \emptyset$.
Later on, Bowen noticed~\cite{bowen} that each piece $\Omega_i$ admits a further decomposition
$\Omega_i=X_{i,1}\cup\dots\cup X_{i,\ell_i}$ into disjoint closed subsets
on each of which $g=f^{\ell_i}$ is \emph{topologically mixing}:
for any non-empty open sets
$U,V$ of $X_{i,j}$, there exists $n_0\geq 1$ such that $f^n(U)\cap V\neq \emptyset$
for any $n\geq n_0$.
\medskip
Let $\operatorname{Diff}^1(M)$ denote the space of $C^1$-dif\-fe\-o\-mor\-phisms
of a connected compact boundaryless manifold $M$ endowed with the $C^1$-topology.
Our goal is to study the recurrence of its generic non-hyperbolic elements. A robust obstruction to the transitivity is the existence of a trapping region, i.e. a
non-empty open set $U\neq M$
such that $f(\overline U)\subset U$.
When this obstruction does not occur, it follows from~\cite{bc}
that the generic dynamics is transitive on the whole manifold.
More precisely, in this case $M$ is a homoclinic class (see the section~\ref{s.period} below)
implying that an iterate $g= f^n$ of $f$ is topologically mixing.
Our goal here is to show that this is also the case for the first iterate $f$:
\begin{theorem}\label{t.transitive}
There exists a dense $G_\delta$ subset $\cG\subset \operatorname{Diff}^1(M)$
such that any transitive diffeomorphism $f\in \cG$ is topologically mixing.
\end{theorem}
A similar statement was obtained in~\cite{aab} for flows but the case of diffeomorphisms is
more difficult:
the proof requires closing and connecting lemmas in order to build segments of orbit
which visit successively two given regions $U,V$.
For technical reasons, the obtained orbits may be shorter than what is expected
(see~\cite{survey})
so that the intersections between $f^n(U)$ and $V$ could occur only at some
particular times $n$, breaking down the topological mixing.
The main point of the present paper is thus a closing lemma with control of the connecting time
(section~\ref{s.closing}).
\medskip
Many examples of non-hyperbolic robustly transitive diffeomorphisms have
been constructed, see for instance~\cite{bd} and \cite{shub}.
Theorem \ref{t.transitive} trivially implies that these dynamics become
topologically mixing \emph{modulo an arbitrarily small $C^1$-perturbation}.
In some particular cases, there is a stronger result: among robustly transitive
partially hyperbolic diffeomorphisms with one-dimensional center bundle,
the set topologically mixing dynamics contains an open and dense subset,
see~\cite{hhu} and~\cite[corollary 3]{bdu}.
As far as we know, all of the known examples of robustly transitive diffeomorphisms are topologically mixing. This raises the following questions:
\begin{questions} \text{ }
\begin{itemize}
\item[1)]
Is \emph{every} robustly transitive diffeomorphism topologically mixing?
\item[2)]
Failing that, is topological mixing at least a $C^1$-open-and-dense condition within the space of all robustly transitive diffeomorphisms?
\end{itemize}
\end{questions}
\medskip
When $M$ is connected and $\omega$ is a volume or a symplectic form,
we denote by $\operatorname{Diff}^1_\omega(M)$ the space of the $C^1$-diffeomorphisms which preserve
$\omega$, endowed with the $C^1$-topology.
The results stated before still hold in the conservative setting and moreover
any $C^1$-generic diffeomorphism is transitive~\cite{abc,bc}.
One thus gets:
\begin{theorem}\label{t.conservatif}
Any diffeomorphism in a dense $G_\delta$ subset $\cG_\omega\subset \operatorname{Diff}^1_\omega(M)$
is topologically mixing.
\end{theorem}
\medskip
We in fact obtain a version of the previous statement for \emph{locally maximal} sets,
i.e. invariant compact sets $\Lambda\subset M$ having a neighborhood $U$
such that $\Lambda=\cap_{i\in \ZZ} f^i(U)$. Conley has proved~\cite{conley}
for homeomorphisms of $\Lambda$ that the non-existence of a trapping region is equivalent to \emph{chain-transitivity}: for any $\varepsilon>0$, there exists a $\varepsilon$-dense periodic sequence $(x_0,\dots, x_n=x_0)$ in $\Lambda$ which is a $\varepsilon$-pseudo orbit,
that is satisfies $d(f(x_i),x_{i+1})<\varepsilon$ for any $0\leq i<n$.
As a consequence of~\cite{bc}, for $C^1$-generic diffeomorphisms, any maximal invariant set which is chain-transitive is also transitive. Generalizing Bowen's result for hyperbolic
diffeomorphisms, we prove:
\begin{theorem}\label{t.main}
There exists a dense $G_\delta$ subset $\cG\subset \operatorname{Diff}^1(M)$
(or $\cG_\omega\subset \operatorname{Diff}^1_\omega(M)$) of diffeomorphisms $f$
such that any chain-transitive locally maximal set $\Lambda$ decomposes uniquely
as the finite union $\Lambda= \Lambda_1\cup\dots\cup \Lambda_\ell,$
of disjoint compact sets on each of which $f^\ell$ is topologically mixing.
\medskip
Moreover, for $1\leq i\leq \ell$,
any hyperbolic periodic $p,q\in \Lambda_i$ with same stable dimension satisfy:
\begin{itemize}
\item[--] $\ell$ is the smallest positive integer such that $W^u(f^\ell(p))\cap
W^s(p)\cap\Lambda\neq \emptyset$,
\item[--] $\Lambda_i$ coincides with
the closure of $W^u(p)\cap W^s(q)\cap \Lambda$.
\end{itemize}
\end{theorem}
Clearly, theorems~\ref{t.transitive} and~\ref{t.conservatif}
follow from theorem~\ref{t.main}.
\section{The period of a homoclinic class}\label{s.period}
Let $f$ be a $C^1$-diffeomorphism and $O$ be a hyperbolic periodic orbit.
We denote by $W\mbox{~$|$\hspace{ -.46em}$\cap$}~ W'$ the set of transversal intersection points between
two submanifolds $W,W'\subset M$.
\paragraph{\bf 2.1\quad Homoclinic class.}
The \emph{homoclinic class} $H(O)$ of $O$
is the closure of the set of transverse intersection
points between the stable and unstable manifolds $W^s(O)$ and $W^u(O)$.
We refer to~\cite{newhouse} for its basic properties, which we now recall:
\begin{itemize}
\item[--] Two hyperbolic periodic orbit $O_1,O_2$ are \emph{homoclinically related}
if $W^s(O_1)$ intersects trans\-ver\-sal\-ly $W^u(O_2)$ and $W^u(O_2)$ intersects transversally
$W^s(O_1)$. This defines an equivalence relation on the set of hyperbolic periodic orbits.
\item[--] $H(O)$ is the closure of the union of the periodic orbits homoclinically related
to $O$.
\item[--] If $O,O'$ are homoclinically related, $H(O)$ coincides with the closure of the set of transversal
intersections between $W^u(O)$ and $W^s(O')$.
\item[--] A homoclinic class is a transitive invariant set.
\end{itemize}
\paragraph{\bf 2.2 \quad Period of a homoclinic class.}
The \emph{period} $\ell(O)\geq 1$ of the homoclinic class $H(O)$ of $O$
is the greatest common divisor of the periods of the hyperbolic periodic points homoclinically related to $O$. The group $\ell(O).\ZZ$ is called the \emph{set of periods} of $H(O)$.
We have the following characterization:
\emph{for $p\in O$,
and $n\in \ZZ$, the manifolds $W^u(f^n(p))$ and $W^{s}(p)$
have a transversal intersection if and only if $n\in \ell(O).\ZZ$.}
More generally:
\begin{proposition}\label{p.intersection}
Consider a hyperbolic periodic point $q$ whose orbit is homoclinically related to $O$ and such that $W^u(p)\mbox{~$|$\hspace{ -.46em}$\cap$}~ W^s(q)\neq \emptyset$.
Then $W^u(f^n(q))\mbox{~$|$\hspace{ -.46em}$\cap$}~ W^{s}(p)\neq \emptyset$ if and only if $n\in \ell(O).\ZZ$.
In particular $W^u(q)$ intersects transversally $W^s(p)$.
\end{proposition}
This proposition is a consequence
of Smale's theorem on transversal homoclinic points~\cite{smale-homoclinic}
and of Palis' inclination lemma~\cite{palis}
(or $\lambda$-lemma).
\begin{smaletheorem}
Consider a local diffeomorphism $f$, a hyperbolic fixed point $p$ and
a transverse homoclinic intersection $x\in W^s(p)\mbox{~$|$\hspace{ -.46em}$\cap$}~ W^u(p)$.
Then, in any neighborhood of $\{p\}\cup\{f^k(x)\}_{k\in \ZZ}$,
there exists, for some iterate $f^n$,
a hyperbolic set $K$ containing $p$ and $x$.
\end{smaletheorem}
\begin{inclinationlemma}
Let $p$ be a hyperbolic fixed point and $N\subset M$ be a submanifold which intersects
$W^s(p)$ transversally. Then for any compact disc $D\subset W^u(p)$
there exists a sequence $(D_k)$ of discs of $N$ and an increasing sequence $(n_k)$ of positive integers
such that $f^{n_k}(D_k)$ converges to $D$ in the $C^1$-topology.
\end{inclinationlemma}
\begin{proof}[Proof of proposition~\ref{p.intersection}]
Let $p,q$ be two hyperbolic periodic points whose orbits are homoclinically related and
assume that $W^u(p)\mbox{~$|$\hspace{ -.46em}$\cap$}~ W^s(q)\neq \emptyset$. Let $G_{p,q}$ be the set of integers $n$
such that $W^u(f^n(q))\mbox{~$|$\hspace{ -.46em}$\cap$}~ W^s(p)\neq \emptyset$.
The set $G_{p,q}$ is invariant by addition. Indeed if $n\in G_{p,q}$,
then $W^u(f^n(q))\mbox{~$|$\hspace{ -.46em}$\cap$}~ W^s(p)$ and $W^u(f^n(p))\mbox{~$|$\hspace{ -.46em}$\cap$}~ W^s(f^n(q))$ are non-empty.
The inclination lemma implies that $W^s(p)$ accumulates on $W^s(f^n(q))$
so that it transversally intersects $W^u(f^n(p))$.
If moreover $m\in G_{p,q}$, we have
$W^u(f^{n+m}(q))\mbox{~$|$\hspace{ -.46em}$\cap$}~ W^s(f^n(p))\neq \emptyset$, so that
$W^s(p)$ intersects transversally $W^u(f^{n+m}(q))$ and $n+m\in G_{p,q}$.
The set $G_{p,q}$ is invariant by subtraction by the period $r$ of $p$.
Hence, for $n\in G_{p,q}$, the opposite
$-n=(r-1).n-r.n$ also belongs to $G_{p,q}$. So $G_{p,q}$ coincides with $G_{q,p}$ and is
a group.
If $q'$ is another hyperbolic periodic point whose orbit is homoclinically related to those of $p,q$
and satisfies $W^u(q')\mbox{~$|$\hspace{ -.46em}$\cap$}~ W^s(p)\neq \emptyset$, then $G_{p,q}=G_{p,q'}$.
Indeed the stable and the unstable manifolds
of $q,q'$ intersect transversally, and the unstable manifolds of $f^n(q)$ and $f^n(q')$
intersect the same stable manifolds.
Consequently the group $G=G_{p,q}$ contains all the periods of the hyperbolic
periodic orbits homoclinically related to the orbit $O$ of $p$.
In particular, $G$ contains $\ell(O).\ZZ$.
Conversely, let us consider $n\in G$ and an intersection point
$x\in W^u(f^n(p))\mbox{~$|$\hspace{ -.46em}$\cap$}~ W^s(p)$.
One defines a local diffeomorphism $g$ which coincides with $f^r$
in a (fixed) neighborhood of $p$ and which sends
an iterate $x^u=f^{-n-k_u.r}(x)\in W^u(p)$ onto an iterate
$x^s=f^{k_s.r}(x)\in W^s(p)$. Since by Smale's homoclinic theorem
the orbits of $x^s, x^u, p$ for $g$
are contained in a hyperbolic set, one can shadow a pseudo-orbit
$p, g^{-m}(x^s),g^{-m+1}(x^s),\dots, g^{m-1}(x^s),p$ by a hyperbolic periodic orbit
that is homoclinically related to $p$.
By construction this orbit is contained in a hyperbolic periodic orbit $O'$
of $f$ that is homoclinically related to $O$ and whose period has the form
$n+k.r$ for some $k\in \ZZ$, where $r$ is the period of $p$.
This implies that $n$ belongs to $\ell(O).\ZZ$, so that $G=\ell(O).\ZZ$.
\end{proof}
\paragraph{\bf 2.3 \quad Pointwise homoclinic class.}
If $p$ is a point of the hyperbolic periodic orbit $O$,
its \emph{pointwise homoclinic class} $h(p)$ is the
closure of the set of transverse intersection
points between the manifolds $W^s(p)$ and $W^u(p)$: this set is in general \textbf{not}
invariant by $f$.
\begin{lemma}\label{l.pointwise}
If the orbit of a hyperbolic periodic point $q$ is homoclinically related to $O$
and $W^u(p),W^s(q)$ have a transverse intersection point, then $h(p)$ coincides with the closure of the set of transversal
intersections between $W^u(p)$ and $W^s(q)$. In particular $h(p)=h(q)$.
\end{lemma}
\begin{proof}
By proposition~\ref{p.intersection} $W^u(q)$ and $W^s(p)$ have a transverse intersection point.
If $n,m$ are the periods of $p$ and $q$, then for $f^{nm}$ the points $p,q$ are fixed,
homoclinically related and their homoclinic class coincide with $h(p), h(q)$ and with the set of transversal intersections between $W^u(p)$ and $W^s(q)$.
\end{proof}
The following proposition decomposes the homoclinic classes in the form
$\Lambda_1\cup \dots \cup \Lambda_\ell$ such that $f^\ell$
is topologically mixing on each piece $\Lambda_i$. However, a priori the pieces are not disjoint.
\begin{proposition}\label{p.mixing}
Let $p\in O$ and $\ell=\ell(O)$ be the period of the homoclinic class.
Then:
\begin{itemize}
\item[--] $H(O)$ is the union of the iterates $f^k(h(p))$;
\item[--] $h(p)$ is invariant by $f^\ell$;
\item[--] the restriction of $f^\ell$ to $h(p)$ is topologically mixing;
\item[--] if $f^j(h(p))\cap f^k(h(p))$
has non-empty interior in $H(O)$, then $f^j(h(p))= f^k(h(p))$.
\end{itemize}
\end{proposition}
\begin{proof}
Let $m,n,k$ be three integers. We claim that the closure of $W^u(f^k(p))\mbox{~$|$\hspace{ -.46em}$\cap$}~ W^s(f^m(p))$
is either empty or coincides with
$f^{m+n\ell}(h(p))$.
Indeed the first set coincides with the image by $f^{m+n.\ell}$ of
the closure of $W^u(f^{k-m-n.\ell}(p))\mbox{~$|$\hspace{ -.46em}$\cap$}~ W^s(f^{-n.\ell}(p))$. If this set is non-empty, one deduces that
$k-m-n.\ell$ belongs to $\ell.\ZZ$, hence $W^u(f^{k-m-n.\ell}(p))$ and $W^u(p)$ accumulate on each other.
Similarly, $W^s(f^{-n.\ell}(p))$ and $W^s(p)$ accumulate on each other.
Consequently the closure of $W^u(f^{k-m-n.\ell}(p))\mbox{~$|$\hspace{ -.46em}$\cap$}~ W^s(f^{-n.\ell}(p))$ coincides with $h(p)$,
proving the claim.
The claim immediately implies that $H(O)$ coincides with the union of the iterates of $h(p)$
and that $f^\ell(h(p))$ coincides with $h(p)$.
Hence the two first items hold.
Let $U,V\subset M$ be two open sets which intersect $h(p)$. We have to show that for any
large $n$, the intersection $f^{n.\ell}(U)\cap V$ intersects $h(p)$.
We first introduce two points $x\in U\cap (W^u(p)\mbox{~$|$\hspace{ -.46em}$\cap$}~ W^s(p))$ and $y\in V\cap (W^u(p)\mbox{~$|$\hspace{ -.46em}$\cap$}~ W^s(p))$.
Let us consider a disc $D\subset W^u(p)\cap U$ containing $x$.
The inclination lemma shows that for $n$ large $f^{n.\ell}(D)$ accumulates on any
disc of $W^u(p)$, and hence on the local unstable manifold of $y$.
As a consequence, for $n$ large $f^{n.\ell}$ intersects transversally in $V$ the local
stable manifold of $y$, which proves the third item in the statement.
Let $A_k$ denote the interior of $f^k(h(p))$ in $H(O)$:
it is non-empty and dense in $f^k(h(p))$.
The open and dense subset $A_0\cup\dots\cup A_{\ell-1}$ of $H(O)$
is the disjoint union of elements of the form $A_{k_1}\cap A_{k_2}\cap\dots\cap A_{k_s}$.
By construction this partition is invariant by $f$.
Since the restriction of $f^\ell$ to each set $A_k$ is topologically mixing,
one deduces that $A_k$ is not subdivided by the partition. This means that
either $A_j=A_k$ or $A_j\cap A_k=\emptyset$. In the latter case one gets $f^j(h(p))=f^k(h(p))$
proving the last item.
\end{proof}
\paragraph{\bf 2.4 \quad Perturbation of the period.}
For any diffeomorphism $g$ close to $f$, one can consider the hyperbolic continuation
$O_g$ of $O$. By the implicit function theorem a given transverse intersection between $W^s(O)$ and $W^u(O)$
will persist, and hence the period of the homoclinic class of $O_g$ depends upper-semi-continously with $g$.
The following perturbation lemma provides a mechanism for the non-continuity of the period.
\begin{proposition}\label{p.cycle}
Let us consider $f\in \operatorname{Diff}^r(M)$, for some $r\geq 1$,
and two hyperbolic periodic orbits $O, O'$ having a \emph{cycle}:
$W^u(O)\cap W^s(O')\neq \emptyset$ and $W^u(O')\cap W^s(O)\neq \emptyset$.
If the period of the orbit $O'$ does not belong to the set of periods $\ell(O).\ZZ$ of the class $H(O)$,
then there exists a diffeomorphism $g$ that is arbitrarily $C^r$-close to $f$
such that $\ell(O_g)<\ell(O)$.
\end{proposition}
\begin{proof}
Let us assume that the stable dimension of $O_1$ is smaller than or equal to that of $O_2$.
Let us choose $p\in O$ and $q\in O'$ so that $W^s(p)$ and $W^u(q)$
have an intersection point $x$ and for some $n\in \ZZ$
the manifolds $W^u(f^n(p))$ and $W^s(q)$
have an intersection point $y$.
One can perturb $f$ in an arbitrarily small neighborhood of
$y$ so that the intersection becomes quasi-transversal
(i.e. $T_yW^u(O)+T_yW^s(O')=T_yM$), hence robust, and we have
not modified the orbits of $p,q,x$.
The inclination lemma ensures that
$W^u(f^n(p))$ accumulates on $W^u(q)$.
This shows that if one fixes a small neighborhood $U$ of $x$,
there exist $x_s\in W^s(p)$ and $x_u\in W^u(f^n(p))$ arbitrarily close to
$x$ whose respective future and past semiorbits avoid $U$.
By a small $C^r$-perturbation it is thus possible to create
a transverse intersection between $W^s(p)$ and $W^u(f^n(p))$ so that after the perturbation
$n$ belongs to the set of periods of the homoclinic class of $O$.
If $n\not\in\ell(O).\ZZ$ we are done.
Otherwise, if $r$ is the period of $O'$ then $W^u(f^{n+r}(p))\cap W^s(q)\neq \emptyset$.
One can thus repeat the same construction replacing $n$ by $n+r\not\in\ell(O)$ and obtain the conclusion.
\end{proof}
\paragraph{\bf 2.5 \quad Relative homoclinic classes.}
If $O$ is contained in an open set $U$, one defines the \emph{relative homoclinic class}
$H(O,U)$ of $O$ in $U$ as the closure of
$W^s(O)\mbox{~$|$\hspace{ -.46em}$\cap$}~ W^u(O)\cap (\bigcap_{n\in\ZZ}f^n(U))$. It is a transitive invariant compact set contained in $\overline U$.
All the results stated in the previous sections remain valid if one considers
hyperbolic periodic orbits and transverse homoclinic/heteroclinic orbits in $U$.
For instance, two hyperbolic periodic orbits
$O_1,O_2\subset U$ are \emph{homoclinically related in $U$}
if both $W^s(O_1)\mbox{~$|$\hspace{ -.46em}$\cap$}~ W^u(O_2)$ and $W^s(O_2)\mbox{~$|$\hspace{ -.46em}$\cap$}~ W^u(O_1)$
meet $\bigcap_{n\in \ZZ}f^n(U)$.
The homoclinic class $H(O,U)$ coincides with the closure of the set
of hyperbolic periodic points whose orbit is homoclinically related to $O$ in $U$.
The relative pointwise homoclinic class $h(p,U)$
is the intersection $h(p)\cap H(O,U)$.
\section{A closing lemma with time control}\label{s.closing}
Pugh's closing lemma~\cite{pugh} allows one to turn any non-wandering point into a periodic point via a small
$C^1$-per\-tur\-ba\-tion of the dynamics.
The proof selects a segment of orbit of the original diffeomorphism which will be closed,
so that it is difficult to control the period of the obtained orbit.
In order to control the period of the closed orbit we propose here a different argument which uses several orbit segments of the original dynamics,
as in the proof of Hayashi's connecting lemma.
A technical condition appears on the periodic points.
\begin{definition}
A periodic point $x$ is \emph{non-resonant} if, for
the tangent map $D_xf^r$ at the period,
the eigenvalues having modulus equal to one are simple
(i.e. their characteristic spaces are one-dimensional)
and do not satisfy relations of the form
$$\lambda_1^{k_1}\lambda_2^{k_2}\dots\lambda_s^{k_s}=1,$$
where the numbers $\lambda_1,\overline{\lambda_1},\dots,
\lambda_s,\overline{\lambda_s}$ are distinct and the $k_i$ are positive integers.
\end{definition}
This is obviously satisfied by hyperbolic periodic points.
Also this condition is generic in $\operatorname{Diff}^1(M)$ and in $\operatorname{Diff}^1_\omega(M)$, see~\cite{kupka,robinson,smale-kupka}.
The statement of the closing lemma with time control is the following.
\begin{theorem}[\bf Closing lemma with time control]\label{t.closing}
Let $f$ be a $C^1$-diffeomorphism, $\ell\geq 2$ be an integer, $x$ be
either a non-periodic point or
a non-resonant periodic point.
Assume that each neighborhood $V$ of $x$ intersects some iterate $f^n(V)$ such that $n$ is not a multiple of $\ell$.
Then, for diffeomorphisms $g$ arbitrarily $C^1$-close to $f$,
$x$ is periodic and its period is not a multiple of $\ell$.
\smallskip
If moreover there exists an open set $U$ such that each small neighborhood
$V$ of $x$ has a forward iterate $f^n(V)$ which intersects $V$ and such that
$f(V),\dots, f^{n-1}(V)$ are contained in $U$, then the orbit of $x$ under $g$ can be chosen in $U$.
If $f$ belongs to $\operatorname{Diff}^1_\omega(M)$, so does $g$.
\end{theorem}
\subsection{Pugh's algebraic lemma and tiled perturbation domains}\label{ss.pugh}
The main connexion results for the $C^1$-topology~\cite{pugh,pugh-robinson,hayashi,bc,abc,approximation}
are obtained by using the two following tools.
The first one allows to perform independent elementary perturbations.
They are usually obtained through
Pugh's ``algebraic" lemma (the name refers to the proof which only involves sequences of linear
maps) and with combinatorial arguments.
\begin{elementary}
For any neighborhood $\cV$
of the identity in $\operatorname{Diff}^1(M)$ (or in $\operatorname{Diff}_\omega^1(M)$),
there exists $\theta\in (0,1)$ and $\delta>0$ such that for any finite collection
of disjoint balls $B_i=B(x_i,r_i)$, with $r_i<\delta$,
and for any collection of points $y_i\in B(x_i,\theta.r_i)$, there exists
a diffeomorphism $h\in \cV$ supported on the union of the $B_i$
which satisfies $h(x_i)=y_i$ for each $i$.
\end{elementary}
Let $d$ be the dimension of $M$.
A \emph{cube} $C$ of $\RR^d$ is the image of the standard cube
$[-1,1]^d$ by a translation and an homothety. For $\lambda>0$ we denote $\lambda.C$
the cube having the same barycenter and whose edges have a length equal to $\lambda$ times
those of $C$. A cube $C$ of a chart $\varphi\colon V\to \RR^d$ of $M$ is the preimage by $\varphi$
of a cube $C'$ of $\RR^d$. The cube $\lambda.C$ is the preimage $\varphi^{-1}(\lambda.C')$.
\begin{algebraic}
For any $f\in \operatorname{Diff}^1(M)$ and any $\eta\in (0,1)$,
there exists $N\geq 1$ and a covering of $M$ by charts $\varphi\colon V\to \RR^d$
whose cubes $C$ have the following property.
For any $a,b\in C$, there is a \emph{connecting sequence} $(a=a_0,a_1,\dots,a_N=b)$ such that for each $0\leq k\leq N-1$ the
point $a_{k}$ belongs to $f^k(5/4.C)$ and the distance $d(a_k, f^{-1}(a_{k+1}))$
is smaller than $\eta$ times the distance between $f^k(5/4.C)$ and
the complement of $f^k(3/2.C)$.
\end{algebraic}
\medskip
In the following, one will fix a $C^1$-diffeomorphism $f$, a neighborhood $\cU\subset \operatorname{Diff}^1(M)$, and:
\begin{itemize}
\item[--] some constants $\theta,\delta$ provided by the elementary perturbation lemma and associated to the neighborhood $\cV=\{h=f^{-1}\circ g, g\in \cU\}$ of the identity;
\item[--] an integer $N\geq 1$ and a finite collection of charts $\{\varphi_s\colon V_s\to \RR^d\}_{s\in S}$ given by Pugh's algebraic lemma and associated to the constant $\eta=(\theta/4)^{4^d}$.
\end{itemize}
\begin{definition}
The collection of charts $\{\varphi_s\}_{s\in S}$ is a \emph{tiled perturbation domain} if we have:
\begin{itemize}
\item[--] the $f^k(V_s)$, with $s\in S$, $0\leq k<N-1$,
have diameter $<\delta$
and are pairwise disjoint;
\item[--] each set $V_k$ is tiled, i.e. is the union of cubes (the \emph{tiles}) with pairwise disjoint interior satisfying:
each tile $C$ intersects (is \emph{adjacent to}) at most $4^d$ other tiles, each of them
having a diameter which differs from the diameter of $C$ by a factor in $[1/2,2]$.
\end{itemize}
\end{definition}
Any point distinct from its $N-1$-first iterates belongs to a tiled
domain, see figure 1 in~\cite{bc}.
Note also that if the interior of $3/2.C$ and $3/2.C'$ intersect, then the tiles $C,C'$
are adjacent.
\subsection{The orbit selection}
The perturbation domain is used to connect together a collection of segments of orbits of $f$.
\begin{definition}
A \emph{pseudo-orbit with jumps in the perturbation domain} is a sequence $(y_i)$ such that
for each $i$, either $f(y_{i})=y_{i+1}$ or the points $y_i$, $f^{-1}(y_{i+1})$ are contained in a same set $V_{s}$.
\end{definition}
\medskip
We are interested by the following additional properties:
\begin{itemize}
\item[1)] When the $y_i$, $f^{-1}(y_{i+1})$ belong to $V_s$, there is
a connecting sequence $(a_{i,0},\dots,a_{i,N})$ with $a_{i,0}=y_i$
and $a_{i,N}=f^{N-1}(y_{i+1})$ such that for each $0\leq k<N$, the ball $B_{i,k}=B(a_{i,k},r_{i,k})$
with $r_{i,k}=\theta^{-1}.d(a_{i,k},f^{-1}(a_{i,k+1}))$ is contained in $f^k(V_s)$.
\item[2)] The balls $B_{i,k}$ are pairwise disjoint.
\end{itemize}
In order to control the periodic pseudo-orbits,
we also introduce an integer $\ell\geq 1$.
\begin{itemize}
\item[3)] The length of the periodic pseudo-orbit
$(y_1,\dots,y_n=y_0)$ is not a multiple of $\ell$.
\end{itemize}
\medskip
When a pseudo-orbit $(y_1,\dots,y_n=y_0)$ satisfies conditions 1), 2)
and $f(y_0)\neq y_1$,
one can apply the elementary perturbation lemma and
build a diffeomorphism $g\in \cU$ by perturbing $f$ in the union of the balls $B_{i,k}$
such that the point $y_0$ belongs to a periodic orbit of length $n$.
\medskip
These conditions can be obtained by the following proposition.
\begin{proposition}\label{p.selection}
Let $(y_i)$ be a periodic pseudo-orbit with jumps in the perturbation domain
such that when $y_{i}$, $f^{-1}(y_{i+1})$ differ, they are contained in a same tile of the perturbation domain.
Then there exists another periodic pseudo-orbit with jumps in the perturbation domain
which satisfies 1) and 2). Moreover if the first pseudo-orbit satisfies 3),
then so does the second one.
\end{proposition}
\begin{proof}
Note that by an arbitrarily small modification of the initial pseudo-orbit,
the points of the pseudo-orbit do not belong to the boundaries of the tiles.
In this way, to each point of the pseudo-orbit which belong
to the perturbation domain is associated a unique tile.
\paragraph{\rm \emph{The shortcut process.}} The new orbit is obtained from the first one by performing successive shortcuts:
if $(y_1,\dots,y_n)$ is a first periodic pseudo-orbit with jumps in the perturbation domain
and if $y_i,y_j$ for some $i<j$ belong to a same set $V_k$, then $(y_1,\dots,y_{i},y_{j-1},\dots,y_n)$ and $(y_{i+1},\dots,y_{j})$ are two new
periodic pseudo-orbits with jumps in the perturbation domain.
In the process, we keep one of them and continue with further shortcuts.
Note that if the initial orbit satisfies 3), i.e. if $n$ is not a multiple
of $\ell$, then the periods of the two new orbits cannot be both multiple of $\ell$:
we can thus choose a new orbit which still satisfies 3).
\paragraph{\rm \emph{Primary shortcuts avoiding accumulations in tiles.}} In a first step, we perform shortcuts
so that the new periodic pseudo-orbit still has jumps in the tiles of
the perturbation domain, but intersect each tile at most once:
we perform a shortcut each time we have a pair $y_i,y_j$ in a same tile of the perturbation domain.
\paragraph{\rm \emph{Construction of connecting sequences.}}
We then consider each jump of the obtained periodic pseudo-orbit
$(y_1,\dots,y_n)$ at the end of the first step: these are the indices $i$ such that
$y_{i}$ is different from $f^{-1}(y_{i+1})$. By definition the two points belong
to a same tile $C_i$ of a domain $V_s$. One can thus use the property given by Pugh's algebraic lemma and build a connecting sequence $(a_{i,0},\dots,a_{i,N})$ with $a_{i,0}=y_{i}$, $x_{i,N}=f^{N-1}(y_{i+1})$, such that for each $0\leq k\leq N-1$,
the distance $d_{i,k}=d(a_{i,k},f^{-1}(a_{i,k+1}))$ is smaller than $\eta$ times the distance
between $f^k(5/4.C_i)$ and the complement of $f^k(3/2.C_i)$.
We then set $r_{i,k}=\theta^{-1}.d_{i,k}$ and introduce the ball
$B_{i,k}=B(a_{i,k},r_{i,k})\subset f^k(V_s)$. By construction the condition 1) is satisfied, but the different balls $B_{i,k}$ may have non-empty intersection
when $i$ varies.
\paragraph{\rm \emph{Secondary shortcuts avoiding ball intersections.}}
Let us now consider the case where two balls $B_{i,k},B_{j,k'}$ intersect.
Note that this has to occur in some domain $f^k(V_s)$ for a given $s\in S$, hence we have $k=k'$.
We then perform the shortcut associated to the pair $y_i,y_j$.
Let us assume for instance that one keeps the orbit $(y_1,\dots,y_{i},y_{j+1},\dots,y_n)$
(the other case is similar). As a new connecting sequence between $y_{i}$ and $f^{N-1}(y_j)$
one introduces
$$(a'_{i,0},\dots,a'_{i,N})= (a_{i,0},\dots,a_{i,k},a_{j,k+1},\dots a_{j,N}).$$
In this way all the balls associated to the new sequence but one coincide with balls
of the former sequences. Only the new ball $B'_{i,k}$ is different: it has the same center as
as $B_{i,k}$ but a larger radius $r'_{i,k}$. Since the distance between $x_{i,k}$
and $f^{-1}(x_{j,k+1})$ is smaller than $2(r_{i,k}+r_{j,k})$, we have
\begin{equation}\label{e.radius}
r'_{i,k}\leq 2\theta^{-1}. (r_{i,k}+r_{j,k}).
\end{equation}
\paragraph{\rm \emph{The process stops.}}
Since the initial length of the pseudo-orbit is finite, the process necessarily stops
in finite time. We have however to explain why along the secondary shortcut procedure each ball
$B_{i,k}$ does not increase too much and does not leave the sets $f^k(V_s)$.
By construction it is centered at a point $a_{i,k}$ associated to a tile $C_i$.
Let us assume that the radius $r_{i,j}$ is a priori bounded by the distance between $f^k(5/4.C_i)$
and the complement of $f^k(3/2.C_i)$.
Since $a_{i,k}\in 5/4.C_i$, the ball $B_{i,k}$ is contained in $3/2.C_i$
and can only intersect the cubes $3/2.C$ such that $C$ and $C_i$ are adjacent tiles.
If $B_{i,k}$ intersects $B_{j,k}$, the point $a_{j,k+1}$
is thus associated to a tile adjacent to $C_i$.
Provided the a priori bound is preserved, the balls centered at $a_{i,k}$
during the process can thus intersect successively at most $4^d$ other balls
coming from adjacent tiles.
The diameter of the tiles adjacent to $C_i$ is at most twice the diameter of $C_i$, hence
from~\eqref{e.radius} after $4^d$ shortcuts, the diameter of the ball centered at $a_{i,k}$
is bounded from above by $(\theta/4)^{4^d}\eta$ times the distance between $f^k(5/4.C_i)$ and the complement of $f^k(3/2.C_i)$.
From our choice of $\eta$ this gives the a priori estimate.
\medskip
When the process stops, all the balls are disjoint, hence properties 1) and 2) are satisfied.
As we already explained, property 3) is preserved.
\end{proof}
\subsection{Proof of theorem~\ref{t.closing}}
Let us introduce as before an integer $N\geq 1$
and a chart $\varphi\colon V\to M$ of a neighborhood $V$ of $x$,
given by Pugh's algebraic lemma.
\paragraph{\bf The non-periodic case.}
Let us first assume that $x$ is non-periodic.
If $V$ is taken small enough, it is disjoint from
its $N-1$ first iterates. It can also be tiled, so that
it defines a perturbation domain and $x$ belongs to the interior of some tile $C$.
By assumption, there exists $z\in C$ and an iterate $f^n(z)\in C$
with $n\geq 1$ which is not a multiple of $\ell$:
the sequence $(z, f(z),\dots f^{n-1}(z))$ thus defines
a periodic pseudo-orbit which satisfies the property 3).
Applying proposition~\ref{p.selection}, there exists
a pseudo-orbit with jumps in the perturbation domain
which satisfies all the properties 1), 2) and 3).
One deduces that there exists a diffeomorphism $g$ in the neighborhood $\cU$ of $f$
having a periodic point in $V$ (close to $x$) whose period is not a multiple of $\ell$. By a new perturbation
(a conjugacy), one can ensure that this periodic point coincides with $x$, as required.
The proof is the same in $\operatorname{Diff}^1_\omega(M)$.
When one gives an open set $U$ containing $\{x, f(x),\dots, f^{N-1}(x)\}$
and $\{z, f(z),\dots f^{n-1}(z)\}$, it also contains the obtained periodic orbit.
\paragraph{\bf The periodic case.}
When $x$ is periodic, it cannot belong to a tiled domain disjoint from a large number of iterates.
However from~\cite[proposition 4.2]{abc}, since $x$ is non-resonant, the orbit $O$ of $x$
satisfies the following property (see~\cite[definition 3.10]{abc}).
\begin{definition}\label{d.circumventable}
A periodic orbit $O$ is \emph{circumventable for $(\varphi, N)$} if there exists
\begin{itemize}
\item[--] some arbitrarily small neighborhoods $W^-\subset W^+$ of $O$,
\item[--] an open subset $V'\subset V$ which is a tiled domain of the chart $\varphi$,
\item[--] some families of compact sets $\mathcal{D}} \def\cJ{\mathcal{J}} \def\cP{\mathcal{P}} \def\cV{\mathcal{V}^-,\mathcal{D}} \def\cJ{\mathcal{J}} \def\cP{\mathcal{P}} \def\cV{\mathcal{V}^+$ contained in the interior
of the tiles of $V'$,
\end{itemize}
such that
\begin{itemize}
\item[--] any finite segment of orbit which connects $ W^-$ to
$M\setminus W^+$ (resp. which connects $M\setminus W^+$ to $W^-$) has a point in a compact set
of $\mathcal{D}} \def\cJ{\mathcal{J}} \def\cP{\mathcal{P}} \def\cV{\mathcal{V}^-$ (resp. of $\mathcal{D}} \def\cJ{\mathcal{J}} \def\cP{\mathcal{P}} \def\cV{\mathcal{V}^+$),
\item[--] for any compact sets $D^+\in \mathcal{D}} \def\cJ{\mathcal{J}} \def\cP{\mathcal{P}} \def\cV{\mathcal{V}^+$, $D^-\in \mathcal{D}} \def\cJ{\mathcal{J}} \def\cP{\mathcal{P}} \def\cV{\mathcal{V}^-$,
there exists a pseudo-orbit with jumps in the perturbation domain $V'$ which connects $D^+$ to $D^-$ and is contained in $W^+$.
\end{itemize}
\end{definition}
Note that one can assume that the period $r$ of $p$ is a multiple
of $\ell$ since otherwise the conclusion of theorem~\ref{t.closing} already holds.
One can consider as before $z\in V\cap W^-$ and an iterate $f^n(z)\in V\cap W^-$ with $n\geq 1$ which is not a multiple of $\ell$.
Let $f^{k^-}(z)$, $f^{k^+}(z)$
be the first and the last iterates $f^k(z)$ of $z$ with $0\leq k\leq n$ which belong to $V\setminus W^+$.
The integers $k^-$ and $n-k^+$ are multiples of $r$, and hence of $\ell$. As a consequence $k^+-k^-$ is not a multiple of $\ell$.
By definition~\ref{d.circumventable},
there exist also some iterates
$z^-,f^{m_1}(z^-)$ of $z$ which belong respectively to
some compact sets $D^-\in \mathcal{D}} \def\cJ{\mathcal{J}} \def\cP{\mathcal{P}} \def\cV{\mathcal{V}^-$ and $D^+\in \mathcal{D}} \def\cJ{\mathcal{J}} \def\cP{\mathcal{P}} \def\cV{\mathcal{V}^+$ respectively
and such that $m_1\geq 1$ is not a multiple of $\ell$.
There also exist a pseudo-orbit
$(y_0,\dots,y_{m_2})$ contained in $W^+$, with jumps in the tiles of $V'$
and such that $y_0\in D^+$ and $y_{m_2}\in D^-$.
In particular, $m_2$ is a multiple of $r$, and hence of $\ell$.
One deduces that the pseudo-orbit
$(f(z^-),\dots,f^{m_1}(z^-),y_1,\dots,y_{m_2})$
has jumps in the tiles of the domain $V'$
and its length $m_2+m_1$ is not a multiple of $\ell$.
Applying proposition~\ref{p.selection}, there exists
a pseudo-orbit with jumps in the perturbation domain
which satisfies properties 1), 2) and 3).
One concludes as in the non-periodic case.
\section{Consequences}
We now give the proof of the theorem~\ref{t.main}, which implies theorems~\ref{t.transitive}
and~\ref{t.conservatif}. It combines the classical generic properties and a standard Baire argument.
\subsection{The non-conservative case}
There exists a dense $G_\delta$ subset $\cG\subset \operatorname{Diff}^1(M)$ of diffeomorphisms $f$ which satisfy:
\begin{enumerate}
\item\label{kupka-smale1} \emph{All the periodic points are hyperbolic.}
\item\label{kupka-smale2}
\emph{Any intersection $x$ between the stable $W^s(O)$
and the unstable manifolds $W^u(O')$ of two hyperbolic periodic orbit is transverse, i.e. $T_xM=T_xW^s(O)+T_xW^u(O')$.}
These two items together form the Kupka-Smale property, see~\cite{kupka,smale-kupka}.
\item\label{homocline}
\emph{Any locally maximal chain-transitive set is a relative homoclinic class.}
\item\label{cycle} \emph{If two hyperbolic periodic orbits $O,O'$ are contained in a same chain-transitive set $\Lambda$,
then by an arbitrarily small $C^1$-perturbation there exists a cycle between $O$ and $O'$
which is contained in an arbitrarily small neighborhood of $\Lambda$.}
\item\label{related} \emph{Any two hyperbolic periodic orbit with the same stable dimension,
contained in a same chain-transitive set $\Lambda$, are homoclinically related in any neighborhood of $\Lambda$.}
The three last items are direct consequences of the connecting lemma
for pseudo-orbits~\cite{bc} (see also~\cite[theorem 6]{approximation} for a local version).
\item\label{semi-continuity} For any $\ell\geq 1$ and any open set $U$,
let $K_{\ell,U}(f)$ denote the closure of the set of periodic points
whose period is not a multiple of $\ell$ and whose orbit is contained in $U$.
\emph{If $g$ is $C^1$-close to $f$, then
$K_{\ell,U}(g)$ is contained in a small neighborhood of $K_{\ell,U}(f)$.}
\begin{proof}
When all the periodic orbits of $f$ are hyperbolic
(or more generally have no eigenvalue equal to $1$),
$f$ is a lower-semi-continuity point of the map
$_{\ell,U}\colon g\mapsto K_{\ell,U}(g)$ for the Hausdorff topology.
One deduces from Baire's theorem
that $K_{\ell,U}$ is continuous in restriction to a dense $G_\delta$
subset $\cG_0\subset \operatorname{Diff}^1(M)$.
If $f\in \cG_0$ does not satisfy the item~\ref{semi-continuity},
then there exists a point $x\not\in K_{\ell,U}(f)$ which is
arbitrarily close to a periodic point $p$ of a diffeomorphism $g$
close to $f$ and whose period is not a multiple of $\ell$.
By a small perturbation, one can assume that the periodic point $p$
has no eigenvalue equal to $1$, and hence one can replace $g$ by any diffeomorphism close: taking $g\in \cG_0$, one contradicts the continuity of $K_{\ell,U}$ on $\cG_0$. This proves the property.
\end{proof}
\item\label{period-robust}
\emph{For any hyperbolic periodic orbit $O$,
any neighborhood $U$ of $O$ and any diffeomorphism
$g$ $C^1$-close to $f$,
the relative homoclinic class $H(O_g,U)$ of $g$ has the same periods as $H(O,U)$.}
Indeed we noticed in section~2.4 that the period map $g\mapsto \ell(O_g)$
is upper-semi-continuous, and hence is locally constant on an open and dense subset of $\operatorname{Diff}^1(M)$.
\end{enumerate}
\bigskip
We now fix $f\in \cG$ and a locally maximal chain-transitive set $\Lambda=\cap_{i\in \ZZ}f^i(U)$
in an open set $U$.
By item~\ref{homocline}, $\Lambda$ is a relative homoclinic class $H(O,U)$.
Then by proposition~\ref{p.mixing},
the set $\Lambda$ admits an invariant decomposition into compact sets
$$\Lambda= \Lambda_1\cup \dots\cup \Lambda_\ell,$$
such that for each $i$, the restriction of $f^\ell$
to $\Lambda_i$ is topologically mixing,
$\Lambda_i$ coincides with the pointwise relative homoclinic
class of a point of $O$ in $U$, and $\ell$ is the period of the relative homoclinic class.
Let us assume by contradiction that $\Lambda_1$ and $\Lambda_i$ intersect for some $1< i\leq\ell$ at a point $x$.
If $x$ is periodic, then it is hyperbolic by item~\ref{kupka-smale1}.
Since $f^\ell$ is transitive in $\Lambda_1$, one deduces that
for any neighborhood $V$ of $x$ there exists $k\geq 0$ and
a segment of orbit $y,f(y),\dots, f^{k\ell+j}(y)$ in $U$ with endpoints in $V$.
One can thus apply theorem~\ref{t.closing} and by a $C^1$-perturbation build a periodic point
arbitrarily close to $x$, whose period is not a multiple of $\ell$
and whose orbit is contained in $U$.
From the item~\ref{semi-continuity}, this shows that $K_{\ell,U}(f)$ contains $x$. Since $\Lambda$ is the locally maximal invariant set in $U$,
this shows that it contains a periodic orbit $O'$ whose period is not
a multiple of $\ell$ and which is hyperbolic by the item~\ref{kupka-smale1}.
From the item~\ref{cycle}, one can create by an arbitrarily small perturbation
a cycle between $O$ and $O'$. From item~\ref{period-robust} and proposition~\ref{p.cycle}
the period of $O'$ is contained in the set of periods
$\ell(O).\ZZ$, a contradiction. The sets $\Lambda_i$ are thus pairwise disjoint.
The uniqueness of the decomposition is easy:
considering any small open set $V$ intersecting $H(O,U)$,
then a large iterate $f^n(V)$ meets $V\cap H(O,U)$
if and only if $n$ is a multiple of $\ell$.
Moreover the closure of
$\bigcup_{k\geq k_0}f^{k\ell}(V\cap H(O,U))$, for $k_0$ large,
coincides with one of the sets $\Lambda_i$.
We have thus obtained the main conclusion of theorem~\ref{t.main}.
Let us consider two hyperbolic periodic points $p,q\in \Lambda_1$ having the same stable dimension. By item~\ref{related}, their orbits are homoclinically related in $U$.
The previous discussion shows that $\Lambda_1=h(p,U)=h(q,U)$
and $\ell$ is the minimal positive integer such that $(W^u(f^\ell(p))\mbox{~$|$\hspace{ -.46em}$\cap$}~ W^s(p))\cap \Lambda$ is non-empty.
By item~\ref{kupka-smale2}, the intersections between $W^u(f^\ell(p))$
and $W^s(p)$ are all transverse, giving the first item of the theorem~\ref{t.main}.
There exists a transverse intersection point in $\Lambda$
between $W^u(p)$ and an iterate $W^s(f^k(q))$.
Using the fact
that the decomposition of the theorem is an invariant partition into disjoint compact sets,
one deduces that $f^k(q)$ belongs to $\Lambda_1$ and that $k$ is a multiple of $\ell$.
By proposition~\ref{p.intersection}, this implies that $(W^u(p)\mbox{~$|$\hspace{ -.46em}$\cap$}~ W^s(q))\cap \Lambda$ is non-empty.
Lemma~\ref{l.pointwise} and item~\ref{kupka-smale2}
now show that $\Lambda_1=h(p)$ is the closure of
$(W^u(p)\mbox{~$|$\hspace{ -.46em}$\cap$}~ W^s(q))\cap\Lambda=W^u(p)\cap W^s(q)\cap\Lambda$, proving the second item of the theorem.
\subsection{The conservative case}
When $\dim(M)\geq 3$ and $\omega$ is a volume form,
the previous proof goes through. In the other cases $\omega$
is a symplectic form and the item~\ref{kupka-smale1} may
fail. However the same proof can be done replacing
the item~\ref{kupka-smale1} by the properties 1' and 1'' below.
\bigskip
Any diffeomorphism in a dense $G_\delta$ subset $\cG_\omega\subset \operatorname{Diff}^1_\omega(M)$ satisfies
the items~\ref{kupka-smale2}-\ref{period-robust} and moreover:
\begin{itemize}
\item[1']\label{kupka-smale}
\emph{All the periodic points are non-resonant.}
\item[1'']
\emph{Any neighborhood of a periodic orbit $O$
contains a hyperbolic periodic orbit $O'$. Consider $\ell\geq 1$.
If the period of $O$ is not a multiple of $\ell$,
then the same holds for $O'$.}
\end{itemize}
\begin{proof}[Proof]
By~\cite{robinson}, there exists a dense $G_\delta$ subset $\cG_\omega'$
of diffeomorphisms satisfying the item 1'.
One may then use similar arguments as in~\cite[proposition 3.1]{newhouse-symplectic}.
Consider any non-hyperbolic
periodic point $x$ of a diffeomorphism $f$,
with some period $r$. Using generating functions, it is possible
to build a diffeomorphism $\tilde f\in \operatorname{Diff}^1_\omega(M)$ that is
$C^1$-close to $f$ such that the dynamics of $\tilde f^r$ in a neighborhood of $x$ is conjugated to a non-hyperbolic
linear symplectic map $A$
which is diagonalizable over $\mathbb{C}$.
Let $\lambda_1,\dots,\lambda_m$ be the eigenvalues
of $A$ with modulus one.
One can assume moreover that they have the form
$e^{2i\pi p_k/q_k}$ where $\ell\wedge q_k=1$.
One deduces that $x$ is the limit of periodic points $y$
whose minimal period is $r.L$ where $L$ is the least common multiple of the $q_k$. The tangent map at $y$ at the period
coincides with the identity on its central part.
Consequently, one can by a small perturbation turn $y$
to a hyperbolic periodic point.
This shows that a diffeomorphism arbitrarily $C^1$-close to $f$
in $\operatorname{Diff}^1_\omega(M)$ has a hyperbolic periodic orbit contained in an arbitrarily small neighborhood of the orbit of $x$ and having a period which is not a multiple
of $\ell$.
We end with a Baire argument.
For $n,\ell\geq 1$, let us denote by $D_{n,\ell}\subset \cG_\omega'$ the subset of diffeomorphisms whose periodic orbits of period less than $n$ which are not a multiple of $\ell$ are
$1/n$-approximated by hyperbolic periodic orbits whose period is not a multiple of
$\ell$.
Since the periodic points of period less than $n$ are finite and vary
continuously with the diffeomorphism, this set is open; it is dense
by the first part of the proof. We then set $\cG_\omega=\bigcap_{n,\ell}D_{n,\ell}$.
\end{proof}
|
\section{Introduction}
The theory of interest rates has gone through two major
developments in recent decades. Following initial
investigations by Merton (1973) and others, the first decisive
advance
culminated in the work of Vasicek (1977) who was able to give
a fairly general characterisation of the arbitrage-free dynamics
of a family of discount bonds, indexed by their maturity. The
well-known model that bears his name appears as an exact
solution obtained with
specialising assumptions. In the wake of Vasicek's work were
a number of other specific interest rate models, of varying
degrees of usefulness and tractability, including, for example,
the CIR model (Cox {\it et al.} 1985) and its generalisations.
The next significant line of development, following the
general martingale
characterisation of arbitrage-free asset pricing by Harrison
$\&$ Kreps (1979) and Harrison $\&$ Pliska (1981), was instigated
with the recognition by Ho $\&$ Lee (1986) that the initial term
structure
might be specified essentially arbitrarily, a feature that has
important practical implications. This insight was
incorporated into the HJM framework (Heath
{\it et al.} 1992), which constituted a major advance in the
subject, providing a general model-independent basis for
the analysis of interest rate dynamics and the pricing of
interest rate derivatives.
Since then there have been numerous further developments. These
include, for example, the infinite dimensional or `string-type'
models of Kennedy (1994), Santa-Clara $\&$ Sornett (1997) and
others, the positive interest rate models of Flesaker $\&$
Hughston (1996), the potential approach of Rogers (1997),
the so-called market models (Brace {\it et al}. 1996,
1997; Jamshidian 1997), and the geometric analysis of the space
of yield curves undertaken by Bj\"ork $\&$ Svensson (1999).
Nevertheless, no criterion has emerged, based on the
extensive econometric evidence available, that allows in
a rational way for the identification of a clearly
preferred class of
models. On these grounds it makes sense to try to cast the
general interest rate framework into a new form, with the
idea that certain models might thus become recognisable
as more natural on mathematical and economic grounds.
With this end in mind, the
purpose of the present article is to propose a
novel application of information geometry to interest rate
theory. The main results are (i)
the construction of a geometric measure for how `different' two
term structures are from one another; (ii) a characterisation of
the evolutionary trajectory of the term structure as a
measure-valued process; (iii) the derivation of dynamics
for the principal moments of the
term structure; and (iv) a reformulation
of arbitrage-free interest rate dynamics in terms
of a class of processes on Hilbert space.
The paper is organised as follows. In \S 2 we review the
basic idea
of information geometry and its role
in estimation theory. The geometry of the normal
distribution is considered in detail as an illustration. In \S 3
a remarkable characterisation of the discount function in terms
of an abstract probability density function is introduced in
Proposition 1. This allows us to apply information geometric
techniques to determine the deviation between different term
structures within a given model. In this connection, in \S 4
we consider a class of flat rate models as examples.
The material of the first four sections of the paper is
essentially static, i.e., set in the present, whereas in \S 5
we investigate the dynamics of
the density function that generates the term structure. This is
carried out in such a way that the resulting dynamics is
manifestly arbitrage-free. Our key result here is formula
(\ref{eq:5.15}), in which we establish that the dynamics of the
term structure can be characterised as a measure-valued process.
This idea is developed further in Proposition 2.
In \S 6 we introduce an analogue of the classical principal
components analysis for yield curves, and in Propositions 3 and
4 we derive formulae for the evolution of the first two moments
of the term structure density process. Then, making use of the
information geometry developed earlier, in \S 7 we map the
dynamics developed in \S 5 to Hilbert space. Our main result
here is Proposition 5, which shows how this can be achieved.
\section{Information geometry}
Because some of the mathematical
techniques we employ here may not be
familiar to those working in finance, it will be appropriate to
begin with a few background remarks. It has long been known
(see, e.g., Amari 1985; Kass 1989; Murray $\&$ Rice 1993) that
a useful
approach to statistical inference is to regard a parametric model
as a differentiable manifold equipped with a metric. The
recognition that a parametric family of probability distributions
has a natural geometry associated with it arose in the work of
Mahalanobis (1936), Bhattacharyya (1943) and Rao (1945) over
half of a century ago.
Suppose, for example, that $X$ is a continuous random variable
taking values on the real line ${\bf R}^{1}$, and that $\rho(x)$ is
a density function for $X$. Because $\rho(x)$ is nonnegative and has
integral unity, it follows that the square-root
likelihood function
\begin{eqnarray}
\xi(x) = \sqrt{\rho(x)}
\label{eq:2.1}
\end{eqnarray}
exists for all $x$, and satisfies the normalisation condition
\begin{eqnarray}
\int_{-\infty}^{\infty} (\xi(x))^{2} {\rm d}x = 1 .
\label{eq:2.2}
\end{eqnarray}
We see that $\xi(x)$ can be regarded
as a unit vector in the Hilbert space
${\cal H}=L^{2}({\bf R}^{1})$. Now let
$\rho_{1}(x), \rho_{2}(x)$ denote a
pair of density functions on ${\bf R}^{1}$, and $\xi_{1}(x),
\xi_{2}(x)$ the corresponding Hilbert space elements.
Then the inner product
\begin{eqnarray}
\cos\phi = \int_{-\infty}^{\infty}\xi_{1}(x)\xi_{2}(x) {\rm d}x
\label{eq:2.3}
\end{eqnarray}
defines an angle $\phi$ which can be interpreted as the {\sl
distance} between the two probability distributions. More
precisely, if we write ${\cal S}$ for the unit sphere in
${\cal H}$, then $\phi$ is the spherical distance
between the points on ${\cal S}$ determined by the vectors
$\xi_{1}(x)$ and $\xi_{2}(x)$.
The maximum possible distance, corresponding to nonoverlapping
densities, is given by $\phi=\pi/2$. This follows from the fact
that $\xi_{1}(x)$ and $\xi_{2}(x)$ are nonnegative functions,
and thus define points on the positive orthant of ${\cal S}$.
We remark that an alternative way of expressing (\ref{eq:2.3})
is
\begin{eqnarray}
\cos\phi = 1 - \frac{1}{2} \int_{-\infty}^{\infty}
(\left( \xi_{1}(x)-\xi_{2}(x)\right)^{2}{\rm d}x ,
\label{eq:2.4}
\end{eqnarray}
which makes it apparent that the angle $\phi$ measures
the extent to which the two distributions are distinct.
The spherical distance of Bhattacharyya introduced above is
applicable in a nonparametric context. In the case of a
parametric family of probability distributions we can develop
matters further. Let us write
$\rho(x,\theta)$ for the parameterised density function. Here
$\theta$ stands for a set of parameters $\theta^{i}$
$(i=1,\cdots,r)$. By varying $\theta$ we obtain an
$r$-dimensional submanifold ${\cal M}$ in ${\cal S}$
determined by the unit vectors $\xi(x,\theta)\in{\cal H}$.
The parameters $\theta^{i}$ are local coordinates
for ${\cal M}$.
The key point that we require in the
following (cf. Dawid 1977) is that the spherical geometry
of ${\cal S}$ induces a Riemannian geometry on
${\cal M}$, for which the metric tensor $g_{ij}(\theta)$ is
given, in local coordinates, by
\begin{eqnarray}
g_{ij}(\theta) = \int_{-\infty}^{\infty} \frac{\partial\xi(x,\theta)}
{\partial\theta^{i}}\frac{\partial\xi(x,\theta)}
{\partial\theta^{j}} {\rm d}x .
\label{eq:2.5}
\end{eqnarray}
By use of definition (\ref{eq:2.1}), we see that an alternative
expression for $g_{ij}(\theta)$ is
\begin{eqnarray}
g_{ij}(\theta) = \frac{1}{4}\int_{-\infty}^{\infty} \rho(x,\theta)
\frac{\partial\ln\rho(x,\theta)}
{\partial\theta^{i}}\frac{\partial\ln\rho(x,\theta)}
{\partial\theta^{j}} {\rm d}x ,
\label{eq:2.6}
\end{eqnarray}
which shows (cf. Brody $\&$ Hughston 1998) that the metric
$g_{ij}$ is, apart from the factor of $\frac{1}{4}$, the Fisher
information matrix, i.e., the covariance matrix of the parametric
gradient of the log-likelihood function (Fisher 1921). We refer
to $g_{ij}(\theta)$ as the Fisher-Rao metric on the
statistical model ${\cal M}$.
The significance of the Fisher-Rao metric in estimation theory
is well known. Suppose that $\tau(\theta)$ is some given function
of the parameters, and that the random variable $T$ represented
by the function $T(x)$ on ${\bf R}^{1}$ is
an unbiased estimator for $\tau(\theta)$ in the sense that
\begin{eqnarray}
\int_{-\infty}^{\infty} \rho(x,\theta) T(x) {\rm d}x =
\tau(\theta) .
\label{eq:2.7}
\end{eqnarray}
The variance of the estimator $T$ is defined, as usual, by
\begin{eqnarray}
{\rm Var}[T] = \int_{-\infty}^{\infty} \rho(x,\theta)
(T(x)-\tau(\theta))^{2}{\rm d}x .
\label{eq:2.8}
\end{eqnarray}
Then a set of fundamental bounds on ${\rm Var}[T]$, independent
of the choice of the estimator $T(x)$, can be obtained
by applying the operator $\sum_{i}\alpha^{i}
\partial_{i}$ to (\ref{eq:2.7}), letting $\alpha^{i}$ be
arbitrary. By use of (\ref{eq:2.1}) and
the Schwartz inequality for $L^{2}({\bf R}^{1})$, we obtain
\begin{eqnarray}
g_{ij} {\rm Var}[T] \geq \frac{1}{4}
\frac{\partial\tau}{\partial\theta^{i}}
\frac{\partial\tau}{\partial\theta^{j}} .
\label{eq:2.9}
\end{eqnarray}
This matrix inequality is interpreted as saying that if
we subtract the right side from the left, the result is
nonnegative definite. It follows that if the
random variables $\Theta^{i}$ $(i=1,\cdots,r)$ are unbiased
estimators for the parameters $\theta^{i}$, satisfying
\begin{eqnarray}
\int_{-\infty}^{\infty} \rho(x,\theta) \Theta^{i}(x) {\rm d}x
= \theta^{i} ,
\label{eq:2.10}
\end{eqnarray}
then the covariance matrix of the estimators is bounded by the
inverse Fisher information matrix:
\begin{eqnarray}
{\rm Cov}[\Theta^{i},\Theta^{j}] \geq \frac{1}{4}g^{ij} .
\label{eq:2.11}
\end{eqnarray}
The Riemannian metric (\ref{eq:2.5}) introduced above can be
used to define a distance measure between two distributions
belonging to a given parametric
family. This measure is invariant in the sense that it is
unaffected by a reparameterisation of the distributions. The
distance is calculated by integrating the infinitesimal line
element ${\rm d}s$ along the geodesic connecting the two points in
the statistical manifold ${\cal M}$, where
\begin{eqnarray}
{\rm d}s^{2} = \sum_{i,j} g_{ij} {\rm d}\theta^{i} {\rm d}
\theta^{j} .
\label{eq:2.12}
\end{eqnarray}
The geodesics with respect to a given metric
$g_{ij}$ are the solutions of the differential
equation
\begin{eqnarray}
\frac{{\rm d}^{2}\theta^{i}}{{\rm d}u^{2}} + \Gamma^{i}_{jk}
\frac{{\rm d}\theta^{j}}{{\rm d}u}
\frac{{\rm d}\theta^{k}}{{\rm d}u} = 0
\label{eq:2.13}
\end{eqnarray}
for the curve $\theta^{i}(u)$ in ${\cal M}$, subject to
the given boundary conditions at the two end points. Here,
we have written
\begin{eqnarray}
\Gamma^{i}_{jk} = \frac{1}{2} g^{il} \left( \partial_{j}g_{kl}
+ \partial_{k}g_{jl} - \partial_{l}g_{jk} \right) ,
\label{eq:2.14}
\end{eqnarray}
where $\partial_{i}=\partial/\partial\theta^{i}$, and the inverse
metric $g^{ij}$, also appearing in (\ref{eq:2.11}), satisfies
$g^{ij}g_{jk}=\delta^{i}_{k}$, where
$\delta^{i}_{k}$ is the Kronecker delta. Note that in
equations (\ref{eq:2.13}) and (\ref{eq:2.14}) above, and elsewhere
henceforth in this article, we employ the standard Einstein
summation convention on repeated indices.
\begin{figure}[t]
\centerline{
\psfig{file=r1fig1.eps,width=12cm,angle=0}
}
\caption{{\it Geodesic curves for normal distributions}. The
statistical manifold ${\cal M}$ in this case is the upper
half plane parameterised by $\mu$ and $\sigma$. We have
$-\infty<\mu<\infty$ and $0<\sigma<\infty$. The
shortest path joining the two normal distributions
${\cal N}(\mu_{1},\sigma_{1})$ and
${\cal N}(\mu_{2},\sigma_{2})$ is given by the unique
semi-circular arc through the given two points and centred
on the boundary line $\sigma=0$.
}
\end{figure}
Let us consider, as an explicit example, the manifold
${\cal M}$ corresponding
to the normal distributions ${\cal N}(\mu,\sigma)$ on
${\bf R}^{1}$, with mean $\mu$ and standard deviation
$\sigma$. For the parameterised density function we have
\begin{eqnarray}
\rho(x,\mu,\sigma) = \frac{1}{\sqrt{2\pi}\sigma}\exp\left(
-\frac{(x-\mu)^{2}}{2\sigma^{2}} \right) .
\label{eq:2.15}
\end{eqnarray}
A straightforward computation, making use of (\ref{eq:2.6}),
gives
\begin{eqnarray}
{\rm d}s^{2} = \frac{1}{\sigma^{2}}({\rm d}\mu^{2} + 2
{\rm d}\sigma^{2})
\label{eq:2.16}
\end{eqnarray}
for the line element, which is defined on the upper half-plane
$-\infty<\mu<\infty$, $0<\sigma<\infty$. The resulting Riemannian
geometry is that of hyperbolic space, which is a homogeneous
manifold with constant negative curvature. The geometry of this
space has been studied extensively, and has many intriguing
properties. For the distance function in the
case of a pair of normal distributions
${\cal N}(\mu_{1},\sigma_{1})$,
${\cal N}(\mu_{2},\sigma_{2})$ we obtain
\begin{eqnarray}
D(\rho_{1},\rho_{2}) = \frac{1}{\sqrt{2}}\log
\frac{1+\delta_{1,2}}{1-\delta_{1,2}} ,
\label{eq:2.17}
\end{eqnarray}
where the function $\delta_{1,2}$, defined by
\begin{eqnarray}
\delta_{1,2} = \sqrt{\frac{(\mu_{2}-\mu_{1})^{2}
+2(\sigma_{2}-\sigma_{1})^{2}}{(\mu_{2}-\mu_{1})^{2}+2
(\sigma_{2}+\sigma_{1})^{2}}} ,
\label{eq:2.18}
\end{eqnarray}
lies between 0 and 1. The geodesics, in particular, are given
in general by semi-circular arcs
centred on the boundary line $\sigma=0$ (this line itself is
not part of the manifold ${\cal M}$). An exceptional situation
arises when $\mu_{1}=\mu_{2}$, for which the geodesic is a
straight line given by constant $\mu$, and we have
\begin{eqnarray}
D(\rho_{1},\rho_{2}) = \frac{1}{\sqrt{2}}\left|\log
\frac{\sigma_{1}}{\sigma_{2}}\right| .
\label{eq:2.19}
\end{eqnarray}
We refer the reader to Burbea (1986), where metric and
distance computations have been carried out explicitly for
other families of distributions.
\section{Discount bond densities}
Our goal now is to make use of the analysis presented in the
previous section to construct a natural metric on the space of
yield curves. In doing so we shall take advantage of a remarkable
`probabilistic' characterisation of discount bonds, which we here
proceed to describe.
Let $t=0$ denote the present, and $P_{0T}$ a smooth
family of discount bonds, where $T$ is the maturity date
$(0\leq T<\infty)$. For positive interest we require
\begin{eqnarray}
0<P_{0T}\leq1, \ \ \ \frac{\partial}{\partial T}
P_{0T} < 0 ,
\label{eq:3.1}
\end{eqnarray}
and we assume that $P_{0T}\rightarrow0$ as $T$ goes to
infinity. A term structure that satisfies these conditions
will be said to be `admissible'. These conditions can,
in fact, be relaxed slightly:
$P_{0T}$ need not be strictly smooth, nor strictly decreasing;
but for most of the present discussion we shall
stick with the assumptions indicated.
\begin{figure}[t]
\centerline{
\psfig{file=r1fig5.eps,width=12cm,angle=0}
}
\caption{{\it The system of admissible term structures}.
A smooth positive interest term structure can be regarded
as a point in ${\cal D}({\bf R}_{+}^{1})$, the convex
space consisting of smooth density functions on
${\bf R}_{+}^{1}$. The points of ${\cal D}({\bf R}_{+}^{1})$
are in one-to-one correspondence with rays lying in the
positive orthant ${\cal S}_{+}$ of the unit sphere ${\cal S}$
in the Hilbert space ${\cal H}=L^{2}({\bf R}_{+}^{1})$.
}
\end{figure}
The interesting point that arises here, of which we shall
make extensive use in the discussions that follow, is that
the discount function $P_{0T}$ can be viewed as a
complementary probability distribution. In other words, we
think of the maturity date as an abstract random variable $X$,
and for its distribution we write
\begin{eqnarray}
{\bf P}[X<T] = 1-P_{0T} .
\label{eq:3.2}
\end{eqnarray}
It should be clear that this can be done if and only if the
positive interest rate conditions given in (\ref{eq:3.1}) hold.
As a consequence we are able to embody the positive interest
property in a fundamental way in the structure of the theory.
Indeed, this basic economic property is essential if we wish
to treat the yield curve consistently and naturally as a kind
of mathematical object in its own right. Now let us introduce
the function $\rho(T)$ defined by
\begin{eqnarray}
\rho(T) = -\frac{\partial}{\partial T} P_{0T} .
\label{eq:3.3}
\end{eqnarray}
Clearly, we have $\rho(T)> 0$ and
\begin{eqnarray}
\int_{0}^{\infty} \rho(T) {\rm d}T = 1 ,
\label{eq:3.4}
\end{eqnarray}
from which we infer that $\rho(T)$ can be consistently viewed as
a probability density function. It follows from the defining
equation (\ref{eq:3.3}) that the term structure density
$\rho(T)$ is the product of the instantaneous forward
rate and the discount function itself. Now clearly if
$\rho_{1}(T)$ and $\rho_{2}(T)$ are admissible term structure
densities, and if $A$ and $B$ are nonnegative constants satisfying
$A+B=1$, then $A\rho_{1}(T)+B\rho_{2}(T)$ is also an admissible
term structure density. Putting these ingredients together, we
see that the term structure of interest rates can be given the
following general characterisation.
\begin{prop}
The system of admissible term structures is isomorphic to the
convex space ${\cal D}({\bf R}_{+}^{1})$ of smooth density
functions on the positive real line.
\label{prop:1}
\end{prop}
At first glance it may seem odd to think of the discount function
in this manner. However, it gives us the advantage of being able to
apply the tools of information geometry in an unexpected way, as
we indicate in what follows.
In particular, there is a one-to-one map from the space
${\cal D}({\bf R}_{+}^{1})$ of such term structure densities to
the positive orthant ${\cal S}_{+}$ of the unit sphere ${\cal S}$
in the Hilbert space ${\cal H}$, as indicated in Figure 2.
Therefore, given two yield
curves we can calculate the distance between them. This can be
carried out either in a nonparametric sense, by use of the
Bhattacharyya spherical distance, or in a parametric sense, by
use of the Fisher-Rao distance. In the former case first we
calculate the corresponding term structure densities $\rho_{1}(T)$
and $\rho_{2}(T)$. These are then mapped to ${\cal S}_{+}$ by
taking the square-roots, and their distance $\phi(\rho_{1},
\rho_{2})$ is given by
\begin{eqnarray}
\phi(\rho_{1},\rho_{2}) = \cos^{-1} \int_{0}^{\infty}
\sqrt{\rho_{1}(T)\rho_{2}(T)}{\rm d}T .
\label{eq:3.41}
\end{eqnarray}
In the parametric case we regard the given parametric family of
yield curves as defining a statistical model ${\cal M}\subset
{\cal S}_{+}$, and the distance between the two yield curves
within the given family is then defined by the Fisher-Rao metric.
\section{Flat term structures}
To provide some illustrations of the principles set
forth in the previous section we consider here properties of
yield curves for which the term structure is {\sl flat}. Such
yield curves, which are of various types, are on the whole
too simple for use in practical modelling. Nevertheless,
they are of interest as examples, because many of the relevant
computations can be carried out explicitly.
In this connection we begin by introducing a representation of
the discount function as a Laplace transform
\begin{eqnarray}
P_{0T} = \int_{0}^{\infty} e^{-rT} \psi(r) {\rm d}r
\label{eq:4.1}
\end{eqnarray}
for some function $\psi(r)$. Thus we think of the discount
function $P_{0T}$ as being given by a weighted superposition
of elementary discount functions, each of the form $e^{-rT}$
for some value of $r$. Taking the limit $T\rightarrow0$, we
find that $\psi(r)$ must satisfy $\int_{0}^{\infty} \psi(r)
{\rm d}r = 1$.
In general the inverse Laplace transform $\psi(r)$ need not
be positive. However, if we restrict our consideration to
nonnegative functions, then $\psi(r)$ can be interpreted as
a density function, and by various
choices of $\psi(r)$ we are led to some interesting
candidates for term structures.
First we consider the case where $\psi(r)$ is
a Dirac $\delta$-function concentrated at a point, that is,
$\psi(r)=\delta(r-R)$. A direct substitution gives
$P_{0T}=\exp(-RT)$,
corresponding to a `flat' term structure with a continuously
compounded rate $R$ for each value of the maturity date $T$.
If the density function $\psi(r)$ is given by an exponential
distribution $\psi(r)=\tau\exp(-\tau r)$, with parameter
$\tau$, then one sees that $\tau$ must have dimensions of
time, and a short calculation gives $P_{0T}=\tau/(\tau+T)$,
which also corresponds to a flat term structure, in this
case with a simple percentage yield of $\tau^{-1}$ for all
maturities. We see that the characteristic time-scale $\tau$
allows us to define an interest rate $R=\tau^{-1}$, which
turns out to be the characteristic interest rate of the
resulting structure, and we can write
$P_{0T}=1/(1+RT)$ for the discount function.
We note that flatness is not a completely unambiguous notion,
because having a uniform continuously compounded yield for all
maturities is not the same thing as having a uniform simple
yield for all maturities. Both define plausible albeit quite
distinct systems of discount bonds. This example illustrates
how by superposing term structures of the elementary form
$\exp(-RT)$ for various maturities, we can obtain other
reasonable looking and well behaved term structures.
We mention one more example, which contains the
previous two examples as special cases. Consider the
standard gamma distribution, with parameters $\kappa$ and
$\lambda$, defined for nonnegative values of $r$ by the
density function
\begin{eqnarray}
\psi(r) = \frac{1}{\Gamma(\kappa)} \lambda^{\kappa}
r^{\kappa-1} \exp(-\lambda r) .
\label{eq:4.6}
\end{eqnarray}
In this case, we can verify that the resulting system of
discount bonds is given by
\begin{eqnarray}
P_{0T} = \left( \frac{\lambda}{\lambda+T} \right)^{\kappa} ,
\label{eq:4.7}
\end{eqnarray}
which assumes a more recognisable form if we set
$\lambda=\kappa\tau$, where $\tau$ again defines a characteristic
time scale, and $\kappa$ is a dimensionless number. Then we have
\begin{eqnarray}
P_{0T} = \left(1+\frac{RT}{\kappa} \right)^{-\kappa} ,
\label{eq:4.8}
\end{eqnarray}
where $R=\tau^{-1}$. The system of discount bonds
arising here can also be interpreted as a flat term structure,
in this case with a constant annualised rate of interest $R$
assuming compounding at the frequency $\kappa$ over the life
of each bond ($\kappa$ need not be an integer). It is not
difficult to check that for $\kappa=1$ this reduces to the
case of a flat rate on the basis of a simple yield, whereas
in the limit $\kappa\rightarrow\infty$ we recover the case of
a flat rate on the basis of continuous compounding.
Now we shall apply the ideas of statistical geometry to make
comparisons between various term structures of the form
(\ref{eq:4.8}). For density function $\rho(T)=-\partial_{T}
P_{0T}$ in this case we obtain
\begin{eqnarray}
\rho(T,R) = R\left( 1+\frac{RT}{\kappa}\right)^{-(\kappa+1)}.
\label{eq:4.9}
\end{eqnarray}
Here we find it convenient to label the density function
by the flat rate $R$. Note that in the limit
$\kappa\rightarrow\infty$ we have
$\rho(T,R)\rightarrow Re^{-RT}$. First consider
the nonparametric separation between
different term structures in this model via spherical distance
of Bhattacharyya given in formula (\ref{eq:3.41}), where in the
present example we write $\rho_{i}(T)=\rho(T,R_{i})$ for
$i=1,2$. A direct integration leads to the expression
\begin{eqnarray}
\phi(\rho_{1},\rho_{2}) = \cos^{-1}\left(
\frac{\sqrt{R_{1}R_{2}}}{R_{1}-R_{2}}\log\frac{R_{1}}{R_{2}}
\right)
\label{eq:4.11}
\end{eqnarray}
for the distance when $\kappa=1$, whereas in the limit
$\kappa\rightarrow\infty$ (continuous compounding) we have
\begin{eqnarray}
\phi(\rho_{1},\rho_{2}) = \cos^{-1}\left(
\frac{2\sqrt{R_{1}R_{2}}}{R_{1}+R_{2}}\right) .
\label{eq:4.12}
\end{eqnarray}
It is interesting to observe that the bracketed term in
(\ref{eq:4.12}) is given by the ratio of the geometric and
arithmetic means of the two rates.
Alternatively, we can view (\ref{eq:4.9}) as a parametric
family of distributions, parameterised by the flat rate $R$.
Then it is natural to consider the Fisher-Rao distance
between the two term structures characterised by $R_{1}$ and
$R_{2}$. A straightforward calculation then leads to a
simple distance formula given by
\begin{eqnarray}
D(R_{1},R_{2}) = \sqrt{\frac{\kappa}{\kappa+2}}\log
\frac{R_{2}}{R_{1}} ,
\label{eq:4.13}
\end{eqnarray}
where we have assumed $R_{2}\geq R_{1}$.
\section{Interest rate dynamics}
The formalism we have developed so far is essentially a static
one, set in the present. Now we turn to the problem of developing
a dynamical theory of interest rates. The idea is that, at each
instant of time, the yield curve is characterised by a term
structure density according to the scheme described in the previous
sections. Then, as time passes, the density function evolves
randomly. As a consequence we obtain a measure-valued process.
In particular, we obtain a process on ${\cal D}({\bf R}_{+}^{1})$.
Our goal in this section is to determine a set of conditions on
this process necessary and sufficient to ensure that the resulting
interest rate dynamics will be arbitrage-free.
We shall assume the reader is familiar with the general theory
of interest rate dynamics as laid out, for example, in Carverhill
(1994), Rogers (1994), Hughston (1996), Baxter (1997), Musiela
$\&$ Rutkowski (1997), Brody (2000) or Hunt $\&$ Kennedy (2000).
For the general discount bond dynamics, let us write
\begin{eqnarray}
{\rm d}P_{tT} = \mu_{tT}{\rm d}t + \Sigma_{tT}\cdot {\rm d}
W_{t} ,
\label{eq:5.1}
\end{eqnarray}
where $\mu_{tT}$ and $\Sigma_{tT}$ are the {\sl absolute drift
and absolute volatility processes}, respectively, for a bond
with maturity $T$. Here,
$W_{t}$ is a vector Brownian motion, and $\Sigma_{tT}$ is a
vector process, and there is an inner product implied between
$\Sigma_{tT}$ and ${\rm d}W_{t}$, signified by a dot. We need not
specify the dimensionality of the Brownian motion, which might
be infinite, and indeed in some respects the infinite dimensional
setting is the most natural one.
In fact, it suffices for our purposes merely to
assume that $P_{tT}$ is a one-parameter family of
continuous semi-martingales on the given probability space, with
respect to the given filtration. However, for simplicity of
exposition we shall stick to the case where the relevant
stochastic basis is generated by a multidimensional Brownian
motion. Here, as in Flesaker $\&$ Hughston (1997a,b), we
regard the discount bond dynamics as the natural
starting position, rather than, say, the instantaneous forward
rate dynamics (Heath {\it et al.} 1992), which we need not
consider here directly. We shall assume nevertheless, as in the
HJM framework, that the
processes $\mu_{tT}$ and $\Sigma_{tT}$ are both smooth in the
variable $T$, and that sufficiently strong technical
conditions are in place to ensure that the instantaneous
forward rate processes are semimartingales.
In order to extend the analysis of the previous section it is
convenient to introduce what is sometimes conveniently referred
to as the `Musiela parameterisation', given by
\begin{eqnarray}
B_{tx} = P_{t,t+x} ,
\label{eq:5.2}
\end{eqnarray}
where $T=t+x$ represents the maturity date of the bond, and
hence $x$ is the time left until maturity. Thus $B_{tx}$ is the
value at time $t$ of a discount bond that has $x$ years left to
mature. This choice of parameterisation has already been shown
to be useful in the geometric analysis of interest rates
(Bj\"ork $\&$ Svensson 1999, Bj\"ork $\&$ Christensen 1999,
Bj\"ork $\&$ Gombani 1999, Bj\"ork 2000). We note that
$B_{t0}=1$ for all
$t$, and that $B_{tx} \rightarrow 0$ as $x\rightarrow\infty$. It
follows that
\begin{eqnarray}
\rho_{t}(x) = -\frac{\partial}{\partial x}B_{tx}
\label{eq:5.4}
\end{eqnarray}
is a measure-valued process in the sense that, for each
value of $t$ the random function $\rho_{t}(x)$ satisfies
$\rho_{t}(x)>0$ and the normalisation condition
\begin{eqnarray}
\int_{0}^{\infty} \rho_{t}(x) {\rm d}x = 1 .
\label{eq:5.5}
\end{eqnarray}
Here we have chosen the
notation $\rho_{t}(x)$ that makes the $x$ dependence more
prominent, to emphasise the fact that, for each value of $t$,
and conditional on information given up to time $t$,
$\rho_{t}(x)$ is a density function, though we might have
written $\rho_{tx}$ instead. As a consequence $\rho_{t}(x)$
describes a process on ${\cal D}({\bf R}_{+}^{1})$. By
consideration of (\ref{eq:5.1}) and (\ref{eq:5.2}) we deduce
for the dynamics of $B_{tx}$ that
\begin{eqnarray}
{\rm d}B_{tx} = \left. {\rm d}P_{tT}\right|_{T=t+x} +
\frac{\partial}{\partial x}B_{tx}{\rm d}t ,
\label{eq:5.6}
\end{eqnarray}
and thus, by use of (\ref{eq:5.1}), that
\begin{eqnarray}
{\rm d}B_{tx} = \left(\mu_{t,t+x}+\partial_{x}B_{tx}\right)
{\rm d}t + \Sigma_{t,t+x}\cdot {\rm d}W_{t} ,
\label{eq:5.7}
\end{eqnarray}
where $\partial_{x}=\partial/\partial x$. Differentiating this
expression with respect to $x$ and introducing the measure-valued
process $\rho_{t}(x)$ according to formula (\ref{eq:5.4}) we
therefore obtain
\begin{eqnarray}
{\rm d}\rho_{t}(x) = \left(-\partial_{x}\mu_{t,t+x}+\partial_{x}
\rho_{t}(x)\right){\rm d}t - \partial_{x} \Sigma_{t,t+x}\cdot
{\rm d}W_{t} .
\label{eq:5.8}
\end{eqnarray}
A further simplification is then achieved by introducing the
notation
\begin{eqnarray}
\beta_{tx} = - \partial_{x}\mu_{t,t+x}
\label{eq:5.9}
\end{eqnarray}
and
\begin{eqnarray}
\omega_{tx} = -\partial_{x} \Sigma_{t,t+x} ,
\label{eq:5.10}
\end{eqnarray}
which gives us
\begin{eqnarray}
{\rm d}\rho_{t}(x) = \left( \beta_{tx} + \partial_{x}
\rho_{t}(x)\right){\rm d}t + \omega_{tx}\cdot {\rm d}W_{t} .
\label{eq:5.11}
\end{eqnarray}
In the foregoing discussion we have not yet imposed the
arbitrage-free condition. This is given by the drift constraint
\begin{eqnarray}
\mu_{tT} = r_{t} P_{tT} + \Sigma_{tT}\cdot \lambda_{t} ,
\label{eq:5.12}
\end{eqnarray}
where $\lambda_{t}$ is the process for the market price of risk.
We note that $\lambda_{t}$, like $\Sigma_{tT}$, is a vector
process. However, $\lambda_{t}$ does not depend on the maturity
$T$. The absence of arbitrage ensures the existence of
$\lambda_{t}$. For our purposes we do not need to insist that
the bond market is complete: all we require is the existence
of a pricing kernel, or equivalently the existence of a
self-financing
`natural numeraire' portfolio with value process $N_{t}$ such
that $P_{tT}/N_{t}$ is a martingale for each value of $T$
(cf. Flesaker $\&$ Hughston 1997c). The numeraire process
satisfies
\begin{eqnarray}
\frac{{\rm d}N_{t}}{N_{t}} = (r_{t}+\lambda_{t}^{2}){\rm d}t
+ \lambda_{t}\cdot {\rm d}W_{t} ,
\label{eq:5.121}
\end{eqnarray}
and the corresponding pricing kernel is given by $1/N_{t}$.
As a consequence of the constraint (\ref{eq:5.12}) we then
have
\begin{eqnarray}
\mu_{t,t+x} = r_{t} B_{tx} + \Sigma_{t,t+x}\cdot \lambda_{t} ,
\label{eq:5.13}
\end{eqnarray}
and therefore, by differentiation of this expression with
respect to $x$, we obtain
\begin{eqnarray}
\beta_{tx} = r_{t}\rho_{t}(x) + \omega_{tx}\cdot\lambda_{t} .
\label{eq:5.14}
\end{eqnarray}
Inserting (\ref{eq:5.14}) in (\ref{eq:5.11}) we are thus able to
express the dynamics of the density function $\rho_{t}(x)$ in the
form
\begin{eqnarray}
{\rm d}\rho_{t}(x) = \left( r_{t}\rho_{t}(x) +
\partial_{x}\rho_{t}(x) \right) {\rm d}t + \omega_{tx}
\cdot \left( {\rm d}W_{t}+\lambda_{t}{\rm d}t\right) .
\label{eq:5.15}
\end{eqnarray}
Before proceeding further, let us verify, as a consistency check,
that the dynamics given by (\ref{eq:5.15}) preserves the
normalisation condition on $\rho_{t}(x)$, given by (\ref{eq:5.5}).
Integrating the right hand side of (\ref{eq:5.15}) with respect to
$x$ and equating the drift and volatility terms separately to zero
leads to the relations
\begin{eqnarray}
r_{t} + \int_{0}^{\infty} \partial_{x}\rho_{t}(x){\rm d}x = 0
\label{eq:5.16}
\end{eqnarray}
and
\begin{eqnarray}
\int_{0}^{\infty} \omega_{tx} {\rm d}x = 0 ,
\label{eq:5.17}
\end{eqnarray}
which must hold for all $t$. Condition (\ref{eq:5.16}) is satisfied
because $\rho_{t}(x)\rightarrow0$ as $x\rightarrow\infty$ and
\begin{eqnarray}
\rho_{t}(0) = r_{t} .
\label{eq:5.18}
\end{eqnarray}
Condition (\ref{eq:5.17}) is satisfied because, by definition,
we have
$\omega_{tx}=-\partial_{x}\Sigma_{t,t+x}$, and the absolute
volatility $\Sigma_{t,t+x}$ vanishes both as $x\rightarrow0$ (a
maturing bond has a definite value and thus has no absolute
volatility), and as $x\rightarrow\infty$ (a bond with infinite
maturity has no value, and hence no absolute volatility).
Summing up matters so far, we see that in (\ref{eq:5.15}) we are
able to cut the standard HJM arbitrage-free interest rate dynamics
in the form of a measure-valued process $\rho_{t}(x)$ subject to
the constraints (\ref{eq:5.16}) and (\ref{eq:5.17}). At first
glance the role of the short rate $r_{t}$ in (\ref{eq:5.15}) seems
anomalous, because it might appear that this has to be specified
separately. However, by virtue of ({\ref{eq:5.18}) we can
incorporate $r_{t}$ directly into the dynamics of $\rho_{t}(x)$.
In fact, there is another way of expressing (\ref{eq:5.15}) which
is very suggestive, and ties in naturally with the Hilbert space
approach to dynamics introduced in \S 7. First
we note that (\ref{eq:5.16}) can be rewritten in the form
\begin{eqnarray}
r_{t} = -\int_{0}^{\infty} \rho_{t}(x) \partial_{x}\ln \rho_{t}(x)
{\rm d}x .
\label{eq:5.19}
\end{eqnarray}
In other words, $r_{t}$ is minus the expectation of the gradient
of the log-likelihood function. Here the expectation is taken
with respect to $\rho_{t}(x)$ itself. Writing $E_{\rho}$ for this
abstract expectation, we have
\begin{eqnarray}
{\rm d}\rho_{t}(x) = \rho_{t}(x) \left( \partial_{x}\ln\rho_{t}(x)
- E_{\rho}[\partial_{x}\ln\rho_{t}(x)]\right)
{\rm d}t + \omega_{tx}\cdot {\rm d}W_{t}^{*} ,
\label{eq:5.20}
\end{eqnarray}
where ${\rm d}W_{t}^{*}={\rm d}W_{t}+\lambda_{t}{\rm d}t$.
We note that $W_{t}^{*}$ has the interpretation of being a
Brownian motion with respect to the risk-neutral measure
associated with the given pricing kernel. In the risk-neutral
measure, for which the term
involving $\lambda_{t}$ effectively disappears, the remaining
drift for $\rho_{t}(x)$ is determined by the deviation of
$\partial_{x}\ln\rho_{t}(x)$ from its abstract mean.
Let us now examine more closely the volatility term
$\omega_{tx}$ appearing in (\ref{eq:5.20}), with a view to gaining
a better understanding of the significance of the volatility
constraint
(\ref{eq:5.17}). Because $\rho_{t}(x)$ must remain positive
for all values of $x$, the coefficient of
${\rm d}W_{t}^{*}$ in (\ref{eq:5.20}) must be of the form
\begin{eqnarray}
\omega_{tx} = \rho_{t}(x) \sigma_{tx}
\label{eq:5.22}
\end{eqnarray}
for some bounded process $\sigma_{tx}$, to ensure that
$\omega_{tx}$ dies off appropriately for values of $x$ such
that $\rho_{t}(x)$ approaches zero. As a consequence, we can
write (\ref{eq:5.15}) in the quasi-lognormal form
\begin{eqnarray}
\frac{{\rm d}\rho_{t}(x)}{\rho_{t}(x)} = (r_{t}+\partial_{x}
\ln\rho_{t}(x)){\rm d}t + \sigma_{tx}\cdot
{\rm d}W_{t}^{*},
\label{eq:5.23}
\end{eqnarray}
and for the constraint (\ref{eq:5.17}) we have
\begin{eqnarray}
E_{\rho}[\sigma_{tx}] = 0,
\label{eq:5.24}
\end{eqnarray}
which can be satisfied by writing
\begin{eqnarray}
\sigma_{tx} = \nu_{tx} - E_{\rho}[\nu_{tx}] ,
\label{eq:5.25}
\end{eqnarray}
where $\nu_{tx}$ is an exogenously specifiable unconstrained
process. Here, for any process $A_{tx}$ we define
$E_{\rho}[A_{tx}] = \int_{0}^{\infty}\rho_{t}(x)A_{tx}{\rm d}x$.
The results established above can then be summarised as follows.
\begin{prop}
The general admissible term structure evolution based on the
information set generated by a multidimensional Brownian motion
$W_{t}$ is given by a measure-valued process $\rho_{t}(x)$ in
${\cal D}({\bf R}_{+}^{1})$ satisfying
\begin{eqnarray}
\frac{{\rm d}\rho_{t}(x)}{\rho_{t}(x)} &=& \left(
\partial_{x}\ln\rho_{t}(x)-E_{\rho}[\partial_{x}\ln\rho_{t}(x)]
\right) {\rm d}t \nonumber \\
& & + \left( \nu_{tx}-E_{\rho}[\nu_{tx}]\right)
\cdot \left( {\rm d}W_{t}+\lambda_{t}{\rm d}t \right) ,
\label{eq:5.26}
\end{eqnarray}
where the processes $\lambda_{t}$ and $\nu_{tx}$ are
specified exogenously, along with the initial term structure
density $\rho_{0}(x)$.
\end{prop}
An advantage of the particular expression (\ref{eq:5.26})
given for the dynamics above is that the preservation of the
normalisation condition on $\rho_{t}(x)$ is evident by
inspection, because this is equivalent to the relation
\begin{eqnarray}
E_{\rho}\left[ \frac{{\rm d}\rho_{t}(x)}{\rho_{t}(x)}
\right] = 0 .
\label{eq:5.27}
\end{eqnarray}
An alternative expression for (\ref{eq:5.26}), which brings
out more explicitly the nonlinearities in the dynamics, is
given by
\begin{eqnarray}
{\rm d}\rho_{t}(x) &=& \left(\partial_{x}\rho_{t}(x)+
\rho_{t}(0)\rho_{t}(x)\right){\rm d}t \nonumber \\
& & + \rho_{t}(x)\left(
\nu_{tx}-\int_{0}^{\infty}\rho_{t}(y)\nu_{ty}{\rm d}y
\right) \cdot {\rm d}W^{*}_{t},
\label{eq:5.28}
\end{eqnarray}
where ${\rm d}W^{*}_{t}={\rm d}W_{t}+\lambda_{t}{\rm d}t$ as
defined earlier.
\section{Principal moment analysis}
The characterisation of the yield curve as an abstract probability
density enables us to develop a rigourous analogue of the classical
`principal component' analysis often used in the study of yield
curve dynamics. To this end we let $\rho_{t}(x)=-\partial_{x}
P_{t,t+x}$ be the density process associated with an admissible
family of discount bond prices, and define the moment processes
\begin{eqnarray}
{\bar x}_{t} = \int_{0}^{\infty} x \rho_{t}(x) {\rm d}x
\label{eq:66.1}
\end{eqnarray}
and
\begin{eqnarray}
{\bar x}_{t}^{\prime(n)} = \int_{0}^{\infty} x^{n} \rho_{t}(x) {\rm d}x
\label{eq:66.2}
\end{eqnarray}
for $n\geq2$, along with the central moment processes
\begin{eqnarray}
{\bar x}_{t}^{(n)} = \int_{0}^{\infty} (x-{\bar x}_{t})^{n}
\rho_{t}(x) {\rm d}x .
\label{eq:66.3}
\end{eqnarray}
It is important to note that in some cases the relevant moments
may not exist. For example, in the case of a continuously
compounded flat yield curve given at $t=0$ by the density
function $\rho_{0}(x)=R{\rm e}^{-Rx}$, we have ${\bar x}_{0}=R^{-1}$,
${\bar x}_{0}^{(2)} = R^{-2}$, ${\bar x}_{0}^{(3)}=3R^{-3}$, and
${\bar x}_{0}^{(4)}=9R^{-4}$ for the first four central moments.
On the other hand, in the example of the simple flat term
structure for which $\rho_{0}(x)=R/(1+Rx)^{2}$ we find that none
of the moments exist, on account of the fatness of the tail of
the distribution. In fact, for the flat rate term structures with
compounding frequency $\kappa$ the moments exist only up to order
$\kappa-1$.
The first four moments, if they exist, are the mean, variance,
skewness and kurtosis of the distribution of the abstract
random variable $X$ characterising the yield curve, and we refer
to these (and other) moments as the
`principal moments' of the given term structure.
At $t=0$ the mean ${\bar x}_{0}$ determines a characteristic
time-scale associated with the given term structure, and
its inverse $1/{\bar x}_{0}$ can be thought of as an associated
characteristic yield. The difference ${\bar x}_{0}^{(2)}
-({\bar x}_{0})^{2}$ then measures the departure of the given
term structure from flatness on a continuously compounded basis.
This is on account of the fact that in the case of an exponential
distribution the variance is given by the square of the mean.
It is legitimate to conjecture that for some purposes the
specification of, e.g., the first three or four moments will be
sufficient to provide an accurate representation of the term
structure. One way of implementing this idea is to introduce the
entropy $S_{\rho}$ of the given distribution, defined by
\begin{eqnarray}
S_{\rho} = -\int_{0}^{\infty}\rho(x)\ln\rho(x){\rm d}x .
\label{eq:66.4}
\end{eqnarray}
Because $\rho(x)$ has dimensions of inverse time, $S_{\rho}$ is
defined only up to an overall additive constant. Therefore, the
difference of the entropies associated with two yield curves has
an invariant significance.
For yield curve calibration we
propose that $\rho(x)$ should be chosen such that $S_{\rho}$ is
maximised subject to the constraints of the data available. For
example, if we are given as data only the mean ${\bar x}_{0}$,
then the maximum entropy term structure is $\rho_{0}(x)=Re^{-Rx}$,
where $R=1/{\bar x}_{0}$.
It is also of great interest to study the dynamics of the
principal characteristics in the case of a general admissible
arbitrage-free term structure. We examine here, in particular,
the mean and the variance processes. For this purpose we
introduce a simplified notation $v_{t}={\bar x}_{t}^{(2)}$ for
the variance process, i.e.,
\begin{eqnarray}
v_{t}=\int_{0}^{\infty}x^{2}\rho_{t}(x){\rm d}x -
({\bar x}_{t})^{2} ,
\label{eq:66.5}
\end{eqnarray}
where the mean process ${\bar x}_{t}$ is given as in (\ref{eq:66.1}).
We assume that both $\rho_{t}(x)$ and the discount bond volatility
$\Sigma_{t,t+x}$ fall off to zero sufficiently rapidly to ensure that
$\lim_{x\rightarrow\infty}x^{n}\rho_{t}(x)=0$ and
$\lim_{x\rightarrow\infty}x^{n}\Sigma_{t,t+x}=0$ for $n=1,2$, and
that the integrals $\int_{0}^{\infty}x^{n}\rho_{t}(x){\rm d}x$ and
$\int_{0}^{\infty}x^{n-1}\Sigma_{t,t+x}{\rm d}x$ exist for
$n=1,2$.
A straightforward calculation then leads us to the following
conclusion:
\begin{prop}
The first principal moment ${\bar x}_{t}$ of an admissible,
arbitrage-free term structure satisfies the dynamical law
\begin{eqnarray}
{\rm d}{\bar x}_{t} = (r_{t}{\bar x}_{t}-1){\rm d}t +
{\bar\Sigma}_{t}\cdot{\rm d}W^{*}_{t} ,
\label{eq:66.6}
\end{eqnarray}
where ${\bar\Sigma}_{t}=\int_{0}^{\infty}\Sigma_{t,t+x}{\rm d}x$.
\end{prop}
{\it Proof}. Starting with (\ref{eq:5.23}) and (\ref{eq:66.1}) we
have
\begin{eqnarray}
{\rm d}{\bar x}_{t} &=& \int_{0}^{\infty}x{\rm d}\rho_{t}(x)
{\rm d}x \nonumber \\
&=& \left( r_{t}{\bar x}_{t}+\int_{0}^{\infty}x\partial_{x}
\rho_{t}(x){\rm d}x\right) {\rm d}t - \left( \int_{0}^{\infty}
x\partial_{x}\Sigma_{t,t+x}{\rm d}x\right)\cdot {\rm d}W_{t}^{*}
\label{eq:66.61}
\end{eqnarray}
by use of (\ref{eq:5.10}). Then,
integrating by parts and using the assumed asymptotic behaviours
for $\rho_{t}(x)$ and $\Sigma_{t,t+x}$, we obtain the desired
result. \hspace*{\fill} $\diamondsuit$
We note that there is a critical level ${\bar x}_{t}^{*}$ for the
first principal moment given by
\begin{eqnarray}
{\bar x}_{t}^{*} = \frac{1}{r_{t}}(1 -
\lambda_{t}\cdot{\bar\Sigma}_{t}).
\label{eq:66.7}
\end{eqnarray}
When ${\bar x}_{t}>{\bar x}_{t}^{*}$ the drift of ${\bar x}_{t}$
is positive, and the drift increases further as ${\bar x}$
increases. On the other hand, when ${\bar x}_{t}<{\bar x}_{t}^{*}$,
the drift of ${\bar x}_{t}$ is negative, and the drift decreases
further as ${\bar x}_{t}$ decreases. For the variance process, we
have:
\begin{prop}
The second principal moment $v_{t}$ of an admissible,
arbitrage-free term structure satisfies the dynamical law
\begin{eqnarray}
{\rm d}v_{t}=\left( r_{t}(v_{t}-{\bar x}_{t}^{2}) -
{\bar\Sigma}_{t}^{2}\right) {\rm d}t + 2 \left(
{\bar\Sigma}_{t}^{(1)}-{\bar x}_{t}{\bar\Sigma}_{t}\right)
\cdot {\rm d}W_{t}^{*} ,
\label{eq:66.8}
\end{eqnarray}
where ${\bar\Sigma}_{t}^{(1)}=\int_{0}^{\infty}x
\Sigma_{t,t+x}{\rm d}x$.
\end{prop}
{\it Proof}. Starting with formula (\ref{eq:66.5}) for $v_{t}$
we have
\begin{eqnarray}
{\rm d}v_{t}=\int_{0}^{\infty} x^{2}{\rm d}\rho_{t}(x) {\rm d}x
- {\rm d}({\bar x}_{t}^{2}) .
\label{eq:66.81}
\end{eqnarray}
For the first term we obtain
\begin{eqnarray}
\int_{0}^{\infty}x^{2}\rho_{t}(x){\rm d}x &=& \left( r_{t}
\int_{0}^{\infty} x^{2}\rho_{t}(x){\rm d}x+\int_{0}^{\infty}
x^{2}\partial_{x}\rho_{t}(x){\rm d}x\right){\rm d}t \nonumber \\
& & - \left(\int_{0}^{\infty}x^{2}\partial_{x}\Sigma_{t,t+x}
\right)\cdot {\rm d}W_{t}^{*} ,
\label{eq:66.82}
\end{eqnarray}
where we have used (\ref{eq:5.10}) and (\ref{eq:5.23}).
As a consequence of the assumed asymptotic behaviour of
$\rho_{t}(x)$ and $\Sigma_{t,t+x}$, this becomes
\begin{eqnarray}
\int_{0}^{\infty}x^{2}\rho_{t}(x){\rm d}x = \left( r_{t}
{\bar x}_{t}^{\prime(2)} - 2{\bar x}_{t} \right){\rm d}t
+ 2 {\bar \Sigma}_{t}^{(1)}\cdot {\rm d}W_{t}^{*} ,
\label{eq:66.83}
\end{eqnarray}
after an integration by parts. For the second term in
(\ref{eq:66.81}) we have
\begin{eqnarray}
{\rm d}({\bar x}_{t}^{2}) = 2{\bar x}_{t}{\rm d}{\bar x}_{t} +
({\rm d}{\bar x}_{t})^{2}
\label{eq:66.84}
\end{eqnarray}
by Ito's lemma, and thus
\begin{eqnarray}
{\rm d}({\bar x}_{t}^{2}) = \left( 2r_{t}{\bar x}_{t}^{2}-
2{\bar x}_{t}+\Sigma_{t}^{2}\right){\rm d}t + 2{\bar x}_{t}
{\bar\Sigma}_{t}\cdot{\rm d}W_{t}^{*}
\label{eq:66.85}
\end{eqnarray}
by use of Proposition 3. Combining (\ref{eq:66.83}) and
(\ref{eq:66.85}), and using the definition (\ref{eq:66.5}) we
obtain (\ref{eq:66.8}). \hspace*{\fill} $\diamondsuit$
In this case we recall that the difference $v_{t}-{\bar x}_{t}^{2}$
acts as a simple measure of the extent to which the distribution
deviates from the `flat' term structure. As a consequence we see
that the effect of the dynamics here is that the second principal
moment of the term structure tends to increase, i.e., has a
positive drift, providing
$v_{t}-{\bar x}_{t}^{2}$ is already above the level given by
\begin{eqnarray}
v_{t}-{\bar x}_{t}^{2} = \frac{1}{r_{t}}\left(
{\bar\Sigma}_{t}^{2}-2\lambda_{t}\cdot ({\bar\Sigma}_{t}^{(1)}-
{\bar x}_{t}{\bar\Sigma}_{t})\right) .
\label{eq:66.9}
\end{eqnarray}
\section{Hilbert space dynamics for term structures}
Now that we have examined some of the advantages of expressing
the arbitrage-free interest rate term structure dynamics as a
randomly evolving density function, let us consider how we
transform to the Hilbert space representation for density
functions considered in \S 2. Denote by $\xi_{tx}$ the process
for the square-root likelihood function, defined by
\begin{eqnarray}
\rho_{t}(x) = \xi_{tx}^{2} .
\label{eq:6.1}
\end{eqnarray}
It follows then, by Ito's lemma, that
\begin{eqnarray}
{\rm d}\rho_{t}(x) = 2\xi_{tx}{\rm d}\xi_{tx} +
({\rm d}\xi_{tx})^{2} ,
\label{eq:6.2}
\end{eqnarray}
and hence
$({\rm d}\rho_{t}(x))^{2} = 4\xi_{tx}^{2}({\rm d}\xi_{tx})^{2}$.
By rearranging (\ref{eq:6.2}) we thus obtain
\begin{eqnarray}
{\rm d}\xi_{tx} = \frac{1}{2\xi_{tx}} {\rm d}\rho_{t}(x) -
\frac{1}{8\xi_{tx}^{3}}({\rm d}\rho_{t}(x))^{2}
\label{eq:6.3}
\end{eqnarray}
for the dynamics of the process $\xi_{tx}$, and hence
\begin{eqnarray}
{\rm d}\xi_{tx} = \left( \partial_{x}\xi_{tx} + \frac{1}{2}
r_{t}\xi_{tx} - \frac{1}{8\xi_{tx}^{3}}\omega_{tx}^{2}
\right) {\rm d}t + \frac{1}{2\xi_{tx}}\omega_{tx}\cdot
{\rm d}W_{t}^{*} ,
\label{eq:6.4}
\end{eqnarray}
where $\omega^{2}_{tx}=\omega_{tx}\cdot\omega_{tx}$. Now
suppose we define $\sigma_{tx}$ by the quotient
\begin{eqnarray}
\sigma_{tx} = \frac{\omega_{tx}}{\xi_{tx}^{2}} ,
\label{eq:6.5}
\end{eqnarray}
as before, and set
$\sigma_{tx}^{2}=\sigma_{tx}\cdot\sigma_{tx}$. Then the
process for the square-root density $\xi_{tx}$ can be
written in the form
\begin{eqnarray}
{\rm d}\xi_{tx} = \left( \partial_{x}\xi_{tx} + \mbox{$\textstyle \frac{1}{2}$}
r_{t}\xi_{tx} - \mbox{$\textstyle \frac{1}{8}$} \xi_{tx}\sigma_{tx}^{2} \right)
{\rm d}t + \mbox{$\textstyle \frac{1}{2}$} \xi_{tx} \sigma_{tx}\cdot {\rm d}W_{t}^{*} .
\label{eq:6.6}
\end{eqnarray}
We recall that the volatility process $\sigma_{tx}$ arising
again in this connection, which is given more explicitly by
the ratio
\begin{eqnarray}
\sigma_{tx} = \frac{\partial_{x}\Sigma_{t,t+x}}
{\partial_{x}B_{tx}} ,
\label{eq:6.7}
\end{eqnarray}
can be specified exogenously, subject only to the
condition that it has mean zero in the measure $\rho_{t}(x)$,
which implies that $\sigma_{tx}$ can be written in the form
(\ref{eq:5.25}).
\begin{figure}[t]
\centerline{
\psfig{file=r1fig4.eps,width=11cm,angle=0}
}
\caption{{\it Interest rate dynamics}. At each instant of
time the interest rate term structure can be represented as
a point on the positive orthant of the unit sphere
${\cal S}$ in the Hilbert space ${\cal H}=
L^{2}({\bf R}_{+}^{1})$. The associated arbitrage-free interest
rate dynamics gives rise to a stochastic trajectory on this
space, which is foliated by hypersurfaces corresponding to
level values of the short-term interest rate.
}
\end{figure}
We would now like to interpret the Hilbert space dynamics in
equation (\ref{eq:6.6}) more directly in a geometrical fashion.
For this purpose we find
it expedient to introduce an index notation, using Greek
letters to signify Hilbert space operations
(cf. Brody $\&$ Hughston 1998).
Thus if the function $\psi(x)$ is an element of
${\cal H}=L^{2}({\bf R}_{+}^{1})$, we denote it by
$\psi^{\alpha}$, and if $\varphi(x)$ belongs to the
dual Hilbert space ${\cal H}^{*}$ we denote this by
$\varphi_{\alpha}$. Furthermore, their inner product
is written
\begin{eqnarray}
\psi^{\alpha}\varphi_{\alpha} = \int_{0}^{\infty}
\psi(x)\varphi(x) {\rm d}x .
\label{eq:6.9}
\end{eqnarray}
There is a preferred symmetric quadratic form $g_{\alpha\beta}$
on ${\cal H}$, given by $g_{\alpha\beta}\psi^{\alpha}\psi^{\beta}
=\int_{0}^{\infty}(\psi(x))^{2}{\rm d}x$, which thus establishes
an isomorphism between ${\cal H}$ and ${\cal H}^{*}$, given by
$\psi^{\alpha}\rightarrow\psi_{\alpha}=g_{\alpha\beta}\psi^{\beta}$.
Intuitively, one can think of $g_{\alpha\beta}$ as corresponding
to the delta function $\delta(x,y)$, and then we have
\begin{eqnarray}
g_{\alpha\beta} \psi^{\alpha}\varphi^{\beta} = \int_{0}^{\infty}
\psi(x)\delta(x,y)\varphi(y) {\rm d}x {\rm d}y.
\label{eq:6.10}
\end{eqnarray}
There are a number of Hilbert space technicalities
that have to be considered for a complete exposition of the
matter, but that is not our immediate concern.
If $\xi(x)>0$ belongs to the positive orthant of
$L^{2}({\bf R}^{1}_{+})$ then the corresponding indexed quantity
$\xi^{\alpha}$ has the interpretation of a `state vector'.
In that case we can think of symmetric quadratic forms as
representing certain classes of random variables.
The expectation of the random variable $H_{\alpha\beta}$ in the
state $\xi^{\alpha}$ is
\begin{eqnarray}
E_{\xi}[H] = \frac{H_{\alpha\beta}\xi^{\alpha}\xi^{\beta}}
{\xi_{\gamma}\xi^{\gamma}} .
\label{eq:6.11}
\end{eqnarray}
Therefore, a state vector determines a mapping from random
variables to real numbers, through (\ref{eq:6.11}). For a
normalised state vector we have $\xi_{\alpha}\xi^{\alpha}=1$,
although for some purposes it is convenient to relax the
normalisation condition. In particular, we notice that the
expectation (\ref{eq:6.11}) only depends on the direction of
$\xi^{\alpha}$.
Now suppose $\xi(x)$ is a positive function. In that case, the
derivative $\partial_{x}$ can be thought of as a linear operator
$D^{\alpha}_{\ \beta}$ on ${\cal H}$, and we have an
endomorphism given by
$\xi^{\alpha} \rightarrow D^{\alpha}_{\ \beta}\xi^{\beta}$.
By making use of this, we can now interpret, in the language of
Hilbert space geometry, the first two terms appearing in the
drift in the dynamical equation (\ref{eq:6.6}).
Let us begin
by noting first that (\ref{eq:5.16}) can be rewritten in the
form
\begin{eqnarray}
\int_{0}^{\infty} \xi_{tx} \partial_{x} \xi_{tx} {\rm d}x
= -\frac{1}{2} r_{t} .
\label{eq:6.12}
\end{eqnarray}
This allows us to interpret the short term interest rate process
$r_{t}$ in terms of the mean of the symmetric part of the operator
$D^{\alpha}_{\ \beta}$ in the state $\xi_{t}^{\alpha}$, i.e.,
\begin{eqnarray}
\frac{D_{\alpha\beta} \xi^{\alpha}_{t}\xi^{\beta}_{t}}
{g_{\alpha\beta}\xi^{\alpha}_{t}\xi^{\beta}_{t}} =
-\frac{1}{2}r_{t} ,
\label{eq:6.13}
\end{eqnarray}
where $D_{\alpha\beta} = g_{\alpha\gamma}D^{\gamma}_{\ \beta}$.
Therefore, if we let $D_{(\alpha\beta)}$ denote the symmetric
part of the operator $D^{\alpha}_{\ \beta}$, then the abstract
random variable in ${\cal H}$ corresponding to the short rate
$r_{t}$ is given by $r_{\alpha\beta}=-2 D_{(\alpha\beta)}$.
Similarly we can represent the abstract random variable $x$
for the time left until maturity in ${\cal H}$ by a symmetric
matrix $X_{\alpha\beta}$. It is interesting to note that the
random variables $X_{\alpha\beta}$ for the maturity date and
$r_{\alpha\beta}$ for the short term interest rate are not
`compatible'. Two random variables $A$ and $B$ are said to be
compatible if the expression $\{\{A,C\},B\} -
\{A,\{C,B\}\}$ vanishes for any random variable $C$, where
$\{A,B\}=AB+BA$ denotes the anticommutator (Segal 1947).
The lack of compatibility here indicates that the abstract
probability system containing both $r_{\alpha\beta}$ and
$X_{\alpha\beta}$ as random variables is not Kolmogorovian.
However, the algebra of random variables generated by
$X_{\alpha\beta}$ is Kolmogorovian.
Now, let
$\eta(x)$ be an arbitrary element of $L^{2}({\bf R}_{+}^{1})$,
and let $\eta^{\alpha}$ be the corresponding Hilbert space
vector. Then clearly we have
\begin{eqnarray}
\int_{0}^{\infty} \eta(x)\left[ \partial_{x} \xi_{tx}
+ \mbox{$\textstyle \frac{1}{2}$} r_{t}\xi_{tx} \right] {\rm d}x = \eta_{\alpha}
\left[ D^{\alpha}_{\ \beta}\xi^{\beta}_{t} -
\left(
\frac{D_{\beta\gamma}\xi^{\beta}_{t}\xi_{t}^{\gamma}}
{g_{\delta\epsilon}\xi_{t}^{\delta}\xi_{t}^{\epsilon}}
\right)
\xi^{\alpha}_{t} \right] .
\label{eq:6.14}
\end{eqnarray}
In other words, the first two terms of the drift in (\ref{eq:6.6})
can be replaced by the expression ${\tilde D}^{\alpha}_{\ \beta}
\xi^{\beta}_{t}$, where
\begin{eqnarray}
{\tilde D}^{\alpha}_{\ \beta} = D^{\alpha}_{\ \beta} -
\left(
\frac{D_{\gamma\delta}\xi^{\gamma}_{t}\xi_{t}^{\delta}}
{g_{\gamma\delta}\xi_{t}^{\gamma}\xi_{t}^{\delta}} \right)
\delta^{\alpha}_{\ \beta} ,
\label{eq:6.15}
\end{eqnarray}
where $\delta^{\alpha}_{\ \beta}$ is the Kronecker delta. Clearly,
we have ${\tilde D}_{\alpha\beta}\xi^{\alpha}\xi^{\beta}=0$.
With this in mind, let us now proceed to the interpretation of the
volatility process $\sigma_{tx}$. Again, $\sigma_{tx}$ has the
character of a linear operator acting on $\xi_{tx}$, subject to
the constraint $E_{\rho}[\sigma_{tx}]=0$. This can be consistently
enforced if there exists a symmetric process
$\nu_{t\alpha\beta}$ such that
\begin{eqnarray}
\int_{0}^{\infty} \eta(x) \xi_{tx} \sigma_{tx}
{\rm d}x = \eta^{\alpha}
\left( \nu_{t\alpha\beta}\xi^{\beta}_{t} -
E_{\xi}[\nu_{t}] \xi_{t\alpha} \right) .
\label{eq:6.16}
\end{eqnarray}
The symmetric operator-valued random process
$\nu_{t\alpha\beta}$, whose existence is
thus implied, is `primitive' in the sense that it is
unconstrained and can be specified exogenously. If we write
\begin{eqnarray}
\sigma_{t\alpha\beta} = \nu_{t\alpha\beta} -
E_{\xi}[\nu_{t}] g_{\alpha\beta} ,
\label{eq:6.17}
\end{eqnarray}
we obtain
\begin{eqnarray}
\int_{0}^{\infty} \eta(x) \xi_{tx} \sigma_{tx}
{\rm d}x = \eta_{\alpha} \sigma_{t\beta}^{\alpha}
\xi^{\beta}_{t} ,
\label{eq:6.18}
\end{eqnarray}
and also
\begin{eqnarray}
\int_{0}^{\infty} \eta(x) \xi_{tx} \sigma_{tx}^2
{\rm d}x = \eta_{\alpha} \sigma_{t\beta}^{\alpha}
\sigma_{t\gamma}^{\beta} \xi^{\gamma}_{t} .
\label{eq:6.19}
\end{eqnarray}
Therefore, putting the various ingredients together, we obtain:
\begin{prop}
The dynamics of the Hilbert space vector $\xi^{\alpha}_{t}$
that characterises the term structure in an admissible,
arbitrage-free interest rate framework is governed by the
stochastic differential equation
\begin{eqnarray}
{\rm d}\xi^{\alpha}_{t} = \left( {\tilde D}^{\alpha}_{\ \beta}
- \mbox{$\textstyle \frac{1}{8}$} \sigma^{\alpha}_{t\gamma}\cdot
\sigma^{\gamma}_{t\beta}\right)\xi^{\beta}_{t}{\rm d}t
+ \mbox{$\textstyle \frac{1}{2}$} \sigma^{\alpha}_{t\beta}\xi^{\beta}_{t}\cdot
\left( {\rm d}W_{t} + \lambda_{t}{\rm d}t\right) ,
\label{eq:6.20}
\end{eqnarray}
where ${\tilde D}^{\alpha}_{\ \beta}$ is given as in
{\rm (\ref{eq:6.15})}, and the adapted operator-valued
process $\sigma_{t\alpha\beta}$ is expressible in the form
\begin{eqnarray}
\sigma_{t\alpha\beta} = \nu_{t\alpha\beta} - \left(
\frac{\nu_{\gamma\delta}\xi^{\gamma}_{t}\xi_{t}^{\delta}}
{g_{\gamma\delta}\xi_{t}^{\gamma}\xi_{t}^{\delta}} \right)
g_{\alpha\beta} ,
\label{eq:6.21}
\end{eqnarray}
where $\nu_{t\alpha\beta}$ is an arbitrary adapted
operator-valued process.
\end{prop}
This result shows that the evolution of the yield curve can be viewed
consistently as a process on the positive orthant of the unit
sphere in Hilbert space, and thus gives rise to an entirely new
way of understanding the dynamics of the term structure. The
purpose of the quadratic term in the drift of (\ref{eq:6.20}) is
to keep the process on the sphere, and in the absence of the
term involving the operator $D^{\alpha}_{\ \beta}$ we would have
a general local martingale on the sphere ${\cal S}$ with respect to
the risk-neutral measure, where the martingale property on
${\cal S}$ is characterised in a standard way by use of the
techniques of stochastic differential geometry (see, e.g., Emery
1989, Ikeda $\&$ Watanabe 1989, Hughston 1996). The term involving
the operator
$D^{\alpha}_{\ \beta}$ splits into a
symmetric and an antisymmetric part. The drift generated by
the antisymmetric part of $D^{\alpha}_{\ \beta}$ is generated
by a symmetry of the sphere ${\cal S}$. The drift generated
by the symmetric part of $D^{\alpha}_{\ \beta}$, on the other
hand, is a negative gradient vector field orthogonal to surfaces
in ${\cal S}$ generated by level values of the short rate $r_{t}$.
This term therefore creates a tendency for the vector $\xi^{\alpha}$
to drift towards a lower interest rate, a property of the negative
gradient field which is then counterbalanced by the effects of the
diffusive term.
\vspace{0.5cm}
\begin{center}
{\bf Acknowledgements}
\end{center}
\vspace{0.2cm}
LPH acknowledges the hospitality of the Finance Department of
the Graduate School of Business of the University of Texas at
Austin, where part of this work was carried out. DCB gratefully
acknowledges financial support from The Royal Society.
|
\section{Introduction}
The increasing number of synoptic surveys are now generating tens to hundreds of
transient events per night, and the rates will keep growing, possibly reaching
millions of transients per night within a decade or so.
Generally, follow-up observations are needed in order to fully exploit
scientifically these data streams. In optical surveys, for
instance, all transients look the same when discovered -- a starlike object
that has changed its brightness significantly -- and yet, they could represent
vastly different physical phenomena. Which ones are worthy of a follow-up?
This is a critical issue for the massive event streams
(e.g., LSST, SKA, etc.), and the sheer volume requires an automated approach
(\cite{Don08}, \cite{Mah10}, \cite{Djo11a}).
The process of scientific measurement and discovery typically operates on time scales
from days to decades after the original measurements, feeding back to a new
theoretical understanding. However, that clearly would not work in the case of
phenomena where a rapid change occurs on time scales shorter than what it takes
to set up the new round of measurements. This results in the need for
real-time systems, consisting of computational analysis and decision engine,
and optimized follow-up instruments that can be rapidly deployed
with immediate analysis and feedback.
These requirements imply a need for an automated classification and decision making.
The classification process for a given transient involves:
(1) obtain available
contextual archival information, and combine it with the measured parameters
from the discovery pipeline,
(2) determine (relative?)
probabilities or likelihoods of it belonging to various classes of transients,
(3) obtain
follow-up observations to best disambiguate competing classes,
(4) use them as a
feedback and repeat for an improved classification. We describe below a few
techniques that help in this process.
Our principal data set is the transient event stream from the
Catalina Real-time Transient Survey (CRTS; http://crts.caltech.edu;
\cite{Dra99}, \cite{Djo11b}, \cite{Mah11}),
but the methodology we are developing is more universally applicable.
\noindent {\bf Bayesian Networks:}
Generally, the available data for any given event would be heterogeneous and
incomplete. That is difficult to accommodate in the standard machine-learning
feature vector approach, but it can be naturally accommodated in a Bayesian
approach, such as Bayesian Networks (BN) (\cite{Mah08}).
We have used three colors obtained from the Palomar 60-inch telescope from
follow-up observations of CRTS transients, and
two contextual parameters: Galactic latitude and proximity to a galaxy.
Priors for six
classes have been used: CVs, SNe, Blazars, other AGN, UV Ceti, and the ``Rest''
(everything else). We are currently adding more parameters and classes.
About 300 objects each have been used for SN and CV, and $\sim$100 for
blazars. The number statistics for other AGN and UV Ceti are still too small.
82\% of the objects classified as SN are indeed SN (79\% for CVs,
69\% for Blazars). The contamination is $\sim10 - 20\%$. Given the fact that a single
set of observations accomplished this, the potential for extending the BN,
and combining its output with other techniques is very promising.
\noindent {\bf Lightcurve Based Classification:}
Structure in a sparse and/or irregular light curves (LC)
can be exploited by automated classification algorithms. This can be done by
collecting LCs for different objects belonging to a class and
representing and encoding the characteristic structure probabilistically in the
form of an empirical probability distribution function (PDF). This can then be used
for subsequent classification of a LC with even a few epochs. Moreover, this
comparison can be made incrementally over time as new observations become
available, with the final classification scores improving with
each additional set of observations. This forms the basis for a real time
classification methodology. Since the observations come in the form of flux at
a given epoch, for each point after the very first one we can form a ($\delta
m, \delta t$) pair. We focus on modeling the joint distribution of all such
pairs of data points for a given LC. By virtue of being increments, the
empirical probability density functions of these pairs are invariant to
absolute magnitude and time shifts, a desirable feature.
Upper limits can also be encoded in this methodology,
e.g., forced photometry magnitudes at a SN location in images
taken before the star exploded. We currently use smoothed 2D histograms to
model the distribution of elementary ($dm,dt$) sets.
In our preliminary experimental
evaluations with a small number of object classes (single outburst like SNe,
periodic variable stars like RR Lyrae and Miras, as well as stochastic
variables like blazars and CVs) we have been able to show that the density
models for these classes are potentially a powerful method for object
classification from sparse/irregular LC data.
Currently we are using the ($dm,dt$) distributions for classification in a binary
mode i.e. successive two-class classifiers in a tree structure
SNe are first
separated from non-SNe (the easiest bit, currently performing at a $\sim$99\%
completness), then
non-SNe are separated into stochastic versus non-stochastic variables, and then each
group further separated into more branches. The most difficult so far has been
the CV-blazar node (based on just the ($dm,dt$) density i.e., without bringing in
the proximity to a radio source since we are also interested in discovering
blazars that were not active when the archival radio surveys were done).
Currently this classifier is performing at a $\sim$71\% completness.
We are also exploring Genetic Algorithms to
determine the optimal ($dm,dt$) bins for different classes. This will in turn
help optimise follow-up observing intervals for specific classes; see, e.g.,
\cite{Mah11} or \cite{Djo11a}.
\noindent {\bf Incorporating the Contextual Information:}
Contextual information can be highly relevant to resolving competing
interpretations: for example, the light curve and observed properties of a
transient might be consistent with it being a cataclysmic variable star, a
blazar, or a SN. If it is subsequently known that there is a galaxy in
close proximity, the SN interpretation becomes much more plausible.
Such information, however, can be characterized by high uncertainty and
absence, and by a rich structure: if there were two galaxies nearby instead of
one then details of galaxy type and structure and native stellar populations
become important, e.g., is this type of SN more consistent with being in
the extended halo of a large spiral galaxy or in close proximity to a faint
dwarf galaxy? The ability to incorporate such contextual information in a
quantifiable fashion is highly desirable. We have been compiling priors for
such information as well. These then get incorporated into the Bayesian network
mentioned earlier.
We are also investigating the use of crowdsourcing (citizen science) as a
means of harvesting the human pattern recognition skills, especially in the
context of capturing the relevant contextual information, and turning them into
machine-processable algorithms. A methodology employing contextual knowledge
forms a natural extension to the logistic regression and classification methods
mentioned above. This will be necessary for larger future surveys where the
data flow will exceed the available human resources, and moreover, it would
make such classification objective and repeatable. It also represents an
example of a human-machine collaborative discovery process.
Transients can also be
found using the technique of image subtraction using a matched
older observation, or a deeper co-added image
(\cite{Dra99}). If the images are properly
matched, transients stand out as a positive residual.
When used with white light as is the case with CRTS, the difference
images tend to have bipolar residuals thus leading to false detections.
We have been experimenting with these to look for
supernovae in galaxies using citizen science where a few amateur astronomers
regularly look at the galaxy images along with the residuals presented to them
A large number of SNe have been found in
this fashion (see \cite{Pri11} for an example, and http://nesssi.cacr.caltech.edu/catalina/current.html
for a list).
A given classifier may not be optimal for all classes, nor to all types of inputs.
That is the primary reason why multiple types of classifiers have to be employed
in the complex task of classifying transients in real time.
Presence of
different bits of information can trigger different classifiers. In some cases more
than one classifier can be used for the same kinds of inputs.
An essential task, then, is to derive an optimal event classification, given inputs
from a diverse set of classifiers such as those described above.
Combining different classifiers with different
number of output classes and in presence of error-bars is a non-trivial task
and is still under development.
{\it Acknowledgments:} This work was supported in part by the NASA grant
08-AISR08-0085, and the NSF grants AST-0909182 and IIS-1118041.
|
Subsets and Splits